GQN: Multi-Agent Deep Reinforcement Learning based on Graph Networks

Cooperation and scalability are crucial to applying multi-agent deep reinforcement learning (MADRL) to Unmanned Aerial Vehicle (UAV) swarm confrontation. However, most existing methods struggle to make real-time planning adjustments when confronted with these challenges. This paper proposes a MADRL algorithm called GQN which utilizes a graphbased method for the construction of node and edge features. GQN converts the latent relations among agents into a graph representation to provide a strong relational inductive bias. Graph network (GN) blocks are composed to implement multilayer graph message passing, abstracting the communication among agents as information aggregation of the nodes in the graph. Our method dynamically captures the feature graph constructed by the local observation of each individual and the interaction with its neighbors, which enables it to adapt well to large-scale scenarios and promote cooperation. Experimental results demonstrate that GQN outperforms existing algorithms in cooperative confrontation scenarios.