Sample efficient reinforcement learning with active learning for molecular design

Abstract

Active learning accelerates the design of molecules during generative reinforcement learning by creating surrogate models of expensive reward functions, obtaining a 4- to 64-fold reduction in computational effort per hit.