Diagnosing and fixing common problems in Bayesian optimization for molecule design

TL;DR

This paper identifies common pitfalls in Bayesian optimization for molecule design, proposes solutions, and demonstrates that addressing these issues significantly improves performance on a standard benchmark, highlighting the potential of BO in molecular machine learning.

Contribution

The paper uncovers three key pitfalls in Bayesian optimization for molecule design and shows how fixing them enhances performance on the PMO benchmark.

Findings

Addressing prior width, over-smoothing, and acquisition maximization improves BO performance.

Basic BO setup can outperform existing methods when pitfalls are fixed.

BO shows promise for molecular design with proper tuning.

Abstract

Bayesian optimization (BO) is a principled approach to molecular design tasks. In this paper we explain three pitfalls of BO which can cause poor empirical performance: an incorrect prior width, over-smoothing, and inadequate acquisition function maximization. We show that with these issues addressed, even a basic BO setup is able to achieve the highest overall performance on the PMO benchmark for molecule design (Gao et al 2022). These results suggest that BO may benefit from more attention in the machine learning for molecules community.