Graph Reinforcement Learning Based Multi-Hotspot Region UAV Dynamic Scheduling in Mobile Edge Computing

With the increasing number of IoT devices, human activities tend to create dynamic multi-hotspot regions, which cannot be adapted to such highly dynamic scenarios due to the immobility of base station edge servers. UAVs serve as an effective solution but have limited resources to carry, thus, solving the dynamic scheduling and collaboration of UAVs in multi-hotspot regions is an important problem. In this paper, we propose a scheme called Dynamic Scheduling based on Graph Reinforcement Learning (DSGR), which aims to maximize the long-term energy efficiency of UAVs by optimizing their flight trajectories and charging timing. First, we abstract the dynamic scheduling problem of UAVs among multiple hotspots as a dynamic topological graph and use graph convolution to learn the temporal and spatial relationships among hotspots. Then, the UAV flight strategies are learned using the neighbor vectors of the graph nodes as action masks as well as the temporal and spatial relationships of the hotspots modeled as inputs to the PPO network. Compared with the benchmark algorithm, our proposed DSGR algorithm achieves the best performance in terms of UAV energy efficiency.