A Distributed Routing Algorithm for LEO Satellite Networks: A Multiagent Transformer-MIX Learning Approach

As a complement to terrestrial networks, low-Earth orbit (LEO) satellite networks are promising to provide ubiquitous and continuous services. To accommodate the dynamic topology of LEO satellite networks and increasing traffic demands, this article proposes a distributed routing approach to optimize the end-to-end delay relying on the multiagent deep reinforcement learning (MADRL), where each satellite node is deployed with an autonomous agent and makes its real-time next-hop decisions independently with local observation information. To promote the inner cooperation between decentralized agents, a centralized training scheme with a unified load-balancing reward is utilized by adopting a novel multiagent Transformer-MIX architecture. Moreover, to derive a better decision for each agent, we design an attention-involved agent network to capture more hidden information, and a Transformer-based parameter recurrent mechanism to generate the joint action-value function is used to enhance a more stable training. The simulation results indicate that our proposed scheme achieves faster convergence and demonstrates superior performance across several key performance metrics compared to existing benchmark schemes. Specifically, when the intersatellite link (ISL) failure rate in the network reaches 18%, our scheme achieves a reduction in end-to-end delay by 13.6% and an increase in packet successful delivery rate by 5.4% compared to the benchmark schemes.