Batched Bayesian Optimization by Maximizing the Probability of Including the Optimum.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Batched Bayesian optimization (BO) accelerates molecular design by efficiently screening large chemical libraries.
• Current acquisition strategies balance exploration and exploitation, often requiring approximations for complex batch functions.

Purpose of the Study:

• To develop a novel acquisition strategy for discrete optimization in batched BO.
• To improve the efficiency and effectiveness of identifying top-performing compounds.

Main Methods:

• Proposed qPO (multipoint Probability of Optimality), an exploitation-motivated acquisition strategy.
• Expressed batch optimality as a sum of individual acquisition scores, simplifying optimization.
• Differentiated qPO from parallel Thompson sampling and analyzed its implicit diversity.

Main Results:

• qPO circumvents the combinatorial challenges of optimizing batch acquisition functions.
• Empirical evidence shows qPO is competitive with and complements state-of-the-art batched BO methods.
• Demonstrated successful application in model-guided exploration of large chemical libraries.

Conclusions:

• qPO offers an effective approach for pure exploitation in discrete batched Bayesian optimization.
• The strategy enhances molecular design by efficiently identifying optimal compounds.
• qPO provides a valuable alternative and complement to existing batched BO acquisition methods.