RL at Harvard

Uniting reinforcement learning researchers to understand

how animals learn from experience, and how machines could learn from experience better.

RL at Harvard is a one-day workshop on reinforcement learning broadly construed. Our goal is to provide a forum for researchers interested in topics related to reinforcement learning—from building agents that beat humans at Atari games, to building robots that can efficiently navigate complex environments, to the neurobiology of dopamine and behavior—to learn about foundational concepts, discuss the latest advances, and make connections between disparate parts of the field.

This year, we will hear from world experts in reinforcement learning like Na Li, Mitsuko Watabe-Uchida, Stephanie Gil, and Samuel Gershman. Our hope is that the roster of speakers showcases the staggering variety of reinforcement-learning-related work that happens at Harvard, and helps strengthen and enrich the connections among Harvard's reinforcement learning researchers.

The latter half of the workshop will be dedicated to a symposium on the neurobiology of causality and contingency. As animals learn to act effectively in complex environments, they face the problem of learning cause-and-effect relationships: if A happens, then B will happen. They must also learn that some environmental transitions and rewards are contingent on other circumstances being true: if A does not happen, then B will not happen. New research from the Namboodiri Lab, Janak Lab, and Uchida Lab suggests that dopamine may provide a neurobiological mechanism for learning these relationships, although it is not completely clear how this role relates to other roles dopamine is thought to have in helping animals learn from experience.