On the automatic design of multi-objective particle swarm optimizers: experimentation and analysis

Research in multi-objective particle swarm optimizers (MOPSOs) progresses by proposing one new MOPSO at a time. In spite of the commonalities among different MOPSOs, it is often unclear which algorithmic components are crucial for explaining the performance of a particular MOPSO design. Moreover, it is expected that different designs may perform best on different problem families and identifying a best overall MOPSO is a challenging task. We tackle this challenge here by: (1) proposing AutoMOPSO, a flexible algorithmic template for designing MOPSOs with a design space that can instantiate thousands of potential MOPSOs; and (2) searching for good-performing MOPSO designs given a family of training problems by means of an automatic configuration tool (irace). We apply this automatic design methodology to generate a MOPSO that significantly outperforms two state-of-the-art MOPSOs on four well-known bi-objective problem families. We also identify the key design choices and parameters of the winning MOPSO by means of ablation. AutoMOPSO is publicly available as part of the jMetal framework.