Multi-Layer Reinforcement Learning Assisted Task Offloading in Satellite Edge Computing
Weichang Wen, Haixia Cui, Tao He
Abstract
Satellite networks can overcome the general geographical constraints, such as deserts and oceans, and offer the universal connectivity worldwide. So, it becomes a popular research topic in recent years. In this paper, we propose a new three-tier satellite edge computing framework which can provide flexible heterogeneous computing for the suitable ground users and thereby enable the global computing services. Additionally, based on the hybrid computing offloading architecture and the computational capability constraints of stable satellite communication links, we present a novel computing task offloading algorithm to minimize the task latency and reduce the energy consumption while ensuring the task success rate. With the help of reinforcement learning, we address the offloading problem for terminal users and develop a novel multi-layer learning extension technique based task offloading algorithm which can enable the higher-level agents to consult the mist node agents. At last, we use the SatEdgeSim simulator to verify the feasibility of our proposed algorithm and simulation results show that the proposed algorithm can significantly reduce the satellite energy consumption and communication latency, and improve the task success rate compared to the existing strategies.