TH MTD G21299 - It works
Transcription
TH MTD G21299 - It works
AN INTERACTIVE DESIRABILITY FUNCTION BASED APPROACH TO GUIDED PARETO-OPTIMAL FRONT A THESIS Submitted in partial fulfilment of the requirements for the award of the degree of DOCTOR OF PHILOSOPHY in MATHEMATICS by AMAR KISHOR DEPARTMENT OF MATHEMATICS INDIAN INSTITUTE OF TECHNOLOGY ROORKEE ROORKEE-247 667 (INDIA) DECEMBER, 2010 ©INDIAN INSTITUTE OF TECHNOLOGY ROORKEE, ROORKEE, 2010 ALL RIGHTS RESERVED INDIAN INSTITUTE OF TECHNOLOGY ROORKEE ROORKEE CANDIDATE'S DECLARATION I hereby certify that the work which is being presented in the thesis entitled AN INTERACTIVE DESIRABILITY FUNCTION BASED APPROACH TO GUIDED PARETO-OPTIMAL FRONT in partial fulfilment of the requirements for the award of the Degree of Doctor of Philosophy and submitted in the Department of Mathematics of the Indian Institute of Technology Roorkee, Roorkee is an authentic record of my own work carried out during a period from July, 2005 to December, 2010 under the supervision of Dr. Shiv Prasad Yadav, Associate Professor, Department of Mathematics and Dr. Surendra Kumar, Assistant Professor, Electrical Engineering Department, Indian Institute of Technology Roorkee, Roorkee. The matter presented in this thesis has not been submitted by me for the award of any other degree of this or any other Institute. (AMAR KISHOR) This is to certify that the above statement made by the candidate is correct to the best of our kno dge. p (Surendra Kumar (Shiv rasad Yadav) Supervisor Supervisor Date: December27, 2010 The Ph.D. Viva-Voce Examination of Mr. Amar Kishor, Research Scholar, has been 20 • i_c9 . 9-o 1~ held on igna4uf f Supervisors : C''`ji.j . Signa ure of xternal Examiner Abstract Decision-making involves the use of a rational proceSs for selecting the best of several alternatives. In real life, decisions are often made on the basis of multiple, conflicting and non-commensurable criteria/objectives in uncertain/imprecise environments. Multiobjective evolutionary algorithm (MOEA) usually attempts to find a good approximation to the complete Pareto-optimal front (POF), which then allows the user to decide, among many alternatives. If a single solution is to be selected in a multiobjective optimization problem (MOOP), at some point during the process, the decision maker (DM) has to reveal his/her preferences. Specifying these preferences a priori, i.e., before alternatives are known, often means to ask too much of the DM. On the other hand, searching for all nondominated solutions as most MOEA (a posteriori) do may result in a waste of optimization efforts to find solutions that are clearly unacceptable to the DM. This study introduces an intermediate approach, that asks for partial preference information from the DM as a priori, and then focus the search (using a posteriori) to those regions of the POF that seem most interesting to the DM. In this way, it is possible to provide a larger number of relevant solutions. The DM or user generally, has at least a vague idea about what kind of solutions might be preferred. If such information (preference) is available, it can be used to focus the search, yielding a more fine-grained approximation of the most relevant (from a DM's perspective) areas (regions) of the POF. A novel approach, named as multiobjective evolutionary algorithm based interactive desirability function approach (MOEA-IDFA), to guide the POF into interesting regions is developed. A set of Paretooptimal solutions is determined via desirability functions (DFs) which reveals DM's preferences regarding different objective regions. The proposed method would be highly effective in generating a compromise solution that is faithful to the DM's preference structure. Theoretical analysis of the methodology is presented to assure the effectiveness of the proposed approach. We apply the proposed approach to numbers of test problems as well as some real life problems having different complexities. It is observed that in almost all cases the proposed approach efficiently guides the population towards the interesting region/regions, allowing a faster convergence and a better coverage of this/these) area/areas of the POF. The idea here is to take the desires of the DM into account more closely when foretelling the biasness onto the set of nondominated solutions. In this way we can create a decision support system (DSS) for the DM to help him/her finding the most satisfactory solution faster. We develop different combination of DFs depending upon the choice of DM and demonstrate these cases with examples. As the approach is MOEA based to validate the proposed approach two different MOEAs: NSGA-II (elitist nondominated sorting genetic algorithm) and MOPSO-CD (multi-objective particle swarm optimization with crowding distance) are presented. An apparent evidence of the efficiency of the proposed ideas is presented via summary of the results of extensive computational tests that have been done in the present thesis. This thesis is described in two parts. The first part deals with development of methodologies (Chapters 2, 3, 4 and 5) and the second part deals with their applications to real world reliability engineering problems (Chapter 6). Conclusions and future scope are summarized in Chapter 7. Acknowledgements First and foremost, I would like to thank my supervisors and mentors Dr. Shiv Prasad Yadav, Associate Professor, Department of Mathematics, Indian Institute of Technology Roorkee and Dr. Surendra Kumar, Assistant Professor, Department of Electrical Engineering, Indian Institute of Technology Roorkee. I feel privileged to express my sincere regards and gratitude to my guides for their valuable guidance and constant encouragement throughout the course of my research work. I express my earnest regards to Prof. & Head Rama Bhargava, Department of Mathematics, Indian Institute of Technology Roorkee for providing valuable advice, computational and other infrastructural facilities during my thesis work. I also would like to thank Prof T. R. Gulati, DRC Chairman, Prof. G. S. Srivastava, former DRC Chairman, my SRC members Dr. N. Sukavanam and Prof. R.S. Anand for their guidance, cooperation, and many valuable comments dedicated to my thesis. The encouragement, support and cooperation which I have received from my friends Komal, Ashok, Jagdish, Deepmala, Gaurav, Monika, Anupam, Kavita, Mohit, Jai Prakash, Sanjeev, Zeyauddin, Alok, Jaideep, Saif, Sangeeta, Mukesh, Abishek, Neeraj, Karunesh, Manjit, Rajni, Suraj and Prabhanjan are beyond the scope of my acknowledgement, yet I would like to express my heartfelt gratitude to them. I owe my sincere thanks to my family members and relatives for their blessings, patience and moral support. I would like to give my special thanks to my brother Lovekush and sisters Sneha and Rakshita. I also want to express my appreciation to my wife Seema for encouraging me to achieve my goals. Above all, I express my deepest gratitude to my Parents and my newly born daughter Aashi, to whom I dedicate this thesis. The financial assistance from Council of Scientific and Industrial Research (CSIR), New Delhi, India, is also gratefully acknowledged. Finally, my greatest regards to the Almighty for bestowing upon me the courage to face the complexities of life and complete this thesis successfully. Kishor) Roorkee December 27, 2010 iv List of Publications Journal Papers (J1) A multi-objective genetic algorithm for reliability optimization problem, International Journal of Performability Engineering, 5 (3), 227-234, April 2009. (J2) Interactive Fuzzy Multi-Objective Reliability Optimization Using NSGA-II, OPSEARCH, Springer publication .46, 214-224, June 2009. (J3) Incorporating Preferences in Multi-objective optimization problems: A novel Approach, Communicated in The Journal of Information & Optimization Sciences. (J4) Guiding MOEA towards interesting regions: A DF based Approach, Communicated in International Journal of Approximate Reasoning. (J5) Introducing Bias among Pareto-optimal Solutions: Reliability Optimization Application, Communicated in International Journal of Quality & Reliability Management (J6) Interactive Trade-off using desirability function and Multi-objective. Evolutionary Algorithm, Communicated in Decision Support Systems. (J7) MOEA-IDFA: Multi-objective evolutionary algorithm based interactive desirability Function Approach, Communicated in Journal of Computational Methods in Sciences and Engineering Conference Papers (C1). Application of a Multi-objective Genetic Algorithm to Solve Reliability Optimization Problem, Conference on Computational Intelligence and Multimedia Applications, 1,458 — 462, held at Sivakasi, Tamilnadu, 13-15 Dec. 2007. ieeexplore.ieee.oreie15/4426318/4426531/04426622.pdf?arnumber=4426622 (C2). Complex bridge system bi-objective reliability optimization problem using NSGA-II, Proceedings of 32th National Systems Conference (NSC-2008), 635-639, organized by I.I.T. Roorkee, Roorkee, 17-19 Dec 2008. ( vi ) Table of Contents Abstract Acknowledgements iii List of Publications Table of Contents vii List of Figures xi List of Tables xvii List of Abbreviations xxi CHAPTER 1 Introduction 1 1.1 MOOP (Basic Concepts and Terminology) 5 1.2 Classification and Review of MOEAs 10 1.2.1Why Evolutionary Approach to MOOP 11 1.2.2A Priori Preference Articulation: (decide -+ search) 13 1.2.3Progressive Preference Articulation: (decide <--> search) 15 1.2.4A Posteriori Preference Articulation: (search —> decide) 16 1.3 DM's Partial Preference Articulation with MOEA -A Review 24 1.3.1Approaches Providing Reference Point 25 1.3.2Approaches Based on Trade-off Information 27 1.3.3Approaches Based on Marginal Contribution 28 1.3.4Approaches Based on Scaling 28 1.3.5Other Approaches 30 1.4 Objectives of the Thesis 32 1.5 Organization of the Thesis 33 CHAPTER 2 Articulation of an a Priori Approach with an a Posteriori Approach 39 2.1 Introduction 39 2.2 DFA as a Priori 41 2.2.1Linear DF 42 2.3 Description of MOEAs Applied as a Posteriori 43 2.3.1NSGA-II or Elitist Nondominated Sorting Genetic Algorithm 43 2.3.2Run Time Complexity of NSGA-II 46 2.3.3MOPSO-CD or Multi-objective Particle Swarm Optimization 47 with Crowding Distance 2.3.4Run Time Complexity of MOPSO-CD - 48 2.3.5Constraint Handling in NSGA-II and MOPSO-CD 49 2.3.6Performance Measure for NSGA-II and MOPSO-CD 49 2.4 Proposed Methodology 50 2.4.1Assumptions: 53 2.4.2Detailed Procedure of the Methodology: MOEA-IDFA 53 2.5 Experimental Suite 55 56 2.6 Results and Discussion 57 2.7 Conclusion CHAPTER 3 Guided POF Articulating Nonlinear DFA with an MOEA 75 (Convex-Concave Combination) 3.1 Introduction 75 3.2 Nonlinear (Convex /Concave) DFA as a Priori 77 80 3.3 Proposed Methodology 3.3.1Detailed Procedure of the Methodology: MOEA-IDFA 82 3.4 Results and Discussion 83 3.4.1Effect of Variations in DF's Key Parameter on POF 85 3.5 Conclusion 85 CHAPTER 4 Guided POF Articulating Nonlinear DFA with an MOEA (All Sigmoidal Combination) 99 4.1 Introduction 99 102 4.2 Nonlinear (Sigmoidal) DFA as a Priori 103 4.3 Proposed Methodology 4.3.1Detailed Procedure of the Methodology: MOEA-IDFA 105 106 4.4 Results and Discussion 4.4.1Effect of Variations in DF's Key Parameter on POF 108 108 4.5 Conclusion CHAPTER 5 Guided POF Articulating Nonlinear DFA with an MOEA (All 121 Convex Combination) 121 5.1 Introduction 123 5.2 Nonlinear (Convex) DFA as a Priori 124 5.3 Proposed Methodology 5.3.1Detailed Procedure of the Methodology: MOEA-IDFA 126 128 5.4 Results and Discussion 5.4.1Effect of Variations in DF's Key Parameter on POF 129 130 5.5 Conclusion CHAPTER 6 Application of the MOEA-IDFA to Reliability Optimization Problems 143 143 6.1 Reliability Optimization (An Overview) 6.1.1Preference Incorporation in Reliability Optimization Problems 146 6.2 Reliability Optimization of a Series System 152 152 6.2.1Problem Description 6.2.2Step-by-Step Illustration of MOEA-IDFA for Series System 152 6.3 Reliability Optimization of Life Support System in a Space Capsule 154 6.3.1Step-by-Step Illustration of MOEA-IDFA for Life Support 154 System in a Space Capsule 6.4 Reliability Optimization of a Complex Bridge System 156 6.4.1Step-by-Step Illustration of MOEA-IDFA for Complex Bridge System 156 ( viii ) 6.5 Residual Heat Removal (RHR) System of a Nuclear Power Plant 157 Safety System 6.5.1Step-by-Step illustration of MOEA-IDFA for RHR System 159 6.6 Reliability Optimization of a Mixed Series-Parallel System 160 6.6.1Step-by-Step Illustration of MOEA-IDFA for Mixed Series161 Parallel System 162 6.7 Conclusion 180 CHAPTER 7 Conclusions and Scope for Future Work 180 7.1 Conclusions 182 7.2 Future Scope 184 BIBLIOGRAPHY ix (x) List of Figures Figure 1.1 Different Ways of Preference Articulation 35 Figure 1.2 An example to travel bigger distances in ways that are more economical 36 Figure 1.3 Description of Dominated and Nondominated Solutions 36 Figure 1.4 Preference Based Classification of MOEA 37 Figure 1.5 Nondominated Ranking of Search Space for all Minimization Case 38 Figure 1.6 Pseudo code for MOPSO 38 Figure 2.1 STB type of DF 58 Figure 2.2 LTB type of DF 58 Figure 2.3 Description of Crowding-Distance 58 Figure 2.4 The Hypervolume Enclosed by the Nondominated Solutions 59 Figure 2.5 Nondominated Sorting of a Population 59 Figure 2.6 Flow Chart Representation of NSGA-II Algorithm 60 Figure 2.7 Flow Chart of MOPSO-CD algorithm 61 Figure 2.8 Flow Chart of the Procedure Proposed: MOEA-IDFA 62 Figure 2.9 POFs of SCH1w.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 65 Figure 2.10 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 66 Figure 2.11 POFs of KUR w.r.t. NSGA-H and MOPSO-CD using MOEA-IDFA (Linear DF Case) 67 Figure 2.12 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 68 Figure 2.13 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 69 Figure 2.14 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 70 Figure 2.15 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 71 Figure 2.16 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 72 Figure 2.17 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 73 Figure 2.18 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 74 Figure 3.1 Shape of a Convex DF 79 Figure 3.2 Shape of a Concave DF 79 Figure 3.3 POFs of SCHlw.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 87 Figure 3.4 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 88 Figure 3.5 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using 89 MOEA-IDFA (Convex-Concave Combination) Figure 3.6 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 90 Figure 3.7 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 91 Figure 3.8 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 92 Figure 3.9 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 93 Figure 3.10 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 94 Figure 3.11 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 95 Figure 3.12 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 96 Figure 3.13 Effect of variations in key parameter of DF on POF for SCH1 Figure 4.1 Proposed STB type of Sigmoidal DF 97 102 Figure 4.2 POFs of SCHlw.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 110 Figure 4.3 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 111 Figure 4.4 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 112 Figure 4.5 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 113 _ Figure 4.6 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 114 Figure 4.7 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 115.. Figure 4.8 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 116 Figure 4.9 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 117 Figure 4.10 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 118 Figure 4.11 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 119 Figure 4.12 Effect of variations in key parameter of DF on POF for SCH1 Figure 5.1 STB type of a Convex DF 120 124 Figure 5.2 POFs of SCH1 w.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 131 Figure 5.3 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 132 Figure 5.4 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 133 Figure 5.5 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 134 Figure 5.6 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 135 Figure 5.7 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 136 Figure 5.8 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 137 Figure 5.9 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 138 Figure 5.10 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 139 Figure 5.11 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 140 Figure 5.12 Effect of variations in key parameter of DF on POF for 141 SCH1 Figure 6.1 Block Diagram of Series System 164 Figure 6.2 Block Diagram of a Life Support System in a Space Capsule 164 Figure 6.3 Block Diagram of Complex Bridge System 165 Figure 6.4 Mixed Series-Parellel System 165 Figure 6.5 Schematic of the RHR of a Nuclear Power Plant 166 Figure 6.6 Simplified Fault Tree of the RHR system (Apostolakis, 167 1974) Figure 6.7 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System (No Preference Case) 167 Figure 6.8 POFs w.r.t. NSGA-II and MOPSO-CD of a Life Support System in a Space Capsule (No Preference Case) ( xiv ) 168 Figure 6.9 POFs w.r.t. NSGA-II and MOPSO-CD of a Complex Bridge System (No Preference Case) 168 Figure 6.10 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences:(a) For Preference 1, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 170 Figure 6.11 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at Different Preferences: (a) For Preference 1, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 172 Figure 6.12 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: Preference 1(a), Preference 2 (b), Preference 3 (c), Preference 4 (d) 174 Figure 6.13 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences:(a) For No Preference, (b) For Preference 1, (c) For Preference 2, (d) For Preference 4 176 Figure 6.14 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: (a) For No Preference, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 178 xv List of Tables Table 2.1 Description of unconstrained Bi-objective Problems 63 Table 2.2 Description of Constrained Bi-objective and Unconstrained Tri-objective Problems 64 Table 2.3 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (Linear DF Case) 65 Table 2.4 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (Linear DF Case) 66 Table 2.5 Parameters and Hypervolumes for KUR using MOEA-IDFA (Linear DF Case) 67 Table 2.6 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (Linear DF Case) 68 Table 2.7 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (Linear DF Case) 69 Table 2.8 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (Linear DF Case) 70 Table 2.9 Parameters and Hypervolumes for TNK using MOEA-IDFA (Linear DF Case) 71 Table 2.10 Parameters and Hypervolumes for VNT using MOEA-IDFA (Linear DF Case) 72 Table 2.11 Parameters and Hypervolumes for MHHM1 using MOEAIDFA (Linear DF Case) 73 Table 2.12 Parameters and Hypervolumes for MHHM2 using MOEAIDFA (Linear DF Case) 74 Table 3.1 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (Convex-Concave Combination) 87 Table 3.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (Convex-Concave Combination) 88 Table 3.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (Convex-Concave Combination) () 89 Table 3„4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (Convex-Concave Combination) 90 Table 3.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (Convex-Concave Combination) 91 Table 3.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (Convex-Concave Combination) 92 Table 3.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (Convex-Concave Combination) 93 Table 3.8 Parameters and Hypervolumes for VNT using MOEA-IDFA (Convex-Concave Combination) 94 Table 3.9 Parameters and Hypervolumes for MHHM1 using MOEAIDFA (Convex-Concave Combination) 95 Table 3.10 Parameters and Hypervolumes for MHHM2 using MOEAIDFA (Convex-Concave Combination) 96 Table 4.1 Parameters and Hypervolumes for SCH1 using MOEA-IDFA 110 (All Sigmoidal Combination) Table 4.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (All Sigmoidal Combination) 111 Table 4.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (All Sigmoidal Combination) 112 Table 4.4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (All Sigmoidal Combination) 113 Table 4.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (All Sigmoidal Combination) 114 Table 4.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (All Sigmoidal Combination) 115 Table 4.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (All Sigmoidal Combination) 116 Table 4.8 Parameters and Hypervolumes for VNT using MOEA-IDFA (All Sigmoidal Combination) 117 Table 4.9 Parameters and Hypervolumes for MHHM1 using MOEAIDFA (All Sigmoidal Combination) 118 Table 4.10 Parameters and Hypervolumes for MHHM2 using MOEAIDFA (All Sigmoidal Combination) 119 Table 5.1 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (All Convex Combination) 131 Table 5.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (All Convex Combination) 132 Table 5.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (All Convex Combination) 133 Table 5.4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (All Convex Combination) 134, Table 5.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (All Convex Combination) 135 Table 5.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (All Convex Combination) 136. Table 5.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (All Convex Combination) 1371 Table 5.8 Parameters and Hypervolumes for VNT using MOEA-IDFA (All Convex Combination) 138 Table 5.9 Parameters and Hypervolumes for MHHM1 using MOEAIDFA (All Convex Combination) 139 Table 5.10 Parameters and Hypervolumes for MHHM2 using MOEAIDFA (All Convex Combination) 140 Table 6.1Data for Mixed Series-Parallel System 161 Table 6.2 Third order minimal cut sets 166 Table 6.3 Initial a Priori Parameters for Series System 169 Table 6.4 Other a Priori Parameters for Series System 169 Table 6.5 A Posteriori Parameters for Series System 169 Table 6.6 Initial a Priori Parameters for Life Support System in a Space Capsule 171 ( xix ) Table 6.7 Other a Priori Parameters for Life Support System in a Space Capsule 171 Table 6.8 A Posteriori Parameters for Life Support System in a Space Capsule 171 Table 6.9 Initial a Priori Parameters for Complex Bridge System 173 Table 6.10 Other a Priori Parameters for Complex Bridge System 173 Table 6.11 A Posteriori Parameters for Complex Bridge System 173 Table 6.12 Initial a Priori Parameters for RHR System 175 Table 6.13 Other a Priori Parameters for RHR System 175 Table 6.14 A Posteriori Parameters for RHR System 175 Table 6.15 Initial a Priori Parameters for Mixed Series-Parallel System 177 Table 6.16 Other a Priori Parameters for Mixed Series-Parallel System 177 Table 6.17 A Posteriori Parameters for Mixed Series-Parallel System 177 XX List of Abbreviations ACO Ant colony optimization CFO Central force optimization DE Differential evolution DF Desirability function DFA Desirability function based approach DM Decision maker DS S Decision support system EA Evolutionary algorithm GA Genetic algorithm GS() Glowworm swarm optimization IDFA Interactive desirability function approach IIMOM Intelligent interactive multi-objective optimization method LTB Larger the better MCDM Multi-criteria decision-making MCS Minimal cut sets MMP Multi-objective mathematical programming MOEA Multi-objective evolutionary algorithm MOEA-IDFA Multi-objective evolutionary algorithm based interactive desirability function approach MOGA Multi-objective genetic algorithm MOLP Multi-objective linear problem MOPSO Multi-objective particle swarm optimization MOPSO-CD Multi-objective particle swarm optimization with crowding distance MOOP Multi-objective optimization problem MOOT Multi-objective optimization technique NLMOOP Nonlinear multi-objective optimization problem NSGA Nondominated sorting genetic algorithm NPGA Niched Pareto genetic algorithm NSGA-II Elitist nondominated sorting genetic algorithm OR Operations research POF Pareto-optimal front PSO Particle swarm optimization RHR Residual heat reservoir SOOP Single objective optimization problem SPEA Strength Pareto evolutionary algorithm STB Smaller the better CHAPTER 1 Introduction MOOP is an important class of problem copiously encountered in engineering and industrial context. Many problems addressed in classical single objective models are actually multi-objective in nature, reason being that the outcomes associated with the decisions are multidimensional. Criticism over the use of a single criterion (objective) as a sole basis for decision-making in such cases necessitates a need for the explicit treatment multiple measures of solution quality or multiple objectives in decisionmaking problems. Multi-criteria decision-making (MCDM) encompasses all the quantitative decision problems characterized by multiple measures of solution quality. There is no actual decision-making involved in single objective optimization problem (SOOP), as the decision is implicit in the measurement of scalar-valued objective function. We can think of decision-making only when multiple objectives, criteria, functions, etc. are involved as the alternatives of choice become more complex and conflicting in nature. Thus, the problem of combining these objectives into a single objective (measure) becomes difficult and some times impractical. If a single measure such as the weighted average of individual, objectives or a value function can be found representing a DM's preference structure, the multi-objective mathematical programming (MMP) can be recasted into a SOOP. Thus, there would be no need for multi-objective optimization techniques (MOOTs) in this case. The weighted average is often difficult to find due to the non-commensurability of individual objective functions and value function is difficult to determine in real situations. To resolve these problems a large number of MOOTs have been developed in literature that can identify the efficient alternatives. The concept of optimality in SOOP is replaced by that of efficiency in MOOP. An efficient solution (Pareto-optimal solution) is one for which there does not exist another feasible solution which does at least as well on every single objective, and better on at least one objective. Generating efficient solutions to MOOP is an area of increased interest by researchers in many disciplines, including Engineering, Operations Research (OR), and Computer Science to name a few. When solving real-world optimization problems, classical methods encounter great difficulty in dealing with the complexity involved in these situations and cannot offer a reliable solution. One can find many real applications in fields such as economics, engineering or science where methods with ample mathematical support (ensuring the optimality of solutions under ideal conditions) are unable to obtain a solution or cannot obtain the solution in a reasonable time. These facts led researchers to develop evolutionary algorithms (EAs) to solve these complex models. The success of EAs produced enormous interest in their study giving rise to an active community and a number of very efficient MOEAs for MOOR The use of EAs to solve problems of this nature has been motivated mainly because of the population-based nature of EAs, which allows the generation of several elements of the Pareto-optimal set in a single run. In addition, the complexity (e.g. very large search space, uncertainty, noise, disjoint POFs, etc.) of some MOOPs may prevent use (or application) of traditional OR solution techniques. Thanks to past two decades of- research work, MOEA is now a well-established and very popular computational research area. Several evolutionary methods are available ensuring full convergence toward the POF in terms of both precision and diversity of solutions (e.g. NSGA-H, MOPSO, MOPSO-CD, SPEA2 etc.). These methods have been widely and deeply tested and compared on different standard test functions. In addition some convergence measuring criteria are also available, being specifically developed for MOOPs (Thanh and Vong, 2000, Parsopoulos and Vrahatis, 2002; Zitzler et al., 2003; Tan et al., 2005; Coello, 2009; Nguyen, 2010). Typically, there are infinitely many Pareto-optimal solutions for a MOOP. Merely, determination of the efficient solutions does not solve the problem completely. Mathematically, every Pareto-optimal point is an equally acceptable solution of the MOOP. However, it is generally desirable to obtain one point as a solution. Selecting one out of the set of Pareto-optimal solutions calls for information that is not contained in the objective functions. Thus, it is often necessary and important issue to incorporate DM's preferences for various objectives in order to determine a suitable (final) solution. The DM is a person (or group of persons) who is supposed to have better insight into the problem and who can express preference relations between different solutions. Solving a MOOP calls for the co-operation between the DM and analyst. By 2 an analyst here, we mean a person or a computer program responsible for the calculation/computation side of the solution process. The analyst generates information for the DM and the solution is selected according to the preferences of DM. In addition, the set is partially ordered and unlike SOOP, no analytical tool can identify the best alternative among these without additional information in the form of subjective preferences of a DM. The preferred solution is then called the best compromise solution. Thus, the two important parts of MCDM problems are: i) An objective part handled by the analyst , and ii) A subjective part handled by the DM. The objective part considers the internal structure of the system characterized by constraints together with the functional relationship between decision variables and decision criteria and based on it the objective part sorts out the efficient alternatives. Then the subjective part takes over using the preferences of the DM to develop a preference ordering relation, which results in a complete ordering of the efficient set, thus, determining the best alternative-best in terms of some criteria of judgment known to the DM. Therefore, interaction with DM is an integral part of the algorithms for MOOP at some point during the optimization process. Following a classification by Horn (1997) and Van Veldhuizen and Lamont (2000) the articulation of preferences with MOEAs may be done either before (a priori), during (progressive), or after (a posteriori) the optimization process. A priori MOEAs involves preference specification prior to the optimization stage, and are traditionally implemented by aggregating objectives into a single fitness function with parameters reflecting the preference of the DM. Interactive MOEA allow the DM to alter parameters during the search, effectively influencing the direction of the search. A posteriori approach is to date the most popular, which allows the DM to choose suitable solution out of the Pareto-optimal solutions presented before DM. In recent years, researchers have started to look into incorporating preference in the search process of an MOEA. Development in this area is interesting and the key element to widespread application of MOEA in practical circumstances where the preferences are incorporated. There are some advantage of knowing the range of each objective for Pareto-optimality and the shape of the POF itself in a problem for an adequate decisionmaking. The task of choosing a preferred single Pareto-Optimal solution is also another 3 extremely important issue. Most of the MOEAs focus on the approximation of the POF without including DM's preferences. However, the determination or approximation of the POF is not enough, and the DM's preferences have to be incorporated in order to determine the solution that better represents these preferences. Some works are found in literature in which DM's preferences are incorporated along with MOEA. Deb and Sundar (2006) in their paper concluded "having been well demonstrating the task of finding multiple Pareto-optimal solutions in MOOPs, the MOEA researchers and applicationists should now concentrate in devising methodologies of solving the complete task of finding preferred Pareto-optimal solutions in an interactive manner with the DM". Hence, researchers entail to work in this very essential area. In this thesis, we articulate desirability function (DF) based approach (a priori) with an MOEA (a posteriori), and demonstrate how, instead of one solution, a preferred set of solutions near the desired region of DM's interest can be found. Thus, a hybrid approach consisting a priori and a posteriori together is proposed in the present work. In other words, we consider an intermediate approach shown through the middle path in Figure 1.1. It may be impractical for a DM to completely specify his/her preferences before any alternatives are known. However, often the DM has at least a vague idea about what kind of solutions might be preferred and can specify partial preferences before the search process. If such information is available, it can be used to focus the search, yielding a more fine-grained approximation of the most relevant (from a DM's perspective) area of the POF and/or reducing computational time. Thus, the goal is no longer to generate a good approximation of all Pareto-optimal solutions, but a small portion of set of Pareto-optimal solution that contains the DM's preferred solution with the highest probability. In our methodology, preferences can be effectively incorporated to an MOEA (a posteriori) with the help of DF as a priori approach (Harrington, 1965; Mehnen et al., 2007). In this thesis we do not lay emphasis on any new efficient set generation algorithms since, we already have a lot (e.g. NSGAII and MOPSO-CD etc.). Aim of the present work is to obtain a guided or partial POF, through an interactive procedure involving DM. Thus, the main core of this thesis is the utilization of an MOEA in finding preferred solution from the POF in the region/regions that are of interest to the DM. We also present a basic theoretical 4 analysis and application of the proposed approach to five different reliability optimization problems. In this chapter, we first give the basic concepts and terminology regarding MOOP (Section 1.1) followed by a classification and review of MOEA, provided in Section 1.2. In section 1.3, the literature dealing with DM's partial preference articulation into MOEAs is reviewed and classified. Section 1.4 presents objectives of the thesis and the organization of the thesis is finally presented in Section 1.5. 1.1 MOOP (Basic Concepts and Terminology) MOOPs can be found everywhere in nature, and we deal with them on a daily basis. In MOOP, there is not a single solution for a given problem; instead, there is a set of solutions from where one can choose. From a person who tries to optimize a budget in a supermarket, trying to get more and better quality products for less amount of motley; industries trying to optimize their production, reducing their production costs. and increasing their quality or people looking for ways that are more economical to travel and covering bigger distance. In the last example, which means of transport we should choose depends on how far we need to go or how cheap we need it to be as shown in Figure 1.2. Although genesis of MOOP is in economics, it has been studied through several disciplines e.g. game theory, OR etc. The idea of solving a MOOP can be understood as helping a human DM in considering the multiple criteria simultaneously and in finding a Pareto-optimal solution that pleases him/her the most. The notion of an efficient solution was introduced by Pareto (1896) hence, the connotation of an efficient solution is also named as Pareto-optimal solution. However, the earliest concept of MCDM appeared with the advances of OR following World War-II. Decision-making in complex environment involves terms such as 'multiple objectives', 'multiple attributes', 'multiple criteria' or 'multiple dimensions' are used to describe different decision situations (Collette and Siarry, 2003). A common feature of these problems is that they consider multiple measures of solution quality. From now, onwards we will use multi-objectives in place of these terms. A general MOOP formulation in standard form is as follows: Minimize (Maximize) f (x){f(x), f2 (x),..., fk (x)} (1.0) subject to : g (x) = 0, j -= 1,..., me ; (1.1) XE R" (1.2) g (x) 5 0, j = me +1, ..., m, where, k 2 is number of objectives in the MOOP; m is the total number of constraints while me is the number of equality constraints, x = (x1 , x2 , x„) is n fromsomeuniverse S2 c le , objective dimensionaldecisionvariable functions f (x), i =1, 2, ..., k where, f, : S2 —> R and the constrained functions g, (x) where, g : S2 —> R are all real valued functions on S2 , f(x) is multi-objective vector or criteria vector of objective functions. When all f 's and gi 's are linear, the problem is called a multi-objective linear problem (MOLP). If at least one of the f,' s is nonlinear the problem is nonlinear multi-objective optimization problem (NLMOOP). Definition 1.1 (Feasible Decision Space): The vector x E le is said to be feasible if and only if Equations 1.1 and 1.2 hold. The set of all feasible vectors is said to be a feasible decision space X (often called feasible design space or constrained set) given as { g ,(x) = 0, j =1,..., me ; and X = x ' g , (x) ... 0, j = me + 1, ..., m. ( 1.2 ) Therefore, we can rewrite the MOOP given by Equations 1.0-1.2 as: x Minimize f (x) xEx P1 which means the solution of constrained optimization problem given by Equations 1.01.2 is just to find a vector x E X, such that the objective function f (x) is minimized (or maximized). From now, onwards we will take P1 as a minimization problem. Definition 1.2 (Feasible Criterion Space): The feasible criterion space Z (also called the attainable set) is defined as the set {f (x) x E X} . 