Surrogate-based Multi-Objective Particle Swarm Optimization

Abstract

This paper presents a new algorithm that approximates real function evaluations using supervised learning with a surrogate method called support vector machine (SVM). We perform a comparative study among different leader selection schemes in a multi-objective particle swarm optimizer (MOPSO), in order to determine the most appropriate approach to be adopted for solving the sort of problems of our interest. The resulting hybrid presents a poor spread of solutions, which motivates the introduction of a second phase to our algorithm, in which an approach called rough sets is adopted in order to improve the spread of solutions along the Pareto front. Rough sets are used as a local search engine, which is able to generate solutions in the neighborhood of the nondominated solutions previously generated by the surrogate-based algorithm. The resulting approach is able to generate reasonably good approximations of the Pareto front of problems of up to 30 decision variables with only 2,000 fitness function evaluations. Our results are compared with respect to the NSGA-II, which is a multi-objective evolutionary algorithm representative of the state-of-the-art in the area.