Project.Report-xxz&mzc

Zhancun Mu, Xizhi Xiao

Keywords: Theory-of-Mind Modeling, Offline Reinforcement Learning, Multi-Agent Reinforcement Learning

Abstract: The Theory of Mind (ToM) ability in multi-agent systems is crucial for coordinating cooperation and understanding communication. ToM involves the capacity to reason about the mental states of other agents, encompassing their beliefs, desires, intentions, and more. However, in modeling ToM, many existing works rely on assumptions like rationality, which may not hold true in real-world scenarios. To tackle this issue, we leverage the sequence modeling capability of Transformers in the offline setting. In this paper, we (i) introduce the multi-agent decision transformer (MADT) for agent modeling and demonstrate its generalization ability with new partners. Additionally, we (ii) propose a framework to enhance online reinforcement learning (RL) policies with ToM modeling using MADT. We evaluate our approach in the Overcooked-AI environment and illustrate its satisfactory generalization ability, even with limited data.

Submission Number: 228