6 Definition 1.3 (Ideal Solution): Let us look at the problem P1 as k SOOPs, each with a different objective function but with the same constraints. The k SOOPs are given by Minimize f (x) subject to x c X Let x ( ,i =1, , i =1,2,...,k (1.3) k be the points where the minimum value of f; (say f,t ) are respectively achieved. We call xt(') the attainable solutions of the original MOOD (i.e. P1) if *(2) = x •(k) = x* , X *(1) = X = That is, if the minimum values of all f are achieved at the same point x* , then x t is called the ideal solution of the problem P1 . An ideal solution is the solution with all the objective functions simultaneously minimized. In a real life problem, such a situation would be very rare. In the absence of an ideal solution, we may prefer a solution regarded best by some other suitable criterion. One such criterion is to look for a point from where the value of any of the objective functions cannot be decreased without increasing the value of at least one of the other objective functions. Such a point is called an efficient point, and the corresponding solution an efficient solution (will be discussed elaborately in Section 1.2). Definition 1.4 (Ideal Objective Vector): The ideal vector f t is the unique k -vector with components Minimize f(x) , i =1, 2, . . , k f = subject to x e X (1.4) obtained by minimizing each objective function separately i.e., f t = (fi ,fkt ) is called the ideal objective vector. Ideal objective vector f t can be used as a reference solution for the algorithms seeking Pareto-optimal solution. Definition 1.5(Anti-ideal Vector): Unlike the ideal objective vector which represents the lower bound of each objective in the entire feasible search space X , the anti-ideal (nadir) objective vector, ztt , represents the upper bound (in case of all minimization 7 problem) of each objective in the entire Pareto-optimal set, and not in the entire search space E-2 * ** ** * * zz'» = f =(f1 , f 2 fk** ), where the maximum solution for the i t h objective function is the decision vector x**(') with the function value f". Definition 1.6 (Dominance): A solution x(I) = (41) , 4),..., x„(1) ) is said to dominate the other solution x(" = (x,2), x?),...., x(2)) if 1. x(I) is not worse than x(2) w.r.t. all objectives i.e., f (x(1) ) f (x(2) ) for all i = 1, 2, ..., k. 2. x(1) is strictly better that x(2) in at least one objective i.e., f (x0)) < f, (x(2) )for at least one i If any of the above condition is violated, the solution x(I) does not dominate the solution x(2) (Deb, 2001). If x(I) dominates x(2), it is also customary to write any of the following: • x(2) is dominated by x(1) ; • x(I) is non-dominated by x(2) , or; • x(1) is non-inferior to x(2) . The depiction of dominance can be visualized using Figure 1.3. Definition 1.7 (Efficient or Pareto-optimal solution): A solution x* E X of P1 is called an efficient solution or Pareto-optimal solution if there is no other point x E X such that for one objective function, say f (x), f ,. (x) < f (x*), and for all other fj (x) < fi (x* ), j 1, 2, ..., k; j The concept of an efficient solution for a MOOP may be regarded as a generalization of the concept of an optimal solution for a SOOP. The inequality signs in the above definition have to be reversed if the problem is to maximize all the objectives. The set P* of all Pareto-optimal solutions of a given MOOP is called the Pareto-optimal set. i.e. P* = (x E X :x is a Pareto-optimal solution) . 8 Definition 1.8 (POF): There are usually a lot (infinite number) of Pareto-optimal solutions for a given MOOP and Pareto-optimal set /3* , the POF (PF* ) is defined as image elements of P* under f, i.e. PF* = If (x) = (fi (x), f 2 (x),..., fk (x)) x E P*1 (1.5) There are generally convex and nonconvex POFs. A POF is said to be convex if and only if V f (1) (x), f (2) (x) E PF* ,V 2 E (0,1), 3 f (3) (x) E PF*such that 2 1f (I) (x)11+ (1 — 2)11f (2) (x)11-f (3) (x) On the contrary a POF is said to be concave if and only if V f (1) (x), f (2) (x) E PF* ,V 2 11-f (" (x)II + (1 2)11f (2) (x) II E (0,1), 3 f (3) (x) E PF*such that f (3) (x)II' Apart from the above two, partially convex, partially concave and discontinuous POFs may also exist. The ultimate goal of a multi-objective algorithm is to identify solutions in the Paretooptimal set. However, identifying the entire Pareto-optimal set for a MOOP is practically impossible due to its large size. In addition, for many problems, especially combinatorial optimization problems, proof of solution optimality is computationally infeasible. Therefore, a practical approach to multi-objective algorithm is to investigate a set of solutions that represent the Pareto-optimal set. With these concerns in mind, a MOOT should achieve the following three conflicting goals: 1. The best-known POF should be as close as possible to the true POF. 2. Solutions in the best-known POF should be uniformly distributed and diverse over of the POF in order to provide the DM a true picture of trade-offs. 3. The best-known POF should capture the whole spectrum of the POF. This requires investigating solutions at the extreme ends of the objective function space. Definition 1.8 (Strongly, Weakly, Inferior and Preferred solution): All Pareto- optimal solutions lie on the boundary of the feasible criterion space Z (Athan and Papalambros, 1996). Often algorithms provide solutions that may not be Pareto-optimal but may satisfy other criteria, making them significant for practical applications. A 9 point x* E X is weakly Pareto-optimal if and only if there does not exist another point x E X , such that f (x) < f (x`) . In other words, a point is weakly Pareto-optimal if there is no other point that improves the entire objective functions simultaneously. In contrast, a point is strongly Pareto-optimal if there is no other point that improves at least one objective function without detriment to another function. A solution which is not weakly efficient is called an inferior solution. A particular efficient solution, which is finally selected by the DM after preference decision-making is known as preferred solution. Pareto-optimal solutions can be divided into improperly and properly Paretooptimal ones according to whether unbounded trade-offs between objectives are allowed or not. Proper Pareto-optimality can be defined in several ways (Yu et al., 1985; Miettinen, 1999). According to Geoffrion (1968) a solution is properly Paretooptimal if there is at least one pair of objectives for which a finite decrement in one objective is possible only at the expense of some reasonable increment in the other objective. Mathematically it is defined as follows: Definition 1.9(Geoffrion's Proper Pareto-Optimality): A point x' e X is proper Pareto-optimal solution if it is Pareto-optimal and there is some real number M > 0 such that for each f;(x) and each x E X satisfying f(x ) < f(x ) , there exists at least one /es, (x) ( j i ) such that fi (x* ) < (x ) and fi (x* ft (x ) < M. f.,(.) - 4(x) If a Pareto-optimal solution is not proper, it is called improper. The quotient defined above is referred to as a trade-off, and it represents the increment in objective function j resulting from a decrement in objective function i (Geoffrion, 1968). 1.2 Classification and Review of MOEAs Before proceeding towards the classification and review of MOEA, one important aspect needs to be discussed given in the following subsection. 10 1.2.1 Why Evolutionary Approach to MOOP ? The major part of earlier mathematical research has concentrated on optimization problems where the functions were linear, differentiable, convex, or otherwise mathematically well behaving. However, in practical problems, objective functions are often nonlinear, non-differentiable, discontinuous, multi-modal etc. and no presumptions can be made about their behaviour. Most traditional optimization methods cannot handle such complexity or do not perform in some cases in which the assumptions, upon which they are based do not hold. For such problems, stochastic optimization methods such as EAs have been implemented effectively because they do not rely upon assumptions concerning the objective and constraint functions. EAs are stochastic search and optimization heuristic derived from the classic evolution theory, working on a population of potential solutions to a problem. The basic idea is that if only those individuals reproduce, which meet a certain selection criteria, thelpopulation will converge to solutions that best meet the selection criteria. If imperfect reproduction occurs, the population can begin to explore the search space and will move to individuals (solutions) that have an increased selection probability. Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Differential Evolution (DE), Ant colony optimization (ACO) and Glowworm swarm optimization (GSO) Central force optimization (CFO) are some well-known EAs found in literature (Holland, 1987; Kennedy and Eberhart, 1995; Barbosa, 1996, 2002; Babu and Chaturvedi, 2000;Castro and Barbosa, 2001;Barbosa and Lemonge, 2002, 2003, 2008; Babu and Chaurasia, 2003; Acan, 2004, 2005; Babu and khan, 2004; Drezner et al., 2005; Salhi et al., 2005; Kennedy, 2006; Krishnanand and Ghose, 2006, 2009; Formato, 2007, 2009, 2010; Salhi and Petch, 2007; Dorigo and StUtzle, 2010). Since, EAs work on a population of solutions instead of a single point (at each iteration), hence are less likely to be trapped in a local minimum. The OR community has developed several approaches to solve MOOPs since the 1950s. Currently, wide varieties of mathematical programming techniques to solve MOOPs are available in the specialized literature (Sawaragi et al., 1985; Steuer, 1986). However, mathematical programming techniques have certain limitations when tackling MOOPs (Coello, 1999). For example, many of them are susceptible to the shape of the POF and may not work when the POF is concave or disconnected. Others require differentiability of the 11 objective functions and the constraints. In addition, most of them only generate a single solution from each run. Thus, several runs (using different starting points) are required in order to generate several elements of the Pareto-optimal set. In contrast, EAs seem particularly suitable to solve MOOPs, because they deal simultaneously with a set of possible solutions (the so-called population) which allows one to find several members of the Pareto optimal set in a single run of the algorithm. Additionally, EAs are less susceptible to the shape or continuity of the POF. This section presents background information to aid the reader understanding of the necessary prior facts supporting this thesis. A brief description of classification of MOEA is presented followed by a discussion of MOEAs and other approaches to solve MOOPs. MOEAs are contemporary algorithms receiving renewed interest from EA researchers to solve MOOPs and are part of the 'soft computing' umbrella of search algorithms.(Coello, 2004) Several substantial reviews are available for classification and solution of MOOP and examined major MOEA approaches (Fonseca and Fleming, 1995; Horn, 1997; Coello, 1999; Van Veldhuizen and Lamont, 2000; Tan et al., 2002; Coello et al., 2006; Lamont and Van Veldhuizen, 2007). Reviews include GA, PSO, DE, Evolution Strategies, Evolutionary Programming, Genetic Programming and their extension to MOEA implementations. Many researchers have attempted to classify algorithms according to various considerations. Since, as discussed earlier preference or priority is an essential part of decision-making in MOOP. Eventually, the DM should finally decide the relative importance of each objective function in order to get a single unique solution to be used as a solution of his original multidisciplinary decision-making problem. The various multiple objective decision-making techniques are commonly classified from a DM's point of view (Hwang and Masud, 1979; Van Veldhuizen and Lamont, 2000). Hwang and Masud (1979) and later Miettinen (1999) fine-tuned the earlier classifications and suggested the following three main classes based on the preference articulation of the DM: • A Priori Preference Articulation: (decide —> search) • Progressive Preference Articulation: (decide <---> search) •A Posteriori Preference ArticulatiOn: (search —> decide) 12 The multi-objective optimization includes numerous different techniques. Thus, it is hard to summarize all of them; however, to give an idea about the most common methods used in the literature, a short overview follows. For a more deep insight the interested reader may consult to (Coello et al., 2007; Coello, 2009) and the references therein included. The main approaches are described below and also shown in Figure 1.5. 1.2.2 A Priori Preference Articulation: (decide -+ search) The DM selects the weights before running the optimization algorithm. In practice, it means that the DM combines the individual objective functions into a scalar cost function (linear or nonlinear combination). This effectively converts a multi-objective problem into a single objective one. In the early stage of multi-objective optimik-ation, objectives were linearly combined into a scalar objective via a predetermined aggregating function to reflect the search for a particular solution on the trade-off surface (Jakob et al., 1992; Wilson and Macleod, 1993). The whole trade-off is then discovered by repeating the process numerous times with different settings for the aggregating function. The drawback of this approach is that the weights are difficult to determine precisely, especially when there is insufficient information or knowledge concerning the optimization problem. Other objective reduction methods include the use of penalty functions for the reduction of multi-objective optimization into a single objective (Ritzel et al., 1994). As mentioned by Fonseca and Fleming (1993)Coello (1996) these conventional multi-objective optimization approaches often have the disadvantage of missing the concave portions of a trade-off curve. An important sub classification of this approach is: • Weighted Sum Approach- In this approach different objectives are combined using weighted coefficients w, = 1,2,..., k. The objective to minimize becomes \ —11c L./J=1 w f (x) . This is one of the most popular approaches for solving MOOP, and may be simplest one (Murty, 1995). Combination used mostly is linear but non linear combination can also be used. 13 • Goal Programming Based Approach- In this approach the user is required to assign targets, or goals, 7; =1,k for each objective. The aim then becomes the minimization of the deviation from the targets to the objectives, or E,_,If,(x) — 7; I (Van Veldhuizen and Lamont, 2000;Knowles et al., 2006;Lamont and Van Veldhuizen, 2007;Jones and Tamiz, 2010). • Goal Attainment Based Approach-The user is required to provide, a vector of weights w, =1,k in addition to the vector of goals, linking the relative under- or over-attainment of the desired goals. Fonseca and Fleming (1993) were probably the first to incorporate preferences from the DM into EA and later discussed by others elaborately (Tan et al., 2002, 2005). • E — Constraint Approach- In this method, the primary objective function is minimized whereas the other objectives are treated .as constraints bound by some allowable levels e, . This technique was developed to alleviate the difficulties faced by the weighted sum approach in solving nonconvex problems (Laumanns et al., 2006; Mavrotas, 2009). • Fuzzy Based Approaches- The concept of fuzzy sets is based on a multi-valued logic where a statement could be simultaneously, partly true and partly false (Zadeh, 1975). In fuzzy logic, a membership function ,U , expresses the degree of truthfulness of a statement, in the range from p = 0 indicating that the statement is false to p =1 for truth. This is in opposite to binary logic where a statement can be only false or true. In an optimization problem, the membership function enables us to associate a normalized value to each objective p1 (f,(x)), which expresses the degree of satisfaction of the considered ith objective. The value of f , (x) is fuzzified by p, to yield a value in the range (0, 1), which quantifies how well a solution satisfies the requirements. Ones the fuzzification has been performed the actual value of each objectives is transformed into logical values. These values have to be aggregated to 14 one in order to get an overall value for the design. In binary logic this is accomplished by the AND operator. However, in fuzzy logic the AND operator could be implemented by several different rules. The most common ones are the min and the product operators. A method for finding numerical compensation for fuzzy multicriteria decision problem is demonstrated by Rao et al. (1988a). A preference structure on aspiration levels in a goal-programming problem based on fuzzy approach is illustrated by Rao et al. (1988b). Mohanty and Vijayaraghavan (1995) presented a multi-objective programming problem and its equivalent goal programming problem with appropriate priorities and aspiration levels based on fuzzy approach. Examples and details on fuzzy approaches to multi-objective optimization could be found in several works available in literature (Zimmermann, 1986, 1987, 2001; Zintmermann, 1990; Fuller and Carlsson, 1996; Wang, 2000). 1.2.3 Progressive Preference Articulation: (decide 4---> search) DM interacts with the optimization program during the optimization process. Typically, the system provides an updated set of solution and let the DM consider whether or not change the weighting of individual objective functions. They rely on progressive information about the DM's preferences simultaneously as they search through the solution space. Interactive methods are very common within the field of operations research. These methods work according to the hypothesis that the DM is unable indicate 'a priori' preferences information because the complexity of the problem. However, the DM is able to give some preference information as the search moves on. The DM then learns about the problem as he/she faces different possible problem solutions. Disadvantages of these types of methods are (Van Veldhuizen and Lamont, 1998; Adra et al.; Rachmawati and Srinivasan, 2009): • The solutions are depending on how well the DM can articulate his preferences. • A high effort is required from the DM during the whole search process. • The solution is depending on the preferences of one DM. If the DM changes his preferences or if there is a change of DM, the process has to be restarted. • The required computational effort is higher than in the previous methods. 15 1.2.4 A Posteriori Preference Articulation: (search -p decide) The DM specifies no weighting before or during the optimization process. The optimization algorithm provides a set of efficient candidate solutions from which the DM chooses the solution to be used. The big advantage of this approach is that the solution is independent of the DM's preferences. The analysis has only to be performed once, as the Pareto-optimal set would not change as long as the problem description is unchanged. However, some of these methods suffer from a large computational burden. Another disadvantage might be that the DM has too many solutions to choose. Present work is an attempt to rectify this very problem. This is the category where most MOEA approaches fall. The main approaches are described below. • Non Pareto Based Approaches- VEGA (Vector Evaluating Genetic Algorithm) was possibly the first multi-objective genetic algorithm proposed by Schaffer which incorporates a special selection operator in which a number of sub-populations were generated by applying proportional selection according to each objective function in turn (Schaffer, 1985). However, it is reported that the. method tends to crowd results at extremes of the solution space, often yielding poor coverage of the POF. Fourman (1985) presented a GA using binary tournaments, randomly choosing one objective to decide each tournament. Kursawe (1991) further developed this scheme by allowing the objective selection to be random, fixed by the user, or to evolve with the optimization process. He also added crowding techniques, dominance, and diploid to maintain diversity in the population. All of these Non-Pareto techniques tend to converge to a subset of the POF, leaving a large part of the Pareto set unexplored. • Pareto Based Approaches (First Generation) - After VEGA, researchers adopted for several years other naïve approaches. Goldberg (1989) first hinted the direct incorporation of the concept of Pareto-optimality into an EA in his seminal book on genetic algorithms (Goldberg, 1989). While criticizing Schaffer's VEGA, Goldberg suggested the use of nondominated ranking and selection to move a population towards the POF in a MOOP. The basic idea is to find the set of solutions in the population that are nondominated by the rest of the population. These solutions are 16 then assigned the highest rank and eliminated from further contention. Another set of nondominated solutions are determined from the remaining population and are assigned the next highest rank. This process continues until all the population are suitably ranked. This procedure of identifying non-dominated sets of individuals is repeated until the whole population has been ranked, as depicted in Figure 1.5. Goldberg also discussed ranking using niching methods and speciation to promote diversity so that the entire POF is covered. Goldberg (1989) did not provide an actual implementation of his procedure, but practically all the MOEAs developed after the publication of his book were influenced by his ideas. From the several MOEAs developed from 1989 to till date, the most representative are the following: The non-dominated sorting genetic algorithm (NSGA) of Srinivas and Deb (1994) implemented Goldberg's thoughts about the application of niching methods. NSGA is based on several layers of classifications of the individuals as suggesteth,by Goldberg. Before selection is performed, the population is ranked based- on nondomination: all nondominated individuals are classified into one category (with a dummy fitness value, which is proportional to the population size, to provide an equal reproductive potential for these individuals). To maintain the diversity of the population, these classified individuals are shared with their dummy fitness values. Then this group of classified individuals is ignored and another layer: , of nondominated individuals is considered. The process continues until all individuals in the population are classified. Since individuals in the first front have the maximum fitness value, they always get more copies than the rest of the population. The algorithm of the NSGA is not very efficient, because Pareto ranking has to be repeated over an over again. Evidently, it is possible to achieve the same goal in a more efficient way. Another approach under this category is multi-objective genetic algorithm (MOGA) in which an individual is assigned a rank corresponding to the number of individuals in the current population by which it is dominated increased by one. All nondominated individuals are ranked one. Fitness of individuals with the same rank is averaged so that all of them are sampled at the same rate. A niche formation method is used to distribute the population over the Pareto-optimal region (Fonseca and Fleming, 1995). Niched Pareto genetic algorithm (NPGA) was proposed by (Coello, 2004). A Pareto dominance-based tournament selection with a 17 sample of the population was used to determine the winner between two candidate solutions. Around ten individuals are used to determine dominance, and the nondominated individual selected. If both the individuals are either dominated or nondominated, then the result of the tournament is decided through fitness sharing. The main lesson learnt from the first generation MOEA was that a successful MOEA had to combine a good mechanism to select nondominated individuals (perhaps, but not necessarily, based on the concept of Pareto-optimality) combined with a good mechanism to maintain diversity (fitness sharing was a choice, but not the only one). • Pareto Based Approaches (Second Generation) - All the first generation Pareto based algorithms, are used as tools to .keep diversity in the population through the whole POF (niching technique), the fitness sharing procedure. They pursue weakness as its dependence on the fitness sharing factor. An important operator that has been demonstrated to improve significantly the performance of multi-objective algorithms is elitism, as can be seen, for example refer Goldberg and Samtani (1986). From the author's perspective, second generation of MOEA started when elitism became a standard mechanism. However, the incorporation of elitism in MOEAs is more complex than its incorporation in single objective optimization. Since, now we have an elite set which size can be important compared to the population. The elitism maintains the knowledge acquired during the algorithm execution and is materialized by preserving the individuals with best fitness in the population or in an auxiliary population. Most authors credit Zitzler and Thiele (1998) with the formal introduction of this concept in a MOEA, mainly because his Strength Pareto Evolutionary Algorithm (SPEA) made a landmark in the field. Needless to say, after the publication of this paper, most researchers of the field started to incorporate external populations in their MOEAs and the use of this mechanism (or an alternative form of elitism) became a common practice. In fact, the use of elitism is a theoretical requirement in order to guarantee convergence of a MOEA and therefore its importance. In the context of multi-objective optimization, elitism usually (although not necessarily) refers to the use of an external population (also called secondary population) to retain the nondominated individuals found 18 along the evolutionary process. The main motivation for this mechanism is the fact that a solution that is nondominated with respect to its current population is not necessarily nondominated with respect to all the populations that are produced by an evolutionary algorithm. Thus, what we need is a way of guaranteeing that the solutions that we will report to the user are nondominated with respect to every other solution that our algorithm has produced. Therefore, the most intuitive way of doing this is by storing in an external memory (or archive) all the nondominated solutions found. If a solution that wishes to enter the archive is dominated by its contents, then it is not allowed to enter. Conversely, if a solution dominates anyone stored in the file, the dominated solution must be deleted. Elitism can also be introduced using a (p+ /1)— selection in which parents compete with their children and those, which are nondominated (and possibly comply with some additional criterion such as providing a better distribution of solutions), are selected for the following generation. Many MOEAs have been proposed during the second generation (which we are still living today). However, most researchers will agree that few of these approaches have been adopted as a reference or have been used by others. The most representative MOEAs of the second generation are the following (Coello, 2004): o Strength Pareto Evolutionary Algorithm (SPEA): This algorithm was introduced by Zitzler and Thiele (1998). This approach was conceived= as a way of integrating different MOEAs. SPEA uses an archive containing nondominated solutions previously found (the so-called external nondominated set). At each generation, nondominated individuals are copied to the external nondominated set. For each individual in this external set, a strength value is computed. It is proportional to the number of solutions to which a certain individual dominates. In SPEA, the fitness of each member of the current population is computed according to the strengths of all external nondominated solutions that dominate it. The fitness assignment process of SPEA considers both closeness to the true POF and even distribution of solutions at the same time. Thus, instead of using niches based on distance, Pareto dominance is used to ensure that the solutions are properly distributed along the POF. Although this approach does not require a niche radius, its effectiveness relies on the size of the external nondominated set. In fact, since 19 the external nondominated set participates in the selection process of SPEA, if its size grows too large, it might reduce the selection pressure, thus slowing down the search. Because of this, the authors decided to adopt a technique that prunes the contents of the external nondominated set so that its size remains below a certain threshold. SPEA forms niches automatically only depending how the individuals are located in relation to each other. o Strength Pareto Evolutionary Algorithm 2 (SPEA2): SPEA2 has three main differences with respect to its predecessor (Zitzler and Thiele, 1998). First one, it incorporates a fine-grained fitness assignment strategy which takes into account for each individual the number of individuals that dominate it and the number of individuals by which it is dominated. Second one, it uses a nearest neighbor density estimation technique which guides the search more efficiently, and third one, it has an enhanced archive truncation method that guarantees the preservation of boundary solutions (Zitzler et al., 2001). o Pareto Archived Evolution Strategy (PAES): Knowles and Come (1999) introduced this algorithm. It stores the solutions of the best POF found in an external auxiliary population (elitism). A new crowding method introduced in this algorithm to promote diversity in the population. 'The objective space is divided into hypercubes by a grid, which determines the density of individuals; the zones with lower density are favored in detriment of the zones with higher density of points. This technique depends only on the parameter of number of grid divisions and is less computationally expensive than niching, avoiding the use of the fitness-sharing factor. Initially conceived as a multiobjective local search method (l+p)-PAES, it has been extended later to the (u+?)-PAES. The rank of each new created individual is set by comparing its dominance or non-dominance to the archive and also by the density of the grid they belong to. o Pareto Envelope-based Selection Algorithm (PESA): It stores the solutions of the best front found in an external auxiliary population (elitism). Not only the crowding mechanism is based on the hypercubes grid division as in PAES, but also the selection criterion is performed by this concept. In a set of test 20 functions competing with PAES and SPEA, PESA is claimed to obtain the best whole results (Come et al., 2000). o Nondominated Sorting Genetic Algorithm II (NSGA-II) or Elitist Nondominated Sorting Genetic Algorithm: It was proposed to resolve the weaknesses of NSGA (Srinivas and Deb, 1994), specially its non-elitist nature. Coello (2009) quoted "although several elitist MOEA exist, few have become widely used and among them, one has become extremely popular called NSGA-II". Deb and Goel (2001) introduced this approach. It maintains the solutions of the best front found including them into the next generation (elitism). The introduction of the controlled elitism operator in the NSGA-II algorithm produces a better equilibrium between exploitation and exploration. In NSGA-II, for each solution one has to determine how many solutions dominate it and the set of solutions to which it dominates. The NSGA-II estimates the density of solutions surrounding a particular solutio' n in the population by computing the average distance of two points on either side of this point along each of the objectives of the problem. This value is the socalled crowding distance. During selection, the NSGA-II uses a crowdedcomparison operator which takes into consideration both the nondomination rank of an individual in the population and its crowding distance (i.e., nondominated solutions are preferred over dominated solutions, but between two solutions with the same nondomination rank, the one that resides in the less crowded region is preferred). The NSGA-II does not use an external memory as the other MOEAs previously discussed. Instead, the elitist mechanism of the NSGA-II consists of combining the best parents with the best offspring obtained (i.e., a (p + — selection). Due to its clever mechanisms, the NSGA-II is much more efficient (computationally speaking) than its predecessor (NSGA), and its performance is so good, that it has become extremely popular in the last few years, becoming a landmark against which other multi-objective EAs have to be compared. There are two version of NSGA-II namely; binary coded and real coded, we are concerned with real coded here. In Chapter 2, it will be discussed in detail. 21 o Multi-objective Particle Swarm Optimization (MOPSO): Kennedy and Eberhart (1995) proposed an approach called "particle swarm optimization (PSO)" which was inspired by the choreography of a bird flock. The way in which PSO updates the particle x, at the generation t is through the formula: (t) = (t —1) + v, (t) (1.6) Where the factor vi (t) is known as velocity and it is given by v,(t) = w*vi (t —1) + Cl * rl * (x pbes,, — xi) + C2* r2 * (xgbe,,, — x1 ) (1.7) In this formula xpbe, is the best solution that x, has viewed, xgbe„, is the best particle (also know as the leader) that the entire swarm has viewed, w is the inertia weight of the particle and controls the trade-off between global and local experience, rl and r2 are two uniformly distributed random numbers in the range [0, 1] and Cl , C2 are specific parameters which control the effect of the personal and global best particles. The approach can be seen as a distributed behavioral algorithm that performs (in its more general version) multidimensional search. In the simulation, the behavior of each individual is affected by either the best local (i.e., within a certain neighbourhood) or the best global individual. The approach uses the concept of population and a measure of performance similar to the fitness value used with EAs. In addition, the adjustments of individuals are analogous to the use of a crossover operator of GA. However, this approach introduces the use of flying potential solutions through hyperspace (used to accelerate convergence) which does not seen to have an analogous mechanism in traditional EAs. Another important difference is the fact that PSO allows individuals to benefit from their experiences, whereas in an EA, normally the current population is the only "memory" used by the individuals. Coello and Lechuga, 2002) found PSO particularly suitable for MOOP mainly because of the high speed of convergence that the PSO presents for single objective optimization problem. The analogy of PSO with EAs makes evident the notion that using a Pareto ranking scheme (Goldberg, 1989) could be the straightforward way to extend this approach to handle MOOPs as well. The historical record of best solutions found by a particle (i.e. an individual) could be used to store 22 nondominated solutions generated in the past (this would be similar to the notion of elitism used in MOEA). The use of global attraction mechanism combined with a historical archive of previously found nondominated vectors would motivate convergence towards globally nondominated solutions. The pseudo-code of general MOPSO is shown in Figure 1.6 (Durillo et al., 2009). After initializing the swarm (Line 1), the typical approach is to use an external archive to store the leaders, which are taken from the non-dominated particles in the swarm. After initialization the leaders archive (Line 2), some quality measure has to be calculated (Line 3) for all the leaders to select usually one leader for each particle of the swarm. In the main loop of the algorithm, using Equations 6&7 the flight of each particle is performed after leader selection (Lines 7-8) and, optionally, a mutation or turbulence operator can be applied (Line 9); then, the particle is evaluated and its corresponding pbest is updated (Lines 10-11). After each iteration, the set of leaders is updated and the quality measure is calculated again (Lines 13-14). After the termination condition, the archive is returned as the result of the search. For further details about the operations contained in the MOPSO pseudo code and detail literature, please refer to (Coello et al (2004), Reyes-Sierra and Coello (2006), del Valle et al (2008), Parsopoulos and Vrahatis (2008) Padhye et al (2009). Raquel and Naval Jr( 2005) proposed another PSO based approach called MOPSO-CD, which extended the algorithm of the single-objective PSO to handle multi-objective optimization problems.. It incorporated the mechanism of crowding distance computation into the algorithm of PSO specifically on global best selection and in the deletion method of an external archive of nondominated solutions. The crowding distance mechanism together with a mutation operator maintains the diversity of nondominated solutions in the external archive. MOPSO-CD also has a constraint handling mechanism for solving constrained optimization problems. Raquel and Naval Jr. (2005) also showed that MOPSO-CD is• highly competitive in converging towards the POF and generated a well-distributed set of nondominated solutions. We discuss this approach in detail in Chapter 2. 23 1.3 DM's Partial Preference Articulation with MOEA -A Review During last few years of research on multi-objective optimization using EAs, it is amply evident that EAs are capable of finding multiple Pareto-optimal solutions in a single simulation run. It is then natural to ask: 'How does one choose a particular solution from the obtained set of Pareto-optimal solutions?' In the following, we first review a few techniques often followed in the context of MCDM. Apart from a priori and a posteriori approaches, Branke (2008) elaborately discussed an intermediate approach (middle path in Figure 1.1) incorporating both of these approaches. Although we agree that, it may be impractical for a DM to specify completely his or her preferences before any alternatives are known, we assume that the DM has at least a vague idea or biasness about which solutions might be preferred, and can indicate partial preferences. The methods discussed in this section aim at integrating such imprecise knowledge into the MOEA approach, biasing the search towards solutions that are considered relevant to the DM. The goal is no longer to generate a good approximation to all Pareto optimal solutions, but a small set (or subsection of the POF) of solution, that contains the DM's preferred solution with the highest probability. This may yield two important advantages: Focus: DM's partial preferences may be utilized to focus the search and generate a subset of all Pareto-optimal alternatives that is particularly interesting to the DM. This avoids overwhelming the DM with a huge set of (mostly irrelevant) alternatives. Speed: By focusing the search onto the relevant part of the search space, one may expect the optimization algorithm to find these solutions more quickly, not wasting computational effort to identify all Pareto-optimal but irrelevant solutions. To reach these goals, the MOEA community can accommodate or be inspired by many of the methods, which generally integrate DM'S preference information into the optimization process. Thus, combining MOEAs and their ability to generate multiple alternatives simultaneously in one run, and methodologies to incorporate user preferences holds great promise. Following brief literature survey contains quite a few techniques to incorporate partial preference information into MOEAs, and previous detailed surveys on this topic include Coello (2000), Rachmawati and Srinivasan (2006), Rachmawati (2009). In the following, we classify the different approaches 24 based on the type of partial preference information asked to the DM, namely a goal or reference point, trade-off information, weighted performance measures (approaches based on marginal contribution), objective scaling and others (Branke, 2008). 