HiComm Improves Multi-Agent RL Communication Efficiency

In cooperative multi-agent reinforcement learning (MARL), effective communication is vital for agents to overcome partial observability. However, most current communication protocols treat messages as unstructured data, ignoring the inherent hierarchical structure often present in agent observations, such as groups and entities. This paper introduces HiComm, a novel communication module designed to leverage this hierarchy. HiComm operates as a receiver-driven system where a querying agent decodes information from a sender's hierarchical observation in three stages: selecting a group, then a specific sender, and finally an entity within that group. This structured information retrieval mechanism, rather than dense vector transmission, drastically reduces communication volume while maintaining or improving performance compared to traditional methods.