Multi-Agent Global Prioritized Experience Learning for UAV Cooperative Jamming in Secure Communication

In unmanned aerial vehicle (UAV) communication networks, the line-of-sight (LoS) propagation link makes the communication information vulnerable to being wiretapped by ground eavesdroppers (GEs). This paper focuses on the maximization of the average secrecy rate with multiple UAV jammers helping multiple UAV transmitters to defend against GEs. We propose a multi-agent global prioritized experience learning (MAGPEL) algorithm. The allocation of UAV transmitters' subchannels, locations, and power levels, along with the allocation of UAV jammers' locations and power levels are jointly optimized. Each UAV takes the role of an agent and uses the global information for training, which comprises specifics on the states and actions of all UAVs. Besides, temporal difference error (TDerror) is used to measure the significance of the experience and calculate the probability that the experience is sampled. Experiences of greater significance can be extracted with a higher probability for training. Simulation results show that the proposed algorithm has better convergence performance and a higher secrecy rate compared with other state-of-the-art methods.