1.3.1 Approaches Providing Reference Point Perhaps the most important way to provide preference information is to provide a reference point, a technique that has a long tradition in MCDM. e.g., Wierzbicki (1979, 1982). A reference point consists of aspiration levels reflecting desirable values for the objective function, i.e., a target the DM is hoping for. Such information can then be used in different ways to focus the search. The use of a reference point to guide the MOEA has first been proposed by Fonseca and Fleming (1993). The basic idea there was to give a higher priority to objectives in which the goal is not fulfilled. Thus, when deciding whether a solution x is preferable to a solution y or not, first only the objectives in which solution x does not satisfy the goal are considered, and x is preferred to y if it dominates y on these objectives. If x is equal to y in . all these objectives, or if x satisfies the goal in all objectives, x is preferred over y either if y does not fulfil some of the objectives fulfilled by x, or if x dominates y on the objectives fulfilled by x. A slightly extended version that allows the DM to.: assign additionally priorities to objectives has been published in Fonseca and Fleming (1998). The work also contains the proof that the proposed preference relation is transitive. The approach by Deb (1999) used an analogy from goal programming. There, the DM can specify a goal in terms of a single desired combination of characteristics t = and the type of goal (e.g. f (x) t f (x), f (x) = t) . Deb (1999) demonstrated how these can be modified to suit MOEAs. The distances from that goal rather than the actual criteria are compared. If the goal for criterion i is to find a solution x with f (x) t theninsteadof consideringthecriterion f ,(x) ,simply f (x) = max {0, f , (x) — t, } is used. If the goal is set appropriately, this approach may indeed restrict the search space to an interesting region. The problem here is to set the goal a priori, i.e. before the POF is known. If the goal vector is outside the feasible range, the method is almost identical to the definition in Fonseca and Fleming (1993). 25 However, if the goal can be reached, the approach from Deb (1999) will lose its selection pressure and stop search as soon as the reference point has been found, i.e., return a solution, which is not Pareto-optimal. The goal-programming idea has been extended in Deb (2001) to allow for reference regions in addition to reference points. Tan et al. (1999) proposed another ranking scheme, which in a first stage prefers individuals fulfilling all criteria, and ranks those individuals according to standard nondominance sorting. What is more interesting, in Tan et al. (1999) is the suggestion on how to account for multiple reference points, connected with AND and OR operations. In Deb and Sundar (2006), the crowding distance calculation in NSGA-II is replaced by the distance to the reference point, where solutions with a smaller distance are preferred. More specifically, solutions with the same non-dominated rank are sorted with respect to their distance to the reference point. Furthermore, to control the extent of obtained solutions, all solutions having are grouped based on distance. Only one randomly picked solution from each group is retained, while all other group members are assigned a large rank to discourage their use. As Fonseca and Fleming (1998) and Tan et al. (1999), this approach is able to improve beyond a reference point within the feasible region, because the non-dominated sorting keeps driving the population to the POF. In addition, as Tan et al. (1999), it can handle multiple reference points simultaneously. Yet another dominance scheme was recently proposed in Molina et al. (2009), where solutions fulfilling all goals and solutions fulfilling none of the goals are preferred over solutions fulfilling only some of the goals. This, again, drives the search beyond the reference point if it is feasible, but it can obviously lead to situations where a solution which is dominated (fulfilling none of the goals) is actually preferred over the solution that dominates it (fulfilling some of the goals). Thiele et al. (2009) integrated reference point information into the indicator based evolutionary algorithm. In brief, the reference direction method allows the user to specify a starting point and a reference point, with the difference of the two defining the reference direction. Then, several points on this vector are used to define a set of achievement scalarizing functions, and each of these is used to search for a point on the POF. In Deb and Kumar (2007), an MOEA is used to search for all these points simultaneously. For this purpose, the NSGA-II ranking mechanism has been modified to focus the search accordingly. The light beam search also uses a reference direction, and additionally 26 asks the user for some thresholds which are then used so find some possibly interesting neighbouring solutions around the (according to the reference direction) most preferred solution. 1.3.2 Approaches Based on Trade-off Information If the user has no idea of what kind of solutions may be reachable, it may be easier to specify suitable trade-offs, i.e., how much gain in one objective is necessary to balance the loss in the other. Greenwood et al. (1997) suggested a procedure, which asks the user to rank a few alternatives, and from this derives constraints for linear weighting of the objectives consistent with the given ordering. Then, these are used to check whether there is a feasible linear weighting such that solution x is preferable to solution y. The authors suggested using a mechanism from White (1984) which removes a minimal, set of the DM's preference statements to make the weight space non-empty. Note that although linear combinations of objectives are assumed, it is possible to identify a concave part of the POF, because the comparisons are only pair-wise. In the guided MOEA proposed in Branke et al. (2001) the user is allowed to specify preferences in the form of maximally acceptable trade-offs like "one unit improvement in objective i is worth at most a`, units in objective F. The basic idea is to modify the dominance criterion accordingly, so that it reflects the specified maximally acceptable trade-offs. A solution x is now preferred to a non-dominated solution y if the gain in the objective where y is better does not outweigh the loss in the other objective. The region dominated by a solution is adjusted by changing the slope of the boundaries according to the specified maximal and minimal trade-offs. This idea can be implemented by a simple transformation of the objectives: It is sufficient to replace the original objectives with two auxiliary objectives S21 and Q2 and use these together with the standard dominance principle, Where S21 (x) = f (x) +f2 (x) a 21 02 (X) = 1 a12 (X) ± 4 27 (x) (1.8) (1.9) Because the transformation is so simple, the guided dominance scheme can be easily incorporated into standard MOEAs based on dominance, and it does not change the complexity nor the inner workings of the algorithm. However, an extension of this simple idea to more than two dimensions seems difficult. Although developed independently and with a different motivation, the guided MOEA can lead to the same preference relation as the imprecise value function approach in Greenwood et al. (1997) discussed above. The guided MOEA is more elegant and computationally efficient for two objectives, the imprecise value function approach works independent of the number of objectives. The idea proposed by Jin and Sendhoff (2002) is to aggregate the different objectives into one objective via weighted summation, but to vary the weights gradually over time during the optimization. The approach runs into problems if the POF is concave, because a small weight change would require the population to make a big "jump". 1.3.3 Approaches Based on Marginal Contribution Several authors have recently proposed to replace the crowding distance as used in NSGA-II by a solution's contribution to a given performance measure, i.e., the loss in performance if that particular solution would be absent from the population (Branke et al., 2004; Zitzler and Kiinzli, 2004; Emmerich et al., 2005), the performance measure used is the hypervolume. The hypervolume is the area (in 2D) or part of the objective space dominated by the solution set and bounded by, a reference point p. The marginal contribution is then calculated only based on the individuals with the same Pareto rank. Zitzler et al. (2007) extended this idea by defining a weighting function over the objective space, and used the weighted hypervolume as indicator. This allows incorporating preferences into the MOEA by giving preferred regions of the objective space a higher weight. 1.3.4 Approaches Based on Scaling All basic MOEA approaches attempt to generate a uniform distribution of representatives along the POF. For this goal, they rely on distance information in the 28 objective space, be it in the crowding distance of NSGA-II or in the clustering of SPEA-2.Many current implementations of MOEAs (e.g., NSGA-II and SPEA) scale objectives based on the solutions currently in the population. While this results in nice visualizations if the front is plotted with a 1:1 ratio, and relieves the DM from specifying a scaling, it assumes that ranges of values covered by the POF in each objective are equally important. Whether this assumption is justified certainly depends strongly on the application and the DM's preferences. In order to find a biased distribution anywhere on the POF, a previous study by Deb (2003) used a biased sharing mechanism implemented on NS GA. In brief, the objectives are scaled according to preferences when calculating the distances. This allows making distances in one objective appear larger than they are, with a corresponding change in the resulting distribution of individuals. Although this allows to focus on one objective or another, the approach does not allow to focus on a compromise region (for equal weighting of the objectives, the algorithm would produce no bias at all). Branke and Deb (2005) applied a biased sharing mechanism extended with a better control of the region of interest and a separate parameter .controlling the strength of the bias. For a solution i on a particular front, the biased crowding distance measure D, is re-defined as follows. Let 7/ be a user-specified direction vector indicating the most probable, or central linearly weighted utility function, and let a be a parameter controlling the bias intensity. Then, a = dH di (1.10) Where d, and d , are the original crowding distance and the crowding distance calculated based on the locations of the individuals projected onto the plane with direction vector 7/. The exponent a controls the extent of the bias, with larger a resulting in a stronger bias. Preference of solutions having a larger biased crowding distance D, will then enable solutions closer to the tangent point to be found. DF based approach also comes under the subsection we are going to discuss below: • DF based Approach- Trautmann and Mehnen (2005) suggested an explicit incorporation of preferences into the scaling. They proposed to map the objectives 29 into the range [0, 1] according to DFs. DFs are analogous to fuzzy membership functions, in fact they are a special case of membership functions (Kim and Lin, 2006) first introduced by Harrington (1965). With one-sided sigmoid (monotone) DFs, the non-dominance relations are not changed. Therefore, the solutions found are always also nondominated in the original objective space. What changes is the distribution along the front. Solutions that are in flat parts of the DF receive very similar desirability values and as MOEAs then attempt to spread solutions evenly in the desirability space, this will result in a more spread out distribution in the original objective space. However, in order to specify the DFs in a sensible manner, it is necessary to at least know the ranges of the POF. Details about DF can be found in numerous papers (Derringer and Suich, 1980; Lee and Park, 2003; Steuer, 2004; Lam and Tang, 2005; Park and Kim, 2005; Trautmann and Mehnen, 2005, 2009; Kim and Lin, 2006; Trautmann and Weihs, 2006; Mandal et al., 2007; Chatsirirungruang and Miyakawa, 2008; Mukherjee and Ray, 2008; Roy and Mehnen, 2008; Heike and Jorn, 2009; Jeong and Kim, 2009; Mehnen, 2009; Noorossana et al., 2009; Trautmann et al., 2009; Lee et al., 2010). We are more concerned with this approach and this work attempts to modify the DFA in a manner so that it can be applied in more general way. 1.3.5 Other Approaches The method by Cvetkovic and Parmee (2000) and others ( Parmee et al., 2000; Parmee, 2001; Cvetkovic and Parmee, 2002, 2003) allowed the DM to articulate fuzzy preferences, like "criterion 1 is much more important than criterion 2". A weight w, , and a minimum level for dominance, z is assigned to each criterion. Then, the concept of dominance is defined as follows: x y <=>w, r (1.11) f,(x)5 f,(x) with inequality in at least one case. To facilitate specification of the required weights, Cvetkovic and Parmee (1999) suggested a method to turn fuzzy preferences into specific quantitative weighting. However, since for every criterion the dominance 30 scheme only considers whether one solution is better than another solution, and not by how much it is better, this allows only a very coarse guidance and is difficult to control. Fuzzy optimization problems also appear in literature with multiple objectives (Hwang and Lai, 1993), and, typically, fuzzy logic has been used by numerous authors to solve MOOPs (Sakawa and Kato, 2009). It is patent that EA (Goldberg, 1989) could be used to solve fuzzy nonlinear programming problems because EAs are solution methods potentially capable of solving general nonlinear programming problems. The association of MOOP with fuzzy logic and evolutionary computation is approached in various ways in the literature. A GA is described in de Moura et al. (2002) to solve MOOPs with fuzzy constraints. In Sakawa and Kato (2009) an interactive fuzzy approach is used to solve nonlinear MOOPs through GAs. A third alternative is described in Jimenez et al. (2004), which describes a multi-objective approach to solve optimization problems with fuzzy constraints using a Pareto-based EA to solve a MOOP associated to the fuzzy problem. In this vein, de Moura et al. (2002) proposed a MOEA to solve optimization problems with costs and constraints using fuzzy approach export—import businesses as a case study. An "a posteriori" decision-making process is described in the work to obtain a crisp solution from fuzzy solution (Jimenez et al., 2001; Jimenez et al., 2006). Hughes (2002) concerned with MOEAs for noisy objective functions only. The: main idea to cope with the noise is to rank individuals by the sum of probabilities of being dominated by any other individual. To consider preferences, the paper proposes a kind of weighting of the domination probabilities. Numerous works including Marler and Arora (2004, 2005) and Marler et al. (2006) evaluated various transformation methods using simple example problems. Viewing these methods as different means to restrict function values sheds light on how the methods perform. In addition, they also demonstrated some transformation methods and advantages of using a simple normalization—modification. Rangarajan et al. (2004) stated that interactive multi-objective optimization methods help focus computational effort to find solutions of interest to the DM. However, most current EAs do not incorporate the expert knowledge of the user. Their paper presented a multiobjective evolutionary optimization framework that interactively incorporated user preference information. Although the weighted sum method is eventually used in 31 the study to depict the Pareto-optimal set, the present analysis is applicable to any MOOP approach. Rachmawati and Srinivasan (2006, 2009) presented a review of preference incorporation in MOEA. It indicates that introducing preference in MOEAs increases the specificity of selection, leading to solutions that are of higher relevance to the DM. When many objectives are involved, a MOEA based on pure Pareto-optimality criterion may not achieve meaningful search. The incorporation of preference addresses this concern. The incorporation of preference is difficult because of uncertainties arising from lack of prior problem knowledge and vagueness of human preference. Coello (2009) in his latest survey paper on MOEA excerpts, "In practical applications of MOEAs, users are normally not interested in a large number of nondominated solutions. Instead, they are usually only interested in a few types of trade-offs among the objectives" under the section 'what else remain to be done' (Coello, 2009). Present work is one of such approach in this direction. 1.4 Objectives of the Thesis DM's preferences can be incorporated into MOEA to make search much more efficient and meaningful. In this way one can zoom in a certain region of the POF and evolve the population only towards the area/areas of interest are incorporated into MOEA. Thus, a decision support system (DSS) can be created for DM to aid decision-making. As discussed earlier, some works are reported in this direction (for example (Cvetkovic and Parmee, 1998, 1999, 2002; Deb, 1999; Branke et al., 2001; Jin and Sendhoff, 2002; Branke and Deb, 2005). It is still relatively infrequent to report outcome of an MOEA that incorporates DM's preferences. Present work is a naïve attempt in this direction to fill the gap pointed by Coello (2009). 32 Objectives of the present work are summarized as • To incorporate the partial user preference with MOEAs. • Development of methodology for MOOP to obtain the interactive (guided) POF. • To hybrid the two approaches namely DFA as a priori and an MOEA (NSGAII/MOPSO-CD) as a posteriori to proVide the guided or biased POF according the wish of the DM. • To develop theoretical analysis of the proposed approach. • To develop new types of DFs to produce required biasness in the POF. • To use the proposed approach for solving reliability optimization problems. 1.5 Organization of the Thesis In this thesis, a hybrid approach consists of DF as a priori and MOEA (NSGAII/MOPSO-CD) as a posteriori is developed to provide the guided or biased POF according the wish of the DM. Apart from theoretical evaluation, the effectiveness of this methodology is demonstrated on a number of benchmark test problems. The developed approach is used to solve mathematical models of some real life engineering optimization problems such as reliability optimization. The Chapter wise summary of the thesis is as follows: In Chapter 2, basics of DF, NSGA-II and MOPSO-CD are. presented. A new approach named MOEA-IDFA (MOEA based IDFA) is proposed in this chapter. Linear DFA is articulated with NSGA-II/MOPSO and theoretical analysis of the approach is provided. Then, the proposed methodology is applied on ten benchmark test problems of different complexities. For each problem performances of both NSGAII and MOPSO-CD are compared using hypervolume metric. 33 In Chapter 3, the mechanism of MOEA-IDFA having convex and concave DFs (used as a priori) to incorporate the DM's preferences is given. It is possible to obtain a desired portion of POF using appropriate combination of convex and concave DFs. Theoretical investigation of the approach is also presented to support the scheme. Then, the methodology is applied on ten standard test problems (provided in Chapter 1) of different complication levels. Finally, performances of both NSGA-II and MOPSO-CD are compared using hypervolume metric for these problems. In Chapter 4, we discuss the functioning of proposed methodology i.e., MOEAIDFA having sigmoidal DFs (used as a priori) to incorporate the DM's preferences. A new type of sigmoidal DF is proposed in this Chapter. It is possible to obtain a desired portion (intermediate region) of POF using appropriate combination of sigmoid DFs. Theoretical study of the approach is provided to support the method. The methodology is applied on ten standard test problems of different difficulty levels. Finally, performances of both NSGA-II and MOPSO-CD are evaluated. In Chapter 5, it is shown that not only single region but multiple regions (of a POF) also can be explored using the suitable DFs. A novel combination of DFs is proposed to achieve this outcome in this chapter. Simultaneously, multiple portions of DF's preference can be explored using the all convex DFs. Theoretical analysis of the approach is provided to support the method. The methodology is applied on ten standard test problems (same as used in Chapter 1-4) of different complication levels. Corresponding each problem, performances of both NSGA-II and MOPSO-CD are also compared. In Chapter 6, five different kinds of problems arising in reliability engineering are modeled as MOOPs and solved using the methodology developed (i.e., MOEAIDFA) in this thesis. These problems are series system, life supported system in a space capsule, complex bridge system, mixed series parallel system and residual heat reservoir (RHR) system for nuclear power plant. This Chapter shows the applicability of the proposed interactive method for the solution of MCDM problems. 34 Chapter 7 is the concluding Chapter of the thesis. In this Chapter the usefulness and performance of the methodology developed in the thesis are critically evaluated. Also, some suggestions for further work in this direction are highlighted. Full preferences e.g. MOEA (NSGA-1 I/ MOPSO) 1\10()P Partial Preferences MOOP+ Partial Preferences (Membership function/ Desirability function) Pareto-front Approximation MOEA U 0 Soop c. Biased Pareto-front Approximation User selection V Solution. Figure 1.1 Different Ways of Preference Articulation 35 a ri rn a Covered Distance Figure 1.2 An example to travel bigger distances in ways that are more economical Dominated solution 12 0 0 Nondominated solution • • • • Pareto-optim al front 0 00 0 0 0 0 0 o........ Minimize Figure 1.3 Description of Dominated and Nondominated Solutions 36 MOEA A Priori (Before) Progressive (During) Aggregation ( Scalarization ) A Posteriori (After) Non Pareto Based Linear fitness combination VEGA Nonlinear fitness combination Pareto Based Goal Programming Based Approach 1st Generation NSGA Goal Attainment Based MOGA NPGA 6 —Constraint IInd Generation SPEA SPEA2 Fuzzy based PAES PESA NSGA-II MOPSO Figure 1.4 Preference Based Classification of MOEA 37 J; 3 • . 1 . 23 • •1 •- • 1 •2 Figure 1.5 Nondominated Ranking of Search Space for all Minimization Case 1: initializeSwarm() 2: initializeLeadersArchive() 3: determineLeadersQuality() 4: generation = 0 5: while generation < maxGenerations do 6: for each particle do selectLeader() 7: updatePosition() // flight (Equations 6 and 7) 8: 9: mutation() evaluation() 10: 11: updatePbest() end for 12: 13: updateLeadersArchive() 14: determineLeadersQuality() 15: generation ++ 16: end while 17: returnArchive() Figure 1.6 Pseudo code for MOPSO 38 CHAPTER 2 Articulation of an a Priori Approach with an a Posteriori Approach In this chapter, a partial user preference approach is proposed. In this method, linear DFA as an a priori and an MOEA (NSGA-II/MOPSO-CD) as a posteriori are articulated together to provide an interactive POF. The approach is analyzed theoretically as well as numerically (on different standard test problems). Application of using linear DF is analyzed elaborately as well as detail description of DF, NSGA-II and MOPSO-CD is presented. 2.1 Introduction Typically, a MOOP has infinitely many Pareto-optimal solutions. However, it is generally desirable to obtain one point as a solution. Selecting one out of the set of Pareto-optimal solutions, calls for information that is not contained in the objective functions. Therefore, interaction with DM is an integral part of the algorithms for MOOP at some point during the optimization process as discussed elaborately in Section 1.1 of Chapter 1. Articulation of preferences may be done either before (a priori), during (progressive), or after (a posteriori) the optimization process. Out of these, a priori and a posteriori are highly used for preference articulation. In the first approach (i.e. a priori), .one declares preferences for various objectives ( f's) before the optimization process, incorporates the preferences into a modified formulation of problem P1 (see Section 1.1) and solves the modified problem to obtain a single Pareto-optimal solution. In a priori method, preferences are often incorporated using a scalarization method in which multiple objectives are aggregated (different aggregation operators may, be used) into a single function. If the resulting solution is acceptable to the DM, the solution process is terminated. However, the most scalarization methods do not transfer preferences from the user to the final solution with complete accuracy. Thus, if the solution is not acceptable to DM, the preferences are altered, and the problem is re-solved to obtain another Pareto-optimal solution. 39 Alternatively, the other approach (a posteriori) entails generating a representation of the entire Pareto-optimal set (without any preference by DM) and then DM choosing from that set a suitable solution point that satisfies his/her preferences. Most of the MOEAs come under this category. This approach provides various options to the DM to choose from. It also has a slight disadvantage of producing a large number of solutions some of them not in the choice range of the DM. The reason for this drawback is the unavailability of the preference by the DM. With the purpose of presenting the DM, a POF in the region of his/her choice one can club the advantages of above two discussed approaches (i.e. a priori and a posteriori). Articulation of preference part can be done as an appropriate a priori approach while the solution part may be obtained using suitable a posteriori approach. In order to provide a DM's preferred (biased) POF a hybrid approach articulating an a priori and an a posteriori approach is proposed in this study combining the goodness of both types of approaches. Providing the biased POF eases the DM's judgment vis-à-vis suitable solution. Jeong and Kim (2009) proposed an interactive DF approach (IDFA) for multirespose optimization to facilitate the preference articulation process. One run of this approach provides just one solution to the DM based on his preference in the form of DF. Present work is an extension of the work done by Jeong and Kim (2009). We present a MOEA based IDFA (MOEA-IDFA) to provide DM a preferred POF rather than providing just a single solution. This approach comes under the category of partial DM's preference incorporation (Branke, 2008). Proposed approach utilizes DFA as a priori while NSGA-II/MOPSO-CD as a posteriori. The rest of the chapter is organized as follows: Section 2.2 describes the essentials of DFA as a priori. Section 2.3 describes the working, computation complexity, constraint handling and performance metric of MOEAs (NSGA-II and MOPSO-CD) used. The methodology of the proposed approach is given in section 2.4. Necessary theorems are stated and proved in order to analyze the methodology theoretically. Results from both MOEAs are discussed and compared in section 2.5. Finally, certain concluding observations are drawn in section 2.6. 40 2.2 DFA as a Priori The concept of desirability is a mean for complexity reduction of a MOOP. The DFA to simultaneously optimizing multiple objectives was originally proposed by Harrington (1965) in the context of multi-objective industrial quality control. DFs are very popular in response surface methodology (Box and Wilson, 1951). To incorporate the DM's vague idea about 'in which region of the objective the optimum should be' the DF is used. Essentially, the approach is to translate the functions to a common scale [0, 1] by the means of mathematical transformations, combine them using the geometric mean and optimize the overall metric. The procedure calls for introducing for each objective (say f , i = 1,k) a function p, (called DF w.r.t. f, ) with a range of values between zero and one that measures how desirable it is that objective f takes on a particular value. In other words, DF maps the objective (say f) to the interval [0, 1], defined by function p : f -+ [0,1] according to the DM's desired values specified as a priori. Once this function is defined for each of the k objectives, an overall objective function is defined by geometric mean of the individual desirabilities. Harrington (1965) introduced two types of DFs. One aims at maxi-or minimization (one-sided specification) where as the other reflects a target value problem (two sided specification). We are concerned with one-sided specification only in this work. Harrington's one-sided DF is a special form of the Gompertz-curve having few parameters governing the kurtosis. Definition of DF reveals its similarity with membership function of fuzzy approach (Kim and Lin, 2006). However, we will retain the name DF in this work as this suits interactive guided approach. The mathematical model of DF was put into a more general form by Derringer and Suich (1980) by introducing LTB (larger the better), STB (smaller the better) and NTB (nominal the better) types of DFs. Depending on whether a particular MOOP is to be maximized (LTB), minimized (STB), or assigned a target value (NTB), appropriate DFs can be used (Nguyen et al., 2009). In this study, we are concerned with only STB type of DF.In general, DF can be further classified into two categories namely: linear DF and nonlinear DF. Before discussing the behaviour of nonlinear DF, it is obligatory to confer the performance of the linear one. Hence, in this chapter we are concentrating on linear DF only. 41 2.2.1 Linear DF By using DF, the goodness of a design objective at different functional values can be well characterized. Since we are concentrating on minimization problem, we are considering here smaller the better (STB) type of DF, where the grade (value of p, ) "zero" quantifies perfect satisfaction for the choice of the most desirable objective function value that corresponds to an ideal target. On the other hand, the grade "one" characterizes a threshold condition pertaining to the most undesirable objective function value that corresponds to a worst scenario to be avoided. Equation (2.1) represents one such formulation of a DF where STB type of DF is used, shown in Figure 2.1. The desirability function ,u, for f is defined by, 0if fi < f:.1 if ff ff c f rr - ) ;i = 1, 2, 3, ..., k (2.1) 1if f - < if fi < f . f. = - A.if f; A ;i = 1, 2, 3, ..., k (2.2) if < f where j;* and f,"" are the i-th components of ideal and anti-ideal vectors of MOOP (i.e., P1). DF can be defined either way too i.e. LTB type, where, "one" quantifies perfect satisfaction (see Equation 2.2 and Figure 2.2). Other types (shapes) of DF can be used too discussed intensely in later chapters. Since in this thesis only minimization problems are taken (maximization problem is converted into minimization one), only STB type of DF is of prime concern. Application of linear DF can be found in various earlier works such as Tang and Paoli (2004) applied linear DFA to represent the technical attributes' values to adapt to different directions of improvement for different technical attributes. Merkuryeva 42 (2005) presented response surface-based simulation meta-modelling procedure using various linear and nonlinear DFs. Nguyen et al. (2009) which solves Multiresponse optimization problem based on the linear DF for a pervaporation process for producing anhydrous ethanol. Most of the works involving linear DF, in a single run, has produced a single solution due to the use of aggregation methods for solving the involved MOOP. In the current Chapter, we are concerned using linear DF as an 'a priori' approach to incorporate the DM's preference. 2.3 Description of MOEAs Applied as a Posteriori Several evolutionary methods are available ensuring full convergence toward the Pareto-optimal solution both in terms of precision and in terms of diversity of solutions. For example, NSGA-II and MOPSO-CD are two such MOEAs. These methods, specifically developed for MOOPs, have been widely and deeply tested and compared on many different test functions and some convergence measuring criteria available. We are taking these two (NSGA-II and MOPSO-CD) prominent MOEAs to act as a posteriori in our approach. We discuss NSGA-II in the subsection 2.3.1 while MOPSOCD in subsection 2.3.2. 2.3.1 NSGA-II or Elitist Nondominated Sorting Genetic Algorithm To solve MOOP in general, simple EA is extended to maintain a diverse set of solutions with the emphasis on moving toward a true Pareto-optimal region. The nondominated sorting genetic algorithm (NSGA) proposed by Srinivas and Deb (1994), is one of the first of such algorithms. It is based on several layers of classification of the individuals. Nondominated individuals get a certain dummy fitness value and then are removed from the population. This process .is repeated until the entire population has been ranked. It is a very effective algorithm but it has been criticized for its computational complexity, lack of elitism and its requirement for specifying sharing parameters in the algorithm. Based:on these issues, a modified version of the NSGA, named NSGA-II Deb et al. (2002) was developed. NSGA-II is a generational MOEA that aims at approximating the POF for a given problem, while keeping high diversity in its result set. It builds on three main modules: 43 • Nondominated Sorting at a certain generation t , partitions the population P, in fronts F , with index i indicating the non-domination rank shared by all individuals contained in such a front. The first front F1 is the actual nondominated front, i.e., it consists of all nondominated solutions in population P, at a certain generation t . The second front F2 consists of all individuals that are non-dominated in the set P, \ F/ , i.e., each member of F2 is dominated by at least one member of F as shown in Figure 2.3. Generally, front Fk comprises all individuals that are nondominated if the individuals in fronts F., with j < k were to be removed from P, (Deb, 2001). • Crowding-distance-assignment calculates a crowding-distance value for each individual within a certain front F as the difference in objective function values between the nearest neighbours at each side of the individual, summed up over all objectives Extreme solutions (i.e., solutions with the smallest and largest function values occurring within the front) are assigned an infinite distance value, which, motivated by the pursuit of diversity, effectively preserves them into the next generation should the front in which they are contained be partially discarded when a new population P,±1 is formed. • Crowded-comparison operator guides the selection process by defining an ordering on I . Each individual has two attributes, a non-domination rank and a crowding-distance value. Between two individuals with differing nondomination ranks, we prefer the individual with the lower rank. Otherwise, with both individuals belonging to the same front, we prefer the individual that is located in the lesser crowded region (i.e., with higher crowding-distance value). Detailed description of the main loop of NSGA-II is given below. Two distinct entities are calculated in the NSGA-II to validate the quality of a given solution. The first is a domination-count where the numbers of solutions that dominate a given solution are tracked. The second keeps track of how many sets of solutions a given solution dominates. In the process, all the solutions in the first nondominated 44 front will have their domination count set to zero. The next step is to select each solution in which the nondomination count is set to zero and visit all other solutions in the solution set and reduce the domination count by one. In doing so, if the domination count of any other solution becomes zero, this solution is grouped in a separate list. This list is flagged as the second nondominated front. This process is then continued with each member of the second list until the next non-dominated front is identified. The process is continued until all fronts are identified. Based on the nondomination count given to a solution, a non-domination level will be assigned. Those solutions that have higher nondomination levels are flagged as non-optimal and will never be visited again. One of the key requirements of a successful solution method is ensuring that a good representative sample from all possible solutions is chosen. Introduction of a density estimation process and a crowded-comparison operator has helped NSGA-II to address the above need. The crowding-distance computation requires sorting of a given population according to each objective function value in ascending order of magnitude. Once this is done, the two boundary solutions with the largest and smallest objective value are assigned distance values of infinity. All other solutions lying in between these two solutions are then assigned a distance value calculated by the absolute normalized distance between each pair of adjacent solutions. After each population member is assigned a crowding-distance value, a crowded-comparison operator is used to compare each solution with the others. This operator considers two attributes associated with every solution which is the nondomination rank and the crowding-distance. Every solution is rated with others based on the non-domination rank. Solutions with lower ranks are deemed better in this attribute. Once solutions that belong to the best front are chosen based on the non-domination rank, the solution that is located in a lessercrowded region is considered better and forms the basis of the NSGA-II algorithm. The flow chart depicting the NSGA-II algorithm is shown in Figure 2.1. For details of NSGA-II, see (Deb et al., 2002, 2000; Deb, 2001, 1999a; Chakraborty et al., 2009; Sinha et al., 2008). The source code for NSGA-II has been taken from the Kanpur Genetic Algorithm Laboratory website http://www.iitk.ac.in/kangal and modified according the application. The code is written in C, and the website is maintained by Prof. Deb and his research group. The motivation for using this algorithm comes from 45'; the very good performance of this algorithm on test functions, and its success in generating the POF. 2.3.2 Run Time Complexity of NSGA-H The run time complexity of an algorithm quantifies the amount of time taken by an algorithm to run, as a function of the size of the input to the problem. The time complexity of an algorithm is commonly expressed using big 0 notation, which suppresses multiplicative constants and lower order terms. Time complexity is commonly estimated by counting the number of elementary operations performed by the algorithm, where an elementary operation takes a fixed amount of time to perform. Thus, the amount of time taken and the number of elementary operations performed by the algorithm differ by at most a constant factor. Since an algorithm may take a different amount of time even on inputs of the same size, the most commonly used measure of time complexity, the worst-case time complexity of an algorithm. In order to identify solutions of the first nondominated front in a population of size N , each solution can be compared with every other solution in the population to find if it is dominated. This requires o(kN) comparisons for each solution, where k is the number of objectives. When this process is continued to find all members of the oCK1.4 2) first nondominated level in the population, the total complexity is! Consider the complexity of one iteration of the entire algorithm. Let population of size be N and k the number of objectives in the MOOP. Then consider the complexity of one iteration of the entire algorithm. The basic operations and their worst-case time complexities are as follows: 1) nondominated sorting is 0(kN 2 ) ; 2) crowding-distance assignment is 0(kN log N) ; 3) sorting in crowded-comparison operator is 0(N log N) . The overall complexity of the algorithm is 0(kN2 ) which is governed by the nondominated sorting part of the algorithm (Deb, 2001; Deb et al., 2002). 