SDRP-DQL: A satellite distributed routing protocol based on double deep Q-learning

Abstract

To address the routing challenges posed by dynamic topology and traffic imbalances in LEO satellite networks, this paper proposes a distributed intelligent routing algorithm, SDRP-DQL, based on DDQN. This algorithm achieves multi-dimensional state awareness and efficient routing decisions by deriving a reward mechanism with congestion trend prediction capabilities based on queuing theory and a distributed deep reinforcement learning framework, significantly optimizing network latency, packet loss rate, and load balancing. Experiments in a real satellite orbit simulation environment validate the superior performance of SDRP-DQL under different network sizes and complex traffic scenarios. Results show that compared with existing mainstream routing algorithms, SDRP-DQL achieves a 34.78% faster convergence speed, lower end-to-end latency and packet loss rate, and demonstrates good robustness and scalability in the face of node failures and link interruptions. This method provides technical support for intelligent routing in future large-scale satellite networks.