Distributed Multi-UAV Trajectory Planning for Downlink Transmission: A GNN-Enhanced DRL Approach

Unmanned aerial vehicle (UAV) swarms hold significant potential for widespread application in disaster rescue scenarios. One typical example involves UAVs acting as aerial access points, providing communication service to terrestrial users. Multi-agent reinforcement learning (MARL) algorithms can be used to plan UAV swarms’ flight trajectories. However, most of the exiting path planning schemes based on MARL encounters two significant challenges: They require a long time of training and struggle to adapt to the dimensional changes of the state space observed by UAVs. To address the above issues, this letter introduces a graph neural network (GNN) enhanced distributed deep reinforcement learning approach integrating a designed transfer learning. All UAVs are trained under the QMix algorithm, GNN introduces itself as the UAV’s observer, and the transfer learning freezes the aggregation layer parameters of the GNN. Simulations highlight that the proposed method can improve the convergence speed of the QMix algorithm by 65% with a 23% reduction in the number of training parameters.