A parallel island model for hypervolume-based many-objective optimization

Abstract

Parallelism arises as an attractive option when Multi-Objective Evolutionary Algorithms (MOEAs) demand an intensive use of CPU or memory. The computational complexity of a MOEA depends on the scalability of its input parameters (i.e., the population size, number of decision variables, objectives, etc.) and on the computational cost of evaluating the objectives of the problem. Nonetheless, current research efforts have focused only on the second case. Therefore, in this chapter, we investigate the performance and behavior of S-PAMICRO, a recently proposed parallelization of SMS-EMOA that inhibits exponential execution time as the number of objectives increases. The idea behind S-PAMICRO is to divide the overall population into several semi-independent subpopulations each of which has very few individuals. Each subpopulation evolves a serial SMS-EMOA with an external archive for maintaining diversity. Our experimental results show that S-PAMICRO outperforms the standard island version of some state-of-the-art MOEAs in most instances of the many-objective Deb-Thiele-Laumanns-Zitzler (DTLZ) and Walking-Fish-Group (WFG) test suites.