A Robust Surrogate-assisted Evolutionary Algorithm based on Maximum Correntropy Criterion⋆

By remarkably reducing real fitness evaluations, surrogate-assisted evolutionary algorithms (SAEAs) have been successfully applied to expensive optimization problems. However, existing SAEAs generally ignore the widespread simulation evaluation noise when constructing surrogate models, which severely limits their robustness and applications. To alleviate this issue, this study proposes a robust SAEA based on maximum correntropy criterion (MCC). MCC can robustly measure the similarity between two random variables by weakening the negative influence of outlier data. With it as the loss function, the trained surrogate model thus could have a good tolerance of the simulation evaluation noise. Taking the radial basis function (RBF) as the basic surrogate model and the differential evolution (DE) algorithm as the optimizer, this study then develops a specific SAEA named MCC-RBF-DE. Comprehensive experimental results on various benchmark functions with evaluation noise show that the introduction of MCC can effectively suppress the influence of noise. Moreover, MCC-RBF-DE shows stronger robustness compared to traditional SAEAs.