46 2.3.3 MOPSO-CD or Multi-objective Particle Swarm Optimization with Crowding Distance Inspired by the emergent motion of a flock of birds and fish searching for food, PSO method was introduced by Kennedy and Eberhart (1995). Due to its simplicity in implementation and high computational efficiency, it has been used in a range of applications. In recent years, applying PSO to MOO has become increasingly popular. Moore and Chapman (1999) attempted to handle MOO by applying Pareto dominance into their approach, although it has been criticized for not adopting any scheme to maintain diversity (Moore and Chapman, 1999). The algorithm of Ray and Liew (2002) uses Pareto dominance for convergence and crowding comparison to maintain diversity as well as a multi level sieve to handle constraints (Ray and Liew, 2002). The algorithm proposed by Parsopoulos and Vrahaits (2008) focuses on addressing the• difficulty of generating the concave portion of the Pareto front by using an aggregation function (Parsopoulos and Vrahatis, 2008). Li (2003) proposes an approach that applies the main techniques of NSGA II to PSO algorithm. Coello et al. (2004) proposed a new version of multi-objective PSO (MOPSO), in which he compares his results to three highly competitive SMO algorithms: NSAG II, PAES and microGA. Agrawal et al. (2008) proposes an interactive particle swarm optimization algorithm (IPSO), which is similar to Coello's MOPSO (Coello et al., 2004). Results show that while MOPSO is superior in converging to the true Pareto front, its diversity mechanism falls behind that of NSGA-II. Raquel and Naval Jr (2005) proposes an algorithm MOPSO-CD which extends PSO in solving MOOPs by incorporating the mechanism of crowding distance computation in the global best selection and the deletion method of the external archive of nondominated solutions whenever the archive is full. The crowding distance mechanism together with a mutation operator maintains the diversity of nondominated solutions in the external archive (Raquel and Naval Jr, 2005). In other words, the algorithm is a modification of PSO that adds an archive of nondominated solutions and uses a crowding distance measure to prevent many similar Pareto optimal solutions from being retained in the archive. MOPSO-CD has drawn some attention recently as it exhibits a relatively fast convergence and well-distributed Pareto front compared with other multi-objective 47 optimization algorithms. The original source code for MOPSO-CD has been taken from http://www.engg.upd.edu.phi--cvmig/mopso.html and modified according the application. The details of MOPSO-CD can be consulted in (Raquel and Naval Jr, 2005). At each generation the following computation takes place: first, the crowding distances of the nondominated solutions in the archive are computed. Guides are selected for each particle in the swarm based on decreasing crowding distance in- the sorted archive. This allows the particles to move towards those nondominated solutions in the external archive which are in the least crowded area in the objective space. To obtain a good distribution of nondomination solution in the archive, MOPSO-CD uses crowding distance in selecting the best particles in its archive. This feature of MOPSO-CD improves its convergence properties and maintains a good spread of the non-dominated solutions. The flow chart of the MOPSO-CD is provided in Figure 2.6. 2.3.4 Run Time Complexity of MOPSO-CD The computational complexity of the algorithm is dominated by the objective function computation, crowding distance computation and the nondominated comparison of the particles in the population and in the archive. If there are k objective functions and N number of solutions (particles) in the population, then the objective function computation has O(kN) computational complexity. The costly part of crowding distance computation is sorting the solutions in each objective function. If there are M solutions in the archive, .sorting the solutions in the archive has 0(kM log M) computational complexity. If the population and the archive have the same number of solutions, say N , the computational complexity for the nondominated comparison is 0(kINT 2 ) . Thus, the overall complexity of MOPSO-CD is 0(kN 2 ) . From Section 2.3.3 and Section 2.3.4 it is clear that the run time complexity is same in case of both NSGA-II and MOPSO-CD if the archive and population size are kw same for MOPSO-CD. Hence, in this work's experimental setup (of standard test problems as well as real world application) archive and population size are set same. 48 2.3.5 Constraint Handling in NSGA-II and MOPSO-CD In order to handle constrained optimization problem, MOPSO-CD adapted the constraint handling mechanism used by NSGA-II due to its simplicity in using feasibility and nondominance of solutions when comparing solutions. A solution i is said to constrained-dominate a solution j if any of the following conditions is true: 1. Solution i is feasible and solution j is not. 2. Both solutions i and j are infeasible, but solution i has a smaller overall constraint violation. 3. Both solutions i and j are feasible and solution i dominates solutions j. When comparing two feasible particles, the particle which dominates the other particle is considered a better solution. On the other hand, if both particles are infeasible, the particle with a lesser number of constraint violations is a better solution. 2.3.6 Performance Measure for NSGA-II and MOPSO-CD The performance of an MOEA can be decomposed into two, interacting, criteria: • Closeness — the nearness of the identified non-dominated solutions to the true POF, and • Diversity — the distribution of the identified solutions across the trade-off surface. This distribution is commonly expressed in criterion-space. Various performance metrics have been proposed to measure accuracy, diversity, and in some cases both simultaneously. A review of performance metrics is provided by Deb (2001). Some of these metrics involve measurements made with respect to the true trade-off surface, whilst others involve a purely relative comparison of two sets of results. The former approach requires that the true surface (POF) be known and can be sampled. While hypervolume measure can be applied without knowing the true surface also it takes care of both - criteria mentioned above. Hence, this study utilizes hypervolume metric as a performance measure for both and MOPSO-CD. 49 r algorithms i.e., NSGA-II Hypervolume Metric Hypervolume metric to evaluate the accuracy of MOEA is the volume-based performance measure. This metric was proposed by Zitzler et al. (2003) which measures the volume of the region that is dominated by the computed solution set. This performance metric does provide very useful information about the dominance as a whole. This metric calculates the volume (in the objective space) covered by the members of the Pareto-optimal set in concern for problems where all objectives are to be minimized. Mathematically, for each solution (e.g., PI , P2 and P3 in Figure 2.7) of the Pareto-optimal set, a hypercube is constructed with a reference point and the solution as the diagonal corners of the hypercube. Thereafter, a union of all hypercubes is found and its Hypervolume calculated. Figure 23 shows the chosen reference point and the three points P , P2 and P3 in the Pareto-optimal set of which hypervolume is to be calculated. The Hypervolume is shown as a hatched region. The hypervolume of a set is measured relative to a reference point, usually the anti-optimal point or "worst possible" point in space. (We do not address here the problem of choosing a reference point, if the anti-optimal point is not known or does not exist: one suggestion is to take, in each objective, the worst value from any of the frOnts being compared.). If a set A has a greater hypervolume than a set B, then A is taken to be a better set of solutions than B. 2.4 Proposed Methodology General MOOP is already described in Section (1.1) given by Equation P1 Minimize f (x) = {f;(x), f2 (x),..., f, (x)} x Eliciting the corresponding DF to each objective (i.e. f,, f2 ,..., (P1) through the interaction of DM a new DF based multi-objective optimization problem (DFMOOP) consisting the DFs is formulated given by Equation P2. Minimize ,u(x) = tui (x),uk (x)} rGx (P2) Where, for each value of an objective function f, , there is a mapping called DF i.e., ,u, to prescribe the variation of vagueness involved discussed in Section 2.2. The overall 50 (resulting) DF's value (say ,u) is also between zero and one. Hence, the DFA works here as a priori. The Pareto-optimal solutions are then determined for this newly formed DFMOOP instead of the original MOOR Solutions of DFMOOP have a unique relationship with the original MOOP of objective functions. In general, DFMOOP is solved using different aggregators, min and product operator are most common, providing a single solution of DFMOOP. This type of approach is repeatedly applied for different degrees of satisfaction values until the DM , is satisfied. Benefit of this technique lies in the fact how well DM's preferences have been incorporated in a priori approach, which is quite rare due to vague nature of human judgment. So in the present approach DFMOOP is solved using purely multi-objective manner (using the discussed algorithms i.e., NSGA-II and MOPSO-CD) without aggregating. Present approach is an attempt to incorporate the benefits of both the a priori and the a posteriori methods together,. For obvious reason, there are some relationships between the solutions of P1 and P2 that can be understood through the following two theorems. Theorem 2.1: The Pareto-optimal solutions • of P2 cones onding the desirability 004T RAL function are also Pareto-optimal solutions of Pl. CYef 1-4/1. z ACC No -C. Date Proof: ROOK Let x* be Pareto-optimal solution for P2. Then by the definition of Pareto-optimal solution Ox e X ,u, (x) < ,u, (x*) ; for i E {1, 2,..., k} and ,u j(x) j(x* ); for j E {1, 2,..., ; j # i ( <=>e X : fi (x)-1 < fif(x*)-*-f̀ * ) ;for i E _f f* fi* 51 (2.3) - (x ) - f.T j< f ( x* - fJ*1 fi ** - , 2, ..., (2.4) E0 Inequality 2.3f,. (x) — f * < f,. (x* )— f,* as f ** f * > 0 <=> f(x) < f (x') ; for i e {1, 2, ..., Similarly Inequality 2.4 <=>f j (x) — fi (x* )— fi* as 4** — f *j > 0 (2.5) (2.6) <=> fi (x* ) ; for j e j# • The proposition holds as Inequalities 2.5 and 2.6 together form the condition that implies that x* is also Pareto-optimal solution of P1 (MOOP). Theorem 2.2: If x* is properly efficient w.r.t. P2 (in Geofferin's sense) then it is also properly efficient w.r.t. P1 in the same sense. Proof: Let x* be properly efficient solution w.r.t. P2. Then by Geoffrion's definition (Definition 1-.9) of Pareto-optimality there must exist some real M > 0 such that for each x E X satisfying ,u, (x) < ,u,(x* ) , there exists at least one ,u j (x) ( j # i) such that (x ) < ,u, (x) and pi( ) — /-gi(x) <M Now ,u,.(x) < Ati (x* ) [f (x) — f: I f ** — f: ] <[f( ) - .4* If:* - f* ] • [f (x) f * f ** f" ]<[f(x*)- f* 1 fi7 - f* ] • [ (x) - f*]<[i;(x )- .1;*] as V** f i* ] > 0; (2.7) f(x) < f(x* ) Similarly Pi (x. ) < (x) if; (/./;* —< Lif (x) — .f,7 /47* — ] (2.8) *)< fi (x) Also, (x* ) — ,u,(x )/P / (X) /di (X) < M 52 [fi(x*)-47f7* — f7]—[4(x)-47fr — fil <111 [f i (x)— fj/f7—f;]—if f (x* )—f.;/f37 — f.;] (f ** cx* )—f* ]—[fcx )—.,*] <M [fi(x)—i5]—[fi(x*)—f;]( [7 — J./ ) or we have, — fe*) fi(x) , m (fi" f(x * ) — (2.9) f j(x)— f j(x* )(f **j — f;) Inequalities 2.7-2.9 together imply that proposition holds, so the condition of properly efficient solution for x* w.r.t. P1 is evident. Corollary 2.3 In Case (f," f * ) is equal to (f7 — f;) then M becomes equal to M . In this situation, the shape of the POF will be same for both P1 and P2 (e.g.,i SCH1 problem to be discussed later). 2.4.1 Assumptions: Throughout this study, we make the following assumptions. 1. The vague or biased goals of the DM can be quantified by eliciting the corresponding DFs through the interaction with the DM. 2. MOEA applied to solve the MOOP provides the global Pareto-optimal solutions. 2.4.2 Detailed Procedure of the Methodology: MOEA-IDFA The methodology consists of five steps. Step 0 is an initialization step. Step 1 and 2 constitute the calculation phase and Steps 3 and 4 the decision-making phase. Figure 1.8 shows the procedure of MOEA-IDFA. In the calculation phase, the DFs are constructed, and then an optimization model of DFs is solved using an MOEA. In the decision-making phase, the DM evaluates the results of the calculation phase, and then articulates his/her preference information. More specifically, if the DM is satisfied with the results on all the objectives, the procedure successfully ends. Otherwise, the DM 53 adjust the parameters (in case of linear DF only bound) of a DF. Then, the procedure goes back to the calculation phase. Each step is described below. Step 0: Initialization of DF's Parameters The DF's parameters (preference parameters on each objective) are to be initialized, to construct the DF for each objective in the first iteration. The, initial bound and goal (target) may be determined based on the DM's subjective judgments. Ideal and antiideal vectors f: and r , of correspondingobjective function f, should be calculated in advance for the given MOOP (P1). Since we are tackling linear DF in this chapter hence, few parameters are less here that will be discussed intensely in the later chapters. Step I: Construction of the DFs As mentioned earlier, the preference parameters initialized in Step 0 are used in constructing the linear DF p, for each f of Pl. DM preference can be utilized in the construction of DF. In case of linear DF the DM can specify any two points within j and r . Thus, a new DFMOOP (see Equation P2) is formed. Step 2: Solving the DFMOOP This newly formed P2 is then solved using an efficient MOEA. As we are discussing general MOOP which can be nonlinear, nonconvex, multimodal and multivariable. Hence, resulting DFMOOP will also be of the same nature. We need some powerful algorithm to solve this optimization problem. NSGA-II and MOPSO-CD are two such competent techniques capable of solving complex MOOPs used here. Step 3: Evaluation of the Solution Present the solution to the DM obtained in Step 2. Theorem 2.1 and 2.2 together imply that the Pareto-optimal solution obtained by solving P2 is also the Pareto-optimal solution of Pl. If the DM is satisfied by the Pareto-optimal solutions, the methodology is terminated. Else, the procedure goes to Step 4. 54 Step 4: Adjusting the Preference Parameters In order to improve the unsatisfactory results, interaction with DM can help modifying (updating) the Step 0 and Stepl. On the other hand, if the DM is fully satisfied by the obtained guided POF the procedure ends successfully. The whole process may be repeated until DM is satisfied. 2.5 Experimental Suite To validate our approach we demonstrate its working on several standard test problems. The set of diverse established test problems is used in this study to verify the proposed approach. The suite consists of ten, tractable, bi objective as well as tri objective functions with varying characteristics as summarized in Table 2.1. The corresponding mathematical definitions, restrictions as well as the characteristics are also provided in the same table. These functions cover many of the features that may be found in realworld problems. Test problems are chosen from a number of significant past studies in this area. Van Veldhuizen and Lamont (2000) cited a number of test problems, which many researches have used in the past. Of them, we choose three problems, we call them SCH1 and SCH2 (Schaffer, 1985), KUR (from Kursawe's study (Kursawe, 1991)). In 1999, Deb suggested a systematic way of developing test problems for MOOPs (Deb, 1999). Following those guidelines Zitzler et al. (2000) proposed six test problems. We choose three of those problems there and call them ZDT1, ZDT2 and ZDT3 detailed in Table 2.1. All the problems mentioned above have two objective functions. None of these problems has any constraint. To seal this gap a constrained MOOP consisting two objectives is considered here, we call it TNK (Tanaka et al., 1995) shown in Table 2.2. To show the comprehensible depiction of the proposed approach, we have considered three unconstrained tri-objective optimization problems also namely VNT, MHHM1 and MHHM2 (Huband et al., 2006) revealed in the Table 2.2. 55 2.6 Results and Discussion The proposed approach in this chapter has been applied to a set of ten standard test problems (SCH1, SCH2, KUR, ZDT1, ZDT2, ZDT3, TNK, VNT, MHHM1 and MHHM2) described in Section 2.5 and which are shown in Tables 2.1-2.2. As conversed earlier in the Section 2.4, linear DF is used as a priori, while NSGAII/MOPSO-CD applied as a posteriori in the proposed approach. As described in the Step 0 of Subsection 2.4.2 of the approach, the parameters of a priori method i.e., linear DFs are obtained with the help of DM corresponding each objective. Initial parameters for DF's used for all test problems are given in Tables 2.3(a)-2. 12 (a). In Step I, preference parameters initialized in Step 0 are used in construction of the linear DFs ( pi s) for each f; , and forming the DFMOOP (i.e., P2). This newly formed P2 is solved using an a posteriori method as shown in Step 2. We have used both NSGA-II and MOPSO-CD as a posteriori method to estimate the effectiveness of the approach on different MOEAs. The values of different set of parameters are tested and fine-tuned through several runs of MOEA (NSGAII/MOPSO-CD) methods used in this approath NSGA-1I and MOPSO-CD parameters used for each problem are given in Tables 2.3(b)-2.13(b). In Step 2, computations have been carried out based on these parameters and results have been reported. To compare the efficiency of NSGA-II and MOPSO-CD, hypervolume metric is used here and the reference points are taken (11, 11) and (11, 11, 11) for bi-objective problems and triobjective problems respectively. Tables 2.3(c)-2.12(c) display the mean, median, standard deviation, best and worst values of hypervolume metrics obtained by using the two algorithms, NSGA-II and MOPSO-CD with the help of 10 runs each. For the steadiness of the results for each problem, 10 runs of each MOEA (NSGA-II and MOPSO-CD) are done under the same parameters and the best POF (based on best hypervolume metric) obtained is reported. The POFs of the problems are depicted in Figures 2.9-2.18. Having a look on the Tables 2.3(c)-2.12(c) reveal that if we consider mean of hypervolume metric as the competence parameter for both NSGA-II and MOPSO-CD, then NSGA-II performed better in case of SCH1, SCH2, KUR, TNK and VNT, while MOPSO-CD was better in the cases ZDT1, ZDT2, ZDT3, MHHM1 and 56 MHHM2 for the used parameters parameter settings. If the DM is satisfied by the POF obtained then the procedure ends successfully as explained in Step 4. We are just showing one iteration of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. However, if the DM is not satisfied then the procedure goes back to Step 1 to accumulate the DM's preferences in a better manner. 2.7 Conclusion Exhaustive description and application of DF, NSGA-II and MOPSO-CD are presented in this chapter. In this chapter, a partial user preference approach named MOEA-IDFA is aldoproposed. In this approach, linear DFA as a priori and NSGA-II/MOPSO-CD as a posteriori are combined together to provide an interactive POF. The theoretical analysis of the proposed MOEA-IDFA is also provided in this chapter.. The performance of the MOEA-IDFA is tested on a set of 10 test problems. Though, application of linear DFA.does not provide any biasness in the POF, still it can be well used as a starting point for the interaction with DM. 57 Figure 2.1 STB type of DF Figure 2.2 LTB type of DF Figure 2.3 Description of Crowding-Distance 58 Reference point . ,. 0 .. . . . .. . .. .. f2 PI N E .. \., ... .,. . \ . .,. . .,. ... \.. \ 2 P3 0 Minimize Figure 2.4 The Hypervolume Enclosed by the Nondominated Solutions Front F3 Front F Front F Front 2 Figure 2.5 Nondominated Sorting of a Population 59 ( Start ) Generate Random Population gen = 0 Find Solution for Each Objective Function Assign Fitness/Nondominated Level For Each Solution Using Fast Sorting Algorithm and Crowding Distance Evaluation End ) 4, Yes Generate Offspring From Crossover/ Mutation/ Elitism Report The Pareto Set of Solutions Combine Population of Parent and Children Sort New Population Based on Nondomination Rank gen < gen_max gen = gen+1 Has The Population Changed From Last Time? Select Nondominated Set with the Best Rank Yes t Select The Next Population For The Next Generation From The Nondominated Set Is The NonDominated Set Smaller Than Initial Population? Yes Go To The Next Best Nondomination Set To Fill The Gap Figure 2.6 Flow Chart Representation of NSGA-H Algorithm 60 Specify The Parameters For MOPSO-CD Randomly Initialize Population Positions, Velocities, pbest and gbest Produce Next Swarm of Particles For Every Particle Evaluate The Fitness Values Insert New Nondominated Solution Into A and Delete All Dominated Solutions From A Compute The Crowding Distance values of Each Nondominated Solution in A and Sort In Descending Randomly Select Particle From a Specified Bottom Portion(e.g. Lower 10% and Replace It With The New Solution r Update pbest and gbest Next Particle ----- Update Particle Position and Velocity 1 Output The Pareto Solution Set End Figure 2.7 Flow Chart of MOPSO-CD algorithm 61 (Step 0) Initialize the preference parameters Calculation Phase (Step 1) Construct the DFs V (Step 2) Solve the DFMOOP using NSGA-II /1VIOPSO-CD 1 DecisionMaking Phase Evaluate the solution (Step 3) Is DM satisfied by the obtained guided front ? Adjust the DF's parameters according the wish of DM's choice of Pareto front Yes No V ( End Figure 2.8 Flow Chart of the Procedure Proposed: MOEA-IDFA 62 Table 2.1 Description of unconstrained Bi-objective Problems Variable Problem Characteristics Objective functions bounds minimize f = (f;(x), f2 (x)), where f,(x)= x 2 , SCH1 x E [0,2] convex, connected x e [-5,10] disconnected x, e [-5, 5] nonconvex, disconnected n=30 x, c [0,1], i =1,2,3,...,30 convex f2(x) = (x — 2)2 SCH2 minimize f = (f, (x), f2 (x)), where f, (x) = —x, if x __ 1, =-2+x, ifl.x3, =4—x, if34, = —4 + x, if x .?-.. 4, f2 (x) = (x —5)2 minimize f = (f, (x), f 2 (x)), where KUR f,(x)=E,_: (-10 exp(-0.2Vx,2 -1- x,2+, )), f2(x) = yii (1,108± 5sin x;) minimize f = (f,(x), f 2 (x)), where f,(x) = x I, ZDT1 f2 (x) = g(x)[1— Vx, / g(x)], 9" g(x) =1 +X x, 11 - 1 i „.2 minimize f = (fi (x), f2 (x)), where f,(x) = x,, ZDT2 n=30 f2(x) = g(x)[1— (x, / g(x))2 I, x, E [0,1], concave i =1,2,3,...,30 9" x. g(x) =X 1+ n-1 r=2 minimize f = (f (x), f 2 (x)), where .1;(x) = x„ ZDT3 n=30 f2 (x) = g(x)[1— V x, 1 g(x) —.-i -1;0-- sin(107TX, )1, X, E [0,1], i =1,2,3,...,30 9" E g(x) =1+ x; n —1 t=2 63 convex, disconnected Table 2.2 Description of Constrained Bi-objective and Unconstrained Tri-objective Problems Problem Variable bounds Objective functions • minimize f = (f, (x), f2 (x)) where f,(x) = x,, f2(x)–x2 TNK Characteristics r x, subject to: x,2 + x22 –1-0.1 cos 16 arctan — ■ .x2 0, x, c [0, a], i =1& 2 nonconvex, disconnected e [-2,2], i=1&2 bended surface, connected x e [0,1] convex x, e [0,1], i = l& 2 convex (x, – 0.5)2 + (x, – 0.5)2 -_ ._ 0.5 minimize f = (f, (x), f2 (x), f3 (x)), where VNT f (x) = x;+ (x2 – 1)2 , f2 (x) = X12 + (x2 + 1)2 ± 1, X f3 (x) = (x, –1)2 + x22 + 2 minimize f = (f (x), f 2 (x), f,(x)), where MHHM1 f;(x) = (x – 0.8)2 , f2 (x) = (x – 0.85)2 , f,(x)= (x –0 .9)2 minimize f = (f (x), f2 (x), f3 (x)), where MHHM2 f(x) = (x, –0 .8)2 + (x2 – 0.6)2 , f2 (x) = (x, – 0.85)2 + (x2 – 0.7)2 , f,(x)= (x, – 0.9)2 ± (x2 – 0.6)2 64 Table 2.3 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (Linear DF Case) Table 2.3 (a) A Priori Parameters for SCH1 .f. A*• 0 4 • 2 J. 4 0 Table 2.3 ( b)A Posteriori Parameters for SCH1 Common Parameters NSGA-H Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 50 0.4 200 0.4 1 1 0.9 5 - Table 2.3 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for SCH1 Mean Median S.D. Best Worst MOPSO-CD 397.58253 397.58254 0.00029 397.58294 397.58212 NSGA-H 416.88258 416.88259 0.00029 416.88299 416.88212 0 NSGA-II MOPSO-CD 3- 1 03 0 f l Figure 2.9 POFs of SCHlw.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 65 Table 2.4 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (Linear DF Case) Table 2.4 (a) A Priori Parameters for SCH2 i fi.• Jr* f2.... -1 1 0 16 Table 2.4 ( b) A Posteriori Parameters for SCH2 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Size rop Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 100 0.4 200 0.4 1 1 0.9 5 pu„-,—(, --Li..4e)e 5 Table 2.4 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for SCH2 Mean Median S.D. Best Worst MOPSO-CD 407.25158 407.25163 0.00029 407.25199 407.25113 NSGA-II 407.97659 407.97659 0.00028 407.97699 407.97617 15 - 0 NSGA-II MC)PSO-CD 129 6 30- 4:11:17ttarmlaurtmtP11.10:.04. • 1 -0.8-0.4 0.00.40.8 f Figure 2.10 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 66 Table 2.5 Parameters and Hypervolumes for KUR using MOEA-IDFA (Linear DF Case) Table 7.2 (a) A Priori Parameters for KUR f,.4 f: 0 4 f; fr 0 4 Table 7.2 ( b)A Posteriori Parameters for KUR Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 100 0.3 200 0.5 1 1 0.8 5 1\49v.4-4. ,.T%—.4ex 5 Table 7.2 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for KUR Mean Median S.D. Best Worst MOPSOCD 1224.80032 1224.80032 0.00028 1224.80079 1224.80000 NSGA-II 1226.60059 1226.60058 0.00030 1226.60099 1226.60013 0 NSGA-1I - • MOPSO-CD -4 -6-8 -10 12 1 •I•1• -20-19-18-17-16 fI Figure 2.11 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 67 Table 2.6 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (Linear DF Case) Table 2.6 (a) A Priori Parameters for ZDT1 fi•• fi 1 0 f2,. A 0 1 Table 2.6 ( b) A Posteriori Parameters for ZDT1 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W Cl c2 Cross. Prob. Cross. Index Z.1,-.4e..p /4,k jr 200 300 0.1 200 0.6 1.1 1.1 0.8 15 15 Table 2.6 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for ZDT1 Mean Median S.D. Best Worst MOPSO-CD 399.59859 399.59860 0.00028 399.59899 399.59816 NSGA-II 399.59554 399.59554 0.00025 399.59592 399.59512 Figure 2.12 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 68 Table 2.7 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (Linear DF Case) Table 2.7 (a) A Priori Parameters for ZDT2 fi. ii." .1; f2" 0 1 0 1 Table 2.7 ( b) A Posteriori Parameters for ZDT2 Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w Cl 200 300 0.1 200 0.5 1.1 c2 1.1 NSGA-H Parameters Cross. Prob. Cross. Index 0.9 15 pl). j =%-...ftx 15 Table 2.7 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT2 Mean Median S.D. Best Worst MOPSO-CD 399.29058 399.29061 0.00026 399.29097 399.29018 NSGA-1I 399.28358 399.28362 0.00028 399.28399 399.28315 0 NSGA-II • 0.8 MOPSO-CD 0.4 - 0.0 0.0 0.4 f1 Figure 2.13 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 69 Table 2.8 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (Linear DF Case) Table 2.8 (a) A Priori Parameters for ZDT3 f.;.. f• • f,* f; 0 8.52 -0.773 1 Table 2.8 ( b) A Posteriori Parameters for ZDT3 Common Parameters NSGA-II Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w Cl 200 300 0.1 200 0.6 1.1 c2 Cross. Prob. Cross. Index 1.1 0.9 15 15 Table 2.8 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for ZDT3 Mean Median S.D. MOPSO-CD 414.75153 414.75151 0.00029 414.75195 414.75110 NSGA-H 414.72459 414.72461 0.00027 414.72410 1.0- Best 414.72500 Worst 0 NS GA-II • 0.5- MOP SO-CD 00- -o. 5 -1 I 0.00.20.4 fl 0 .60.8 Figure 2.14 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 70 Table 2.9 Parameters and Hypervolumes for TNK using MOEA-IDFA (Linear DF Case) Table 2.9 (a) A Priori Parameters for TNK 1 .4.. f2 0 1.05 0 f2. 1.05 Table 2.9 ( b)A Posteriori Parameters for TNK Common Parameters MOPSO-CD Parameters Pop - Size Max. Gen. Mutation Prob. Arch. Size w 200 200 0.5 200 0.4 cl 1 c2 1 NSGA-II Parameters Cross. Prob. 0.8 cross. .A.u::,. . 4% Index --,-tg-3, 5 5 Table 2.9 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for TNK Mean MOPSO-CD NSGA-II Median S.D. 399.22257 399.22257 397.32452 397.32453 0.00027 Best Worst 399.22299 399.22212 0.00028 397.32496 397.32412 0 NSGA-II OPS 0-CD 1.0 - 0.8 - 0.20.0 - 0.00.2• 0.40.6 f 0.81.0 l Figure 2.15 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 71 Table 2.10 Parameters and Hypervolumes for VNT using MOEA-IDFA (Linear DF Case) Table 2.10 (a) A Priori Parameters for VNT fi. fi.,.. 0 4 f2• .f2 5 1 f3.• A 2 4 Table 2.10 ( b) A Posteriori Parameters for VNT Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W 200 250 0.5 200 0.4 . NSGA-II Parameters CI c2 Cross. Prob. Cross. Index 1 1 0.9 10 ,444.J, - .04,4.4.4.10 Table 2.10 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for VNT Mean Median S.D. Best Worst MOPSOCD 6093.40050 6093.40053 0.00032 6093.40099 6093.40000 NSGA-II 6095.10050 6095.10051 0.00032 6095.10099 6095.10000 Figure 2.16 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 72 Table 2.11 Parameters and Hypervolumes for MHHM1 using MOEA-IDFA (Linear DF Case) Table 2.11 (a) A Priori Parameters for MHHM1 .A. f" f2 f2" f3 f3" 0 0.01 0 0.0025 0 0.01 Table 2.11 ( b)A Posteriori Parameters for MHHM1 Common Parameters MOPSO-CD Parameters . Pop -Si ze Max. Gen. Mutation Prob. Arch. Size w 200 250 0.6 200 0.4 cl 1 c2 1 NSGA-II Parameters Cross. Prob. Cross. Index 0.9 10 --si.1.44cr,211 - 10 Table 2.11 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for MHHM1 Mean Median S.D. Best Worst MOPSOCD 7999.00051 7999.00052 0.00032 7999.00097 7999.00000 NSGA-1I 7999.00040 7999.00038 0.00014 7999.00066 7999.00011 MOPS 0-CD 0 NSC.+A -II 1-n/ 0.0025 11 114/41,TPU 1:0 PP 5.Z1 0.000 0.0020 0.002 0.0015 0.004 0.0010 0.006 0.0005 0.008 0.0000 0.010 Figure 2.17 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 73 Table 2.12 Parameters and Hypervolumes for MHHM2 using MOEA-IDFA (Linear DF Case) Table2.12 (a) A Priori Parameters for MHHM2 f .... A. 0 A A 0.0125 0 0.0125 0.0125 0 Table 2.12 ( b) A Posteriori Parameters for MHHM2 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W CI c2 Cross. Prob. Cross. Index AAA.A4, 200 300 0.5 200 0.4 1 1 0.9 15 15 r1,...t10.0 Table 2.12 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for MHHM2 Mean Median S.D. Best Worst MOPSOCD 7995.10051 7995.10051 0.00031 7995.10095 7995.10000 NSGA-1I 7995.00051 7995.00052 0.00031 7995.00095 7995.00000 MOPSO-CD NSC+A-II 0214 0.008 0.006 0.004 0.002 0.000 4/002 0..0 0.061120 0.004 0.001 000 .0002 Figure 2.18 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Linear DF Case) 74 CHAPTER 3 Guided POF Articulating Nonlinear DFA with an MOEA (Convex-Concave Combination) In this chapter, we articulate nonlinear (convex and concave) DF as a priori with an MOEA (NSGA-II/MOPSO-CD) as a posteriori, and demonstrate how in a single run, instead of one solution, a preferred set of solutions near the desired region of DM's interest can be found. Consequence of combination of convex-concave DF is analyzed theoretically as well as numerically (on ten different benchmark problems). A set of Pareto-optimal solutions is determined via DFs of the objectives, which incorporates DM's preferences regarding different objectives. 3.1 Introduction Having been well demonstrated the task of finding multiple POFs in MOOPs, the MOEA researchers and applicationists should now concentrate in devising methodologies of solving the complete task of finding preferred and POFs* in an interactive manner with a DM. Although the ultimate target in such an activity is to come up with a single solution, the use of an MOEA procedure can be well applied with a decision-making strategy in finding a set of preferred solutions in regions of interest to the DM, so that the solutions in a region collectively bring out properties of the solutions there. Such an activity will then allow the DM to first make a higher-level search of choosing a region of interest on the POF, rather than using a single solution to focus on a particular solution. In general, solution of MOOPs can be classified vis-à-vis at which stage the preferences of the DM are integrated, discussed earlier in Section 1.1 of Chapter 1. A priori optimization methods, e.g. desirability index (Harrington, 1965) utilize DM's knowledge that has to be specified upfront. These techniques generate single point solution only. A posteriori methods, however, create a set of solutions (i.e., . Paretooptimal solutions) or Pareto set. Most MOEA approaches can be classified as a posteriori. They attempt to discover the whole set of Pareto-optimal solutions or at least a well-distributed set of representatives. The DM then looks at the set (possibly 75 very large) of generated alternatives and makes final decision based on his/her preferences. However, if the Pareto-optimal solutions are too many, their analysis to reach the final decision is quite a challenging and burdensome process for the DM. In addition; in a particular problem, the user may not be interested the complete Pareto set; instead, the user may be interested in a certain region of the Pareto set. Such a bias can arise if not all objectives are of equal importance to the user. Finding a preferred distribution in the region of interest is more practical and less subjective than finding one biased solution in the region of interest. Although it is usually tough for a DM to completely specify his or her preferences before any alternatives are known, the DM often has a rough idea of the preferential goals towards which the search should be directed, so that he/she may be able to articulate vague, linguistic degrees of importance or give reasonable trade-offs between the different objectives. Such information should be integrated into the MOEA to bias the search towards solutions that are preferential for the DM. This would in principle yield two important advantages: (1) Instead of a diverse set of solutions, many of which irrelevant to the DM, a search biased towards the DM's preferences will yield a more fine-grained, and thus more suitable, selection of alternatives; (2) By focusing the search onto the relevant part of the search space, the optimization algorithm is expected to find these solutions more quickly. To achieve this task a methodology is proposed in this chapter. In the most practical cases generally, DM have at least a vague idea in which region/regions of the objective the 'optimums' should be. For this purpose,. DFs that map the objectives to the interval [0, 1] according to their desired values can be specified as a priori to state a relative preference of the objectives. The Pareto-optimal solutions are then determined for the DFs instead of the objective functions using an MOEA. Thus, the proposed method is a fusion of an a priori and an a posteriori approach. Furthermore, this application in practice is quite straightforward. In addition, in spite of the focus on the desired region/regions, the MOEA search can be carried out without any additional constraints to be included for the decision variables and/or objectives. Therefore, no individuals are 'wasted' during the optimization process in case they do not fit into the constrained intervals. The MOEA is not touched at all 76 which facilitates the use of the Method in practice as existing optimization tools can be used for optimization process as a posteriori. In other words, this approach is independent of the MOEA used and can be easily coupled to any of them without any deep modification to the main structure of the method chosen. For comparison purpose we are using two of them namely NSGA-II and MOPSO-CD. Once the DM has agreed on the parameters of the DFs, which is the key step, MOEAs can be applied on the transformed objective space (i.e., DFMOOP) without modification (Trautmann and Mehnen, 2009). By means of DFs, the solutions concentrate in the desired region/regions which facilitate the solutions selection process and support the MOEAs in finding relevant solutions. If DM has single/multiple region/regions as preference/preferences for an objective, modification in the DF should be done accordingly. For example, if DM is interested in finding the any corner portion of the POF of a bi-objective minimization problem the DF corresponding the objectives can be adapted using convex and concave DFs. We shall discuSs the consequence of using convex and concave DFs together in this chapter. Jeong and Kim (2009) proposed an interactive desirability function approach (IDFA) for multirespose optimization to facilitate the preference articulation process. One run of this approach provides just one solution to the DM based on his preference in the form of DF. Present work is an extension of the work done by Jeong and Kim (2009). We present a MOEA based IDFA (MOEA-IDFA) to provide DM a preferred portion of the POF rather than providing just a single solution. Section 3.2 describes the prerequisites of nonlinear (convex-concave) DFA as a priori. The methodology of the proposed approach is provided in Section 3.3. Necessary theorems are also provided in this section in order to analyze the methodology theoretically. Results corresponding to the test problems from both MOEAs (NSGA-II and MOPSO-CD) are compared and discussed in Section 3.4. Finally, concluding observations are drawn in Section 3.5. 3.2 Nonlinear (Convex /Concave) DFA as a Priori If a DM has some preference (i.e., biasness) towards one or more objectives, a nonlinear DF may fulfill the purpose. Different shapes of DF will provide different type of biasness towards different objective region of the POF. The form of the DF 77 originally proposed by Harrington (1965) was based on the exponential function of a linear transformation of the A 's. Derringer and Suich (1980) found the specification not very flexible in the sense that the DFs cannot assume a variety of shapes. Derringer and Suich (1980) introduced a modified (alternative) version of DF for both one sided and two sided case based on a power of a linear transformation of the f t 's. One-sided case (STB type) formulation of a DF is given in Equation 3.1, as we are concerned with one sided specification shown in Figure 3.1. if f, < f . if f: < A 5 f," ;i = 1, 2, 3, ..., ke R+ (3.1) if f .. < f Where parameters f" andare minimal and maximal acceptable levels of f, respectively. The DFs u, 's are on the same scale and are discontinuous at the points f,* , A and f** . The values of n, (a kind of weighting factor) can be chosen so that the DF is easier or more difficult to satisfy. Parameter ni is a positive constant whose increasing magnitude creates a correspondingly more convex desirability curve. For example, if n, is cliosen to be greater than one in Equation 3.1, p, is near 0 even if the A is not low, making the DF more easier to satisfy in terms of desirability. As values of n, move closer to 0, the desirability reflected by Equation 3.1 becomes lower making it harder to satisfy in terms of desirability. Linear DF (i.e. for ni =1) has been analyzed already in Chapter 1. Use of Linear DF does not provide biasness towards objectives. In this chapter, we are going to discuss the effect of combination of convex and concave DFs for MOOP so the Equation 3.1 will be modified slightly for the further uses, as convex DF (Equation 3.2) and as concave DF (Equation 3.3) shown in Figures 3.1-3.2. Both of these DFs are monotonic between the lower (f ' ) and upper ( ) bound of their definition. To the authors best knowledge this is the first attempt to incorporate convex and concave DF together. In both the cases n, is the key parameter. 