Decision-making Method Based on Multi-agent Deep Reinforcement Learning

Based on the decision-making architecture of information pooling and sharing in the hidden layer, the communication protocol is set manually, and the pooling method is used to integrate the information. Although the problem of communication and extension between agents is solved, it is difficult for tasks lacking prior knowledge to design effective communication protocols. The centralized decision- making architecture based on two-way RNN communication uses the information storage characteristics of two-way RNN structure. It can self learn the communication protocol between agents, which overcomes the rigid requirement of task prior knowledge in communication protocol design. The action distribution of a single agent is used as the output of the multi- agent network to replace the joint action distribution, and the global state information in the environment is used as the input instead of simply inputting the local information to different agents. The effectiveness of the method is verified by an example.