Multi-Agent Cooperative Target Search Based on Reinforcement Learning

This paper considers the problem of cooperative search for multiple stationary targets by multi-agent with limited sensing and communication capabilities. An integrated learning algorithm for cooperative search based on reinforcement learning is proposed. In particular, we consider a state containing local probability map and neighbouring agents map which provides the agent with information to plan routes and search collaboratively. In addition, a reward function consisting of target found reward, time consumption reward and guiding reward is designed to guide agents to explore and learn efficiently. To ensure the stability of training process, the policies of agents are frozen and shared periodically in a distributed training framework. The proposed method is tested under simulated scenarios compared with coverage control methods and random strategies. Multiple simulation results show considerable advantages.