Multi-agent Learning without rewards

Title

Multi-agent Learning without rewards

Speaker

Giorgia Ramponi - University of Zurich

Abstract

This talk studies how to learn equilibrium behavior in multi-agent systems without relying on hand-designed rewards. Instead, it focuses on learning from demonstrations and human knowledge through multi-agent imitation learning. I will present recent results in mean-field games and Markov games showing that standard single-agent imitation methods are insufficient for recovering equilibria, and that equilibrium-aware objectives and interaction can be necessary. In particular, I will discuss improved guarantees for mean-field settings, hardness results for non-interactive learning in Markov games, and a reward-free interactive method, MAIL-WARM, that achieves rate-optimal learning of equilibria from data.

Bio

Giorgia Ramponi is an Assistant Professor at the University of Zurich and affiliated faculty at the ETH AI Center and UZH.ai hub. Her research focuses on reinforcement learning, imitation learning, and multi-agent learning, with an emphasis on principled approaches to sequential decision-making in interactive environments. Previously, she was a postdoctoral researcher at the ETH AI Center and Google Brain, advised by Niao He and Andreas Krause at ETH Zurich.

When

Monday, March 30th, 14:30

Where

Room 322, UniGe DIBRIS/DIMA, Via Dodecaneso 35