A comparative study of variation operators used for evolutionary multi-objective optimization

Abstract

In this paper, several variation operators based on Pareto efficiency, extracted from Differential 
Evolution, Estimation of Distribution Algorithms, Evolution Strategies and Evolutionary Programming, 
are compared in order to determine whether or not they increase the performance of the non-Pareto 
based versions. Firstly, we compare the selected variation operators in pairs, each operator with a 
modification of itself, in which we remove those elements related to the Pareto efficiency. Then, 
in a second experiment we compare among the selected operators, the variation operators used 
in the NSGA-II algorithm and the ones presented by the authors, PBCO and RBMO. In all the 
experiments the variation operators are incorporated in a well-known algorithm usually considered 
as a reference for making the comparisons, the NSGA-II algorithm. The experiments show that the 
Pareto based variation operators selected from the literature do not usually present a better behavior 
than their non-Pareto based versions; and none of them presents a better performance than the one 
reached by the variation operators defined by the authors, which were entirely built around the Pareto 
information of the individuals. These facts suggest that more effort should be placed in the design of 
variation operators devoted to multi-objective algorithms in order to achieve superior results to those 
obtained by means of general variation operators. (C) 2014 Elsevier Inc. All rights reserved.