Throughput and Link Utilization Improvement in Satellite Networks: A Learning-Enabled Approach
Hao Wu
TL;DR
This work tackles throughput limitations in satellite networks caused by congested ISLs in densely populated regions. It introduces a learning-enabled framework that couples topology-isomorphism-based link load prediction with an MDP model of store-and-forward routing, and offers two policy-generation methods: model-based value iteration and model-free Q-learning. Evaluation on a TSN-1ATSN constellation shows improvements in throughput and link utilization, while requiring less than 20% of the computation time of constraint-based routing, indicating strong scalability for mega-constellations. Overall, the approach provides a practical, scalable path to better utilization of satellite networks under dynamic topologies and traffic patterns.
Abstract
Satellite networks provide communication services to global users with an uneven geographical distribution. In densely populated regions, Inter-satellite links (ISLs) often experience congestion, blocking traffic from other links and leading to low link utilization and throughput. In such cases, delay-tolerant traffic can be withheld by moving satellites and carried to navigate congested areas, thereby mitigating link congestion in densely populated regions. Through rational store-and-forward decision-making, link utilization and throughput can be improved. Building on this foundation, this letter centers its focus on learning-based decision-making for satellite traffic. First, a link load prediction method based on topology isomorphism is proposed. Then, a Markov decision process (MDP) is formulated to model store-and-forward decision-making. To generate store-and-forward policies, we propose reinforcement learning algorithms based on value iteration and Q-Learning. Simulation results demonstrate that the proposed method improves throughput and link utilization while consuming less than 20$\%$ of the time required by constraint-based routing.