High-Dimensional Bayesian Optimization via Tree-Structured Additive Models

Bayesian Optimization (BO) has shown significant success in tackling expensive low-dimensional black-box optimization problems. Many problems of interest are high-dimensional, and scaling BO to such settings remains an important challenge. In this paper, we consider generalized additive models which functions with overlapping subsets variables composed model a high-dimensional target function. Our goal is lower the computational resources required facilitate faster learning by reducing complexity while retaining sample-efficiency existing methods. Specifically, constrain underlying dependency graphs tree structures order both structure acquisition For former, propose hybrid graph algorithm based on Gibbs sampling mutation. addition, novel zooming-based that permits be employed more efficiently case continuous domains. We demonstrate discuss efficacy our approach via range experiments synthetic real-world datasets.