Reinforcement learning systems are often trained in simulators or from data collected under a nominal model, yet deployed in environments with different dynamics, observations, or interacting agents. This talk examines how to make reliable decisions when the model itself is uncertain. I will begin with robust reinforcement learning, where policies optimize performance against an explicit set of plausible models, and show how duality connects robustness to adversarial reweighting. I will then focus on risk-averse quantal response equilibria, which combine risk sensitivity and bounded rationality in multi-agent decision-making. These ideas provide a framework for studying when equilibria can be computed and learned efficiently, including in games where standard equilibrium notions pose computational challenges. I will also discuss extensions to distributionally robust cooperative multi-agent reinforcement learning.
Kishan Panaganti is a researcher at Tencent Hunyuan working on reinforcement learning for foundation models. He was previously a PIMCO Postdoctoral Fellow at Caltech and received his Ph.D. in Electrical and Computer Engineering from Texas A&M University in 2023 and his master’s degree from the Indian Institute of Science in 2017. His research develops theoretical foundations and algorithms for reinforcement learning and multi-agent systems, with a focus on learning reliably under model uncertainty.