Fri, Dec 03, 2021 @ 01:00 PM - 02:30 PM
Thomas Lord Department of Computer Science
Time: 1:00-2:30pm, December 3rd
Committee: Haipeng Luo (host), Rahul Jain, David Kempe, Vatsal Sharan, Jiapeng Zhang
Title: Robust and Adaptive Online Reinforcement Learning
Abstract: Online reinforcement learning (RL) studies how an agent learns to behave in an unknown environment from scratch. In this thesis, I focus on the theoretical foundations of this learning paradigm, with emphasis on designing algorithms that are robust to the non-stationarity of the environment, where the non-stationarity may come from natural drift, adversarial manipulation, or the existence of other agents. While being robust, most of our algorithms are also \"adaptive\" at the same time in the sense that they do not sacrifice nice performance guarantees if the environment happens to be stationary. More broadly speaking, the performance of our algorithms automatically scale with some intrinsic properties that reflect the difficulty of the problem.
For future work, I plan to characterize the fundamental limit of RL in large state space, a central topic in theoretical RL. We hope to answer the following questions: \"what are the minimal assumptions to be made so that RL algorithms can find near-optimal policies with polynomial number of samples\", and the similar question under the restriction of \"polynomial computational time\".
Audiences: Everyone Is Invited
Contact: Lizsl De Leon