Efficient Path Selection Design for Large Scale LEO Satellite Constellations Using Graph Embedding-Based Reinforcement Learning

Yuhan Kang, Yifei Zhu, Dan Wang, Zhu Han

Abstract: The rapid expansion of large-scale Low Earth Orbit (LEO) satellite constellations marks a new phase in global connectivity and communication. However, efficient path selection among numerous satellites connected by inter-satellite links (ISL) in such dynamic networks poses substantial challenges. This paper introduces a novel Path Selection Mechanism (PSM-LEO) utilizing Graph Embedding-Based Reinforcement Learning (GERL). Our GERL-based PSM-LEO method employs Graph Embedding (GE) to model the satellite network in a simplified low-dimensional space, simplifying the analysis of complex satellite relationships. Combining Reinforcement Learning (RL) with GE allows for dynamic adaptation of path choices in response to dynamic network conditions and communication needs. To the best of our knowledge, we are the first to propose GERL for PSM-LEO. We evaluate our method's performances through simulations in the Ansys Systems Tool Kit (STK), focusing on diverse LEO scenarios. Our results reveal that our approach surpasses several benchmarks in convergence speed, end-to-end latency, and energy consumption, demonstrating enhanced data transfer efficiency, reduced latency, and improved reliability. Additionally, we assess the scalability of the proposed method by analyzing its performance with increasing satellite constellation sizes. Our results confirm the framework's high scalability, demonstrating its suitability for addressing path selection challenges in future larger satellite networks.

External IDs:dblp:journals/tnse/KangZWH25