Abstract:

Low Earth orbit satellite networks serve as crucial components of space communication networks, effectively addressing challenges encountered by terrestrial communication networks in edge computing, such as limited coverage of ground-based stations and vulnerability to natural disasters. Therefore, this study introduces low Earth orbit satellite networks into the context of ground edge computing, forming a satellite-ground cooperative network for edge computing scenarios and investigating key tasks offloading in edge computing. Given that factors like long communication distances, limited satellite resources, and rapid changes in node positions in satellite-ground networks can lead to increased system latency and energy consumption, this paper aims to minimize system latency and energy consumption by modeling them as a Markov game. A reinforcement learning algorithm based on multi-agent deep deterministic policy gradient specifically tailored for satellite-ground cooperative network edge computing scenarios is proposed, aiming to reduce overall system overhead through centralized training and distributed execution. Finally comparative analysis with other algorithms demonstrates that, compared to local task processing, deep deterministic policy gradient algorithm, and double-delay deep deterministic policy gradient algorithm, the proposed algorithm achieves an average reduction in system overhead of 37.25%, 9.420%, and 6.756%, respectively, while exhibiting superior stability and convergence.

Key words: satellite-ground collaborative network, edge computing, deep reinforcement learning, task offloading decision