Research on Efficient Multiagent Reinforcement Learning for Multiple UAVs' Distributed Jamming Strategy

To support Unmanned Aerial Vehicle (UAV) joint electromagnetic countermeasure decisions in real time, coordinating multiple UAVs for efficiently jamming distributed hostile radar stations requires complex and highly flexible strategies. However, with the nature of the high complexity dimension and partial observation of the electromagnetic battleground, no such strategy can be generated by pre-coded software or decided by a human commander. In this paper, an initial effort is made to integrate multiagent reinforcement learning, which has been proven to be effective in game strategy generation, into the distributed airborne electromagnetic countermeasures domain. The key idea is to design a training simulator which close to a real electromagnetic countermeasure strategy game, so that we can easily collect huge valuable training data other than in the real battle ground which is sparse and far less than sufficient. In addition, this simulator is able to simulate all the necessary decision factors for multiple UAV coordination, so that multiagents can freely search for their optimal joint strategies with our improved Independent Proximal Policy Optimization (IPPO) learning algorithm which suits the game well. In the last part, a typical domain scenario is built to test, and the use case and experiment results manifest that the design is efficient in coordinating a group of UAVs equipped with lightweight jamming devices. Their coordination strategies are not only capable of handling given jamming tasks for the dynamic jamming of hostile radar stations but also beat expectations. The reinforcement learning algorithm can do some heuristic searches to help the group find the tactical vulnerabilities of the enemies and improve the multiple UAVs' jamming performance.