MARIC: an efficient multi-agent real-time intention-based communication model for team cooperation

Effective communication is crucial for promoting team cooperation in multi-agent reinforcement learning tasks. Human intention plays a key role in facilitating effective communication, driving individuals to communicate quickly and accurately with team members. In this paper, we propose an efficient multi-agent real-time intention-based communication model for team cooperation called MARIC. In MARIC, an intention correlation network is proposed to drive the agent’s belief about who to communicate with, which considers several influencing factors, e.g. the agent’s situation awareness about each team member and the agent’s purpose inference of other teammates. This network maps the agent’s local observation to real-time communication intention, and quantifies the correlation between the agents regarding each influencing factor to label the necessity of communication. To improve the stability of the training process and avoid redundant information, a message simplicity loss is proposed as a regularization technique. MARIC is evaluated against several existing baselines in two typical multi-agent cooperative tasks: predator–prey and Traffic Junction, and a mixed cooperative-competitive task: Google Research Football. The experimental results show that MARIC outperforms the baselines.