78 if f < if f:;i = 1, 2, 3, ..., k ;n1 ; n1 E R+ (3.2) if < .1; CE91,1 Ite__ Figure 3.2_Shape of a -Goirrex-DF if f, < f . if ;i = 1, 2,3, ..., k ; /I;E R+ , if f .* < f; Figure 31:Shape of a-Concave DF 79 (3f.3) 3.3 Proposed Methodology MOOP is already described in Section (1.1) given by Equation P1 Minimize f (x)= {f1 (x), f2 (x),..., fk (x)} xeX (P1) Eliciting the corresponding DF to each objective (i.e. f,fk ) through the interaction of DM a new DF based multi-objective optimization problem (DFMOOP) consisting the DFs is formulated given by Equation P2. Minimize ,u(x) = Ipi (x), p2 (x),...„ uk (x)} rGx (P2) Where, for each value of an objective function f; , there is a mapping called DF i.e., to prescribe the variation of vagueness involved discussed in Section 3.2. The overall (resulting) DF's value (say ,u) is also between zero and one. The Pareto-optimal solutions are then determined for this newly formed DFMOOP instead of the original MOOP. Solutions of DFMOOP have a unique relationship with the original MOOP of objective functions. In general, DFMOOP is solved using different aggregators, min and product operator are most common, providing a single solution of DFMOOP. This type of approach is repeatedly applied for different degrees of satisfaction values until the DM is satisfied. Benefit of this technique lies in the fact how well DM's preferences have been incorporated in a priori approach, which is quite rare due to vague nature of human judgment. So in the present approach DFMOOP is solved using purely multi-objective manner (using the discussed algorithms i.e., NSGA-II and MOPSO-CD) without aggregating. Present approach is an attempt to incorporate the benefits of both the a priori and the a posteriori methods together. Now we are going to prove a theorem, which establish the relationship between the Pareto-optimal solutions of P2 and Pl. 80 Theorem 3.1: The Pareto-optimal solutions*Of P2 corresponding the DFs in Equation 3.2 and Equation 3.3 are Pareto-optimal solutions of Pl. Proof: Let x* be Pareto-optimal solution for P2. Now depending on the DM's priority p, (ormay be defined by either Equation 3.2 or Equation 3.3. By the definition of Pareto-optimal solution E X pi (X) < (X ) ; for i {1,2,...,1c} and ,u (x) ,u, (x* ); for j E {1, 2, k}; j # i f(x) — f* <( f(x*)— f* f, —f ,f** —f <=> E X for i E {1, 2,...,14;n, EU + ;n, 1 or n, 5 1, depending on the shape of as1 or Ili 1, and f,* S f,(x), f(x )z f,' ( f (x)— f * <(f,(x* )— f,* g f** — f* f** f* ;usin the monotonicities of the DFs Again using the monotonicities of the DFs for j n, also n,or 11, 1 and f *j fi (x), f., (x* )fr indicates (3.4) or n, j # i .n, ( fi (x)— f, < [f,(x* )— f; vil ,..•,.• — •• • J." - - - -1.1i fi - fi f ,(x)— 1 ** * JeJ [f J (x*)— < ** *f J7 (3.5) f l f l Inequality (3.4)f, (x)f * < f(x) — f* as f ** <:=> f (x) < fi (x* ); for i E {1, 2, ..., Similarly, Inequality (3.5) <=> fi (x)— f;f; as f;* — <=> 1,(x)f '); for j E {1,2,..., lc} j # i > 0 (3.6) > 0, j (3.7) The proposition holds as Inequalities 3.6 and 3.7 together form the condition for Pareto-optimal solution of Pl. 81 3.3.1 Detailed Procedure of the Methodology: MOEA-IDFA The methodology consists of five steps. Step 0 is an initialization step. Steps 1 and 2 constitute the calculation phase and Steps 3 and 4 the decision-making phase. Figure 1.8 shows the procedure of MOEA-IDFA. In the calculation phase, the DFs are constructed, and then an optimization model of DFs is solved using an MOEA. In the decision-making phase, the DM evaluates the results of the calculation phase, and then articulates his/her preference information. More specifically, if the DM is satisfied with the results on all the objectives, the procedure successfully ends. Otherwise, the DM adjusts the parameters (shape and bound) of a DF. Then, the procedure goes back to the calculation phase. Each step is described below. Step 0: Initialization of DF's Parameters The DF's parameters (preference parameters on each objective) are to be initialized, to construct the DF for each objective in the first iteration. The initial bound and goal (target) may be determined based on the DM's subjective judgments. Ideal and antiideal vectors f* and r , of corresponding ith objective function f. should be calculated in advance for the given MOOP (Equation P 1). The initial bound and shape may be determined based on the DM's subjective judgments. Step 1: Construction of the DFs From this step calculation phase starts. As mentioned earlier, the preference parameters initialized in Step 0 are used to construct the nonlinear (convex or concave) DF Ai, for each f of P1 . DM preference can be utilized in the construction of DF via the value of n, (a kind of weighting factor determines shape of DF). Thus, a new DFMOOP (see Equation P2) is formed. Step 2: Solving the DFMOOP This newly formed P2 is then solved using an efficient MOEA. As we are discussing general MOOP which can be nonlinear, nonconvex, multimodal and multivariable. Hence, resulting DFMOOP will also be of the same nature. We need some powerful 82 algorithm to solve this optimization problem. NSGA-II and MOPSO-CD are two such competent techniques capable of solving complex MOOPs used here. Step 3: Evaluation of the solution Present the solution to the DM obtained in Step 2. Theorem 3.1 and 3.2 together imply that the Pareto-optimal solution obtained by solving P2 is also the Pareto-optimal solution of Pl. However, the POF of P2 is different from P1 due to different function formulation. Therefore, they produce a section of POF of. Pl, guided according the choice of the DF. For example, for a bi-objective minimization problem using a convex and a concave DF produces biasness towards one objective. If the DM is satisfied by the Paretooptimal solutions, the methodology is terminated. Else, the procedure goes to Step 4. Step 4: Adjusting the Preference Parameters In order to improve the unsatisfactory results, interaction with DM can help modifying (updating) the Step 0 and Stepl. On the other hand, if the DM is fully satisfied by the obtained guided POF the procedure ends successfully. The whole process may be repeated until DM is satisfied. 3.4 Results and Discussion The proposed approach in this chapter has been applied to a set of ten standard test problems (SCH1, SCH2, KUR, ZDT1, ZDT2, ZDT3, TNK, VNT, MHHM1 and MHHM2). As conversed earlier in the Section 3.2, convex-concave combination of DFs is used as a priori, while NSGA-II/MOPSO-CD applied as a posteriori in the proposed approach. As described in the Step 0 of Subsection 3.3.1 of the approach, the parameters of a priori method i.e., nonlinear (convex or concave) DFs are obtained with the help of DM corresponding each objective. Initial parameters of DF are used for all test problems are given in Tables 3.1(a)-3.10(a). In Step 1, preference parameters initialized in Step 0 are used in construction of the convex or concave DFs (,uis ) for each f with the help of DM, and forming the 83 DFMOOP (i.e., P2). This newly formed P2 is solved using an a posteriori method as shown in Step 2. Since our methodology is based on MOEA used, we have used both NSGA-II and MOPSO-CD as a posteriori method to estimate the effectiveness of the approach on different MOEAs. The values of different set of parameters are tested and fine-tuned through several runs of MOEA (NSGA-II/MOPSO-CD) methods used in this approach. NSGA-II and MOPSO-CD parameters used for each problem are given in Tables 3.1(b)-3.10(b). In Step 2, computations have been carried out based on these parameters and results have been reported. To compare the efficiency of NSGA-II and MOPSO-CD, hypervolume metric is used here, for them reference points are taken (11, 11) and (11, 11, 11) for bi-objective problems and tri-objective problems respectively. Tables 3.1(c)- 3.10(c) display the mean, median, standard deviation, best and worst values of hypervolume metrics obtained by using NSGA-II and MOPSO-CD with the help of 10 runs each. For the steadiness of the results for each problem, 10 runs of each MOEA (NSGA-II and MOPSO-CD) are done under the same parameters and the best POF (based on best hypervolume metric) obtained is reported. The POFs of the problems are depicted in Figures 3.3-3.12, which clearly imply that use of convex-concave combination of DF guides the POF towards one particular end of the POF. Having a look on the Tables 3.1(c)-3.10(c) reveal that if we consider mean of hypervolume metric as the competence parameter for both NSGA-H and MOPSO-CD, then NSGA-II performed better in case of KUR, ZDT1, ZDT2, VNT, MHHM1 and MHHM2 while MOPSO-CD was better in the cases SCH1, SCH2, ZDT3 and TNK for the used parameters parameter settings. If the DM is satisfied by the POF obtained then the procedure ends successfully as explained in Step 4. We are just showing one iteration of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. However, if the DM is not satisfied then the procedure goes back to Step 1 to accumulate the DM's preferences in a better manner. 84 3.4.1 Effect of Variations in DF's Key Parameter on POF If the DM is unsatisfied by the outcome of Step 4 of the proposed approach then analyst calls for the variation in DF. Thus, it is significant to understand the effect of the parameters of DF on POF. It is obvious from the results and discussions in Section 3.4 that use of convex-concave combination of DFs produces biased Pareto-optimal c@vic_ci v-e_ solutions towards the objective having,sonw4K-DF. To investigate the effect of the DF's key parameter further, we are taking SCH1 as our problem and NS.GA-II as the MOEA having the same a posteriori setting as used in Table 3.1 (b). As the key parameter in case of a convex or concave DF p1 is n, we are using rest of the parameters same as taken in Table 3.1 (a) and will vary n to observe the effect of it. Different combination of n have yielded different portions of POF of SCH1 as shown in Figures 3.13. The key parameters n1 and n2 used for Figure 3.13 (a) are having values 5 and 0.1 respectively. Similarly the values of n1 and n2 used for Figure 3.13 (b) are taken 10 and 0.01 respectively. Figure 3.13 (c) depicts the POF in case of n1 and n2 are taken 15 and 0.001 respectively. 25 and 0.0001 are the value used for n1 and n2 respectively in Figure 3.13 (d). It is clear from the Figures 3.13 that increment in the value of n1 while simultaneous decrement in the value of n2 contracts POF further towards the biased portion. 3.5 Conclusion In this chapter, a partial user preference approach named MOEA-IDFA having nonlinear DF has been proposed. In this approach, nonlinear (convex-concave combination of DFs) DFA as a priori and NSGA-II/MOPSO-CD as a posteriori are combined together to provide an interactive POF. The theoretical analysis of the proposed MOEA-IDFA is also provided in this chapter. It has been observed that use of convex DF together with concave DF produces a definite bias among the Paretooptimal solutions of original MOOP. This type of approach is useful if the DM prefers 85 a particular objective over another objective.The performance of the MOEA-IDFA is tested on a set of 10 test problems. Similar approach to find a bias in the POF proposed by Branke et al. (2001) reported results in convex POF only. This approach is better than the method proposed by Branke et al. (2001) in terms of its ability to produce biasness in the disconnected and non-convex POFs as well. According the author's best knowledge, there is no works till date, in literature that comprise the biasness among tri-objective problems. In this chapter, we successfully apply the MOEA-IDFA using sigmoidal DFs to three standard tri-objective problems (VNT, MHHM1 and MHHM2). al9 86 Table 3.1 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (Convex-Concave Combination) Table 3.1 (a)A Priori Parameters for SCH1 A... f 0 4 jr .r f, ; n1 n2 0 2. 0.001 4 Table 3.1 ( b)A Posteriori Parameters for SCH1 Common Parameters NSGA-II Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 150 0.5 200 0.5 1 1 0.8 5 Mme. -,,.,„,k. „5, 5 Table 3.1 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for SCH1 Mean Median S.D. Best Worst MOPSO-CD 392.62053 392.62060 0.00024 392.62087 392.62011 NSGA-I1 392.30585 392.30585 0.00001 392.30586 392.30583 4- O 0 NSGA-II - 1\IOPSO-CD 3- ri 2 - 4-a 1 0 VO 0 .." 4 01 Figure 3.3 POFs of SCH1w.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave. Combination) 87 Table 3.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (Convex-Concave Combination) Table 3.2 (a) A Priori Parameters for SCH2 fi..• -1 f,... f; A 1 0 16 ni n2 10 0.001 Table 3.2 ( b) A Posteriori Parameter.s for SCH2 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. 200 150 0.4 200 0.4 1 1 0.9 Cross. 4A4..2 Index ..G---A.>. 10 10 Table 3.2 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for SCH2 Mean Median S.D. Best Worst MOPSO-CD 403.58635 403.58636 0.00005 403.58643 403.58622 NSGA-H 402.90365 402.90364 0.00002 402.90368 402.90362 O 15 - NSGA-II MOPSO-CD 12 963Ctittli:Muazammo 0 0 1 Figure 3.4 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 88 Table 3.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (Convex-Concave Combination) Table 3.3 (a) A Priori Parameters for KUR 11 .4". -20 -14.4 f; ./2.. -11.6 0 ni n2 10 0.002 Table 3.3 ( b) A Posteriori Parameters for KUR Common Parameters MOPSO-CD Parameters Pop • Size Max. Gen. Mutation Prob. Arch. Size w 200 150 0.3 200 0.5 NSGA-II Parameters Cl c2 Cross. Prob. Cross. Index 1 1 0.8 10 0-1—x—i- • '5-- 5 T' Table 3.3 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for KUR Mean Median S.D. Best Worst MOPSO-CD 1219.10046 1219.10043 0.00025 1219.10086 1219.10005 NSGA-1I 1223.10006 1223.10006 0.00002 1223.10008 1223.10002 I•1I• -20-19-18-17 f1 Figure 3.5 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 89 Table 3.4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (Convex-Concave Combination) for ZDT1 Table 3.4 (a) AParameters Priori fi* J1`• 0 1 ' f2 s n n2 0 1 12 0.01 Table 3.4 ( b)A Posteriori Parameters for ZDT1 Common Parameters Pop Size 200 Max. Gen. Mutation Prob. 300 0.1 MOPSO-CD Parameters Arch. Size 200 w 0.6 Cl C. 1.1 1.1 NSGA-H Parameters Cross. Prob. Cross. Index frtwt. 0.8 15 15 D...44 Table 3.4 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT1 Mean Median S.D. Best Worst MOPSO-CD 399.53662 399.53655 0.00024 399.53698 399.53625 NSGA-H 399.63846 399.63846 0.00002 399.63846 399.63842 1.0 0.8 0.6 0.4 - 0 O 0 NSGA-II " MOPSO-CD 0 oo 0 0.2 - 0.0 0I00.20.4 f1 0. 6 0.8 110 Figure 3.6 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 90 Table 3.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (Convex-Concave Combination) Table 3.5 (a) A Priori Parameters for ZDT2 f, A 0 f2... f2 1 1 0 ni n2 20 0.001 Table 3.5 ( b) A Posteriori Parameters for ZDT2 Common Parameters MOPSO-CD Parameters NSGA-II Parameters -• Pop , Size Max. Gen. Mutation Prob. Arch. Size W Cl C2 Cross. Prob. 200 350 0.2 200 0.5 1.1 1.1 0.8 Cross. iii,,A,1. Index T-i...:415a 15 15 Table 3.5 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT2 Mean Median S.D. Best Worst MOPSO-CD 399.28417 399.28372 0.00163 399.28878 399.28344 NSGA-H 399.29071 399.29076 0.00016 399.29078 399.29024 NSGA-II 0 • 0 0 o MOPSO-CD o 0.4 0.2 0.0 - 0.00.20.40.60.81.0 f1 Figure 3.7 POFs of ZDT2 w.r.t. NSGA-H and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 91 Table 3.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (Convex-Concave Combination) Table 3.6 (a) A fi• 0 .r. 8.52 Priori Parameters for ZDT3 f2.• f2 1 -0.773 ni n2 10 0.00 1 Table 3.6 ( b) A Posteriori Parameters for ZDT3 Common Parameters NSGA-II Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W CI c2 Cross. Prob. Cross. Index MA.,),--. 200 350 0.1 200 0.6 1.1 1.1 0.8 15 15 D---ett->o, Table 3.6 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT3 Mean Median S.D. Best Worst MOPSO-CD 414.65639 414.65627 0.00023 414.65688 414.6562 NSGA-II 414.58246 414.58246 0.00003 414.58249 414.58242 Figure 3.8 POFs of ZDT3 w.r.t. NSGA-H and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 92 Table 3.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (Convex-Concave Combination) Table 3.7 (a) A Priori Parameters for TNK f; fi .4 0 f2" 1.05 0 1.05 nI n2 10 ` 0.1 Table 3.7 ( b) A Posteriori Parameters for TNK Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 200 0.5 .200 0.5 cl 1 C. 1 NSGA-II Parameters Cross. Prob. Cross. Index --1-1--4-o. 0.8 10 10 Alt,--1. Table 3.7 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for TNK Mean Median S.D. Best Worst MOPSO-CD 397.29493 397.29492 0.00002 397.29495 397.29491 NSGA-II 397.04872 397.04882 0.00025 397.04886 397.04802 NSGA -II 1 .0 - Q4 • MOPS 0-CD e. 0.80_.6 w1 0.4 0.2 0 - 0.00.20.40.60.81.0 f l Figure 3.9 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 93 Table 3.8 Parameters and Hypervolumes for VNT using MOEA-IDFA (Convex-Concave Combination) Table 3.8 (a) A Priori Parameters for VNT I f:* f; f2* f2 ic n, n, n, 0 4 1 5 2 4 5 0.1 0.1 Table 3.8 ( b) A Posteriori Parameters for VNT Common Parameters MOPSO-CD Parameters NSGA-H Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w Cl c2 Cross. Prob. Cross. Index Ayet).41 StAvg, 200 250 0.6 200 0.4 1 1 0.8 10 10 Table 3.8 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for VNT Mean Median S.D. Best Worst MOPSO-CD 6069.00002 6069.00001 0.00001 6069.00005 6069.00001 NSGA-I1 6080.00004 6080.000045 0.00002 6080.000074 6080.00000 Figure 3.10 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 94 Table 3.9 Parameters and Hypervolumes for MHHMI using MOEA-IDFA (Convex-Concave Combination) Table 3.9 (a)A Priori Parameters for MHHMI fi• fi.. J; .1.2... 0 0.01 0 0.0025 J 3" f;* n, n, ri, 0 0.01 5 5 0.1 Table 3.9 ( b) A Posteriori Parameters for MHHM1 Common Parameters NSGA-II Parameters MOPSO-CD Parameters . • Pop Size Max. Gen. Mutation Prob. Arch. Size w -i c2 Cross. Prob. Cross. Index 200 250 0.5 200 0.5 1 1 0.9 15 15 Table 3.9 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for MHHMI Mean Median S.D. Best Worst MOPSO-CD 7999.00452 7999.00344 0.00326 7999.00990 7999.00132 NSGA-II 7999.00574 7999.00567 0.00178 7999.00866 7999.00267 ■ MOPSO-CID 0 NSGA-II slelat.1=trtetin-J13XY1''.27'1' 0.040 0.002 0.004 0.006 0.008 0.0025 0.0020 0.0015 0.0010. 0.0005•.r, 0.0000 0.010 Figure 3.11 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 95 Table 3.10 Parameters and Hypervolumes for MHHM2 using MOEA-IDFA (Convex-Concave Combination) Table 3.10 (a) A Priori Parameters for MHHM2 f3• fi* I. f; f2* f3 0 0.0125 0 0.0125 0 0.0125 'I, n2 n3 5 5 0.1 Table 3.10 ( b) A Posteriori Parameters for MIIHM2 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W cl c2 Cross. Prob. Cross. Index Aveni- 200 300 0.4 200 0.5 1 1 0.8 10 10 11/4-A.115G Table 3.10 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for MHHM2 Mean Median S.D. Best Worst MOPSO-CD 7995.00005 7995.00006 0.00004 7995.00009 7995.00001 NSGA-H 7995.00007 7995.000078 0.00002 7995.00009 7995.00003 Figure 3.12 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (Convex-Concave Combination) 96 43- O 0 4- in I -- .1 0 1-12 o° o f 3- O O O O 9) 0 0 O 0- O clmlizsramia• 00 1 f 0 f 1 (a) (b) 4- 4- -= 25 0 0.006 I .)-12___=•- 00 1 3- 1 31I 10 0- O 0 03111END 0- 0123 01 f 1 34 f1 (c) (d) Figure 3.13 Effect of variations in key parameter of DF on POF for SCH1 97 98 CHAPTER 4 Guided POF Articulating Nonlinear DFA with an MOEA (All Sigmoidal Combination) In this chapter, we articulate sigmoidal DF as a priori with an MOEA (NSGAHIMOPSO-GD)—as---a—posteriorh—and—demonstrate—ho-w;instead of one solution3—apreferred set of solutions near the desired portion of DM's interest can be found. A new type of sigmoidal DF is proposed in this Chapter that can guide the Pareto-optimal solutions close to the intermediate region of the original POF. Consequence of sigmoidal DFA as a priori is analyzed theoretically as well as numerically (on ten different standard test problems). 4.1 Introduction If a single solution is to be selected in a MOOP, at some point during the process, the DM has to reveal his/her preferences. Specifying these preferences a priori, i.e., before alternatives. are known, often means to ask too much of the DM. On the other hand, searching for all nondominated solutions as most MOEA do may result in a waste of optimization efforts to find solutions that are clearly unacceptable to the DM. This chapter introduces an intermediate approach, that asks for partial preference information from the DM a priori; and then focus the search to those regions of the Pareto optimal front that seem most interesting to the DM. That way, it is possible to provide a larger number of relevant solutions. It seems intuitive that this should also allow to reduce the computation time, although this aspect has explicitly only been shown in Branke and Deb (2005) and Thiele et al. (2007). In general, solution of MOOPs can be classified regarding at which stage the preferences of the DM are integrated, discussed earlier in Section 1.1 of Chapter 1. A priori optimization methods, e.g. desirability index (Harrington, 1965) utilize DM's knowledge that has to be specified upfront. These techniques generate single point solution only. A posteriori methods, however, .create a set of solutions (i.e., Paretooptimal solutions) or Pareto set. Most MOEA approaches can be classified as a posteriori. They attempt to discover the whole set of Pareto-optimal solutions or at 99 least a well-distributed set of representatives. The DM then looks at the set (possibly very large) of generated alternatives and makes final decision based on his/her preferences. However, if the Pareto-optimal solutions are too many, their analysis to reach the final decision is quite a challenging and burdensome process for the DM. In addition, in a particular problem, the user may not be interested the complete Pareto set; instead, the user may be interested in a certain region/regions of the Pareto set. Finding a preferred distribution in the region of interest is more practical and less subjective than finding one biased solution in the region of interest. Although it is usually tougher for a DM to completely specify his or her preferences before any alternatives are known, the DM often has a rough idea of the preferential goals towards which the search should be directed, so that he/she may be able to articulate vague, linguistic degrees of importance or give reasonable trade-offs between the different objectives. Such information should be integrated into the MOEA to bias the search towards solutions that are preferential for the DM. This would in principle yield two important advantages: (1) Instead of a diverse set of solutions, many of which irrelevant to the DM, a search biased towards the DM's preferences will yield a more fine-grained, and thus more suitable, selection of alternatives; (2) By focusing the search onto the relevant part of the search space, the optimization algorithm is expected to find these solutions more quickly. To achieve this task a methodology is proposed in this chapter. In the most practical cases generally, DM have at least a vague idea in which region/regions of the objective the 'optimums' should be. For this purpose, DFs that map the objectives to the interval [0, 1] according to their desired values can be specified as a priori to state a relative preference of the objectives. The Pareto-optimal solutions are then determined for the DFs instead of the objective functions using an MOEA. Thus, the proposed method is a fusion of an a priori and an a posteriori approach. Furthermore, this application in practice is quite straightforward. In addition, in spite of the focus on the desired region/regions, the MOEA search can be carried out without any additional constraints to be included for the decision variables and/or objectives. Therefore, no individuals are 'wasted' during the optimization process in case they do not fit into the constrained intervals. The MOEA is not touched at all 100 which facilitates the use of the method in ptactice as existing optimization tools can be used for optimization process as a posteriori. In other words, this approach is independent of the MOEA used and can be easily coupled to any of them without any deep modification to the main structure of the method chosen. For comparison purpose we are using two of them namely NSGA-II and MOPSO-CD. Once the DM has agreed on the parameters of the DFs, which is the key step, MOEAs can be applied on the transformed objective space (i.e., DFMOOP) without modification (Trautmann and Mehnen, 2009). By means of DFs, the solutions concentrate in the desired region/regions which facilitate the solutions selection process and support the MOEAs in finding relevant solutions. If DM has single/multiple region/regions as preference/preferences for an objective, modification in the DF should be done accordingly. For example, if DM is interested in finding -the intermediate portion of the POF of a bi-objective minimization problem the , DF corresponding the objectives can be modified using sigmoidal DFs. We propose a new kind of DF based on sigmoidal function and discuss the consequence of using sigmoidal DFs in this chapter. Jeong and Kim (2009) proposed an interactive DF approach (IDFA) for multirespose optimization to facilitate the preference articulation process. One run of this approach provides just one solution to the DM based on his preference in the form of DF. Present work is an extension of the work done by Jeong and Kim (2009). We present a MOEA based IDFA i.e., MOEA-IDFA to provide DM a preferred portion of the POF rather than providing just a single solution. Section 4.2 describes the prerequisites of sigmoidal DFA as a priori. The methodology of the proposed approach is provided in Section 4.3. Necessary theorems are also provided in this section in order to analyze the methodology theoretically. Results corresponding to the ten test problems (Section .2.5) from both MOEAs (NSGA-II and MOPSO-CD) are compared and discussed in Section 4.4. Finally, concluding observations are drawn in Section 4.5. 101 4.2 Nonlinear (Sigmoidal) DFA as a Priori If a DM has some preference (i.e., biasness) towards one or more objectives, a nonlinear DF may fulfill the purpose. Different shapes of DF will provide different type of biasness towards different objective region of the POF. The form of the DF originally proposed by Harrington (1965) was based on the exponential function of a linear transformation of the je,' s. Derringer and Suich (1980) found the specification not very flexible in the sense that the DFs cannot assume a variety of shapes. Derringer and Suich (1980) introduced a modified (alternative) version of DF for both one sided and two sided case based on a power of a linear transformation of the f, 's. In this chapter we propose a new kind of DF based on sigmoidal function and defined as f <.f; 111= (fP7t) .)) 1- F Z A* fi** ;a, E (4.1) f" < Figure 4.1 Proposed STB type of Sigmoidal DF One-sided case (STB type) formulation of a sigmoidal DF is given in Equation 4.1, as we are concerned with one sided specification, shown in Figure 4.1. Where parameters 102 f:and f:* are minimal and maximal acceptable levels of f, are on the same scale and are discontinuous at the points respectively. The DFss f: , f and f ** . The value of a, (a kind of weighting factor) can be chosen so that the DF is easier or more difficult to satisfy. Here, F+2f,- ) is the crossover point of the sigmoidal DF, shown in Figure 4.1. In case of sigmoidal based DF a, is the key parameter. hinear--DF—and-cornbination_o_Lconvex,_-con.came..DF have_been_analyzed already in Chapters 2 and Chapter 3 respectively. Use of Linear DF does not provide biasness towards objectives. The combination of convex-concave DF does provide biasness towards a corner of the POF. In this chapter, we will about to discuss the effect of sigmoidal DFA for MOOP. 4.3 Proposed Methodology MOOP is already described in Section 1.1 given by Equation P1 Minimize f (x)= {ft (x), f2 (x),..., fk (x)} (P1) XEX Eliciting the corresponding DF to each objective (i.e. f2,..., fk ) through the interaction of DM a new DF based multi-objective optimization problem (DFMOOP) consisting the DFs is formulated given by Equation P2. Minimize ,u(x) = {,u,(x), p,(x),..., Pk (x)} (P2) xeX Where, for each value of an objective function j; , there is a mapping called DF i.e., p, to prescribe the variation of vagueness involved discussed in Section 2.2. The overall (resulting) DF's value (say ,u) is also between zero and one. The Pareto-optimal solutions are then determined for this newly formed DFMOOP instead of the original MOOP. Solutions of DFMOOP have a unique relationship with the original MOOP of objective functions. In general, DFMOOP is solved using different aggregators, min and product operator are most common, providing a single solution of DFMOOP. This type of approach is repeatedly applied for different degrees of satisfaction values until the DM is satisfied. Benefit of this technique lies in the fact how well DM's preferences have been incorporated in a priori approach, which is quite rare due to vague nature of human judgment. So in the present approach DFMOOP is 103 solved using purely multi-objective manner (using the discussed algorithms i.e., NSGA-II and MOPSO-CD) without aggregating. Present approach is an attempt to incorporate the benefits of both the a priori and the a posteriori methods together. Now we are going to prove a theorem, which establishes the relationship between the Pareto-optimal solutions of P2 and Pl. Theorem 4.1: The Pareto-optimal solutions of P2 corresponding the DF in Equation 4.1 are also Pareto-optimal solutions of P1. Proof: Let x* be Pareto-optimal solution for P2. Then by the definition of Pareto-optimal solution Ox e X ,u, (x) < du (x ) ; for i e {1, 2,..., k} and ,ti j (x)i (x* ); for j e {1, 2,..., k} ; j I > x c X 11+ 2`Ea ( f (x)(1.9)) <11+ e "i(j;(x)-(f.' 9; for i e {1, 2, ..., lc}; a, E R+ 1--a,if,(x) (-4* (- 11a,rf,(x)(-1 +2 7 f 1 <=> e` > as a, e R+ and .f,* f(x),f,(x*) .f" <=> f(x) < f(x*) \f'Zf- ) (4.2) <=> f. (x)< f(x* ) for i E {1, 2, ..., 1+e for j E <=> e [ E aJ (x.) ; j ai aj [ f (x) ( 4 217 ))) > e [ a i [ f (x* ( fi+2fjJJJ <=>(x) -1;+2f7(x* ) f-7+2 f7 ) as a c R+ and f; < f i (x), fi (x <=> f f (x) f f (x* ) for j e i (4.3) The proposition holds as Inequalities 4.3 and 4.4 together form the condition for Pareto-optimal solution of Pl. 104 4.3.1 Detailed Procedure of the Methodology: MOEA-IDFA The methodology consists of five steps. Step 0 is an initialization step. Step 1 and 2 constitute the, calculation phase and Steps 3 and 4 the decision-making phase. Figure 1.8 shows the procedure of MOEA-IDFA. In the calculation phase, the DFs are constructed, and then an optimization model of DFs is solved using an MOEA. In the decision-rnaking_phase,the DM evaluates the results of the calculation phase, and then articulates his/her preference information. More specifically, if the DM is satisfied with the results on all the objectives, the procedure successfully ends. Otherwise, the DM adjusts the parameters (shape and bound) of a DF. Then, the procedure goes back to the calculation phase. Each step is described below. Step 0: Initialization of DF's Parameters The DF's parameters (preference parameters on each objective) are to be initialized, to construct the DF for each objective in the first iteration. The initial bound and goal (target) may be determined based on the DM's subjective judgments. Ideal and antiideal vectors f: and 4. , of correspondingobjective function f should be calculated in advance for the given MOOP (Equation P1). The initial bound and shape may be determined based on the DM's subjective judgments Step 1: Construction of the DFs From this step calculation phase starts. As mentioned earlier, the preference parameters initialized in Step 0 are used to construct the nonlinear (convex or concave) DF ,u for each f, of P1 . DM preference can be utilized in the construction of DF via the value of a, (a kind of weighting factor determines shape of DF). Thus, a new DFMOOP (see Equation P2) is formed. Step 2: Solving the DFMOOP This newly formed P2 is then solved using an efficient MOEA. As we are discussing general MOOP which can be nonlinear, nonconvex, multimodal and multivariable. Hence, resulting DFMOOP will also be of the same nature. We need some powerful 105 algorithm to solve this optimization problem. NSGA-II and MOPSO-CD are two such competent techniques capable of solving complex MOOPs used here. Step 3: Evaluation of the Solution Present the solution to the DM obtained in Step 2. Theorem 4.1 and 4.2 together imply that the Pareto-optimal solution obtained by solving P2 is also the Pareto-optimal solution of Pl. However, the POF of P2 is different from. P1 due to different function formulation. Therefore, they produce a section of POF of P1, guided according the choice of the DF. For example, for -a bi-objective minimization problem using sigmoidal DF produces biasness towards intermediate region of the POF. If the DM is satisfied by the Paretooptimal solutions, the methodology is terminated. Else, the procedure goes to Step 4. Step 4: Adjusting the Preference Parameters In order to improve the unsatisfactory results, interaction with DM can help modifying (updating) the Step 0 and Stepl . On the other hand, if the DM is fully satisfied by the obtained guided POF the procedure ends successfully. The whole process may be repeated until DM is satisfied. 4.4 Results and Discussion The proposed approach in this chapter has been applied to same set of ten standard test problems used in Chapters 2 and 3. As conversed earlier in the Section 4.2, sigmoidal combination of DFs is used as a priori, while NSGA-II/MOPSO-CD applied as a posteriori in the proposed approach. As described in the Step 0 of Subsection 4.3.1 of the approach, the parameters of a priori method i.e., nonlinear (sigmoidal) DFs are obtained with the help of DM corresponding each objective. Initial parameters of DF's used for all test problems are given in Tables 4.1(a)-4.10(a). In Step 1, preference parameters initialized in Step 0 are used in construction of the sigmoidal DFs (,u,s ) for each f with the help of DM, and forming the DFMOOP (i.e., P2). This newly formed P2 is solved using an a posteriori method as shown in 106 Step 2. Since our methodology is based on MOEA used, we have used both NSGA-II and MOPSO-CD as a posteriori method to estimate the effectiveness of the approach on different MOEAs. The values of different set of parameters are tested and fine-tuned through several runs of NSGA-II and MOPSO-CD in this approach. NSGA-II and MOPSO-CD parameters used for each problem are given in Tables 4.1(b)-4.10(b). In Step 2, computations have been carried out based on these parameters and results have been reported. To compare the efficiency of NSGA-II and MOPSO-CD, hypervolume metric is used here, for them reference points are taken (11, 11) and (11, 11, 11) for biobjective problems and tri-objective problems respectively. Tables 4.1(c)- 4.10(c) display the mean, median, standard deviation, best and worst values of hypervolume metrics obtained by using NSGA-II and MOPSO-CD with the help of 10 runs each. For the steadiness of the results for each problem, 10 runs of each MOEA (NSGA-II and MOPSO-CD) are done under the same parameters and the best POF (based on best hypervolume metric) obtained is reported. The POFs of the problems are depicted in Figures 4.2-4.11, which clearly imply that use of all sigmoidal combination of DF guides the POF towards intermediate region of the POF. Having a look on the Tables 4.1(c)-4.10(c) reveal that if we consider mean of hypervolume metric as the competence parameter for both NSGA-II and MOPSO-CD, then NSGA-H performed better in case of TNK, MHHM1 and MHHM2 while MOPSO-CD was better in the cases SCH1, SCH2, KUR, ZDT1, ZDT2, VNT and ZDT3 for the used parameters parameter settings. If the DM is satisfied by the POF obtained then the procedure ends successfully as explained in Step 4. We are just showing one iteration of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. However, if the DM is not satisfied then the procedure goes back to Step 1 to accumulate the DM's preferences in a better manner. 107 4.4.1 Effect of Variations in DF's Key Parameter on POF If the DM is unsatisfied by the outcome of Step 4 of the proposed approach then analyst calls for the variation in DF. Thus, it is significant to understand the effect of the parameters of DF on POF. It is obvious from the results and discussions in Section 4.4 that use of sigmoidal combination of DFs yields Pareto-optimal solutions in the mid portion of the POF. To investigate the effect of the DF's (i.e., ,u, 's) key parameter ( a; ) further, we are taking SCH1 as our problem and NSGA-II as the MOEA having the same a posteriori setting as used in Table 4.1 (b). As the key parameter in case of a sigmoidal DF ,u, is a, we are using rest of the parameters same as taken in Table 4.1 (a) and will vary a, to observe the effect of it. Different combination of a,' s have yielded different portions of POF of SCH1 as shown in Figures 4.12. The key parameters a, and a2 used for Figure 4.12 (a) are having equal values 2. Similarly the values of al and a2 used for Figure 4.12 (b) are taken 3. Figure 4.12 (c) depicts the POF in case of al and a2 are taken 5 each. 7 is the value used for both al and a2 in Figure 4.12 (d). It is clear from the Figures 4.12 that increment in the values of the parameters of the DF to a certain level (depending upon the problem) shrinks this mid portion even more. 4.5 Conclusion In this chapter, a partial user preference approach named MOEA-IDFA having nonlinear DF is proposed. In this approach, nonlinear (sigmoidal combination of DFs) DFA as a priori and NSGA-II/MOPSO-CD as a posteriori are combined together to provide an interactive POF. The theoretical analysis of the proposed MOEA-IDFA is also provided in this chapter. It has been observed that use of sigmoidal combination of DFs produces a definite bias (intermediate region of POF) among the Pareto-optimal solutions of original MOOP. This type of approach is useful if the DM does not prefer a particular objective over another objective.The performance of the MOEA-IDFA is tested on a set of 10 test problems. Effect of DF's key parameter on POF has also been investigated. Similar approach to find a bias in the POF proposed by Branke et al. 108 (2001) reported results in convex POF only. This approach is better than the method proposed by Branke et al. (2001) in terms of its ability to produce biasness in the disconnected and non-convex POFs as well. According the author's best knowledge, there are no works till date, in literature that comprise the biasness among tri-objective problems. In this chapter, we successfully apply the MOEA-IDFA using sigmoidal DFs to three standard tri-objective problems (VNT, MHHM1 and MHHM2). 