Automatic Design of Multi-objective Particle Swarm Optimizers

Multi-objective particle swarm optimizers (MOPSOs) have been widely used to deal with optimization problems having two or more conflicting objectives. As happens with other metaheuristics, finding the most adequate parameters settings for MOPSOs is not a trivial task, and it is even harder to choose structural components that determine the algorithm’s design. Thus, it is an open question whether automatically-designed MOPSOs can outperform the best human-designed MOPSOs from the literature. In this paper, we first design and develop a component-based architecture and an algorithmic template, called AMOPSO, for the auto-design and auto-configuration of MOPSOs using jMetal and we integrate it with irace, an automatic-configuration tool. Second, by taking as our starting point two algorithms (OMOPSO and SMPSO), we conduct a study focused on automatically generating three AMOPSO variants by using different well-known multi-objective benchmarking problem families (ZDT, DTLZ, and WFG) as training problems for automatic design, and then we analyze whether they improve upon the initial versions of the algorithms and how their components differ. Experiments show that the two AMOPSO variants obtained from using, respectively, the ZDT and DTLZ problems for training are able to statistically outperform the SMPSO and OMOPSO algorithms in all three benchmark families previously indicated.