An Investigation on Preference Order Ranking Scheme for Multiobjective Evolutionary Optimization

It may be generalized that all Evolutionary Algorithms (EA) draw their strength from two sources: exploration and exploitation. Surprisingly, within the context of multiobjective (MO) optimization, the impact of fitness assignment on the exploration-exploitation balance has drawn little attention. The vast majority of multiobjective evolutionary algorithms (MOEAs) presented to date resort to Pareto dominance classification as a fitness assignment methodology. However, the proportion of Pareto optimal elements in a set P grows with the dimensionality of P. Therefore, when the number of objectives of a multiobjective problem (MOP) is large, Pareto dominance-based ranking procedures become ineffective in sorting out the quality of solutions. This paper investigates the potential of using preference order-based approach as an optimality criterion in the ranking stage of MOEAs. A ranking procedure that exploits the definition of preference ordering (PO) is proposed, along with two strategies that make different use of the conditions of efficiency provided, and it is compared with a more traditional Pareto dominance-based ranking scheme within the framework of NSGA-II. A series of extensive experiments is performed on seven widely applied test functions, namely, DTLZ1, DTLZ2, DTLZ3, DTLZ4, DTLZ5, DTLZ6, and DTLZ7, for up to eight objectives. The results are analyzed through a suite of five performance metrics and indicate that the ranking procedure based on PO enables NSGA-II to achieve better scalability properties compared with the standard ranking scheme and suggest that the proposed methodology could be successfully extended to other MOEAs

I. Introduction

It May Be generalized that all evolutionary algorithms (EAs) draw their strength from two sources: exploration and exploitation. Genetic algorithms (GAs) are no exception. When exploration and exploitation are unbalanced, either the search stalls around unpromising areas of the objective space or it prematurely converges to some local optima.