An adaptive model selection strategy for surrogate-assisted particle swarm optimization algorithm
Haibo Yu, Yin Tan, Chaoli Sun, Jianchao Zeng, Yaochu Jin · 2016
Computationally expensive problems pose a serious challenge to the successful application of evolutionary algorithms to complex engineering optimization. To address this challenge, surrogate models, also known as metamodels, are commonly used in lieu of the expensive fitness function for the computational cost of optimization. However, it is nontrivial to choose an appropriate metamodel for properly replacing the expensive fitness function. In this paper, an adaptive model selection strategy based on fitness landscape analysis is proposed for a surrogate-assisted particle swarm optimization algorithm. The structure of the sampling space is learned, based on which the more suited surrogate model, either a polynomial regression model or a radial basis function network, will be chosen to estimate the fitness value. Simulation results on seven widely used benchmark functions demonstrate the efficacy of the proposed algorithm.