A Scaled Prediction Information Product Criterion for Bayesian Optimization

Bayesian optimization (BO) is an efficient method for solving expensive black-box optimization problems, but the existing BO could be more efficient and have higher optimal accuracy to improve the greed characteristic of infill criteria in the optimization stage. Considering the greedy characteristics of the Bayesian infill criterion, a novel scaled prediction information product (SPIP) criterion is proposed. In the optimization process, the Gaussian process (GP) model provides the predicted information, and the SPIP criterion conducts the surrogate model management and sample updating, which guarantees the efficient BO process continues until terminated together. Experimental results of six benchmark functions and the aircraft wing design problem show that the proposed method can improve optimization efficiency and reduce the expensive simulation cost while enhancing optimization accuracy.