Cohere For AI - Jacob Adamczyk, PhD student at University of Mass.

Jacob will present three recent papers from his research group exploring innovative approaches to reward shaping in reinforcement learning. Building on the foundational work of "Policy invariance under reward transformations" by Ng, Harada, and Russell, He'll show how fresh perspectives on classic reward shaping theory can lead to new analytical results in RL. After generalizing this foundational work, we will discuss two ideas for bootstrapping inspired by these findings. Specifically, He'll show how their results enable the adaptive construction of (1) tight value function bounds and (2) potential functions for reward shaping, without requiring prior knowledge. Experimental results will be presented to illustrate the practical benefits of these methods, and He'll conclude by discussing opportunities for future work in this domain.


Bio: Jacob is an Applied Physics PhD student at the University of Massachusetts Boston, focusing on the intersection of statistical mechanics and reinforcement learning. He is interested in advancing deep RL from several directions, typically through online value-based methods, with results on: reward shaping, value function bounds, transfer learning, average-reward RL, and applications to biological and quantum control. In March 2025, He'll be joining Sony AI's deep RL team as an intern, where he'll be working on the average-reward problem for continuing tasks in high-dimensional video games.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo