[2J1-GS-2] Machine learning: Gaussian process model

[2J1-GS-2-02] Bayesian Optimization Based on Meta Learning with Neural Process

〇Makoto Kawano¹, Wataru Kumagai², Kota Matsui², Yusuke Iwasawa¹, Yutaka Matuso¹ (1. The University of Tokyo, 2. RIKEN AIP)

Keywords:Bayesian Optimization, Neural Process, Deep Learning

Bayesian optimization is a technique that optimizes the black box function based on a probabilistic model with a few observation points as possible. In this study, we consider Bayesian optimization in a situation where similar functions other than the target function to be evaluated can be accessed at a low cost. In this paper, we propose BONP using neural processes (NPs), a deep generation model considering the uncertainty of prediction, as a surrogate model. Although NPs can be used for meta-learning, it often ignores given observations and causes under-fitting. To avoid this issue, we also propose a new Dot-CNP that maps observation points to function space and apply it to BONP. In experiments, we dealt with the regression problem with the 1d-synthetic function and the Bayesian optimization problem with the three types of acquisition functions, and demonstrated the effectiveness of the proposed method.