[2D6-GS-2] Machine learning: Bayesian estimation

[2D6-GS-2-04] Risk-averse Heteroscedastic Bayesian Optimization Using Monte Carlo Acquisition Functions

〇Takehiro Sano¹, Kazuma Nakagawa¹, Takashi Ikeuchi² (1. SCREEN Advanced System Solutions Co., Ltd., 2. SCREEN Holdings Co., Ltd.)

Keywords:heteroscedastic Bayesian optimization, Monte Carlo acquisition function, input-dependent noise

Risk-averse heteroscedastic Bayesian optimization (RAHBO), which optimizes a black-box function while minimizing input-dependent noise based on a mean-variance objective, is a practical approach for various black-box optimization tasks. RAHBO can obtain a solution with a desired function value and low noise variance via heteroscedastic acquisition functions defined by two predictive distributions from a heteroscedastic Gaussian process. However, existing methods are incapable of parallel optimization and cannot save the time required for black-box optimization tasks. In this paper, we propose a novel parallel method for RAHBO. In other words, constructing a predictive distribution corresponding to the mean-variance objective enables us to employ Monte Carlo acquisition functions. We also demonstrate experimentally that the performance of the proposed parallel method is comparable to that of conventional non-parallel methods.