Kai Shao: Reinforcement Learning for Finite Space Mean-Field Type Games

Abstract: Mean Field Type Games (MFTGs) model strategic interactions among large coalitions. Although the theoretical foundations of MFTGs are well established, efficient and scalable computational methods remain limited. In this talk, I will present reinforcement learning approaches for solving MFTGs in finite state and action spaces with general dynamics and reward structures. I will begin by formalizing the notion of MFTGs and showing that their solutions provide approximate Nash equilibria for finite-population coalition games. Then, I will introduce two algorithms: the first leverages quantization of the mean-field space combined with Nash Q-learning, and the second employs a deep reinforcement learning framework. Both theoretical analysis and numerical experiments will be discussed to illustrate the effectiveness of the proposed methods. 

This is joint work with Jiacheng Shen and Mathieu Laurière from NYU Shanghai.