Research talk: Successor feature sets: Generalizing successor representations across policies

Опубликовано 25 января 2022, 1:40

Speaker: Kianté Brantley, PhD Student, University of Maryland

Successor-style representations have many advantages for reinforcement learning. For example, they can help an agent generalize from experience to new goals. However, successor-style representations are not optimized to generalize across policies—typically, a limited-length list of policies is maintained and information shared among them by representation learning or generalized policy iteration. Join University of Maryland PhD candidate Kianté Brantley to address these limitations in successor-style representations. With collaborators from Microsoft Research Montréal, he developed a new general successor-style representation, which brings together ideas from predictive state representations, belief space value iteration, and convex analysis. The new representation is highly expressive. For example, it allows for efficiently reading off an optimal policy for a new reward function or a policy that imitates a demonstration. Together, you’ll explore the basics of successor-style representation, the challenges of current approaches, and results of the proposed approach on small, known environments.

Learn more about the 2021 Microsoft Research Summit: Aka.ms/researchsummit