Joint Partitioned Link-Building and Deep Reinforcement Learning Aided Routing for LEO Satellite Networks

In this paper, a partitioned link-building and deep reinforcement learning (DRL) aided routing collaborative algorithm is proposed, aiming to improve the transmission performance of large-scale satellite networks via the synergy of appropriate link-building and intelligent routing decisions. Firstly, the algorithm establishes a differentiated regional constellation link-building (RCL) mechanism, dynamically adjusting link establishment strategies based on the satellite geographical distribution and topological characteristics to extend link duration and improve topological stability, laying a stable topological foundation for routing decisions. Furthermore, it introduces a multi-agent double deep Q-network (DQN) architecture and designs a joint reward function that integrates hop count and distance metrics to guide efficient packet forwarding to the destination. Simulation results reveal that the proposed algorithm improves the average convergence rate by 21% compared to baseline schemes, significantly reduces reward volatility after convergence, and enhances end-to-end latency performance by an average of 22%, verifying its effectiveness and stability in dynamic satellite networks.