Surrogate-assisted PSOs are one of the most popular black-box optimizers for computationally expensive optimization problems, but those employ approximation models less scalable for the increase of the problem dimension on a restricted number of fitness evaluations. This paper proposes a hybrid surrogate-assisted PSO (HyPSO), which utilizes approximation and classification surrogate models for computationally expensive optimization problems. A basic idea of HyPSO is in the hybridization of surrogate model types, although existing works only consider the approximation model compatible with the PSO framework. HyPSO intends to manage a trade-off between approximation/classification models in terms of the model accuracy and the screening capacity; this contributes to hedge the risk of the over-fitting issue in building surrogates under a restriction of fitness evaluations. In particular, HyPSO constructs an approximation model with Radial Basis Function (RBF) and a classification model with Support Vector Machine (SVM). Then, it estimates a global best solution and a personal-best solution of a particle with an RBF model and an SVM model, respectively. Experimental results show that HyPSO significantly outperforms an alternative approach, i.e., OPUS and the standard PSO on a set of single-objective benchmark functions. Especially, HyPSO has a good scalability against the increase of the problem dimension.
View full abstract