Federated Graph Reinforcement Learning-Driven Adaptive Routing Strategy for Mega LEO Satellite Networks

This letter addressed the routing selection problem in mega low Earth orbit (mLEO) satellite constellations, where the large-scale of satellite network and highly dynamic environment result in an expansive routing space and necessitate frequent adjustments based on real-time network conditions. To tackle these challenge, the routing selection problem is modeled as an integer optimization problem and subsequently reformulated as a Markov Decision Process (MDP). A federated graph reinforcement learning-based routing strategy (FGRLR) is proposed, which combines graph neural networks (GNNs) with traditional reinforcement learning to capture dynamic changes in satellite networks, and leverages a federated learning-based training approach to accommodate diverse user demands. Simulation results demonstrate that the proposed method significantly reduces the computational complexity while maintaining superior routing performance of users.