Multi-Agent Mean Field Predict Reinforcement Learning

The study of multi-agent reinforcement learning can solve many problems in real life. The current research can be divided into two aspects: one is adding the information of other agents into the critic-network to form a global critic-network, as MADDPG; the other is putting them into the actor-network, like CommNet, which takes the actions or observations from other agents into consideration. However, the two methods are faced with these problems: the action space is huge when the number of agents increases; In reality, due to the limitation of bandwidth and delay, communication often cannot perform well or even work normally. Inspired by MFRL, we design our algorithm MFPRL to solve this problem. The neighbors’ average action is predicted by a separate MFP network. The experiment shows that our method achieves better results than MFRL.