Distributed multi-agent deep reinforcement learning for trajectory planning in UAVs-assisted edge offloading

To deal with the diverse computing tasks generated by Internet of Things devices (IoTDs), unmanned aerial vehicles (UAVs)-assisted edge offloading technology has emerged. However, there are many challenges in edge offloading, such as the dynamic change of channel state and the limited computing resources. Therefore, it is crucial to plan the trajectories of UAVs intelligently to improve the offloading efficiency. In this paper, we propose an individual-inference-based distributed deep deterministic policy gradient (IID-DDPG) algorithm for multi-UAV trajectory planning. The proposed IID-DDPG algorithm adopts a distributed training method, which only involves the interactions of neighbor agents. Consequently, the IID-DDPG algorithm has strong scalability and small communication burden. In addition, the causal inference method is first used to measure the importance of neighbor information. Secondly, an information aggregation network is designed to assist UAV agents to infer global knowledge, so as to enhance the stability of distributed training. Finally, the effectiveness and superiority of the proposed IID-DDPG algorithm are verified by extensive simulations.