In this project, we propose a distributed reinforcement learning method in a multi-agent environment that reflects policies based on the latent representations of other agents through actions. Recent advancements in deep reinforcement learning have led to the development of methods in multi-agent systems where agents continuously learn and influence each other's policies. However, many real-world multi-agent systems include various agents with pre-learned behaviors or agents with strategies that are pre-encoded and do not change during interactions, requiring coordination with each of them individually. Our proposed method uses neural network-based encoders and decoders to identify the latent policies of other agents from high-dimensional observations in the early stages of coexisting interactions, and then determines the agent's own actions based on this information. This allows for rational decision-making that adapts to the behavior patterns of other agents. Experiments show that the proposed method improves coordination among agents compared to existing methods.
