Adaptive sampling Bayesian algorithm for constrained black-box optimization problems

Constrained black-box optimization (CBBO) has become increasingly popular in process optimization. Algorithms often encounter difficulties in balancing feasibility and optimality, with some even failing to find feasible solutions. This article introduces an adaptive sampling Bayesian optimization algorithm (ASBO) to solve CBBO problems effectively. The developed infill sampling criterion introduces an adaptive acquisition function to facilitate multistage optimization. The three stages consist of exploring feasible solutions, balancing feasibility and optimality, and optimizing. Furthermore, a hybrid method is proposed for complex problems. A gradient-based optimizer (GBO) aids in constructing the posterior distribution, thereby enhancing the identification of feasible regions. Additionally, four auxiliary strategies are developed to enhance stability and accelerate convergence in simulation-based optimization. The effectiveness of the proposed algorithm is validated through three benchmark problems and two process optimization cases. Comparative analysis against state-of-the-art algorithms demonstrates better iteration efficiency of ASBO algorithms.