109 Table 4.1 Parameters and Hypervolumes for SCH1 using MOEA-IDFA (All Sigmoidal Combination) Table 4.1 (a) A Priori Parameters for SCH1 fi* .1;*• 1.2. 0 4 0 ./2. a1 a2 4 3.9 3.9 Table 4.1 ( b)A Posteriori Parameters for SCH1 NSGA-II Parameters MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w c1 c2 Cross. Prob. Cross. Index 200 200 0.4 200 0.5 1 1 0.9 10 10 Table 4.1 (c) Hypervolumes w.r.t. NSGA-I1 and MOPSO-CD for SCH1 Mean Median S.D. Best Worst MOPSO-CD 397.24474 397.24475 0.00003 397.24478 397.24470 NSGA-II 397.13946 397.13945 0.00003 397.13949 397.13941 Figure 4.2 POFs of SCH1w.r.t NSGA4I and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 110 Table 4.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (All Sigmoidal Combination) Table 4.2 (a) A Priori Parameters for SCH2 fi.• A -1 1 .12 i-;- " al a2 0 16 10 10 TiaTuTurir-past r or! Parameters-for—SCH2 MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 250 0.5 200 0.5 NSGA-II Parameters Cl C2 Cross. Prob. Cross. Index 1 1 0.7 10 10 Table 4.2 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for SCH2 Mean Median S.D. Best Worst MOPSO-CD 406.68519 406.68534 0.00055 406.68539 406.68354 NSGA-H 405.89094 405.89092 0.00002 405.89097 405.89091 O 15- 0 NSGA-II • 12- MOPSO-CD 9630 0.8-0.40.0OA0.8 1 Figure 4.3 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 111 Table 4.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (All Sigmoidal Combination) Table 4.3 (a) A Priori Parameters for KUR A -20 fi*" -14.4 A -11.6 f2. al a2 0 10 10 Table 4.3 ( b) A Posteriori Parameters for KUR Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 150 0.2 200 0.6 Cl 1.1 NSGA-II Parameters c2 1.1 Cross. Prob. Cross. Index 0.6 10 /14,LAJ, ..v../Afie 5 Table 4.3 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for KUR Mean Median S.D. Best Worst MOPSO-CD 1222.30007 1222.30002 0.00012 1222.30043 1222.30001 NSGA -II 1220.00173 1220.00167 0.00014 1220.00197 1220.00144 -12 ,iI1, -20-19-18-17-16-15-14 f 1 Figure 4.4 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 112 Table 4.4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (All Sigmoidal Combination) Table 4.4 (a) A Priori Parameters for ZDT1 f," f" /2 0 1 0 r f 1 al a2 15 15 Table-4:4{ b)-,4-Posteriori—Parameters for ZDT1 MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size 200 300 0.1 200 Cl 0.5I 1.1 W- NSGA-II Parameters c2 1.1 Cross. Prob. Cross. Index 0.8 10 filA.J., :z.4-9c--,0 10 Table 4.4 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT1 Mean Median S.D. Best Worst MOPSO-CD 399.66342 399.66083 0.00642 399.67865 399.65346 NSGA-II 399.65731 399.65787 0.00241 399.65978 399.65312 1.0 - 0 0 NS GA-II " MOPSC)-CD 0.0 0.81.0 0.00.20.40.6 f 1 Figure 4.5 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using. MOEA-IDFA (All Sigmoidal Combination) 113 Table 4.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (All Sigmoidal Combination) Table 4.5 (a) A Priori Parameters for ZDT2 fi f: 0 1 f;, f; 0 Table 4.5 ( b) A '1 al az 8.0 8.0 Posteriori Parameters for ZDT2 NSGA-II Parameters MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 350 0.1 200 0.5 . Ci c2 Cross. Prob. Cross. Index ,4'tuJ. 1-t-airx 1.1 1.1 0.6 10 15 Table 4.5 (c) Hypervolumes w.r.t.NSGA-II and MOPSO-CD for ZDT2 Mean Median S.D. Best Worst MOPSO-CD 406.97954 406.97943 0.00034 406.97998 406.97912 NSGA-I1 406.81763 406.81757 0.00017 406.81798 406.81734 0 N ° MOPSO-CD 040.1 - 0. 0 0.00.20.40.60.81.0 f1 Figure 4.6 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 114 Table 4.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (All Sigmoidal Combination) Table 4.6 (a) A Priori Parameters for ZDT3 fi 0 1,..,•• jrz.• A 8.52 -0.773 1 al a2 10 10 Table 4.6 ( b)A Posteriori Parameters for ZDT3 Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W CI C. 200 350 0.02 200 0.6 1.1 1.1 NSGA-II Parameters Cross. Prob. Cross. Index 1,4-AA-41 ..f—v-44.c. ' 0.6 15 15 Table 4.6 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for ZDT3 Mean Median S.D. Best Worst MOPSO-CD 415.29906 415.29905 0.00003 415.29909 415.29900 NSGA-II 414.60759 414.60765 0.00028 414.60798 414.60712 Figure 4.7 POFs of ZDT3 w.r.t. NSGA-If and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 115 Table 4.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (All Sigmoidal Combination) Table 4.7 (a) A Priori Parameters for TNK 11. .f 0 f2• .1 1.05 1.05 0 al a2 20 20 Table 4.7 ( b)A Posteriori Parameters for TNK Common Parameters NSGA-II Parameters MOPSO-CD Parameters . Pop. Size Max. Gen. Mutation Prob. Arch. Size w 200 300 0.5 200 0.5 C1 c2 Cross. Prob. Cross. Index Ai,t.I - 1 1 0.9 10 10 A- Table 4.7 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for TNK Mean Median S.D. Best Worst MOPSO-CD 397.41494 397.41452 0.00183 397.41877 397.41132 NSGA-II 399.19856 399.19855 0.00003 399.19859 399.19852 0 NSGA-II . 0 - 1\10PSO-CD 1.0-a 0.8 0.6C,1 0.40.2 - 0.0 - O I•IIII• 0.00.20.4• 0.60.81.0 f l Figure 4.8 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 116 Table 4.8 Parameters and Hypervolumes for VNT using MOEA-IDFA (All Sigmoidal Combination) Table 4.8 (a)A Priori Parameters for VNT J1• fi f; J 2. f2 lc• a1 a2 a3 0 4 1 5 2 4 5 5 5 Table 4.8 ( b) A Posteriori Parameters for VNT Pop Size 200 Max. Gen. 250 NSGA-H Parameters MOPSO-CD Parameters Common Parameters Mutation Prob. 0.5 Arch. Size 200 w 0.5 cl c2 1 1 Cross. Prob. 0.8 Cross. Index 10 ,,. _r_.e,Pi 10 Table 4.8 (c) Hypervolumes w.r.t: NSGA-II and MOPSO-CD for VNT Mean Median S.D. Best Worst MOPSO-CD 6093.50480 6093.50561 0.00237 6093.50678 6093.50009 NSGA-1t 6087.20061 6087.20065 0.00023 6087.20099 6087.20022 Figure 4.9 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 117 Table 4.9 Parameters and Hypervolumes for MHHM1 using MOEA-IDFA (All Sigmoidal Combination) Table 4.9 (a) A Priori Parameters for MHHMI ..f 0.01 0 f2,.. f2 fl 0 f; .f;* al az a3 0 0.01 7000 7000 7000 0.0025 Table 4.9 ( b) A Posteriori Parameters for MHHM1 Common Parameters MOPSO-CD Parameters - Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 350 0.5 200 0.5 NSGA-II Parameters cl C .. Cross. Prob. Cross. Index 1 1 0.9 10 .1-1"-410, 10 Table 4.9 (c) Hypervolumes w.r.t. NSGA-H and MOPSO-CD for MHHM1 Mean Median S.D. Best Worst MOPSO-CD 7999.04947 7999.04343 0.02916 7999.09897 7999.01132 NSGA-II 7999.05225 7999.05431 0.02476 7999.09538 7999.01453 • MOPSO-CD 0 NSOA -II 0,000 0.002 0.004 0.006 0.008 0.0025 0,0020 0.0015 0.0010 0.0005 0.0100.0000 Figure 4.10 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 118 Table 4.10 Parameters and Hypervolumes for MHHM2 using MOEA-IDFA (All Sigmoidal Combination) Table 4.10 (a) A Priori Parameters for MHHM2 .. .. f3,.. A f. f; f2 A 0.0 0.0125 0.0 0.0125 0.0 0.0125 al a2 a3 2000.0 2000.0 2000.0 Table 4.10 ( b) A Posteriori Parameters for MHHM2 Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 300 0.3 200 0.5 NSGA-II Parameters Cl c2 Cross. Prob. Cross. Index 1 1 0.8 10 )44,,u4,-- <4.-t-Pir 10 Table 4.10 (c) Hypervolumes w.r.t. NS1-1A-11 and MOPSO-CD for MHHM2 Mean Median S.D. Best Worst MOPSO-CD 7995.00055 7995.00056 0.00022 7995.00087 7995.00012 NSGA-II 7995.00075 7995.00077 0.00021 7995.00098 7995.00036 • 110PSO-CD 0 NSCIA-11 Figure 4.11 POFs of MHHM2 w.r.t. NSGA-I1 and MOPSO-CD using MOEA-IDFA (All Sigmoidal Combination) 119 4- 4- 0 32- 1, 001 r:0(.41 3 I de1I 0 o 0123 4 4 f fl (a) 4- (b) ' 0 4- 3- 32- O 0 0 0 0 0 2 f1 34 fi (d)• (c) Figure 4.12 Effect of variations in key parameter of DF on POF for SCH1 120 CHAPTER 5 Guided POF Articulating Nonlinear DFA with an MOEA (All Convex Combination) In this chapter, a method articulating sigmoidal DF as a priori with an MOEA (NSGAII/MOPSO-CD) as a posteriori is proposed. DM can have multiple regions together as the preference in one POF. Convex DFA can help exploring the multiple regions of a POF in a single run. Consequence of convex DFA as a priori is analyzed theoretically as well as numerically (on ten different standard test problems). 5.1 Introduction Most MOEA approaches can be classified as a posteriori. They attempt to discover the whole set of Pareto-optimal solutions or at least a well-distributed set of representatives. The DM then looks at the set (possibly very large) of generated alternatives and makes final decision based on his/her preferences. However, if the Pareto-optimal solutions are too many, their analysis to reach the final decision is quite a challenging and burdensome process for the DM. In addition, in a particular problem, the user may not be interested the complete Pareto set; instead, the user may be interested in a certain region/regions of the Pareto set. Such a bias can arise if not all objectives are of equal importance to the user. Finding a preferred distribution in the region/regions of interest is more practical and less subjective than finding one biased solution in the region of interest. Although it is usually tougher for a DM to completely specify his or her preferences before any alternatives are known, the DM often has a rough idea of the preferential goals towards which the search should be directed, so that he/she may be able to articulate vague, linguistic degrees of importance or give reasonable trade-offs between the different objectives. Such information should be integrated into the MOEA to bias the search towards solutions that are preferential for the DM. This would in principle yield two important advantages: 121 (1) Instead of a diverse set of solutions, many of which irrelevant to the DM, a search biased towards the DM's preferences will yield a more fine-grained, and thus more suitable, selection of alternatives; (2) By focusing the search onto the relevant part of the search space, the optimization algorithm is expected to find these solutions more quickly. In the most practical cases generally, DM have at least a vague idea in which region/regions of the objective the 'optimums' should be. For this purpose, DFs that map the objectives to the interval [0, 1] according to their desired values can be specified as a priori. The Pareto-optimal solutions are then determined for the DFs instead of the objective functions using an MOEA. Proposed method is a fusion of an a priori and an a posteriori approach. Furthermore, this application in practice is quite straightforward. In addition, in spite of the focus on the desired region, the MOEA search can be carried out without any additional constraints to be included for the decision variables and/or objectives. Therefore, no individuals are 'wasted' during the optimization process in case they do not fit into the constrained intervals. The MOEA is not touched at all which facilitates the use of the method in practice as existing optimization tools can be used for optimization process as a posteriori. In other words, this approach is independent of the MOEA used and can be easily coupled to any of them without any deep modification to the main structure of the method chosen. For comparison purpose we are using two of them namely NSGA-II and MOPSO-CD. Once the DM has agreed on the parameters of the DFs, which is the key step, MOEAs can be applied on the transformed objective space (i.e., DFMOOP) without modification (Trautmann and Mehnen, 2009). By means of DFs, the solutions concentrate in the desired region/regions which facilitate the solutions selection process and support the MOEAs in finding relevant solutions. If DM has multiple regions as preference/preferences for a MOOP, modification in the DF should be done accordingly. For example, if DM is interested in finding the end portions of the POF of a bi-objective minimization problem, the DF corresponding both the objectives can be transformed as convex DF. We shall discuss the consequence of using convex DFs in this chapter. Jeong and Kim (2009) proposed an interactive DF approach (IDFA) for multirespose optimization to facilitate the preference articulation process. One run of 122 this approach provides just one solution to the DM based on his preference in the form of DF. Present work is an extension of the work done by Jeong and Kim (2009). We present a MOEA based IDFA i.e., MOEA-IDFA to provide DM a preferred portion of the POF rather than providing just a single solution. Section 5.2 describes the prerequisites of convex DFA as a priori. The methodology of the proposed approach is provided in Section 5.3. Necessary theorems are also provided in this section in order to analyze the methodology theoretically. Results corresponding to the ten test problems (Section 2.5) from both MOEAs (NSGA-II and MOPSO-CD) are compared and discussed in Section 5.4. Finally, concluding observations are drawn in Section 5.5. 5.2 Nonlinear (Convex) DFA as a Priori If a DM has some preference/preferences (i.e., biasness) towards one or more objectives, a nonlinear DF may fulfill the purpose. Different shapes of DF will provide different type of biasness towards different objective region of the POF. The thrm of the STB type of convex DF is already discussed in Chapter 3 proposed by Derringer and Suich (1980), defined as - f, f . _ f. y if f, < if .1;.* ;i =1,2,3,..., E (5A) if f < f, 1 Where parameters /and A** are minimal and maximal acceptable levels of A respectively (see Figure 5.1). The DFs p,' s are on the same scale and are discontinuous at the points f, , A and . The values of n, (a kind of weighting factor) can be chosen so that the DF is easier or more difficult to satisfy. The key parameter of a convex DF p, is n, . Use of Linear DF discussed in Chapter 2 does not provide any biasness for DM. The combination of convex-concave DF does provide biasness towards a corner of the POF, discussed in Chapter 3. Sigmoidal DF guides towards the intermediate region of the POF elaborated in Chapter 4. 123 Figure 5.1 STB type of a Convex DF Till now we were discussing single region (either towards any corner or any intermediate portion of the POF) of biasness. In this chapter, we discuss the effect of convex DFA for MOOP that can guide the POF into multiple regions. 5.3 Proposed Methodology MOOP is already described in Section (1A) given by Equation P1 Minimize f (x) = {f,(x), fk (x)} xe X (P1) Eliciting the corresponding DF to each objective (i.e. f2 ,..., fk ) through the interaction of DM a new DF based multi-objective optimization problem (DFMOOP) consisting the DFs is formulated given by Equation P2. Minimize ,u(x) {Pi(x), 12 (x), • • •, k (X) } xeX 124 (P2) Where, for each value of an objective function }; , there is a mapping called DF i.e., p, to prescribe the variation of vagueness involved discussed in Section 2.2. The overall (resulting) DF's value (say p) is also between zero and one. The Pareto-optimal solutions are then determined for this newly formed DFMOOP instead of the original MOOP. Solutions of DFMOOP have a unique relationship with the original MOOP of objective functions. In general, DFMOOP is solved using different aggregators, min and product operator are most common, providing a single solution of DFMOOP. This type of approach is repeatedly applied for different degrees of satisfaction values until the DM is satisfied. Benefit of this technique lies in the fact how well DM's preferences have been incorporated in a priori approach, which is quite rare due to vague nature of human judgment. So in the present approach DFMOOP is solved using purely multi-objective manner (using the discussed algorithms i.e., NSGA-II and MOPSO-CD) without aggregating. Present approach is an attempt to incorporate the benefits of both the a priori and the a posteriori methods together. Now we are going to prove two theorems, which establish the relationship between the Pareto-optimal solutions of P2 and Pl. Theorem 5.1: The Pareto-optimal solutions of P2 corresponding the DF in Equation 3.2 are also Pareto-optimal solutions of Pl. Proof: Let x* be Pareto-optimal solution for P2. Then by the definition of Pareto-optimal solution ,Zfx E X (X) < (X* ) ; for i E {1,2,...,k} and p (x) p j(x+ ); for j j # i for i E {1,2,...,0;n, >1 '=, OX E X rf(x)-f. <if(x )-f*" f** f* )f** -f* ■ 125 (5.2) asand f: f (x), f(x* )< f: (5.3) f - f f* and [ fj(x)— f; jni <( fi(x.)— * jnj ;for j fss — -1 .;47.-4 G ni >1; jai f.(x)— t( f 7\ < Jf-(x*)** as n,and f; < f (x), f i (x* ) f " j (5.4) Inequality (5.3) <=> f,(x)— f s < f,(x )— f: as f" — f, > 0 <=> fi (x)< f(xl ); for i E {1,2,..., lc} (5.5) Similarly, Inequality (5.4) <=> f i (x)— f f (x* )— fJ asf;* — f; > 0, j # i <=> f j(x)_. fi (x j); for j # i (5.6) The proposition holds as Inequalities 5.5 and 5.6 together form the condition for Pareto-optimal solution of P 1. 5.3.1 Detailed Procedure of the Methodology: MOEA-IDFA The methodology consists of five steps. Step 0 is an initialization step. Step 1 and 2 constitute the calculation phase and Steps 3 and 4 the decision-making phase. Figure 1.8 shows the procedure of MOEA-IDFA. In the calculation phase, the. DFs are constructed, and then an optimization model of DFs is solved using an MOEA. In the decision-making phase, the DM evaluates the results of the calculation phase, and then articulates his/her preference information. More specifically, if the DM is satisfied with the results on all the objectives, the procedure successfully ends. Otherwise, the DM adjusts the parameters (shape and bound) of a DF. Then, the procedure goes back to the calculation phase. Each step is described below. Step 0: Initialization of DF's Parameters The DF's parameters (preference parameters on each objective) are to be initialized, to construct the DF for each objective in the first iteration. The initial bound and goal (target) may be determined based on the DM's subjective judgments. Ideal and antiideal vectors f* and r , of corresponding 126 1Ih objective function j; should be calculated in advance for the given MOOP (Equation P1). The initial bound and shape may be determined based on the DM's subjective judgments Step 1: Construction of the DFs From this step calculation phase starts. As mentioned earlier, the preference parameters initialized in Step 0 are used to construct the nonlinear (convex or concave) DF At, for each f of P1. DM preference can be utilized in the construction of DF via the value of n. (a kind of weighting factor determines shape of DF). Thus, a new DFMOOP (see Equation P2) is formed. Solving the DFMOOP This newly formed P2 is then solved using an efficient MOEA. As we are discussing Step 2: general MOOP which, can be nonlinear, nonconvex, multimodal and multivariable. Hence, resulting DFMOOP will also be of the same nature. We need some powerful algorithm to solve this optimization problem. NSGA-II and MOPSO-CD are two such competent techniques capable of solving complex MOOPs used here. Step 3: Evaluation of the Solution Present the solution to the DM obtained in Step 2. Theorem 5.1 and 5.2 togetherimply that the Pareto-optimal solution obtained by solving P2 is also the Pareto-optimal solution of Pl. However, the POF of P2 is different from P1 due to different function formulation. Therefore, they produce a section of POF of Pl, guided according the choice of the DF. For example, for a bi-objective minimization problem using convex DF produces biasness towards both end of the POF. If the DM is satisfied by the Pareto-optimal solutions, the methodology is terminated. Else, the procedure goes to Step 4. Step 4: Adjusting the Preference Parameters In order to improve the unsatisfactory results, interaction with DM can help modifying (updating) the Step 0 and Step]. On the other hand, if the DM is fully satisfied by the 127 obtained guided POF the procedure- ends successfully. The whole process may be repeated until DM is satisfied. 5.4 Results and Discussion The proposed approach in this chapter has been applied to the same set of ten standard test problems used earlier (Chapters 2-4). As conversed earlier in the Section 5.2, all convex combination of DFs is used as a priori, while NSGA-II/MOPSO-CD applied as a posteriori in the proposed approach. As described in the Step 0 of Subsection 5.3.1 of the approach, the parameters of a priori method i.e., nonlinear (sigmoidal) DFs are obtained with the help of DM corresponding each objective. Initial parameter of DF's used for all test problems are given in Tables 5.1(a)-5.10(a). In Step 1, preference parameters initialized in Step 0 are used in construction of the sigmoidal DFs ( ,u,s ) for each f with the help of DM, and forming the DFMOOP (i.e., P2). This newly formed P2 is solved using an a posteriori method as shown in Step 2. Since our methodology is based on MOEA used, we have used both NSGA-II and MOPSO-CD as a posteriori method to estimate the effectiveness of the approach on different MOEAs. The values of different set of parameters are tested and fine-tuned through several runs of NSGA-II and MOPSO-CD in this approach. NSGA-II and MOPSO-CD parameters used for each problem are given in Tables 5.1(b)-5.10(b). In Step 2, computations have been carried out based on these parameters and results have been reported. To compare the efficiency of NSGA-II and MOPSO-CD, hypervolume metric is used here, for them reference points are taken (11, 11) and (11, 11, 11) for biobj ective problems and tri-objective problems respectively. Tables 5.1(c)- 5.10(c) display the mean, median, standard deviation, best and worst values of hypervolume metrics obtained by using NSGA-II and MOPSO-CD with the help of 10 runs each. For the steadiness of the results for each problem, 10 runs of each MOEA (NSGA-II and MOPSO-CD) are done under the same parameters and the best POF (based on best hypervolume metric) obtained is reported. The POFs of the problems are depicted in Figures 5.2-5.11, which clearly imply that use of all convex combination of DF guides the POF towards multiple regions of the POF. 128 Having a look on the Tables 5.1(c)-5.10(e) discloses that if we consider mean of hypervolume metric as the competence parameter for both NSGA-II and MOPSO-CD, then NSGA-II performed better in case of SCH2, KUR, ZDT1, ZDT2, MHHM1 and MHHM2 while MOPSO-CD was better in the cases SCH1, TNK, VNT and ZDT3 for the used parameters parameter settings. If the DM is satisfied by the POF obtained then the procedure ends successfully as explained in Step 4. We are just showing one iteration of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. However, if the DM is not satisfied then the procedure goes back to Step I to accumulate the DM's preferences in a better manner. 5.4.1 Effect of Variations in DF's Key Parameter on POF If the DM is unsatisfied by the outcome of Step 4 of the proposed approach then analyst calls for the variation in DF. Thus, it is significant to understand the •effect of the parameters of DF on POF. It is obvious from the results and discussions in Section 5.4 that use of convex combination of DFs yields Pareto-optimal solutions in the extreme portions of the POF. To investigate the effect of the DF's key parameter further, we are taking SCH1 as our problem and NSGA-II as the MOEA having the same a posteriori setting as used in Table 5.1 (b). As the key parameter in case of a convex DF p, is n, we are using rest of the parameters same as taken in Table 5.1 (a) and will vary n, to observe the effect of it. Different combination of n, s have yielded different portions of POF of SCH1 as shown in Figures 5.12. The key parameters ni and n2 used for Figure 5.12 (a) are having equal values 2. Similarly the values of ni and n2 used for Figure 5.12 (b) are taken 4. Figure 5.12 (c) depicts the POF in case of n1 and n2 are taken 5 each. 20 is the value used for both ni and n2 in Figure 5.12 (d). It is clear from the Figures 5.12 that increment in the values of both the key parameters contracts POF towards the extreme portions. Thus, if the DM wants the extreme portions of POF (in case of bi-objective problem) in single run then all convex type of combination of DF can be used. 129 5.5 Conclusion The main crux of this chapter is application of DFA to exploit the population approach of an MOEA procedure in finding more than one solutions not on the entire POF, but in the regions of Pareto-optimality which are of interest to the DM. In this chapter, a partial user preference approkh named MOEA-IDFA having nonlinear DF is proposed. In this approach, nonlinear (convex combination of DFs) DFA as a priori and NSGA-II/MOPSO-CD as a posteriori are combined together to provide an interactive POF. The theoretical analysis of the proposed MOEA-IDFA is also provided in this chapter. It has been observed that use of convex combination of DFs produces multiple biases (towards extreme regions of POF) among the Pareto-optimal solutions of original MOOR This type of approach is useful if the DM does not prefer a intermediate region of the POF. The performance of the MOEA-IDFA is tested on a set of 10 test problems. It is observed and displayed that multiple (two) regions can also be explored using suitable combination of DFs. Effect of the key parameters of DF on POF is also presented According the author's best knowledge, there are no works till date, in literature that comprise the biasness among tri-objective problems. In this chapter, we successfully apply the MOEA-IDFA using convex DFs to three standard tri-objective problems (VNT, MHHM1 and MHHM2). 130 Table 5.1 Parameters and Hypervolumes for SCHI using MOEA-IDFA (All Convex Combination) Table 5.1 (a) A Priori Parameters for SCH1 ii. fi.. f; 0 4 0 A nt n2 4 10 10 Table 5.1 ( b) A Posteriori Parameters for SCH1 Common Parameters MOPSO-CD Parameters NSGA-H Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W c1 c2 Cross. Cross. Prob. Index 200 250 0.4 200 0.4 1 1 0.9 10 ...0,, 10 Table 5.1 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for SCHI Mean Median S.D. Best Worst MOPSO-CD 396.88250 396.88254 0.00015 396.88269 396.88221 NSGA-11 393.58259 393.58257 0.00015 393.58287 393.58232 Figure 5.2 POFs of SCH1 w.r.t NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 131 Table 5.2 Parameters and Hypervolumes for SCH2 using MOEA-IDFA (All Convex Combination) Table 5.2 (a) A Priori Parameters for SCH2 J1• fi** f; 127 n-1 n, -1 1 0 16 11.0 11.0 Table 5.2 ( b)A Posteriori Parameters for SCH2 NSGA-II Parameters MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w 200 300 0.5 200 0.6 Ci c2 Cross. Prob. Cross. Index 1 1 0.7 10 4. j,,..4-9,1 10 Table 5.2 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for SCH2 Mean Median S.D. Best Worst MOPSO-CD 405.25125 405.25123 0.00021 405.25154 405.25101 NSGA-II 406.97663 406.97662 0.00022 406.97689 406.97623 15- NSGA-II MOPSO-CD 129- 63• 0 ,'TIED:unurro 0 1 tl Figure 5.3 POFs of SCH2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 132 Table 5.3 Parameters and Hypervolumes for KUR using MOEA-IDFA (All Convex Combination) Table S.3 (a) A Priori Parameters for KUR 1 -20 fi.. f; -14.4 -11.6 f2... n, n2 0 12.0 12.0 Table 5.3 ( b)A Posteriori Parameters for KUR Common Parameters Pop size 200 MOPSO-CD Parameters Max. Gen. Mutation Prob. Arch. Size w 200 0.2 200 0.6 cl 1.1 c2 1.1 NSGA-II Parameters Cross. Prob. 0.5 Cross. iti,..JIndex J_'O 10 10 Table 5.3 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for KUR Mean Median S.D. Best Worst MOPSO-CD 1223.60011 1223.60000 0.00013 1223.60025 1223.6000 NSGA-H 1224.80046 1224.80060 0.0004 1224.8009 1224.800 -19-18-17-16-15 f1 Figure 5.4 POFs of KUR w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 133 Table 5.4 Parameters and Hypervolumes for ZDT1 using MOEA-IDFA (All Convex Combination) Table 5.4 (a) A Priori Parameters for ZDT1 f.. ii 1 0 f2 f2" ni n2 0 1 10 10 Table 5.4 ( b) A Posteriori Parameters for ZDTI NSGA-H Parameters MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 350 0.15 200 0.5 1.1 1.1 0.7 10 6/wei- ItAz--te--x 10 Table 5.4 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDTI Mean Median S.D. Best Worst MOPSO-CD 399.59552 399.59555 0.0004 399.59578 399.59522 NSGA-H 399.59883 399.59555 0.0002 399.59578 399.59522 1.0 - 0 NS GA-II ° MOPS 0-CD 0.8 0.6 0.4 - 0. 0.2 0.0 - 0.001.20.40.60.81.0 f l Figure 5.5 POFs of ZDT1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 134 Table 5.5 Parameters and Hypervolumes for ZDT2 using MOEA-IDFA (All Convex Combination) Table 5.5 (a) A Priori Parameters for ZDT2 fi* fi f; 0 1 0 f2.. 1 ni n2 20.0 20.0 Table 5.5 ( b) A Posteriori Parameters for ZDT2 NSGA-H Parameters MOPSO-CD Parameters Common Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. Cross. Index 200 350 0.15 200 0.5 1.1 1.1 0.6 15 .1.1,..049c 15 Table 5.5 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT2 Mean Median S.D. Best Worst MOPSO-CD 399.28380 399.28378 0.00005 399.28388 399.28374 NSGA-1I 399.59579 399.595814 0.00012 399.59595 399.59554 0 NSGA-II 1.0 - ' MOPSO-CD 0.80.6(D o 040.2 0.0 - I•I•I 0:60.81.0 0.00.20.4 fl Figure 5.6 POFs of ZDT2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 135 Table 5.6 Parameters and Hypervolumes for ZDT3 using MOEA-IDFA (MI Convex Combination) Table 5.6 (a) A Priori Parameters for ZDT3 fi* fi... 0 8.52 f2...., f2 -0.773 1 ni n2 2.5 2.5 Table 5.6 ( b) A Posteriori Parameters for ZDT3 Common Parameters MOPSO-CD Parameters NSGA-H Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w Cl c2 Cross. Prob. 200 350 0.02 200 0.5 1.1 1.1 0.4 Cross. /141.A.1 Index 4).--ef eyi 15 15 Table 5.6 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for ZDT3 Mean Median S.D. Best Worst MOPSO-CD 414.72488 414.724914 0.00014 414.72499 414.72454 NSGA-II 414.65120 414.65117 0.00006 414.65128 414.65112 1.0 0.8 - 0 NSGA-II 0.6 - • MOPSO-CD 0.4 0.2 0.0 -0.2 -0.4 -0.6 -0.8 - I•I•I 0.00.20.40.60.8 fI Figure 5.7 POFs of ZDT3 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 136 Table 5.7 Parameters and Hypervolumes for TNK using MOEA-IDFA (All Convex Combination) Table 5.7 (a) A Priori Parameters for TNK ,1 i fi'" 0 1.05 99 12 f2 ni n2 0 1.05 8.0 8.0 Table 5.7 ( b)A Posteriori Parameters for TNK Common Parameters MOPSO-CD Parameters NSGA-II Parameters Max. Gen. Mutation Prob. Arch. Size W cl c2 Cross. Cross. A.4.4...41,- Size Prob. Index apex 200 350 0.5 200 0.6 1 1 0.7' 10 10 Pop Table 5.7 (c) Hypervolumes w.r.t. NSGA-If and MOPSO-CD for TNK Mean Median S.D. Best Worst MOPSO-CD 397.32435 397.32435 0.00003 397.32438 397.32431 NSGA-H 399.22263 399.22265 0.00018 399.22278 399.22212 O NSGA-II • MOPSO-CD 0.4 0.2 0.0 0.00.20.40.60.81.0 f l Figure 5.8 POFs of TNK w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 137 Table 5.8 Parameters and Hypervolumes for VNT using MORA-IDFA (All Convex Combination) Table 5.8 (a) A Priori Parameters for VNT f: .1— f; . f2 * .12.. .f2." n, n2 n, 0 4 1 5 2 4 5.0 5.0 5.0 Table 5.8 ( b) A Posteriori Parameters for VNT Common Parameters Pop Size 206 MOPSO-CD Parameters NSGA-II Parameters Max. Gen. Mutation Prob. Arch. Size w Cl CZ Cross. Prob. Cross. Index 250 0.5 200 0.5 1 1 0.9 10 M.A..A.3-._T v..46 10 Table 5.8 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for VNT Mean Median S.D. Best Worst MOPSO-CD 6085.10015 6085.10014 0.00009 6085.10025 6085.10004 NSGA-II 6070.40029 6070.40013 0.00031 6070.40098 6070.400012 Figure 5.9 POFs of VNT w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 138 Table 5.9 Parameters and Hypervolumes for MHHM1 using MOEA-IDFA (All Convex Combination) Table 5.9 (a) A Priori Parameters for MIIHM1 ft.. fi.. .i.; f2.• ✓3 0 0.01 0 0.0025 0 1 0.01 fl n, n, 5.5 5.5 5.5 Table 5.9 ( b) A Posteriori Parameters for MHHM1 Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w Cl c2 Cross. Prob. 200 350 0.4 200 0.5 1 1 0.9 Cross. /1/1.w4Index ..L.1...dlc2. 10 10 'Table 5.9 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for MHHM1 Mean Median S.D. Best Worst MOPSO-CD 7999.00091 7999.00050 0.00170 7999.00565 7999.00001 NSGA-II 7999.00100 7999.00056 0.00166 7999.00565 7999.00004 ■ MOPSQ-CD 0 NSGA-II 0.000 0.002 0.00$ 0.006 ✓ 0.00g 0.0025 0.0020 0.0015 0.0010. 0.0005 \ 0.0000 0.010 Figure 5.10 POFs of MHHM1 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 139 Table 5.10 Parameters and Hypervolumes for MHHM2 using MOEA-IDFA (All Convex Combination) Table 5.10 (a) A Priori Parameters for MHHM2 J f. f; 0.0 0.0125 0.0 ..f2. 0.0125 ./3.* f3** nl n2 n3 0.0 0.0125 5.0 5.0 5.0 Table 5.10 ( b) A Posteriori Parameters for MHHM2 Common Parameters NSGA-II Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size w cl c2 Cross. Prob. 200 350 0.3 200 0.5 1 1 0.7 Cross. /..,L4— Index 10 10 Table 5.10 (c) Hypervolumes w.r.t. NSGA-II and MOPSO-CD for MHHM2 Mean MOPSO-CD NSGA-H 7999.10011 7999.10010 7995.00144 Best S.D. Median 7999.00056 0.00002 Worst 7999.10015 7999.10010 0.00191 7999.00565 7999.00004 - MOPSO-CD 0 NSGA-II 0.014 0.0112 .° V.008 0,006 ✓ 0.004 0.002 0.000 -0.002 O. 0.01 0.008 0.006 0.004c 0.002 000 -0.002 Figure 5.11 POFs of MHHM2 w.r.t. NSGA-II and MOPSO-CD using MOEA-IDFA (All Convex Combination) 140 4- 4- 3- 3- w' 2- -I 2 0 0- 0- 0 1 01 34 11 (a) 1-• f 4- 1 (b) 4- Vi = _5 1'I 2_--- 5 3- 4 3 H =26 3- 1- 0 Irt"ANI 0- 234 0 f1 '34 01 OCEENNID 0- XIO0I; 1 (d) (c) Figure 5.12 Effect of variations in key parameter of DF on POF for SCH1 141 142 CHAPTER 6 Application of the MOEA-IDFA to Reliability Optimization Problems In this Chapter, we exemplify application of the proposed MOEA based IDFA i.e., MOEA-IDFA discussed in Chapters 2-5, via the five well-known reliability optimization problems of MOOP category. In Section 6.1 an overview as well as the review of preference incorporation in reliability optimization problem is specified. In the later sections (Sections 2-6) application of the proposed approach is shown on five reliability optimization problems. 6.1 Reliability Optimization (An Overview) The reliability of a system is generally measured in terms of probability that the system will not fail during the delivery of the service. A system can be designed for optimal reliability either by adding redundant components or by increasing the reliability of components. In both cases, an increased demand for the applied resources of cost, volume and weight must be observed. Therefore, a balance is required between resources and reliability. Thus, it is worth considering the MOOP techniques to solve this kind of problem. In most practical situations involving reliability optimization, there are several mutually conflicting goals such as maximizing system reliability and minimizing cost, weight, volume and constraints required to be addressed simultaneously. Some main objectives can be expressed as Objective 1 The most important objective is the maximization of system reliability (R s ). It enables the system to function satisfactorily throughout its intended service period. Max R, As in our approach we are considering all minimization problems. Hence, the above objective is equivalent to minimization of system unreliability (Q, =1— R s ), can be expressed equivalently as follows Min 143 Objective 2 The addition of the redundant components increases not only the system reliability but also its overall cost ( C ). A manufacturer has to balance these conflicting objectives, keeping in view the importance of reducing the overall cost. This objective can be expressed as Min Cs Objective 3 As with cost, every added redundant component increases the weight of the system. Usually, the overall weight of a system needs to be minimized along with its cost even as reliability is maximized (or unreliability is minimized) i.e., Min Ws When two or more elements compose a system, the reliability of the latter depends upon the reliability of the formers as well as the functional interactions amongst them. The basic modes of interaction are series and parallel: in the former case a failure in one component results in a failure of the whole system, thus the reliability of a series system is expressed by Rs = n r,;i =1,n , where r; is the component reliability of 1=1 the ith component of the system, n is the total number of components in the systems. Conversely, in parallel systems, all the components must fail to make the system fails; therefore the Tenability is assessed as Rs = fJ (1- r1 )=1-1q=;i = 1,n, where q is i=i the unreliability of thecomponent of the system. Some systems with many components are not single series nor parallel but a combination of both types. In such a case systems are modeled as a group of subsystems, each of which might be parallel, series or a combination of both. Some systems do not bear series-parallel structures and therefore cannot be assessed directly using the previous formulae (for example complex bridge system etc.). Instead, the analyst should resort to other analytic tools like Fault Tree Analysis (see (Andrews and Moss, 1993) for details) in order to identify minimal combinations of elements whose simultaneous failure doom the whole system to failure or minimal cut sets (MCS). Every system can be seen as a collection or series of MCS. Some components of the system can be present in many MCS so that the probability of failure of such MCS is not independent but rely upon the replicated components. When the MCS are disjoint, i.e. when no element is present in more than one MCS, the failure 144 probabilities of the MCS are independent. Hence, the sum of all disjoint MCS reliabilities gives the system reliability. However, in practice this addition might bear too many terms and therefore some simplifications are done, like considering MCS composed of up to m elements, where m gives the order of the cut set. There is an extensive literature on reliability the reader can consult for more details, e.g. (Andrews and Moss, 1993; Verma et al., 2010). Optimization of reliability of complex systems is an extremely important issue in the field of reliability engineering. Over the past three decades, reliability optimization problems have been formulated as non-linear programming problems within either single-objective or multi-objective environment. Tillman et al. (1980) provides an excellent overview of a variety of optimization techniques applied to solve these problems. However, he reviewed the application of only derivative-based optimization techniques, as metaheuristics were not applied to the reliability optimization problems by that time. Reliability optimization problems can be classified into three sections namely component reliability problem (nonlinear problem), redundancy allocation problem (integer nonlinear problem), component reliability and redundancy allocation problem (mixed integer nonlinear problem) (Gopal et al., 1978; Govil and Agarwal, 1983; Dhillon, 1986; Misra, 1991, 2009; Misra and Sharma, 1991a, 1991b; Mohamed Lawrence, 1992; Singh and Misra,, 1994; Prasad and Kuo, 2000; Kuo and Prasad, 2000; Amari et al., 2002, 2004; Amari and McLaughlin, 2005; Kapur and Verma, 2005; Ha and Kuo, 2006; Misra et al., 2006; Kuo and Wan, 2007; Levitin and Amari, 2008; Coelho, 2009; Aggarwal et al., 2009b, 2009a; Kuo et al.; Amari and Dill, 2010) In this thesis we are concerned with the first one (component reliability problem) only and hence will use the term reliability optimization problem in this sense here after. Over the last decade, metaheuristics have also been applied to solve the reliability optimization problems. To list a few of them, Coit and Smith (1996) were the first to employ a GA to solve reliability optimization problems. Later, Ravi et al. (1997) developed an improved version of non-equilibrium simulated annealing called INESA and applied it to solve a variety of reliability optimization problems. Further, Ravi et al. (2000) first formulated various complex system reliability optimization problems with single and multi objectives as fuzzy global optimization problems. They also developed 145 and applied the non-combinatorial version of another meta-heuristic viz. threshold accepting to solve these problems. Recently, Shelokar et al. (2002) applied the ant colony optimization algorithm to these problems and obtained compared results to those reported by Ravi et al. (1997). Vinod et al. (2004) applied Gas to Risk Informed In-Service Inspection (RI-ISI) which aims at prioritising the components for inspection within the permissible risk level thereby avoiding unnecessary inspections. A new fuzzy multi-objective optimization method is introduced and it is used for the optimization decision-making of the series and complex system reliability with two objectives is presented by Mahapatra and Roy (2006). Mahapatra (2009) considered a series-parallel system to find out optimum system reliability with an additional entropy objective function. Marseguerra et al. (2006) applied GA to solve the reliability problem. Salazar et al. (2006, 2007) solved the system reliability optimization problem by using several EAs and MOEAs. More recently, Ravi (2007) developed an extended version of the great deluge algorithm and demonstrated its effectiveness in solving the reliability optimization problems. 