Fmnet: Multi-Agent Cooperation By Communicating With Featured Message Network

Multi-agent systems are taking great part in nowadays industries. Cooperative environments may help agents perform better, but the training process is usually time consuming and costly. Although the training process can be speeded up by communicating with important information (such as observations, action sequence, model data and other potential data) between agents, redundant information is still a disturbance term. In this paper, we present a method, called Featured Message Network, which uses a fixed random network to extract features of information from agents as featured message. By communicating with featured messages, agents can cooperate more efficiently while not suffering from useful information loss or redundant information disturbing. We optimized a traditional Deep Q-learning system with our method. We tested our method on Meeting in a Grid problem and experimental results show that training process becomes faster and more stable with our method.