A Study About Meta-Optimizing the NSGA-II Multi-Objective Evolutionary Algorithm.

Abstract

The automatic design of multi-objective metaheuristics is an active research line aimed at, given a set of problems used as training set, to find the configuration of a multi-objective optimizer able of solving them efficiently. The expected outcome is that the auto-configured algorithm can be used of find accurate Pareto front approximations for other problems. In this paper, we conduct a study on the meta-optimization of the wellknown NSGA-II algorithm, i.e., we intend to use NSGA-II as an automatic configuration tool to find configurations of NSGA-II. This search can be formulated as a multi-objective problem where the decision variables are the NSGA-II components and parameters and the the objectives are quality indicators that have to be minimized. To develop this study, we rely on the jMetal framework. The analysis we propose is aimed at answering the following research questions: RQ1 - how complex is to build the meta-optimization package?, and RQ2 - can accurate configurations be found? We conduct an experimentation to give an answer to these questions.