Project Grant 2410981

This $569,138 Project Grant awarded by the National Science Foundation's (NSF) Computer and Information Science and Engineering (CISE) Federal Grant Program (CFDA 47.070) supports research to develop new methods for active testing of autonomous decision-making in reinforcement learning agents. The project aims to identify the most informative situations to test and focus data collection on such consequential data, in order to produce higher confidence evaluations of AI systems before real-world deployment. Specifically, the research consists of three thrusts: deriving an optimal policy for evaluating a given policy, developing novel adaptive sampling algorithms to reduce inefficiency in data collection for policy evaluation, and creating a methodology to adaptively set a policy's initial state to most accurately evaluate it. These new methods are expected to advance foundational reinforcement learning knowledge and enable effective policy evaluation in realistic domains. The award was granted to the University of Wisconsin System's University of Wisconsin - Madison division, with a performance period from September 1, 2024 to August 31, 2027.