Batch Bayesian optimization with adaptive batch acquisition functions via multi-objective optimization

Jixiang Chen, Fu Luo, Genghui Li, Zhenkun Wang

Abstract: Bayesian optimization (BO) is a powerful method for solving expensive black-box optimization problems, and it determines the candidate solutions for expensive evaluation via optimizing the acquisition function. Benefiting from the development of the hardware resources, batch Bayesian optimization (BBO) approaches, characterized by selecting a batch of solutions for expensive evaluation in each iteration, have attracted more and more attention. However, existing BBO methods, using a single or fixed combination of acquisition functions, suffer from lousy flexibility and low robustness when facing various problems. To deal with these issues, this paper proposes a BBO method with adaptive batch acquisition functions via multi-objective optimization (called BBO-ABAFMo). Specifically, multiple acquisition functions are adaptively chosen to form a multi-objective optimization problem (MOP). Its Pareto-optimal solutions provide the candidate solutions for expensive evaluation according to a minimum-diverse-exploitative (MDE) strategy. The experimental results show the advantages of the proposed BBO-ABAFMo over some state-of-the-art methods.