Abstract: Laser inter-satellite link (LISL) is an effective method to achieve high speed, high bandwidth, and low power consumption in low Earth orbit (LEO) satellite networks, and there are many routing algorithms for LISL, but most of the algorithms assume that the LISL is stable, and do not take into account the packet loss due to the LISL jitter. In this paper, we propose a Caching Based Dynamic Routing algorithm (CBDR) for interstellar laser links based on the combination of caching and deep reinforcement learning (DRL). The algorithm integrates a distributed multi-intelligence deep reinforcement learning framework with a double-rate synchronous dynamic random memory (DDR) caching mechanism, which significantly reduces the packet loss rate and delay during link interruption and achieves load balancing by calculating the primary and backup paths and combining with cache retransmission. Simulation results show that CBDR performs well in simple scenarios. Compared with traditional deep Q learning and shortest path algorithms, CBDR improves packet loss rate, average delay, and load balancing by 15.14%, 7.95%, and 58.25%, respectively, in complex scenarios, and its performance is basically consistent with that in simple scenarios, effectively enhancing the robustness of LEO constellation networks under LISL.

Key words: laser inter-satellite link, LEO satellite networks, deep reinforcement learning, caching