Bayesian Optimization with Fixed Constraints using Acceptance Functions
Gabriel D. Uribe-Guerra, Danny Múnera, Julián D. Arias-Londoño · 2022
Bayesian optimization (BO) is a statistical approach that allows efficient modeling and optimization of black-box models. Most approaches using BO are focused on unconstrained context. In the case of BO with fixed constraints, the default solution is to include them during the optimization of the acquisition function. However, several alternatives could be used for the same purpose. In this work, three schemes for BO with constraints are compared using two synthetic objective functions combined with linear and nonlinear constraints. The results suggest that an approach based on including an acceptance function into the acquisition function formulation obtains the best results.