Task Migration in LEO Satellite Networks: A Multi-Agent Reinforcement Learning Approach

Low Earth orbit satellite edge computing (LEO-SEC) networks, equipped with edge servers, provide computational and storage services to users globally, thereby reducing service latency. However, the stochastic nature of tasks and the heterogeneity of resources can result in load imbalances among satellites, adversely affecting user experience and system stability. Therefore, studying the issue of task migration in LEO-SEC networks is crucial. This paper investigates the task migration problem between LEO satellites, considering load imbalance and the dynamic nature of LEO-SEC networks. The problem is modeled as an integer programming (IP) problem. To solve it, we develop a multi-agent Markov decision process (MAMDP) and propose a migration approach based on the multi-agent proximal policy optimization (MAPPO) algorithm. Simulation results demonstrate the convergence performance of the proposed method and show that it outperforms baseline schemes in terms of efficiency and effectiveness.