Abstract

We consider the multi-agent reinforcement learning setting with imperfect information. The reward function depends on the hidden goals of both agents, so the agents must infer the other players’ goals from their observed behavior in order to maximize their returns. We propose a new approach for learning in these domains: Self Other-Modeling (SOM), in which an agent uses its own policy to predict the other agent’s action and update its belief of their hidden goal in an o line manner. We evaluate this approach on three different tasks and show that the agents are able to learn better policies using their estimate of t other players’ goals, in both cooperative and competitive settings.