Truncation-learning-driven surrogate assisted social learning particle swarm optimization for computationally expensive problem

Abstract

Surrogate-assisted evolutionary optimization greatly slashes the computational burden of evolutionary algorithms for computationally expensive problems. However, new issues arise concerning the compatibility and fault tolerance of surrogates, evolutionary learning operators, and problem property. To this end, this paper proposes a truncation-learning-driven surrogate assisted social learning particle swarm optimizer (TL-SSLPSO) to coordinate these three ingredients. For avoiding and correcting the deceptions induced by the low confidence exemplars due to the surrogate in behavior learning, TL-SSLPSO equally segments the iterative population into multiple sub-populations with different fitness levels and selects exemplars from the randomly selected high-level sub-populations for the behavior learning of low-level sub-population, while truncating the behavior learning of the highest-level sub-population composed of some of the best approximated or real evaluated particles and retaining the sub-population directly to the next generation. Besides, a greedy sampling strategy is employed to find promising solutions with better fitness versus the global best to complement the truncation learning. Extensive experiments on twenty-four widely used benchmark problems and a stepped cantilever beam design problem with 17 steps are conducted to assess the effectiveness of cooperation between truncation learning and greedy sampling and comparisons with several state-of-the-art algorithms demonstrate the superiority of the proposed method.