6.1.1 Preference Incorporation in Reliability Optimization Problems It is very difficult for DM to specify accurately their preference on the goals a priori in multi-objective reliability optimization problems. There are three key issues in the interactive multi-objective optimization methods (1) how to elicit preference information from the DM over a set of candidate solutions, (2) how to represent the DM's preference structure in a systematic manner, (3) how to use the DM's preference structure to guide the search for improved solutions. A brief review of DM's preference articulation for reliability optimization problem is presented below. To accommodate the preference of DM in reliability optimization problem Sakawa (1978) considered a multi-objective formulation to maximize reliability and minimize cost for reliability allocation by using the surrogate worth trade-off method. Dhingra (1992) and Rao and Dhingra (1992) researched the reliability and redundancy apportionment problem for a four-stage and a five-stage overspeed protection system, 146 using crisp and fuzzy multi-objective optimization approaches respectively. Ravi et al. (2000) modeled the problem of optimizing the reliability of complex systems as a fuzzy multi-objective optimization problem and studied it. Huang et al. (2007) reported a new effective multi-objective optimization method, Intelligent Interactive multi-objective optimization method (IIMOM), which is characterized by the way the DM's preference structure model is built and used in guiding the search for improved solutions. IIMOM is applied to the reliability optimization problem of a multistage mixed system. Five different value functions are used to simulate the DM in the solution evaluation process. The results illustrate that IIMOM is effective in capturing different kinds of preference structures of the DM, and it is an effective tool for the DM to find the most satisfying solution (Huang et al., 2005, 2007). Pandey et al. (2007) proposed an enhanced particle swarm optimization (EPSO) to simulate the DM's opinion in the solution process. Inagaki et al. (2009) solved another problem to maximize reliability and minimize cost and weight by using an interactive optimization method. During review, we found that incorporation of DM's preference in reliability optimization problems is infrequent and of growing interest. The main crux of the proposed approach is exploitation of the population approach of an MOEA procedure in finding more than one solutions not on the entire POF, but in the regions of Pareto-optimality which are of interest to the DM. Our proposed approach can be used in an interactive manner to guide DM towards the preferred region of POF. Depending on the type of choice (preference) available from the DM, we define following cases of preference. We also present the appropriate combination of DFs needs to implement the particular preference of DM. • No Preference No choice made by DM. Though in this case DM don't make any preference but this Case is very important to start the interaction with DM. Before we concentrate in more detail in other cases, we must say something about how to start the interactive methodology's solution process. It is possible that we ask even the first choice from the DM. In this case, it is typically useful to first show the ideal (Qs , CS" , Ws* ) and the antiideal( Q:* , C:* Ws** , ) objective values to her/him in order to give some understanding of 147 what kind of solutions are feasible. Alternatively, we can calculate a so-called neutral compromise solution as the first solution (Wierzbicki, 1999). The result of this case is a good starting point when no preference information is yet available. The DFs (,uQ „uc,andcorresponding the objectives Qs , Cs and Ws is constructed based upon initialized parameters of Step 0 of the proposed approach. In case of No Preference, no extra information is needed from the DM to construct the DFs. In fact, in this case, linear DFA described in Chapter 2 may be applied; expressions of DFs are given below. if a <Cc (6.1) Ata Q.: <a if C < if C: S C (6.2) if C.:* < 0 kris if<ws ** W, - Ws* 11/;" 1 if • (6.3) if kris * Ws** < Preference 1 Next we assume that the DM is able to rank the relative importance of objectives. In this case, the DM prefers a region closer to the maximum of the reliability (one of his/her objective). In other words, for a bi-objective problem, visually DM wants to explore the top right portion of the POF shown him/her in the No Preference case. In case of bi-objective problem having objectives Qs and C„ the DFs (PQ, „uc., ) can be constructed through a concave and convex representations discussed in Chapter 3 as below 148 If Q, <Q:* ni if Q = R+ (6.4) Q:* if Cs <C*,* if C: 5 C5. < C:* ; n2; n2 E Pc, = (6.5) if C:* <Cs • Preference 2 This case is similar to the previous case; the difference here is the choice of the objective made by the DM. In this case the DM prefers a region closer to the minimum of the Cost (one of his/her objective). In other words, for a bi-objective problem, visually DM wants to explore the bottom left portion of the POF shown him/her in the No Preference case. In case of bi-objective problem having objectives Qs and Cs , the DFs (N ,N ) can be constructed through a convex and concave representations discussed in Chapter 3 as below 0 = if Q, <Q.:* if V, Qs;ni E R, (6.6) if Q.:* <Qs if Cs <Cs" Pc, if C; S C < C*; ; n2 S 1 ; n2 e ❑ + (6.7) if Cs- <C, In case of tri-objective problem the situation is a bit different than the bi-objective problem. In this case the convex concave combinations also give rise to the both convex combination for any two objectives. Out of many cases of tri-objective problem we consider a case that is comparable here. If DM wants a bottom left portion of the POF shown him/her in the No Preference case. DFs (N „tic.) corresponding objectives 149 Q, and Cs can be constructed through a convex and concave representation shown above in •Equations (6.6-6.7) while DF (,uw ) corresponding objective weight Ws is given by a concave representation as 0 if W. <W.:" W,-WS "' fikv = W _ w* j S -W . 1 • if Ws* _147.,; n3; n3 E R+ (6.8) if Ty.* <ws Preference 3 In this case, the DM prefers an intermediate portion of the POF shown him/her in the No Preference case. The DFs„uc,and pw ) corresponding the objectives Qs , CS and Ws can be constructed through all sigmoid representation discussed in Chapter 4, also shown below Zf * Qs<Q, if Vs. Qs Q.:* ;061E R+ (6.9) if Q.:' if C < C:* +e (—a f7))) C: Cs ; a2 c R+ (6.10) if <cs if W. <Ws** TV: +2W:* A j if w: <wy<w,";a3 e R+ if W.,-** <W., 150 (6.11) • Preference 4 DM can have multiple regions together as the preference in one POF. This case, in one way is complementary to the Preference3 case. In this case, DM wants only extreme portions of the POF and no intermediate regions. In other words, for a bi-objective problem, visually DM wants to explore simultaneously top right and bottom left portions of the POF shown him/her in the No Preference case. If DM is interested in finding multiple regions in POF and prescribes Preference 4 as the choice then all convex representation (examined in Chapter 5) of DFs (,u Q, „tic, and ,u s ) corresponding the objectives Q„ Cs and Ws can be used shown below 0if a <Q;* p Q, = - . * Q, Qs 5_ Q: ;n1 .1;n1 E 11+ , a -V, 1if Q*,* <a (6.12) <C: C: C Cs* ; n2; n2 e .1Z+ (6.13) C:* < 0 if ws<ws** jn3 if Ws:* Ws W: ; n3 __1; n3 E (6.14) Through numerical experiments on, five reliability optimization problems, in this Chapter, we try to make sure that a specific combination of DFs will result in a corresponding specific preferred portion of a POF as per wish of the DM. 151 6.2 Reliability Optimization of a Series System 6.2.1 Problem Description Here a series system having five components, shown in Figure 6.1, is considered, each having component reliability r; , i =1,2,...,5. The system reliability Rs , unreliability Qs and system cost Cs are given by 5 Rs = n r, or 5 Qs = (1— Rs )=1—Fir, E 5 Cs = 1=1 1 a, log [1_ r +b,) The problem is to find the decision variables r,, i= 1, 2,..., 5 which minimize both Qs and Cs Subject to 0.50.99; i =1,2,...,5, In other words, the problem can be posed as a MOOP given by Minimize (Qs ,Cs ) subject to 0.5 < re < 0.99, i =1, 2, ..., 5; where vectors of coefficients a, and b, are (6.15) a = {24,8,8.75,7.14,3.33} and b = {120,80,70,50,30} respectively. Huang (1997) solved the problem given by Equation 6.1-6.2 using fuzzy approach and reported only three Pareto-optimal solutions using aggregation method. We are going to solve this problem according the preference of DM. 6.2.2 Step-by-Step Illustration of MOEA-IDFA for Series System Step 0: Intialization of Priori Parameters The DF's priori preference parameters need initialized to construct the DFs (,uQ, and pc, ) for each objective (i.e., Qs and Cs ), shown in Table 6.3. 152 Step 1: Construction of DFs The DFs (PQ and pc., ) corresponding the objectives Q5 and Cs is constructed based upon the DM's preference parameters shown in Table 6.4 and the initialized parameters of Step 0 (as discussed in Subsection 6.1.1) ■ in case of No Preference, Equations (6.1-6.2) are used, ▪ in case of Preference 1, Equations (6.4-6.5) are used, ■ in case of Preference 2, Equations (6.6-6.7) are used, ■ in case of Preference 3, Equations (6.9-6.10) are used and ■ in case of Preference 4, Equations (6.12-6.13) are used. Step 2: Solving the DFMOOP The following DFMOOP is formed using the DFs obtained in Step 1 that needs to be solved by any MOEA. Minimize (pQs , pc.$ ) subject to 0.50.99, i =1,2,...,5; (6.16) For each case (preference), the combination of DF will be different and so the problem will be different for each case shown above. We are using NSGA-II and MOPSO-CD to solve the DFMOOP obtained in Equation 6.16. The optimal parameters setting used for both of these algorithms are provided in Table 6.5. Step 3: Evaluation of the Solution As shown in Chapters 2-5 Pareto-optimal solution of DFMOOP (Equation 6.16) is also the Pareto-optimal solution of MOOP (Equation 6.15). We have plotted the best POF (on the basis of max hypervolume metric) obtained through various run of the algorithms of the series system for each case. The POF of No Preference case is shown in Figure 6.7. For rest of the cases the plots are shown in Figures 6.10. Step 4: Adjustment of the Parameters If the DM is satisfied by the solution obtained in Step 3 the approach stops successfully. Other wise the key preference parameters shown in Table 6.4 can be altered to meet the DM's choice and the method again go back to Step 2. The process is repeated until DM is satisfied. We are just showing one run of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. 153 6.3 Reliability Optimization of Life Support System in a Space Capsule This application concerns the reliability design of a life-support system in a space capsule (Shelokar et al., 2002; Salazar et al., 2006) its configuration is presented in Figure 6.2. The system, which requires a single path for its success, has two redundant subsystems each comprising components 1 and 4. Each of the redundant subsystems is in series with component 2 and the resultant pair of series-parallel arrangement forms two equal paths. Component 3 is inserted as a third path and backup for the pair. The continuous optimization models that were originally formulated for the reliability design of this system approached the problem in two different ways: Sheloker et al. (2002) adopted a single criterion methodology in which a cost function of component reliability was minimised, subject to constraints on system and components' reliabilities. On the other hand, Salazar et al (2006) used a bi-criterion approach using a number of heuristic algorithms such as ant colony optimization, tabu search, and NSGA-II. The block-diagram is presented in Figure 6.2. The system reliability Rs , unreliability Qs and system cost Cs are given by (Tillman et al., 1980): Maximze Rs = 1— r3 [(1 - 10(1 — r4)] 2 — (1— r3)[1— r2{1— (1— TO(1— r4)}i 2 or Minimize Qs =1— Rs ; Minimize Cs =+ 2K2r2"2 + K3r3"3 + 2K4r4a4; subject to 0.5=1,2,3,4 In other words, the problem can be posed as a MOOP given by Minimize (Q s ,Cs ) subject to 0.5 r,i =1,2,3,4; where vectors of coefficients (6.17) K, and a, are K = {100,100,200,150} and a = {0.6, 0.6,0.6, 0.6} respectively. 6.3.1 Step-by-Step Illustration of MOEA-IDFA for Life Support System in a Space Capsule Step 0: Intialization of Priori Parameters 154 The DF's a priori preference parameters need initialized to construct the DFs (,uQ and) for each objective (i.e., Qs and CS ), shown in Table 6.6. Step 1: Construction of DFs The DFs (,uQ, and pc., ) corresponding the objectives Qs and Cs is constructed based upon the DM's preference parameters shown in Table 6.7 and the initialized parameters of Step 0 • in case of No Preference, Equations (6.1-6.2) are used, • in case of Preference 1, Equations (6.4-6.5) are used, • in case of Preference 2, Equations (6.6-6.7) are used, • in case of Preference 3, Equations (6.9-6.10) are used and • in case of Preference 4, Equations (6.12-6.13) are used. Step 2: Solving the DFMOOP The following DFMOOP is formed using the DFs obtained in Step 1 that needs, to be solved by any MOEA. Minimize (p Qs , pc, ) subject to 0.5 r,i =1,2,3,4; (6.18) For each case (preference), the combination of'DF will be different and so the problem will be different for each case shown above. We are using NSGA-II and MOPSO-CD to solve the DFMOOP obtained in Equation 6.18. The optimal parameters setting used for both of these algorithms are provided in Table 6.8. Step 3: Evaluation of the Solution As shown in Chapters 2-5 Pareto-optimal solution of DFMOOP (Equation 6.17) is also the Pareto-optimal solution of MOOP (Equation 6.18). We have plotted the best POF obtained through various run of the algorithms of the series system for each case. The POF of No Preference case is shown in Figure 6.8. For rest of the cases the plots are shown in Figures 6.11. Step 4: Adjustment of the Parameters If the DM is satisfied by the solution obtained in Step 3 the approach stops successfully. Other wise the key preference parameters shown in Table 6.7 can be altered to meet the DM's choice and the method again go back to Step 2. The process is 155 repeated until DM is satisfied. We are just showing one run of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. 6.4 Reliability Optimization of a Complex Bridge System Here a bridge network system as shown in Figure 6.3, has been considered, each having a component reliability r,, i=1,2,...,5. Misra and Agnihotri (2009) have investigated the peculiarities associated with the system. The system reliability Rs , unreliability Qs and system cost Cs are given by (Tillman et al., 1980): Rs = r1r4 + r2 rs + r2 r3 r4 + r1 r3 rs + 2r, r2 r3r4r5 r1r2r4r5 r1r2r3r4 r1r3r4r5 r2 r3 r4 rsr2 r3 r5 5 Cs = Ea, exp ,=11—r, The problem is to find the decision variables i =1,2,...,5 which minimize both Qs and C Subject to 0 < r,i =1,2,...,5, In other words, the problem can be posed as a MOOP given by Minimize (Qs ,Cs ) subject to 0 r,i =1,2,...,5; (6.19) where a, =1 and b, = 0.0003, Vi, i = 1, 2,...,5 . 6.4.1 Step-by-Step Illustration of MOEA-IDFA for Complex Bridge System Step 0: Intialization of Priori Parameters The DF's a priori preference parameters need initialized to construct the DFs (,uQ, and) for each objective (i.e., Qs and Cs ), shown in Table 6.9. Step 1: Construction of DFs The DFs (//Q‘ and) corresponding the objectives Qs and C, is constructed based upon the DM's preference parameters shown in Table 6.10 and the initialized parameters of Step 0 • in case of No Preference, Equations (6.1-6.2) are used, • in case of Preference 1, Equations (6.4-6.5) are used, • in case of Preference 2, Equations (6.6-6.7) are used, 156 w in case of Preference 3, Equations (6.9-6.10) are used and w in case of Preference 4, Equations (6.12-6.13) are used. Step 2: Solving the DFMOOP The following DFMOOP is formed using the DFs obtained in Step 1 that needs to be solved by any MOEA. Minimize (1.1c ,,uc.$ ) subject to 0i =1,2, (6.20) For each case (preference), the combination of DF will be different and so the problem will be different for each case shown above. We are using NSGA-II and MOPSO-CD to solve the DFMOOP obtained in Equation 6.20. The optimal parameters setting used for both of these algorithms are provided in Table 6.11. Step 3: Evaluation of the Solution As shown in Chapters 2-5 Pareto-optimal solution of DFMOOP (Equation 6.20) is also the Pareto-optimal solution of MOOP (Equation 6.19). We have plotted the best POF obtained through various run of the algorithms of the series system for each case. The POF of No Preference case is shown in Figure 6.9. For rest of the cases the plots are shown in Figures 6.12. Step 4: Adjustment of the Parameters If the DM is satisfied by the solution obtained in Step 3 the approach stops successfully. Other wise the key preference parameters shown in Table 6.10 can be altered to meet the DM's choice and the method again go back to Step 2. The process is repeated until DM is satisfied. We are just showing one run of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. 6.5 Residual Heat Removal (RHR) System of a Nuclear Power Plant Safety System This problem considers the robust design of a Nuclear Power Plant safety system: The Residual Heat Removal system (RHR) is a low pressure system (400 psi) directly connected to the primary system which is at higher pressure (1200 psi), where psi stands for pounds per square inch. The RHR constitutes an essential part of the low- 157 pressure core flooding system which is part of the Emergency Core Cooling System of a nuclear reactor. Its objectives are: • to remove decay and residual heat from the reactor so that refueling and servicing can be performed, • to supplement the spent fuel cooling capacity, • to condense reactor steam so that decay and residual heat may be removed if the main condenser is unavailable following a reactor scram In Figure 6.5, a schematic of the system is shown (Marseguerra et al., 2002)). The unreliability of the system Q, can be obtained from the simplified fault tree shown in Figure 6.6 in which the 16 most important (third-order) cut sets (see Table 5.2), originated by the combination of eight basic events, are considered (Apostolakis, 1974; Marseguerra et al., 2002). The simplified expression for a yields. Qs = q1 q3q5 + qlq3q6 (1- q5) + q1 q3q7 (1-q5 )(1-q6 ) + ql q3q8 (1- q5 )(1-q6 )(1-q,) + q1q4q5 (1- q3 ) + ql q4q6 (1-q3 )(1-q5 )+ qiq4q7 (1-q3 )(1-q5 )(1-q6 ) + qi q4q8 (1-q3 )(1-q5 )(1-q6 )(1-q7 )+ q2q3q,(1- q1 ) + q2q3q6 (1-q1 )(1- q,) + q2q3q7 (1-q1 )(1-q5 )(1-q6 )+ q2q3q8 (1- q1 )(1- q,)(1- q6 )(1- q7 ) + q2q4q5 (1- q1 )(1-q3 )+ q2q4q6(1- q1 )(1- q3 )(1-q5 ) + q2q4q7 (1-q1 )(1-q3 )(1-q5 )(1-q6 )+ q2q4q8(1 - q1 )(1 - q3)(1 - q5 )(1 - q6 )(1 - q,), where q, and (1— q,)=r,, are respectively the component unreliability and component reliability of the ith basic event are subject to constraint 0 ro q S 1. Furthermore, normally the system design problem seeks also to constrain the system cost, here taken equal to a nonlinear combination of the components reliabilities Cs = 2E K,/,`; In this problem, the vector of coefficients of the nonlinear combination is K = {100,100,100,150,100,100,100,150} and the exponents area, = 0.6, Vi, i =1,2,...,8. ) is considered as an In the approach proposed in the work, the system cost (Cs ) objective to be minimized instead of a constraint, and this, together with the minimization of the system unreliability (Qs ), leads to a MOOP Minimize (Qs ,Cs ) subject to 0 5 r,i =1,2,...,8; 158 (6.21) 6.5.1 Step-by-Step illustration of MOEA-IDFA for RHR System Step 0: Intialization of Priori Parameters The DF's priori preference parameters need initialized to construct the DFs (,uQand) for each objective (i.e., Qs and Cs ), shown in Table 6.12. Step 1: Construction of DFs The DFs (/./Q and /JO corresponding the objectives Qs and Cs is constructed based upon the DM's preference parameters shown in Table 6.13 and the initialized parameters of Step 0 • in case of No Preference, Equations (6.1-6.2) are used, • in case of Preference 1, Equations (6.4-6.5) are used, • in case of Preference 2, Equations (6.6-6.7) are used and • in case of Preference 4, Equations (6.12-6.13) are used. We were not able to find the parameter setting for this problem corresponding Preference 3. Step 2: Solving the DFMOOP The following DFMOOP is formed using the DFs obtained in Step 1 that needs to be solved by any MOEA. Minimize (uQ, , tics ) subject to 0i =1,2,...,8; (6.22) For each case (preference), the combination of DF will be different and so the problem will be different for each case shown above. We are using NSGA-II and MOPSO-CD to solve the DFMOOP obtained in Equation 6.22. The optimal parameters setting used for both of these algorithms are provided in Table 6.14. Step 3: Evaluation of the Solution As shown in Chapters 2-5 Pareto-optimal solution of DFMOOP (Equation 6.22) is also the Pareto-optimal solution of MOOP (Equation 6.21). We have plotted the best POF obtained through various run of the algorithms of the series system for each case. The plots are shown in Figures 6.13. Step 4: Adjustment of the Parameters If the DM is satisfied by the solution obtained in Step 3 the approach stops successfully. Other wise the key preference parameters shown in Table 6.13 can be 159 altered to meet the DM's choice and the method again go back to Step 2. The process is repeated until DM is satisfied. We are just showing one run of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. 6.6 Reliability Optimization of a Mixed Series-Parallel System The multi-objective reliability optimization problem is taken from Sakawa (1978) and Ravi et al. (2000). The block diagram for this problem is in Figure 6.4. Relevant data for this problem is given in the Table 6.1. A multistage mixed system is considered, where the problem is to allocate the optimal reliabilities r„ i =1,2,3,4 of four components whose redundancies are specified in order to achieve the following three goals Maximize Rs ,minimize Cs ,minimize Ws or Minimize Qs,minimize q,minimize Ws Subject to Vs =65; Ps :5_12000; i=t 0.5 j =1,2,3,4; In other words, the problem can be posed as a MOOP of three objectives given by Minimize (Qs ,Cs ,Ws ) subject to 4 V = V,n, (6.23) 65; Ps 5_12000; 0.5<r, <1,i=1,2,3,4; where V, are the reliability, unreliability, cost, weight and volume of the system respectively. Here, r, represents the reliability of the system. In addition, we have, Ps =Ws *Vs ; 4 4 4 R = 11[1 — ( 1;C, = EC= EkVnj; J .1 >=i 1=1 160 component of the )6.; 171 ;W./ =[logi,3 ( 1 gio (1— ri ) (1— rj ) V= 7; P;t.j) login (0 /4 )1 Table 6.1Data for Mixed Series-Parallel System a; = 8.0,a7 = 6.0, a; = 2.0; y; = 2.0,7: = 0.5, y; = 0.5; N1` = 2.0, x=10.0, /3; = 3.0, 16: =18.0; fir = 3.0,)621v =2.0, 13; =10.0, JO: = 8.0; IA" = 2.0, )6; = 2.0, )57 = 6.0, 13: = 8.0; n1 = 7, n2 = 8, n3 = 7, n4 = 8; 6.6.1 Step-by-Step Illustration of MOEA-IDFA for Mixed Series-Parallel System Step 0: Intialization of Priori Parameters The DF's priori preference parameters need initialized to construct the DFs (uQ,and pw, ) for each objective (i.e., Qs 7 Cs and W. ), shown in Table 6.15. Step 1: Construction of DFs The DFs (,uo,and) corresponding the objectives Q, and Cs is constructed with the help of DM's preference parameters shown in Table 6.16 and the initialized parameters of Step 0 • in case of No Preference, Equations (6.1-6.3) are used, ■ in case of Preference 2, Equations (6.6-6.8) are used, • in case of Preference 3, Equations (6.9-6.11) are used and • in case of Preference 4, Equations (6.12-6.14) are used. Step 2: Solving the DFMOOP The following DFMOOP is formed using the DFs obtained in Step I that needs to be solved by any MOEA. 161 Minimize (N s , Pcs , fiws ) (6.24) subject to Vs 4 = EV,n, 65; Ps ..12000; 0.51=1,2,3,4; For each case (preference), the combination of DF will be different and so the problem will be different for each case shown above. We are using NSGA-II and MOPSO-CD to solve the DFMOOP obtained in Equation 6.24. The optimal parameters setting used for both of these algorithms are provided in Table 6.17. Step 3: Evaluation of the Solution As shown in Chapters 2-5 Pareto-optimal solution of DFMOOP (Equation 6.24) is also the Pareto-optimal solution of MOOP (Equation 6.23). We have plotted the best POF obtained (having maximum hypervolume metric) through various run of the algorithms of the series system for each case. The plots are shown in Figures 6.14. Step 4: Adjustment of the Parameters If the DM is satisfied by the solution obtained in Step 3 the approach stops successfully. Other wise the key preference parameters shown in Table 6.16. can be altered to meet the DM's choice and the method again go back to Step 2. The process is repeated until DM is satisfied. We are just showing one run of the approach here as we assume that in this problem DM is satisfied by the results obtained in Step 3. 6.7 Conclusion We have suggested a novel way of taking preference information coming from the DM more closely into account in DF based methods developed for multi-objective optimization. Our goal is to be able to produce portions (subsets) of Pareto-optimal set that are necessary to the DM than the ones produced with standard MOEAs. We have carried out several computational tests in order to compare the outputs of MOEA-IDFA approach using different combination (all linear, convex-concave, all sigmoidal and all convex) of DFs. With five reliability optimization problems of multiobjective nature, we have tested all the four cases (types of combination of DFs used) with real DMs. Four different set of DFs are utilized to simulate the DM's preference in the solution process in order to illustrate the effectiveness of the proposed methodology 162 in capturing different kinds of preference structures (different portions of the POF) of the DM. In other words, we have replaced the responses of the DM by DFs. The results are encouraging and suggest the applicability of the proposed approach to more complex and real-world engineering problems. 163 IN 0- 2 1 5 [---• OUT Figure 6.1 Block Diagram of Series System I- r 4 -4) OUT --I 4I-Figure 6.2 Block Diagram of a Life Support System in a Space Capsule 164 OUT Figure 6.3 Block Diagram of Complex Bridge System H1 IN • OUT 2 x 2 X Figure 6.4 Mixed Series-Parellel System 165 n rPRIMARY CONTAINMENTF022 —I E:] LEG 1 tag I I /1-01 I REACTOR F019 i. I F023 PRESSURE I. • I VESSEL - MOTOR OPERATED - AIR OPERATED - PANEL LIGHT 375 PSI \ ntotyro .F060 ire I DRY WELL SPRAY I HEADER F005 F015F017 RHR•HX SULRESSION SPRAY HEADER FO0B 1 F0019 71 1 1 .1 11 11 F047 STRAINER SUPPRESSION POOL (TORUS) COO 2A: F03/A F034A Figure 6.5 Schematic of the RHR of a Nuclear Power Plant Table 6.2 Third order minimal cut sets Check valve F019 (1 — 1) Cut set Valve clapper stuck open (4 =I) I. Valve clapper fails at design pressure (j = 2) 2 3 Valve clapper fails at design pressure Valve clappet stuck open 4 Valve clapper fails at. design pressure 5 Valve dapper, stuck open 6 7 Valve clapper stuck open Valve clappet fails at design pressure 8 Valve clappet stuck open 9 Valve clappet fails at. design pressure 10 Valve clapper stuck open 11 . Valve clappet fails at design pressure 12 13 Valve clappet fails at design pressure 14 Valve clappet stuck open Valve clapper fails at design pressure 1.5 Valve clappet stuck open 16 Valve F022 opened (1 = 2) Left open by mistake (j =1) Left open by mistake Mechanical failure (j = 2) Mechanical failure Left open by mistake Left- open by mistake. Mechanical failure Mechanical failure Mechanical failure Left open by mistake Left open by mistake Left open by mistake Mechanical failure Mechanical failure Mechanical failure Mechanical failure 166 Valve F023 opened i = 3) Internal valve failure (j = 1) Internal valve failure Internal valve failure Casing rupture. (j = 2) Casing rupture Internal valve failure Casing rupture Casing rupture Inadvertently left open (j =3) Inadvertently left open Left open. for future _maintenance (j = 4) Left open for future maintenance Inadvertently left open Left open for future maintenance Left open for future maintenance Inadvertently left open System failure Valve F023 opened Valve 1022 opened Media nical failorc Check- valve F019 fails to close Valve cluppet stuck open !mortal value1:Awe Valve cluppet rails at design pressure Valve let op en by mistake Casing, rupture Valve inadvertent left open 'Valve left open for.. .future • . I , m aintenance Figure 6.6 Simplified Fault Tree of the RHR system (Apostolakis, 1974) 6000 NS GA-II 550 - • MOPS O-CD 500 o 450 - 400 0.00.2OA0.60.81.0 Reliability Figure 6.7 POFs w.r.t. NSGA-H and MOPSO-CD of a Series System (No Preference Case) 167 700- 0 NSGA-II NIOPSO-CD 680- 660 - 640- • •• 0.93 0.96 Reliability • •I• 0.99 Figure 6.8 POFs w.r.t. NSGA-H and MOPSO-CD of a Life Support System in a Space Capsule (No Preference Case) 0N 5.016 SG A-II M OP SO-CD • 5.012 C 5.008 - 5.004 - ., 5.000, ,, 0.00.20.40.60.81.0 Reliability Figure 6.9 POFs w.r.t. NSGA-II and MOPSO-CD of a Complex Bridge System (No Preference Case) 168 Table 6.3 Initial a Priori Parameters for Series System Parameters initialized in Step 0 V Q:. C: C:. 0.031 0.9509 385.5 585.88 Table 6.4 Other a Priori Parameters for Series System Preference 1 Preference 2 Preference 3 Preference 4 nt n2 ni n2 a2 a2 nt n2 10 0.001 0.001 10 20 0.1 15 15 Table 6.5 A Posteriori Parameters for Series System Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation P rob . Arch. Size W c] c2 Cross. Prob. Cross. Index Cross. Prob. 200 300 0.4 200 0.4 1 1 0.8 10 10 200 300 0.4 200 0.5 1 1 0.8 15 15 200 350 0.5 200 0.5 1 1 0.9 15 15 200 350 0.5 200 0.5 1 1 0.9 10 15 200 350 0.5 200 0.6 1.1 1.1 0.9 10 15 4..: 4., ct 2 a. (-4 4.' M 4." v. V a 169 600 600 - 0 NS GA-II 0 NS GA-II MO PSO-C7D 550 - 550 - O O ° MOPS C)-CD O 500 500 o o 450 - 400 - 450 - o.r) 0 400 0.00.0.40.60.81.0 0.00.20.40.60.81.0 Reliability Reliability (a) (b) 600 - 0 NS • 550 - 0 NS GA-II MO PS 0-CD 0 ° MO PS O-CD 500 .4 450 400 - o" 0.00.20.40.60.8 Reliability 1 1.0 I I I I I 0.40,60.81.0 0 . 0 0.7 Reliability (d) (c) Figure 6.10 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: (a) For Preference 1, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 170 Table 6.6 Initial a Priori Parameters for Life Support System in a Space Capsule Parameters initialized in Step 0 Q: Q:* C: C:* 0.01 0.1 641 700 Table 6.7 Other a Priori Parameters for Life Support System in a Space Capsule Preference 1 Preference 2 Preference 3 Preference 4 ni n2 ni n2 a2 a2 ni n2 10 0.001 0.001 10 170 170 10 10 Table 6.8 A Posteriori Parameters for Life Support System in a Space Capsule Common Parameters Pop Size Max. Gen. at'-'0 z 200 400 L 0,.. N tj c.,. 200 Mutati- MOPSO-CD Parameters NSGA-H Parameters Arch. Size W Cl c2 Cross. Prob. Cross. Index Cross. Prob. 0.6 200 0.4 1 1 0.8 10 10 400 0.7 200 0.5 1 1 0.8 15 10 200 400 0.7 200 0.5 1 1 0.9 15 15 M "cd 200 450 • 0.7 200 0.6 1 1 0.9 10 10 t; 200 450 0.8 200 0.6 1.0 1.0 0.9 10 10 Pro onb. ,.:. IM. A. 171 700- 0 NSGA-II ' MOPSO-CD 680O O 660- 660 - 640- 640 • I ,•• 0.90 f • •I....... 0.900.930.960.99 Reliability 700- 0 NSGA-II • Reliability • •• 0.99 (b) (a) 700 - 0.930.96 MOPSO-CD 0 NSGA-II " MOPSO-CD 680 - 680 . 660 - 660 - 640 - 640 0.93 • • 0.96 6 Reliability Reliabi • • ,•••• 0.900.930.96 0.99 Reliability • 0.99 (d) (c) Figure 6.11 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at Different Preferences: (a) For Preference 1, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 172 Table 6.9 Initial a Priori Parameters for Complex Bridge System Parameters initialized in Step 0 Q, QS- C' s C: 0.0160 1.0 5.0015 5.01686 Table 6.10 Other a Priori Parameters for Complex Bridge System Preference 1 Preference 2 Preference 3 Preference 4 ni n2 ni n2 az a, ni n2 5 0.1 0.1 5.0 1000 1000 5 5 Table 6.11 A Posteriori Parameters for Complex Bridge System Common Parameters MOPSO-CD Parameters Pop Size Max. Gen. Mutation Prob. Arch. Size W 200 200 0.6 200 200 200 0.7 200 250 200 200 NSGA-H Parameters Cl c2 Cross. Prob. Cross. Index Cross. Prob. 0.5 1 1 0.8 10 10 200 0.5 1 1 0.8 10 10 0.7 200 0.5 1 1 0.9 10 10 250 0.7 200. 0.5 1 1 0.6 10 10 250 0.7 200 0.6 1.0 1.0 0.6 10 10 %-1 or It Z' a: '16 a: M L a. a. 173 5.016 - 0 NSGA-II 5.012 - 5.012 - a.: a 5.008 - 5.008 - 5.004 - 0 00 ,, 5.000 0.00.20.40.60.81.0 5.000 0'00.20.40.60.81.0 Reliability Reliability (a) (b) 0 NSGA-II 5.016 - O O ° IvIOPSO-CD IklOPSO-CD 5.004 - 0 0 NSGA-II 5.016 - 5.016- ' MOPSO-CD 0 NSGA-I1 MOPSO-CD 5.012- 5.012 -1 0 a 1 5.008 - 5.008 - . CI 0 0 5.004 - 5.004 4 4..) , ,, 5 000, 0.00.20.40.60.81.0 ,. 5.000 0.00.20.40.60.81.0 Reliability Reliability (c) (d) Figure 6.12 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: Preference 1(a), Preference 2 (b), Preference 3 (c), Preference 4 (d) 174 Table 6.12 Initial a Priori Parameters for RHR System Parameters initialized in Step 0 Q,. Q, C,. C„7" 0 1 0.413412 412.0398 Table 6.13 Other a Priori Parameters for. RHR System Preference 1 ni 15 Preference 2 n 2i 0.001 Preference 4 ni n2 ni n2 0.001 15 .15 15 Table 6.14 A Posteriori Parameters for RHR System Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. M uta tion Prob. Arch. Size w Cl c2 Cross. Prob. Cross. Index Cross. Prob. 4-■ 200 300 0.3 200 0.6 1 1 0.9 10 10 L u. 200 300 0.3 200 0.6 1 1 0.8 10 10 ' 200 4.<s. 200 L a, 350 0.1 200 0.6 1 1 0.9 10 10 350 0.1 200 0.6 1 1 0.9 10 15 ,..4 4, z'' 175 450 - o 0 NSGA-II 450 NSGA-II • ' MOPSO-CD 300 - MOPSO-CD 300 - .) ef)• 150- 150- 0- 0 Reliability Reliability (a) (b) O NSGA-II • 300 - 0 0.00.20.40.60.81.0 0.00.20.40.60.81.0 450 - 0 MOPSO-CD 0, 0 0 0 150 - 0- • 0.40.60.8 1.0 0.0 0.2 Reliability (c) (d) Figure 6.13 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: (a) For No Preference, (b) For Preference 1, (c) For Preference 2, (d) For Preference 4 176 Table 6.15 Initial a Priori Parameters for Mixed Series-Parallel System Parameters initialized in Step 0 C: C, 131, W:* 317.552 381.935 174.872 186.665 Q: 0.0025 0.023 Table 6.16 Other a Priori Parameters for Mixed Series-Parallel System Preference 2 Preference 3 Preference 4 n1 n2 n3 a2 a2 a, ril n2 n3 10 0.001 0.001 800 0.35 2 10 10 10 Table 6.17 A Posteriori Parameters for Mixed Series-Parallel System Common Parameters MOPSO-CD Parameters NSGA-II Parameters Pop Size Max. Gen. Mutation Prob . Arch. Size w cl c2 Cross. Prob. Cross. Index Cross. Prob. ft Z:' 200 200 0.4 200 0.4 1 1 0.8 10 10 ..; 200 350 0.4 200 0.5 1 1 0.9 15 15 200 350 0.4 200 0.5 1.1 1.1 0.9 10 10 200 400 0.1 200 0.6 1 1 0.9 10 10 ,..1 eu f.) ti at, L a.. 177 (a) (b) (c) (d) Figure 6.14 POFs w.r.t. NSGA-II and MOPSO-CD of a Series System at different preferences: (a) For No Preference, (b) For Preference 2, (c) For Preference 3, (d) For Preference 4 178 179 CHAPTER 7 Conclusions and Scope for Future Work The fundamental theme of this study is the development of an efficient strategy to provide DM the guided or interactive POF of his/her interest. Various benchmark problems as well as real-world engineering problems (reliability optimization problems) are examined to ensure the efficacy of the approach proposed. This study proposes a novel scheme for user preference incorporation in MOEAs. The results on the single and multiple target scenarios indicate the ability of our scheme to efficiently explore and focus the search on the regions of interest to the DM. The main advantages of our scheme are its general applicability and the straightforward integration of decision-making and optimization. The expert establishes his/her preference in light of the feasible alternatives rather than specifying trade-offs a priori. This chapter is organized as follows: Section 7.1 derives the conclusions based on the present study and Section 7.2 enlists the suggestions for further work in this direction. 7.1 Conclusions In this study, we have addressed an important task of combining MOEA with a DFA to not find a single Pareto-optimal solution, but to find' a set of solutions near the desired regions of DM's interest on the POF. With a number of trade-off solutions in the region of interests we have argued that the DM would be able to make a better and more reliable decision than that with a single solution. A new interactive approach named as MOEA-IDFA, is proposed to guide the POF in the region/regions of interest of DM. The proposed approach is better than IDFA in terms of the fact that the MOEA-IDFA provides guided region/regions unlike single solution (in case of IDFA), in one run of the approach. Guided POF provides DM flexibility (by providing more than one solution) in choosing final solution as well as reduces the burden of the DM by not providing unnecessary (non preferred) solutions. It is concluded that use of different type of DF exerts different kind of biasness in POF. We have categorized the DF on the 180 basis of the choice of DM. To investigate the effectiveness of the approach two types of MOEAs (NSGA-II and MOPSO-CD) have been employed to solve the DFMOOP and their results have been compared. This methodology links ideas and contributions that span the following areas: • The preference is the main character of MOOP, which produces the different solution suitable to the various requirement of DM. The preference is often expressed by goal, importance, priority, and weights. The structure of DF is an alternate way to denote preference of the DM. • It attempts to find a guided POF according the wish of the DM. • Structure of the method is simple and mathematical complexities are very few. Hence, it can be easily programmed and implemented. • MOEA-IDFA is robust and does not require assurance from the user regarding the mathematical properties (such as continuity, differentiability and convexity, etc.) of the objective functions and constraints. • One of the most important practical advantage of this approach is that the mathematical models of real life optimization problems can be solved. • Basic theoretical analysis of the approach is also presented in this work. • Apart from single region of interest, multiple regions of interest of the DM are also incorporated, which is a unique achievement of this study. • Both bi-objective and tri-objective optimization problems are examined here in perspective of MOEA-IDFA. • Two very popular MOEAs (NSGA-II and MOPSO-CD) have been utilized to explain the functioning of MOEA-IDFA. • Several standard test problems as well as five real world problems corresponding reliability optimization have been solved using MOEA-IDFA, which provided the usefulness of the approach. 