A Distance Metric for Evolutionary Many-Objective Optimization Algorithms Using User-Preferences

Abstract

In this paper we propose to use a distance metric based on user-preferences to efficiently find solutions for many-objective problems. In a user-preference based algorithm a decision maker indicates regions of the objective-space of interest, the algorithm then concentrates only on those regions to find solutions. Existing user-preference based evolutionary many-objective algorithms rely on the use of dominance comparisons to explore the search-space. Unfortunately, this is ineffective and computationally expensive for many-objective problems. The proposed distance metric allows an evolutionary many-objective algorithm’s search to be focused on the preferred regions, saving substantial computational cost. We demonstrate how to incorporate the proposed distance metric with a user-preference based genetic algorithm, which implements the reference point and light beam search methods. Experimental results suggest that the distance metric based algorithm is effective and efficient, especially for difficult many-objective problems.