Bayesian Optimization with Setup Switching Cost

Abstract

Bayesian Optimization (BO) in its classical form is cost-unaware. However, many real-world problems are resource-constrained and hence incur a cost whenever such resources are needed, such as when a new setup is used. We are then looking at adapted cost-aware solution methods that are improving the performance of BO over cost-constrained problems. We find that parameter-free algorithms can yield comparable results to fine-tuned algorithms used in constrained optimization