181 7.2 Future Scope There may exist several interesting directions for further research and development based on the work in this thesis. Some of the suggestions for future work in this direction are: • Since the proposed approach is a MOEA based approach, a comparative study of MOEA-IDFA with other multi-objective computational algorithms such as SPEA2, DE and ACO type methods need to be carried out. • Another type of DFs can be developed to guide POF in other interesting regions of DM's interest. • In this study, maximum two regions of a POF have been obtained on the basis of DM's interest. In future more than two regions simultaneously can be obtained developing proper combination of DFs. • Theoretical analysis of this approach is a major area in which a lot of work needs to be done. POFs of MOOP and DFMOOP have interesting relationship that need to be investigated theoretically. • Research is needed to be carried out to choose the applicability of different combinations of DFs in different applications. • Relationship between membership function and DF needs to be explored theoretically as well numerically. • Effect of MOEA-IDFA needs to be investigated on MOOP having more than three objectives. 182 183 BIBLIOGRAPHY [1] Acan, A. (2004). An external memory implementation in ant colony optimization. Ant Colony, Optimization and Swarm Intelligence, 247-269. [2] Acan, A. (2005). An external partial permutations memory for ant colony optimization. Evolutionary Computation in Combinatorial Optimization, 1-11. [3] Adra, S., Griffin, I., and Fleming, P. (1993). A comparative study of progressive preference articulation techniques for multiobjective optimisation (Springer Berlin / Heidelberg). [4] Aggarwal, K. K., Gupta, J. S., and Misra, K. B. (2009a). A new heuristic criterion for solving a redundancy optimization problem. Reliability, IEEE Transactions on 24, 86-87. [5] Aggarwal, K. K., Misra, K. B., and Gupta, J. S. (2009b). A fast algorithm for reliability evaluation. Reliability, IEEE Transactions on 24, 83-85. [6] Agrawal, S., Panigrahi, B. K., and Tiwari, M. K. (2008). Multiobjective particle swarm algorithm with fuzzy clustering for electrical power dispatch. IEEE Transactions on Evolutionary Computation 12, 529-541. [7] Amari, S. V., and Dill, G. (2010). Redundancy optimization problem with warm-standby redundancy. In Reliability and Maintainability Symposium (RAMS), 2010 Proceedings-Annual (IEEE), 1-6. [8] Amari, S. V., Dugan, J. B., and Misra, R. B. (2002). Optimal reliability of systems subject to imperfect fault-coverage. Reliability, IEEE Transactions on 48, 275-284. 184 [9] Amari, S. V., and McLaughlin, L. (2005). Optimal design of a condition-based maintenance model. In Reliability and Maintainability, 2004 Annual Symposium-RAMS (IEEE), 528-533. [10] Amari, S. V., Pham, H., and Dill, G. (2004). Optimal design of k-out-of-n: G subsystems subjected to imperfect fault-coverage. Reliability, IEEE Transactions on 53, 567-575. [11] Andrews, J. D., and Moss, T. R. (1993). Reliability and risk assessment (Longman Group, UK). [12] Apostolakis, G. E. (1974). Mathematical methods of probabilistic safety analysis. [13] Athan, T. W., and Papalambros, P. Y. (1996). A note on weighted criteria methods for compromise solutions in multi-objective optimization. Engineering - Optimization 27, 155 176. [14] Babu, B. V., and Chaturvedi, G. (2000). Evolutionary computation strategy for optimization of an alkylation reaction. In Proceedings of International Symposium & 53rd Annual Session of IIChE (CHEMCON-2000) (Citeseer), 18-21. [15] Babu, B. V., and Chaurasia, A. S. (2003). Optimization of pyrolysis of biomass using differential evolution approach. In Second International Conference on Computational Intelligence, Robotics, and Autonomous Systems (CIRAS2003), Singapore. [16] Babu, B. V., and Jehan, M. M. L. (2004). Differential evolution for multiobjective optimization. In Evolutionary Computation, 2003. CEC'03. The 2003 185 Congress on (IEEE), 2696-2703. [17] Barbosa, H. J. C. (2002). A coevolutionary genetic algorithm for constrained optimization. In Evolutionary Computation, 1999. CEC 99. Proceedings of the 1999 Congress on (IEEE). [18] Barbosa, H. J. C. (1996). A genetic algorithm for min-max problems. In Goodman, editors, Proceedings of the First International Conference on Evolutionary Computation and Its Applications, 99-109. [19] Barbosa, H. J. C., and Lemonge, A. C. C. (2003). A new adaptive penalty scheme for genetic algorithms. Information Sciences 156, 215-251. [20] Barbosa, H. J. C., and Lemonge, A. C. C. (2008). An adaptive penalty method for genetic algorithms in constrained optimization problems. Frontiers in Evolutionary Robotics. Vienna: I-Tech Education and Publishing 1, 9-34. [21] Barbosa, H. J. C., and Lemonge, A. C. C. (2002). An adaptive penalty scheme in genetic algorithms for constrained optimization problems. In Proceedings of the Genetic and Evolutionary Computation Conference (Morgan Kaufmann Publishers Inc.), 287-294. [22] Box, G. E. P., and Wilson, K. B. (1951). On the experimental attainment of optimum conditions. Journal of the Royal Statistical Society. Series B (Methodological) 13, 1-45. [23] Branke, J. (2008a). Consideration of partial user preferences in evolutionary multiobjective optimization. Multiobjective Optimization, 157-178. [24] Branke, J., and Deb, K. (2005). Integrating user preferences into evolutionary multi-objective optimization. Knowledge Incorporation in Evolutionary 186 Computation, 461-477. [25] Branke, J., Deb, K., Dierolf, H., and Osswald, M. (2004). Finding knees in multi-objective optimization,1-10. [26] Branke, J., KauBler, T., and Schmeck, H. (2001). Guidance in evolutionary multi-objective optimization. Advances in Engineering Software 32, 499-507. [27] Castro, R. E., and BARBOSA, H. (2001). Otimizacao de estruturas corn multiobjetivos via algoritmos geneticos. Rio de Janeiro 206. [28] Chakraborty, J., Konar, A., Nagar, A., and Das, S. (2009). Rotation and translation selective Pareto optimal solution to the box-pushing problem by mobile robots using NSGA-II. In Evolutionary Computation, 2009. CEC'09. IEEE Congress on (IEEE), 2120-2126. [29] Chatsirirungruang, P., and Miyakawa, M. (2008). Application of genetic algorithm to numerical experiment in robust parameter design for signal multiresponse problem. In Proceedings of The 13th International Conference on Industrial Engineering Theory, Applications & Practice, 7-10. [30] Coelho, L. S. (2009). An efficient particle swarm approach for mixed-integer programming in reliability-redundancy optimization applications. Reliability Engineering & System Safety 94, 830-837. [31] Coello Coello, C. A. (2009). Evolutionary multi-objective optimization: some current research trends and topics that remain to be explored. Frontiers of Computer Science in China 3, 18-30. [32] Coello, C. A. (2000). An updated survey of GA-based multiobjective optimization techniques. ACM Computing Surveys (CSUR) 32, 143. 187 [33] Coello, C. A. C. (1999). A comprehensive survey of evolutionary-based multiobjective optimization techniques. Knowl. Inf. Syst. 1, 129-156. [34] Coello, C. A. C. (1996). An empirical study of evolutionary techniques for multiobjective optimization in engineering design (Ph.D. Thesis). [35] Coello, C. A. C. (2009). Evolutionary multi-objective optimization: some current research trends and topics that remain to be explored. Frontiers of Computer Science in China 3, 18-30. [36] Coello, C. A. C. (2004). List of references on evolutionary multiobjective optimization. Laboratorio Nacional De Informatica Avanzada (LANIA). [37] Coello, C. A. C., de ComputaciOn, S., and Zacatenco, C. S. P. (2006). Twenty years of evolutionary multi-objective optimization: A historical view of the field. IEEE Computational Intelligence Magazine 1, 28-36. [38] Coello, C. A. C., Lamont, G. B., and Van Veldhuizen, D. A. (2007). Evolutionary algorithms for solving multi-objective problems (Springer-Verlag New York Inc). [39] Coello, C. A. C., and Lechuga, M. S. (2002). MOPSO: A proposal for multiple objective particle swarm optimization. Proceedings of the Evolutionary Computation on, 1051-1056. [40] Coello, C. A. C., Pulido, G. T., and Lechuga, M. S. (2004). Handling multiple objectives with particle swarm optimization. IEEE Trans. Evolutionary Computation 8, 256-279. [41] Coit, D. W., and Smith, A. E. (1996). Reliability optimization of series-parallel 188 systems using a genetic algorithm. IEEE Transactions on Reliability 45, 254260. [42] Collette, Y., and Siarry, P. (2003). Multiobjective optimization: principles and case studies (Springer Verlag). [43] Come, D., Knowles, J., and Oates, M. (2000). The Pareto envelope-based selection algorithm for multiobjective optimization. In Parallel Problem Solving from Nature PPSN VI (Springer), 839-848. [44] Cvetkovic, D., and Parmee, I. (2003). Agent-based support within an interactive evolutionary design system. AI EDAM 16, 331-342. [45] Cvetkovic, D., and Parmee, I. C. (1998). Evolutionary design and multi— objective optimisation. In 6th European Congress on Intelligent Techniques and Soft Computing EUFIT (Citeseer), 397-401. [46] Cvetkovic, D., and Parmee, I. C. (2002). Preferences and their application in evolutionary multiobjectiveoptimization. IEEE Transactions on Evolutionary Computation 6, 42-57. [47] Cvetkovic, D., and Parmee, I. C. (1999). Use of Preferences for GA—based Multi—objective Optimisation. In GECCO-99: Proceedings of the Genetic and Evolutionary Computation Conference (Citeseer), 1504-1509. [48] Deb, K. (2003). Multi-objective evolutionary algorithms: Introducing bias among Pareto-optimal solutions. Advances in Evolutionary Computing: Theory and Applications, 263-292. [49] Deb, K. (1999). Multi-objective genetic algorithms: Problem difficulties and construction of test problems. Evolutionary computation 7, 205-230. 189 [50] Deb, K. (2001). Multi-objective optimization using evolutionary algorithms (Wiley). [51] Deb, K., Agrawal, S., Pratab, A., and Meyarivan, T. (2000). A Fast Elitist NonDominated Sorting Genetic Algorithm for Multi-Objective Optimization: NSGA-II (KanGAL report 200001). Indian Institute of Technology. [52] Deb, K., and Goel; T. (2001). Controlled elitist non-dominated sorting genetic algorithms for better convergence. In Evolutionary Multi-Criterion Optimization (Springer), 67-81. [53] Deb, K., and Kumar, A. (2007). Light beam search based multi-objective optimization using evolutionary algorithms. In Proc. of the Congress on Evolutionary Computation (CEC 2007), 2125-2132. [54] Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE transactions on evolutionary computation 6, 182-197. [55] Deb, K., and Sundar, J. (2006). Reference point based multi-objective optimization using evolutionary algorithms. In Proceedings of the 8th annual conference on Genetic and evolutionary computation (ACM), 642. [56] Derringer, G., and Suich, R. (1980). Simultaneous optimization of several response variables. Journal of quality technology 12, 214-219. [57] Dhillon, B. S. (1986). Reliability apportionment/allocation: a survey. Microelectronics Reliability 26, 1121-1129. [58] Dhingra, A. K. (1992). Optimal apportionment of reliability and redundancy in 190 seriessystems under multiple objectives. IEEE Transactions on Reliability 41, 576-582. [59] Dorigo, M., and Stutzle, T. (2010). Ant Colony Optimization: Overview and Recent Advances. Handbook of Metaheuristics, 227-263. [60] Drezner, T., Drezner, Z., and Salhi, S. (2005). A multi-objective heuristic approach for the casualty collection points location problem. Journal of the Operational Research Society 57, 727-734. [61] Durillo, J. J., Garcia-Nieto, J., Nebro, A. J., Coello, C. A., Luna, F., and Alba, E. (2009). Multi-Objective Particle Swarm Optimizers: An Experimental Comparison. In Proceedings of the 5th International .Conference on Evolutionary Multi-Criterion Optimization (Springer-Verlag), pp. 495-509. [62] Emmerich, M., Beume, N., and Naujoks, B. (2005). An EMO algorithm using the hypervolume measure as selection criterion. In Evolutionary Multi-Criterion Optimization (Springer), pp. 62-76. [63] Fonseca, C. M., and Fleming, P. J. (1995). An overview of evolutionary algorithms in multiobjective optimization. Evolutionary computation 3, 1-16. [64] Fonseca, C. M., and Fleming, P. J. (1993). Genetic algorithms for multiobjective optimization: Formulation, discussion and generalization. In Proceedings of the fifth international conference on genetic algorithms (Citeseer), pp. 416-423. [65] Fonseca, C. M., and Fleming, P. J. (1998). Multiobjective optimization and multiple constraint handling with evolutionary algorithms. II. Application example. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans 28, 38-47. 191 [66] Formato, R. A. (2009). Central force optimization: A new deterministic gradient-like optimization metaheuristic. OPSEARCH 46, 25-51. [67] Formato, R. A. (2007). Central force optimization: A new metaheuristic with applications in applied electromagnetics. Progress In Electromagnetics Research 77, 425-491. [68] Formato, R. A. (2010). Improved CFO algorithm for antenna optimization. Progress In Electromagnetics Research 19, 405-425. [69] Fourman, M. P. (1985). Compaction of symbolic layout using genetic algorithms. In Proceedings of the 1st International Conference on Genetic Algorithms (L. Erlbaum Associates Inc.), 141-153. [70] Fuller, R., and Carlsson, C. (1996). Fuzzy multiple criteria decision making: Recent developments. Fuzzy Sets and Systems 78, 139-153. [71] Geoffrion, A. (1968). Relaxation and the dual method in mathematical programming (California Univ Los Angeles Western Management Science Inst). [72] Goldberg, D. E. (1989). Genetic Algorithms in Search and Optimization (Addison-wesley). [73] Goldberg, D. E., and Samtani, M. P. (1986). Engineering optimization via genetic algorithm. In Proceedings of the ninth Conference on Electronic Computation, 471-482. [74] Gopal, K., Aggarwal, K. K., and Gupta, J. S. (1978). An improved algorithm for reliability optimization. IEEE Transactions on Reliability 27, 325-328. 192 [75] Govil, K. K., and Agarwala, R. A. (1983). Lagrange multiplier method for optimal reliability allocation in a series system. Reliability Engineering 6, 181190. [76] Greenwood, G. W., Hu, X. S., and D'Ambrosio, J. G. (1997). Fitness functions for multiple objective optimization problems: Combining preferences with Pareto rankings. Foundations of genetic algorithms 4, 437. [77] Ha, C., and Kuo, W. (2006). Reliability redundancy allocation: An improved realization for nonconvex nonlinear programming problems. European Journal of Operational Research 171, 24-38. [78] Harrington, E. C. (1965). The desirability function. Industrial Quality Control 21, 494-498. [79] Heike, T., and Jorn, M. (2009). Preference-based Pareto optimization in certain and noisy environments. Engineering Optimization 41, 23-38. [80] Holland, J. H. (1987). Genetic algorithms and classifier systems: foundations and future directions (Michigan Univ., Ann Arbor (USA)). [81] Horn, J. (1997). Multicriterion decision making. Handbook of evolutionary computation (Oxford University Press). [82] Huang, H. Z. (1997). Fuzzy multi-objective optimization decision-making of reliability of series system. Microelectronics and Reliability 37, 447-449. [83] Huang, H. Z., Tian, Z., and Zuo, M. (2007). Intelligent Interactive Multiobjective Optimization of System Reliability. Computational intelligence in reliability engineering, 215-236. 193 [84] Huang, H. Z., Tian, Z., and Zuo, M. J. (2005). Intelligent interactive multiobjective optimization method and its application to reliability optimization. IIE Transactions 37, 983-993. [85] Huband, S., Hingston, P., Barone, L., and While, L. (2006). A review of multiobjective test problems and a scalable test problem toolkit. IEEE Transactions on Evolutionary Computation 10, 477-506. [86] Hughes, E. J. (2002). Multi-Objective Evolutionary Guidance for Swarms. In CEC'02: proceedings of the 2002 Congress on Evolutionary Computation: May 12-17, 2002, Hilton Hawaiian Village Hotel, Honolulu, Hawaii (IEEE), 2, 1127-1132. [87] Hwang, C. L., and Lai, Y. J. (1993). ISGP-II for multiobjective optimization with imprecise objective coefficients. Computers & Operations Research 20, 503-514. [88] Hwang, C. L., and Masud, A. S. M. (1979). Multiple objective decision making, methods and applications: a state-of-the-art survey (Springer). [89] Inagaki, T., Inoue, K., and Akashi, H. (2009). Interactive optimization of system reliability under multiple objectives. Reliability, IEEE Transactions on 27, 264-267. [90] Jakob, W., Gorges-Schleuter, M., and Blume, C. (1992). Application of genetic algorithms to task planning and learning. In Parallel Problem Solving from Nature, 2nd Workshop, Lecture Notes in Computer Science, 291-300. [91] Jeong, I. J., and Kim, K. J. (2009). An interactive desirability function method to multiresponse optimization. European Journal of Operational Research 195, 194 412-426. [92] Jimenez, F., Gomez-Skarmeta, A. F., Roubos, H., and Babuska, R. (2001). A multi-objective evolutionary algorithm for fuzzy modeling. In IFSA World Congress and 20th NAFIPS International Conference. [93] Jimenez, F., Sanchez, G., Cadenas, J. M., Gomez-Skarmeta, A. F., and Verdegay, J. L. (2004). A multi-objective evolutionary approach for nonlinear constrained optimization with fuzzy costs. In 2004 IEEE International Conference on Systems, Man and Cybernetics, 6 , 5771-5776. [94] Jimenez, F., Cadenas, J. M., Sanchez, G., Gomez-Skarmeta, A. F., and Verdegay, J. L. (2006). Multi-objective evolutionary computation and fuzzy optimization. International Journal of Approximate Reasoning 43, 59-75. [95] Jin, Y., and Sendhoff, B. (2002). Incorporation Of Fuzzy Preferences Into Evolutionary Multiobjective Optimization. In Proceedings of the Genetic and Evolutionary Computation Conference (Morgan Kaufmann Publishers Inc.), 683. [96] Jones, D., and Tamiz, M. (2010). Practical Goal Programming (Springer). [97] Kapur, P. K., and Verma, A. K. (2005). An Optimization of Integrated Reliability Model with Multiple Constraints. Quality, reliability and information technology: trends and future directions, 180. [98] Kennedy, J. (2006). Swarm intelligence. Handbook of Nature-Inspired and Innovative Computing, 187-219. [99] Kennedy, J., and Eberhart, R. (1995). Particle swarm optimization. In IEEE International Conference on Neural Networks, 1995. Proceedings., 1942-1948. 195 [100] Kim, K. J., and Lin, D. K. J. (2006). Optimization of multiple responses considering both location and dispersion effects. European Journal of Operational Research 169, 133-145. [101] Knowles, J., and Come, D. (1999). The pareto archived evolution strategy: A new baseline algorithm for pareto multiobjective optimisation. In Congress on Evolutionary Computation (CEC99 (Citeseer), 98-105. [102] Knowles, J., Come, D., and BioCentre, M. I. (2006). Evolutionary M ultiobjective Optimization. In (Loire Valley,France). [103] Krishnanand, K. N., and Ghose, D. (2006). Glowworm swarm based optimization algorithm for multimodal functions with collective robotics applications. Multiagent and Grid Systems 2, 209-222. [104] Krishnanand, K. N., and Ghose, D. (2009). Glowworm swarm optimisation: a new method for optimising multi-modal functions. International Journal of Computational Intelligence Studies 1, 93-119. [105] Kuo, W., and Prasad, V. R. (2000). An annotated overview of system-reliability optimization. IEEE Transactions on Reliability 49, 176-187. [106] Kuo, W., Prasad, V. R., Tillman, F. A., and Hwang, C. Optimal reliability design. 2001 (Cambridge University Press, Cambridge). [107] Kuo, W., and Wan, R. (2007). Recent advances in optimal reliability allocation. Computational Intelligence in Reliability Engineering, 1-36. [108] Kursawe, F. (1991). A variant of evolution strategies for vector optimization. Parallel Problem Solving from Nature, 193-197. 196 [109] Lam, S. W., and Tang, L. C. (2005). A graphical approach to the dual response robust design problems. In Reliability and Maintainability Symposium, 2005. Proceedings. Annual, 200-206. [110] Lamont, G. B., and Van Veldhuizen, D. A. (2007). Evolutionary Algorithms for Solving Multi-Objective Problems (Springer). [111] Laumanns, M., Thiele, L., and Zitzler, E. (2006). An efficient, adaptive parameter variation scheme for metaheuristics based on the epsilon-constraint method. European Journal of Operational Research 169, 932-942. [112] Lee, D., Jeong, I., and Kim, K. (2010). A posterior preference articulation approach to dual-response-surface optimization. IIE Transactions 42, 161-171. [113] Lee, M., and Park, J. (2003). More efficient consideration of dispersion effect by a probability-based desirability function in multiresponse problem (Ph.D. Thesis). [114] Levitin, G., and Amari, S. V. (2008). Multi-state systems with multi-fault coverage. Reliability Engineering & System Safety 93, 1730-1739. [115] Li, X. (2003). A non-dominated sorting particle swarm optimizer for multiobjective optimization. In Genetic and Evolutionary ComputationGECCO 2003 (Springer), 198-198. [116] Mahapatra, G. S., and Roy, T. K. (2006). Fuzzy multi-objective mathematical programming on reliability optimization model. Applied Mathematics and Computation 174, 643-659. [117] Mahapatra, G. (2009). Reliability Optimization of Entropy Based Series- 197 Parallel System Using Global Criterion Method. [118] Mandal, A., Johnson, K., Wu, C. F. J., and Bornemeier, D. (2007). Identifying promising compounds in drug discovery: Genetic Algorithms and some new statistical techniques. J. Chem. Inf. Model 47, 981-988. [119] Marler, R. T., and Arora, J. S. (2005). Function-transformation methods for multi-objective optimization. Engineering Optimization 37, 551-570. [120] Marler, R. T., and Arora, J. S. (2004). Survey of multi-objective optimization methods for engineering. Structural and multidisciplinary optimization 26, 369395. [121] Marler, R. T., Kim, C. H., and Arora, J. S. (2006). System identification of simplified crash models using multi-objective optimization. Computer Methods in Applied Mechanics and Engineering 195, 4383-4395. [122] Marseguerra, M., Zio, E., and Bosi, F. (2002). Direct Monte Carlo availability assessment of a nuclear safety system with time-dependent failure characteristics. Proceedings of MMR. [123] Marseguerra, M., Zio, E., and Martorell, S. (2006). Basics of genetic algorithms optimization for RAMS applications. Reliability Engineering & System Safety 91, 977-991. [124] Mavrotas, G. (2009). Effective implementation of the [epsilon]-constraint method in Multi-Objective Mathematical Programming problems. Applied Mathematics and Computation 213, 455-465. [125] Mehnen, H. T. J. (2009). Preference-based Pareto optimization in certain and noisy environments. Engineering Optimization 41.1-10. 198 [126] Mehnen, J., Wagner, T., Kersting, P., Tipura, I., and Rudolph, G. (2007). Evolutionary Five Axis Milling Path Optimization, GECCO'07, July 7-11 2007, London, UK, 2122-2128. [127] Merkuryeva, G. (2005). Response Surface-Based Simulation Metamodelling Methods. Supply Chain Optimisation, 205-215. [128] Miettinen, K. (1999). Nonlinear multiobjective optimization (Springer). [129] Misra, K. B. (1991). An algorithm to solve integer programming problems: An efficient tool for reliability design. Microelectronics Reliability 31, 285-294. [130] Misra, K. B. (2009). Reliability optimization of a series-parallel system. Reliability, IEEE Transactions on 21, 230-238. [131] Misra, K. B., and Sharma, U. (1991a). An efficient approach for multiple criteria redundancy optimization problems. Microelectronics Reliability 31, 303-321. [132] Misra, K. B., and Sharma, U. (1991b). Applications of a search algorithm to reliability design problems. Microelectronics Reliability 31, 295-301. [133] Misra, N., van der Meulen, E. C., and Vanden Branden, K. (2006). On estimating the scale parameter of the selected gamma population under the scale invariant squared error loss function. Journal of Computational and Applied Mathematics 186, 268-282. [134] Misra, R. B., and Agnihotri, G. (1979). Peculiarities in Optimal Redundancy for a Bridge Network. Reliability, IEEE Transactions on 28, 70-72. 199 [135] Mohamed Lawrence, M. (1992). Optimization techniques for system reliability: a review. Reliability Engineering & System Safety 35, 137-146. [136] Mohanty, B. K., and Vijayaraghavan, T. A. S. (1995). A multi-objective programming problem and its equivalent goal programming problem with appropriate priorities and aspiration levels: a fuzzy approach. Computers & Operations Research 22, 771-778. [137] Molina, J., Santana, L. V., Hernandez-Diaz, A. G., Coello Coello, C. A., and Caballero, R. (2009). g-dominance: Reference point based dominance for multiobjective metaheuristics. European Journal of Operational Research 197, 685-692. [138] Moore, J., and Chapman, R. (1999). Application of particle swarm to multiobjective optimization. Department of Computer Science and Software Engineering, Auburn University. [139] de Moura, L., Yamakami, A., and Bonfim, T. R. (2002). A genetic algorithm for multiobjective optimization problems with fuzzy constraints. In Second international workshop on Intelligent systems design and application (Dynamic Publishers, Inc.), 142. [140] Mukherjee, I., and Ray, P. K. (2008). Optimal process design of two-stage multiple responses grinding procesSes using desirability functions and metaheuristic technique. Applied Soft Computing 8, 402-421. [141] Murty, M. (1995). The Analytic Rank of JO (AO (Q). In Number theory: Fourth Conference of the Canadian Number Theory Association, July 2-8, 1994, Dalhousie University, Halifax, Nova Scotia, Canada (Canadian Mathematical Society), 263. 200 [142] Nguyen, H. H., Jang, N., and Choi, S. H. (2009). Multiresponse optimization based on the desirability function for a pervaporation process for producing anhydrous ethanol. Korean Journal of Chemical Engineering 26, 1-6. [143] Nguyen, N. T. (2010). Intelligent Information and Database Systems: Second International Conference, Aciids, Hue City, Vietnam, Mar. 24-26, 2010, Proceedings. [144] Noorossana, R., Davanloo Tajbakhsh, S., and Saghaei, A. (2009). An artificial neural network approach to multiple-response optimization. The International Journal of Advanced Manufacturing Technology 40, 1227-1238. [145] Padhye, N., Branke, J., and Mostaghim, S. (2009). Empirical comparison of MOPSO methods: guide selection and diversity preservation. In Proceedings of the Eleventh conference on Congress on Evolutionary Computation (Institute of Electrical and Electronics Engineers Inc., The), 2516-2523. [146] Pandey, M. K., Tiwari, M. K., and Zuo, M. J. (2007). Interactive enhanced particle swarm optimization: a multi-objective reliability application. Proceedings of the Institution of Mechanical Engineers, Part 0: Journal of Risk and Reliability 221, 177-191. [147] Pareto, V. (1896). Cours d'Economie Politique, volume I and II. F. Rouge, Lausanne 250. [148] Park, K. S., and Kim, K. J. (2005). Optimizing multi-response surface problems: How to use multi-objective optimization techniques. IIE Transactions 37, 523-532. [149] Parmee, I. C. (2001). Evolutionary and adaptive computing in engineering design (Springer Verlag). 201 [150] Parmee, I. C., Cvetkovic, D., Watson, A. H., and Bonham, C. R. (2000). Multiobjective satisfaction within an interactive evolutionary design environment. Evolutionary Computation 8, 197-222. [151] Parsopoulos, K. E., and Vrahatis, M. N. (2008). Multi-Objective Particles Swarm Optimization Approaches. Multi-objective optimization in computational intelligence: theory and practice. [152] Parsopoulos, K. E., and Vrahatis, M. N. (2002). Recent approaches to global optimization problems through particle swarm optimization. Natural Computing 1, 235-306. [153] Prasad, V. R, and Kuo, W. (2000). Reliability optimization of coherent systems. IEEE Transactions on Reliability 49, 323-330. [154] Rachmawati, L. (2009). Incorporation of human decision making preference into evolutionary multi-objective optimization (Ph.D. Thesis). [155] Rachmawati, L., and Srinivasan, D. (2009). Multiobjective evolutionary algorithm with controllable focus on the knees of the Pareto front. IEEE Transactions on Evolutionary Computation 13, 810-824. [156] Rachmawati, L., and Srinivasan, D. (2006). Preference incorporation in multiobjective evolutionary algorithms: A survey. In IEEE Congress on Evolutionary Computation, 2006. CEC 2006, 962-968. [157] Rangarajan, A., Ravindran, A. R., and Reed, P. (2004). An interactive multiobjective evolutionary optimization algorithm. In In Proceedings of the 34th International Conference on Computers & Industrial Engineering, 277-282. 202 [158] Rao, J. R., Tiwari, R. N., and Mohanty, B. K. (1988a). A method for finding numerical compensation for fuzzy multicriteria decision problem. Fuzzy Sets and Systems 25, 33-41. [159] Rao, J. R., Tiwari, R. N., and Mohanty, B. K. (1988b). A preference structure on aspiration levels in a goal programming problem--A fuzzy approach. Fuzzy sets and systems 25, 175-182. [160] Rao, S. S., and Dhingra, A. K. (1992). Reliability and redundancy apportionment using crisp and fuzzy multiobjective optimization approaches. Reliability Engineering & System Safety 37, 253-261. [161] Raquel, C. R., and Naval Jr, P. C. (2005). An effective use of crowding distance in multiobjective particle swarm optimization. In Proceedings of the 2005 conference on Genetic and evolutionary computation (ACM), 264. [162] Ravi, V. (2007). Modified great deluge algorithm versus other metaheuristics in reliability optimization. Intelligence in Reliability Engineering, 21-36. [163] Ravi, V., Murty, B. S. N., and Reddy, P. J. (1997). Nonequilibrium simulated annealing-algorithm applied to reliability optimization of complex systems. IEEE Transactions on Reliability 46, 233-239. [164] Ravi, V., Reddy, P. J., Zimmermann, H. J., fuer Untemehmensforschung, L., and Aachen, T. H.. (2000). Fuzzy global optimization of complex system reliability. IEEE Transactions on Fuzzy Systems 8, 241-248. [165] Ray, T., and Liew, K. M. (2002). A swarm metaphor for multiobjective design optimization. Engineering Optimization 34, 141-153. [166] Reyes-Sierra, M., and Coello, , C. A. C. (2006). Multi-objective particle swarm • 203 optimizers: A survey of the state-of-the-art. International Journal of Computational Intelligence Research 2, 287-308. [167] Ritzel, B. J., Wayland Eheart, J., and Ranjithan, S. (1994). Using genetic algorithms to solve a multiple objective groundwater pollution containment problem. Water Resources Research 30, 1589-1589. [168] Roy, R., and Mehnen, J. (2008). Technology Transfer: Academia To Industry. Evolutionary Computation in Practice, 263. [169] Sakawa, M. (1978). Multiobjective optimization by the surrogate worth tradeoff method. IEEE Transactions on Reliability 27, 311-314. [170] Sakawa, M., and Kato, K. (2009). An interactive fuzzy satisficing method for multiobjective nonlinear integer programming problems with block-angular structures through genetic algorithms with decomposition procedures. Fuzzy Sets and Systems, 81-99. [171] Salazar, A., Daniel, E., Rocco, S., and Claudio, M. (2007). Solving advanced multi-objective robust designs by means of multiple objective evolutionary algorithms (MOEA): A reliability application. Reliability Engineering & System Safety 92, 697-706. [172] Salazar, D., Rocco, C. M., and Galvan, B. J. (2006). Optimization of constrained multiple-objective reliability problems using evolutionary algorithms. Reliability Engineering & System Safety 91, 1057-1070. [173] Salhi, S., Drezner, T., and Drezner, Z. (2005). A Multi-Objective Heuristic Approach for the Casualty Collection Points Location Problem. [174] Salhi, S., and Petch, R. J. (2007). A GA based heuristic for the vehicle routing 204 problem with multiple trips. Journal of Mathematical Modelling and Algorithms 6, 591-613. [175] Sawaragi, Y., Nakayama, H., and Tanino, T. (1985). Theory of multiobjective optimization (Orlando Academic Press). [176] Schaffer, J. D. (1985). Multiple objective optimization with vector evaluated genetic algorithms. In Proceedings of the 1st International Conference on Genetic Algorithms (L. Erlbaum Associates Inc.), pp. 93-100. [177] Shelokar, P. S., Jayaraman, V. K., and Kulkarni, B. D. (2002). Ant algorithm for single and multiobjective reliability optimization problems. Quality and Reliability Engineering International 18, 497-514. [178] Singh, H., and Misra, N. (1994). On redundancy allocations in systems. Journal of Applied Probability 31, 1004-1014. [179] Sinha, N., Purkayastha, B. S., and Purkayastha, B. (2008). Optimal Combined Non-convex Economic and Emission Load Dispatch Using NSDE. ' In Conference on Computational Intelligence and Multimedia Applications, 2007. International Conference on (IEEE), 473-480. [180] Srinivas, N., and Deb, K. (1994). Muiltiobjective optimization using nondominated sorting in genetic algorithms. Evolutionary computation 2, 221248. [181] Steuer, D. (2004). Multi-criteria-optimisation and desirability indices. HT014602036. [182] Steuer, R. E. (1986). Multiple criteria optimization: Theory, computation, and application (Wiley, New York). 205 [183] Tan, K. C., Khor, E. F., and Lee, T. H. (2005). Multiobjective evolutionary algorithms and applications (Springer Verlag). [184] Tan, K. C., Lee, T. H., and Khor, E. F. (2002). Evolutionary algorithms for multi-objective optimization: performance assessments and comparisons. Artificial intelligence review. 17, 251-290. [185] Tan, K. C., Lee, T. H., and Khor, E. F. (1999). Evolutionary algorithms with goal and priority information for multi-objective optimization. In Proceedings of the 1999 Congress on Evolutionary Computation: CEC99: July 6-9, 1999, Mayflower Hotel, Washington, DC, USA (IEEE), 106. [186] Tanaka, M., Watanabe, H., Furukawa, Y., and Tanino, T. (1995). GA-based decision support system for multicriteria optimization. In IEEE International Conference on Systems, Man and Cybernetics, 1995. Intelligent Systems for the 21st Century. [187] Tang, L. C., and Paoli, P. (2004). A spreadsheet-based multiple criteria optimization framework for quality function deployment. International Journal of Quality and Reliability Management 21, 329. [188] Thanh, N. H., and Vong, N. T. (2000). Determination of cropping pattern by the multi-objective optimization model. Khoa Hoc Ky Thuat Nong Nghiep (Viet Nam); Journal of Agricultural Sciences and Technology. [189] Thiele, L., Miettinen, K., Korhonen, P. J., and Molina, J. (2009). A preferencebased evolutionary algorithm for multi-objective optimization. Evolutionary Computation 17, 411-436. [190] Tillman, F. A., Hwang, C. L., and Kuo, W. (1980). Optimization of systems 206 reliability (Marcel Dekker Inc). [191] Trautmann, H., and Mehnen, J. (2005). A method for including a-prioripreference in multicriteria optimization (University of Dortmund, Germany). [192] Trautmann, H., and Mehnen, J. (2009). Preference-based Pareto optimization in certain and noisy environments. Engineering Optimization 41, 23-38. [193] Trautmann, H., Wagner, T., Naujoks, B., Preuss, M., and Mehnen, J. (2009). Statistical methods for convergence detection of multi-objective evolutionary algorithms. Evolutionary Computation 17, 493-509. [194] Trautmann, H., and Weihs, C. (2006). On the distribution of the desirability index using Harrington's desirability function. Metrika 63, 207-213. [195] del Valle, Y., Venayagamoorthy, G. K., Mohagheghi, S., Hernandez, J., and Harley, R. G. (2008). Particle swarm optimization: Basic concepts, variants and applications in power systems. IEEE Transactions on Evolutionary Computation 12,171. [196] Van Veidhuizen, D. A., and Lamont, G. B. (1998). Multiobjective evolutionary algorithm research: A history and analysis. Air Force Inst. Technol., Dayton, OH, Tech. Rep. TR-98-03. [197] Van Veidhuizen, D. A., and Lamont, G. B. (2000). Multiobjective evolutionary algorithms: Analyzing the state-of-the-art. Evolutionary computation 8, 125148. [198] Verma, A. K., Ajit, S., and Karanki, D. R. (2010). Reliability and Safety Engineering (Springer). 207 [199] Vinod, G., Kushwaha, H. S., Verma, A. K., and Srividya, A. (2004). Optimisation of ISI interval using genetic algorithms for risk informed inservice inspection. Reliability Engineering & System Safety 86, 307-316. [200] Wang, H. F. (2000). FuZzy multicriteria decision makingan overview. Journal of Intelligent and Fuzzy Systems 9, 61-83. [201] White, C. C. (1984). A generalized model of sequential decisionmaking under risk. European Journal of Operational Research 18,19-26. [202] Wierzbicki, A. P. (1982). A mathematical basis for satisficing decision making. Mathematical modelling 3, 391-405. [203] Wierzbicki, A. P. (1999). Multicriteria decision making: advances in MCDM models, algorithms, theory, and applications (Kluwer Netherlands). [204] Wierzbicki, A. P. (1979). The Use of Reference Objectives in Multiobjective Optimization-Theoretical Implications and Practical Experience. Int. Inst. Applied System Analysis, Laxenburg, Austria, Working Paper WP-79-66. [205] Wilson, P. B., and Macleod, M. D. (1993). Low implementation cost IIR digital filter design using genetic algorithms. In IEE/IEEE Workshop on Natural Algorithms in Signal Processing, 4. [206] Yu, P. L., Lee, Y. R., and Stam, A. (1985). Multiple-criteria decision making: concepts, techniques, and extensions (Plenum Publishing Corporation). [207] Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoning. Information sciences 8, 199-249. [208] Zimmermann, H. J. (1990). Decision making in ill-structured environments and 208 with multiple criteria. Readings in multiple criteria decision aid, 119-151. [209] Zimmermann, H. J. (2001). Fuzzy set theory--and its applications (Springer Netherlands). [210] Zimmermann, H. J. (1987). Fuzzy sets, decision making, and expert systems (Springer). [211] Zimmermann, H. J. (1986). Multicriteria decision making in crisp and fuzzy environments. Fuzzy sets theory and applications. NATO ASI Series 177, 233256. [212] Zitzler, E., Brockhoff, D., and Thiele, L. (2007). The Hypervolume Indicator Revisited: On the Design of Pareto-compliant Indicators Via Weighted Integration. In In: Obayashi, S., Deb, K., Poloni, C., Hiroyasu, T., Murata, T. (eds.) EMO 2007 (Springer, Heidelberg), 862-876. [213] Zitzler, E., Deb, K., and Thiele, L. (2000). Comparison of multiobjective evolutionary algorithms: Empirical results. Evolutionary computation 8, 173195. [214] Zitzler, E., and Kanzli, S. (2004). Indicator-based selection in multiobjective search. In Parallel Problem Solving from Nature-PPSN VIII (Springer), 832842. [215] Zitzler, E., Laumanns, M., and Thiele, L. (2001). SPEA2: Improving the strength Pareto evolutionary algorithm. In EUROGEN (Citeseer), 95-100. [216] Zitzler, E., and Thiele, L. (1998). An evolutionary algorithm for multiobjective optimization: The strength pareto approach. Swiss Federal Institute of Technology, TIK-Report 43. 209 [217] Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C. M., and da Fonseca, V. G. (2003). Performance assessment of multiobjective optimizers: An analysis and review. IEEE Transactions on Evolutionary Computation 7, 117-132. 210