Analysis on population size and neighborhood recombination on many-objective optimization

Abstract

This work analyzes population size and neighborhood recombination in the context of many-objective optimization. Large populations might support better the evolutionary search to deal with the increased complexity inherent to high dimensional spaces, whereas neighborhood recombination can reduce dissimilarity between crossing individuals and would allow us to understand better the implications of a large number of solutions that are Pareto-optimal from the perspective of decision space and the operator of variation. Our aim is to understand why and how they improve the effectiveness of a dominance-based many-objective optimizer. To do that, we vary population size and analyze in detail convergence, front distribution, the distance between individuals that undergo crossover, and the distribution of solutions in objective space. We use DTLZ2 problem with m = 5 objectives in our study, revealing important properties of large populations, recombination in general, and neighborhood recombination in particular, related to convergence and distribution of solutions. © 2012 Springer-Verlag.