posted on 2022-06-16, 19:13authored byAlexis
W. Mills, Joshua J. Goings, David Beck, Chao Yang, Xiaosong Li
Reinforcement
machine learning is implemented to survey a series
of model potential energy surfaces and ultimately identify the global
minima point. Through sophisticated reward function design, the introduction
of an optimizing target, and incorporating physically motivated actions,
the reinforcement learning agent is capable of demonstrating advanced
decision making. These improved actions allow the agent to successfully
converge to an optimal solution more rapidly when compared to an agent
trained without the aforementioned modifications. This study showcases
the conceptual feasibility of using reinforcement machine learning
to solve difficult environments, namely, potential energy surfaces,
with multiple, seemingly, correct solutions in the form of local minima
regions. Through these results, we hope to encourage extending reinforcement
learning to more complicated optimization problems and using these
novel techniques to efficiently solve traditionally challenging problems
in chemistry.