Surrogate-assisted PSO with archive-based neighborhood search for medium-dimensional expensive multi-objective problems

Thousands of real function evaluations are not burdensome when a surrogate-assisted evolutionary algorithm (SAEA) is used to solve expensive multi-objective optimization problems (MOPs). To reduce the computational overhead, this paper studies a surrogate-assisted multi-objective particle swarm optimization algorithm, named SaMOPSO_NS, in which an external archive-based neighborhood search, as a local search, and a pbest-dominance-based infill criterion are newly developed. The local search works hard to refine exploitation around the current non-dominated individuals once the trigger mechanism is activated, while the pbest-dominance-based infill criterion chooses non-dominated individuals predicted by an ensemble surrogate for actual evaluations. With the collaborative efforts between the local search strategy and the infill criterion, computing resources are more efficiently allocated. Three types of benchmark test instances with different dimensions as well as an engineering expensive MOP are used to examine the proposed algorithm. Experimental results demonstrate that the proposed algorithm significantly outperforms its rivals with fewer real evaluations on most medium-dimensional MOPs. Moreover, the optimized electromagnetic acoustic transducers achieved an amplitude and amplitude ratio of 9.666E-07 mm and 0.1328 respectively, markedly outpacing previously reported results.