Space-filling experimental design for efficient Bayesian optimization
Shigeru Kinoshita, Yuta Inoue, Tetsuro Watanabe, Kosuke Ikeda, Seiwa Nishio, Atsushi Teruya, Naoki Sakai, Takashi Goda · 2024
In this study, we propose a space-filling Latin hypercube design with small fill distance and large separation radius based on simulated annealing for efficient Bayesian optimization. Through an example focused on optimizing high aspect ratio hole etching, we demonstrate that initial sampling with the proposed design is advantageous compared to conventional designs in terms of the convergence of Bayesian optimization with fewer experiments.