An adaptive multi-agent deep reinforcement learning routing algorithm for large-scale LEO constellation

With the widespread application of satellite communications in LEO, satellite routing has become an essential component for ensuring efficient and reliable information transmission. However, the excessive number of satellite nodes and the highly dynamic topology characteristics make the design of LEO satellite routing that can guarantee low end-to-end latency challenging. To address this challenge, this paper proposes a fully distributed MA-DRL framework, treating each satellite as an independent agent. The framework accelerates the learning process of agents through an adaptive reward mechanism. each satellite makes decisions based on its own observations of its neighbors, fully utilizing the local information among satellite nodes. Finally, experimental results demonstrate that AMAR outperforms baseline algorithms in terms of average delivery time, packet loss rate, and throughput.