A Service-Oriented Multipath Routing Scheme Based on GNN and RL for Multilayer Satellite Networks

The research focus in satellite networks is gradually shifting from single-layer communication to exploring core network services in multilayer satellite networks (MLSNs), focusing on solving the synergy and optimization problems of diversified functions such as computation, communication, caching, awareness, and security. In this paper, we map service requests to multivariate satellite functional nodes, construct the set of Medium Earth Orbit (MEO) satellites that must be passed through, and propose a multipath routing scheme based on graph neural networks (GNNs) and reinforcement learning (RL), aiming to improve network utilization. We model the service demand-oriented multipath routing as a stochastic optimization problem, which is decomposed into two subproblems: cross-layer path planning and traffic splitting. For the first subproblem, we design a cross-layer service path planning (CLSP) algorithm based on domain head satellites to compute all reachable paths between node pairs. For the second subproblem, we design a GNN-based reinforcement learning traffic engineering (GRL-TE) algorithm that dynamically allocates traffic ratios by evaluating path quality using GNNs. The experimental results based on the MLSN show that the CLSP algorithm can realize the routing computation of various types of requests with a request success rate of up to 60%. Compared with the baseline algorithm, the GRL-TE algorithm has superior performance in terms of network utilization, average throughput, path delay, packet loss rate, and Jain fairness index.