UAV Relay Selection and Resource Scheduling System Based on Hierarchical Graph Attention Reinforcement Learning
Abstract
Uncrewed aerial vehicles (UAV) are currently widely deployed as aerial relays to provide network coverage for consumer electronics like smart terminals in emergency scenarios. Operating these UAV networks must balance data transfer rates and device power consumption while avoiding computational overload caused by frequent network structure changes. To solve these problems, we propose a novel Hierarchical Graph Attention Reinforcement Learning (HGATRL) framework specifically designed for UAV networks. This framework adopts a spatio-temporal decoupling architecture and includes two independent control layers. The topology management layer applies a graph attention network to map the distribution of nodes into a dynamic graph and generates robust UAV flight positions and optimal relay paths. The channel control layer utilizes a decoupled dueling double Q-Network to dynamically allocate frequency bands, transmit power levels, and communication channels for different user devices. Extensive simulation results show that the proposed HGATRL method performs significantly better than existing baseline methods in improving connection rates, saving energy, and reducing communication outages. These findings strongly confirm its practical value for building reliable UAV communication systems in the future.