Decentralized MARL for SDN Ground Station Cluster Selection in LEO Constellations Under Stochastic Weather

Abstract

Low Earth Orbit (LEO) satellite constellations enable global low-latency connectivity but face challenges due to weather-dependent link variability, orbital dynamics, and heterogeneous ground infrastructure. In this paper, we propose a hybrid LEO-terrestrial architecture that integrates Software-Defined Networking (SDN) ground station clusters with repeater-assisted reception and a structured fallback mechanism. We develop a stochastic model that captures weather-driven reliability, correlated repeater behavior, and multi-path reception, and formulate an optimization problem to minimize the expected communication delay. To address the intractability of this problem in large-scale, partially observable environments, we design a fully decentralized Multi-Agent Reinforcement Learning (MARL) framework based on Proximal Policy Optimization (PPO), where each satellite makes decisions using only local observations. The reward is aligned with the analytical delay objective, ensuring consistency between the model and learned policies. Simulation results across diverse scenarios demonstrate that the proposed approach reduces the mean delay by 40-60% and significantly decreases fallback usage compared to baseline methods. These results highlight the effectiveness of the proposed framework in enabling adaptive and delay-efficient control in next-generation LEO satellite systems.