Effects of the Use of Scalar Fitness Functions in Evolutionary Many-Objective Optimization

Abstract

Recently, a powerful algorithm called MOEA/D (Multiobjective Evolutionary Algorithm Based on Decomposition) has been proposed for multiobjective optimization. This algorithm is based on a number of scalar fitness functions with uniformly distributed weight vectors. In this paper, we examine the effect of scalar fitness functions on the scalability of MOEA/D to many-objective problems. Experimental results show advantages and disadvantages of MOEA/D over Pareto-based algorithms such as NSGA-II. Based on experimental results, we suggest a modified version of MOEA/D to improve its scalability.