Seminars

View all Seminars  |  Download ICal for this event

Modern Perspectives in Reinforcement Learning: Performativity, Robustness, and Multi-Agency

Series: Bangalore Theory Seminars

Speaker: Debmalya Mandal, University of Warwick

Date/Time: Apr 23 12:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Abstract:
The rise of human-centric LLMs has ushered in an era of experience, presenting an opportunity to revisit and refine classic concepts in reinforcement learning (RL). In the first part of this talk, I will present our work on modeling performative aspects of reinforcement learning. We propose a game-theoretic solution concept that captures how the underlying environment changes in response to the deployed policy, and how to optimize in the presence of such performative effects. In the second part, I will discuss our ongoing work on robustness in reinforcement learning from human feedback (RLHF). I will show how to design robust RLHF methods when human feedback may be adversarially corrupted or the distribution of prompts shifts at deployment time. The talk will conclude with some reflections on multi-agent and strategyproof aspects of RLHF.


Microsoft Teams link:

Link

We are grateful to the Kirani family (Link and the Walmart Center for Tech Excellence (Link for generously supporting this seminar series


Hosts: Rameesh Paul, Debajyoti Kar, KVN Sreenivas, Nirjhar Das, Rahul Madhavan