The influence of using Adaptive Operator Selection in a Multiobjective Evolutionary Algorithm Based on Decomposition

Abstract

Recently, a new algorithm has arisen for solving Multiobjective Optimization Problems (MOPs), called Multiobjective Evolutionary Algorithms based on Decomposition (MOEA/D). This algorithm decom poses a MOP into a number of single optimization subproblems and optimizes them simultaneously using an evolutionary process. Adaptive Operator Selection (AOS) mechanisms are added to MOEA/D to improve this evolutionary process by choosing the right mutation operators from a pool to be applied at the reproduction phase. This thesis provides an overview of existing MOEAs, AOS mechanisms and the combination of these two. Also, a new AOS mechanism is proposed. It combines a ?tness-rate-rank based (FRR) credit assignment from existing research with a probability based operator selection mechanism that uses tournament selection. Unlike existing AOS methods using FRR, the selection probabilities of the operators are solely based on the relative order of the estimated rewards. The purpose of this is to improve the balance between the exploration and exploitation of the operators, and therefore improve the performance. This AOS is then used within the MOEA/D framework to form a new algorithm: MOEA/D-FRR-TS. Several experiments were conducted to test the performance of this algorithm. Results of these experiments show that there is no signi?cant difference between MOEA/D-FRR-TS and other probability based AOS + MOEA/D combinations. Furthermore, it is shown that MOEA/D-FRR-TS using a pool of operators can improve MOEA/D that uses a single operator.