The multi-agent reinforcement learning (MARL) suffers from several issues when it is applied to large-scale environments. Specifically, the communication among the agents is limited by the communication distance or bandwidth. Besides, the interactions among the agents are complex in large-scale environments, which makes each agent hard to take different influences of the other agents into consideration and to learn
a stable policy. To address these issues, a soft graph attention reinforcement learning (SGA-RL) is proposed. By taking the advantage of the chain propagation characteristics of graph neural networks, stacked graph convolution layers can overcome the limitation of the communication and enlarge the agents’ receptive field to promote the cooperation behavior among the agents. Moreover, unlike traditional multi-head attention mechanism which takes all the heads into consideration equally, a soft attention mechanism is designed to learn each attention head’s importance, which means that each agent can learn how to treat the other agents’ influence more effectively during large-scale
environments. The results of the simulations indicate that the agents can learn stable and complicated cooperative strategies with SGA-RL in large-scale environments.