Basic Bayesian Optimization is Underrated for Molecule Design

Recently, the most popular algorithms for molecule design are based on techniques like reinforcement learning, which explore randomly. This paper asks whether Bayesian optimization (BO), a technique which explores deliberately, can do better. We explain some common pitfalls of BO and how to address them. Notably, we show that a correctly-tuned basic BO setup is able to achieve the highest overall performance on the PMO benchmark for molecule design (Gae et al, 2022). We end with a discussion of the strengths and limitations of BO, arguing why it is overall underrated by the ML community.