Abstract
With the vision of supporting global internet access, satellite communication has emerged as a key component for sixth-generation (6G) communication networks. A pivotal trend for satellite communication is using Low Earth Orbit (LEO) satellites to construct a space-based backbone network, which necessitates establishing multi-hop communication paths between satellites with certain inter-satellite routing strategies. However, the vast number of satellites and the fast-varying network topology, combined with the diversified traffic demands, pose significant challenges for inter-satellite routing. In this paper, we propose a novel inter-satellite routing scheme that employs Graph Neural Networks (GNN) and Deep Reinforcement Learning (DRL) to address the above-mentioned challenges. The overall objective is minimizing network-wide average end-to-end delay, which can be achieved by selecting paths with low aggregated propagation delay and balancing the load distribution among various links. Based on this fact, we formulate the inter-satellite routing problem, in which the satellite network is modeled as a graph. Then, we design a GNN and DRL integrated solution framework, where GNN is employed to learn the complex relationship between satellite nodes in the graph and DRL is employed to make routing decisions that are adaptive to the features of satellite networks. Simulation results demonstrate that our solution can adapt to the time-varying network dynamics and diverse traffic demands, outperforming benchmark schemes in terms of average end-to-end delay, reliability, and link utilization.