A Graph-Based Soft Actor Critic Approach in Multi-Agent Reinforcement Learning

Multi-Agent Reinforcement Learning (MARL) is widely used to solve various real-world problems. In MARL, the environment contains multiple agents. A good grasp of can guide agents learn cooperative strategies. Centralized Training Decentralized Execution (CTDE), a centralized critic strategies learning. However, having in leads curse dimensionality and influence other agents’ strategies, resulting difficulties for critics We propose graph-based approach overcome above It uses graph neural network, which partial observations as input, information between aggregated by methods extract about whole environment. this way, improve their understanding overall state while avoiding dimensional explosion. Then we combine dual dynamic decomposition method with soft actor-critic train policy. The former individual global rewards learning, latter help an optional policy better. call Graph-based Actor-Critic (MAGAC). compare our proposed several classical MARL algorithms under Multi-agent Particle Environment (MPE). experimental results show that achieve faster learning speed better