CISE Seminar: Xuezhou Zhang, Boston University
Date: Friday, February 2, 2024
Time: 3:00pm – 4:00pm
Location: 8 Saint Mary’s Street, PHO 203

Xuezhou Zhang
Assistant Professor
Boston University
Representation Learning for Efficient RL
Reinforcement Learning (RL) has been considered a promising paradigm to solve decision making tasks. However, over the past years, RL has only found limited successes in applications with high quality data or a near-perfect simulator at disposal. The main drawback of existing RL algorithms is their poor sample complexity, i.e. it takes too many trials and errors to learn a good policy. In this talk, I will discuss recent attempts to solve this problem through the paradigm of representation learning, which aims to learn low-dimensional embedding of the observation and perform efficient RL in the embedding space instead of the raw observation space. This approach has achieved up to 1000x improvement in sample complexity over existing methods on the Atari Game benchmarks.
Faculty Host: Alex Olshevsky
Student Host: Mehdi Kermanshah