Multi-Agent Reinforcement Learning With Distributed Targeted Multi-Agent Communication

Centralized training distributed execution (CTDE) in multi-agent reinforcement learning (MARL) is a commonly used application paradigm. This paradigm usually assumes that the global state of the environment can be obtained during training, which is often difficult to satisfy in various scenarios due to constraints such as data transfer and processing power. Fully distributed multi-agent reinforcement learning algorithms do not depend on the knowledge of global state, with each agent trained independently and treating the remaining agents as part of the environment. However, applying single-agent algorithms to multi-agent systems faces the problem of non-smoothness of the environment and difficulty in forming effective collaborative strategies. In this paper, we propose a new method, Distributed Targeted Multi-Agent Communication (DTMAC), which makes each agent generate messages and pass them to other agents, explicitly enhancing the collaboration among individual agent and facilitating the formation of collaborative strategies. Experiments are given to illustrate the effectiveness of the method.