Abstract

Multi-agent reinforcement learning, despite its popularity and empirical success, faces significant scalability challenges in large-population dynamic games, especially with heterogeneous players. This talk will use fundamental linear-quadratic games as an example and present recent frameworks that provide principled designs for efficient and scalable learning algorithms in multi-agent systems with heterogeneous players. In the first part, we will introduce the Graphon Mean Field Game approach and present provably convergent policy gradient algorithms for large-population games in which agents interact weakly through a symmetric graph. The second part of the talk will focus on the Alpha-Potential Game framework, which enables the development of efficient learning algorithms for asymmetric network games that go beyond mean-field approximations. This talk is based on joint work with Yufei Zhang.