The article introduces Fact-based Agent modeling (FAM), a new method in multi-agent reinforcement learning. FAM uses ‘facts’ – rewards and observations obtained by agents after taking actions – as a reconstruction target to learn the policy representation of other agents. This method overcomes the challenges faced in non-stationary environments where all agent policies are learned simultaneously. In contrast to traditional methods that assume access to local information of other agents during execution or training, FAM only relies on its local information, making it more feasible in unknown scenarios. Experimental results show FAM outperforms baseline methods in efficiency of agent policy learning and achieving higher returns in complex competitive-cooperative mixed scenarios.
Publication date: 20 Oct 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2310.12290
