Learning structured communication for multi-agent reinforcement learning

This work explores the large-scale multi-agent communication mechanism for reinforcement learning (MARL). We summarize general topology categories structures, which are often manually specified in MARL literature. A novel framework termed Learning Structured Communication (LSC) is proposed by a flexible and efficient (hierarchical structure). It contains two modules: structured module communication-based policy module. The learns to form hierarchical structure maximizing cumulative reward of agents under current policy. adopts graph neural networks generate messages, propagate information based on learned structure, select actions. In contrast existing mechanisms, our method has learnable structure. Experiments battle scenarios show that LSC high efficiency global cooperation capability.