Multi-fidelity Bayesian optimization based on optimal-value entropy

Abstract

Bayesian optimization is an effective approach for an expensive black-box function optimization problem. Bayesian optimization aims for an efficient optimization with a fewer number of function evaluations. On the other hand, for example, although a simulated physical value is optimized in a material development using numerical simulation, the numerical accuracy and the computational cost of the simulation often have a trade-off relationship. Multi-fidelity Bayesian optimization aims for cost-efficient optimization using such multi-fidelity information sources. In this paper, we propose multi-fidelity Bayesian optimization based on optimal-value entropy, which does not require a hyperparameter and can be computed efficiently. Furthermore, we show the extensions for parallel queryings. Finally, we demonstrate the effectiveness of the proposed methods via numerical experiments.