Computing-Aware Routing for LEO Satellite Networks Based on Multi-Step DQN

The large-scale coverage of LEO satellite networks and the enhancement of onboard satellite computing resources have emerged as a pivotal solution for meeting the computing and routing demands of remote sensing (RS) tasks in areas without ground network coverage. Nonetheles, efficiently leveraging LEO satellite network resources effective computation offloading and relay routing presents substantial challenges. In this study, we delve into the joint optimization of task offloading and routing path selection within LEO satellite networks, aiming to maximize long-term user quality of service while adhering to energy constraints. We propose a novel integrated modeling scheme for relay routing and computation offloading of RS tasks in LEO satellite networks, effectively reducing the scale of system decision problems. Furthermore, we devise a multi-step reward aggregation Deep Q-Learning (DQN-MCAR) algorithm for computing-aware routing, effectively addressing the sequential learning problem among the computing-aware routing decisions of subtasks. Finally, theoretical analysis confirms that our proposed algorithm has lower complexity, and the superiority of our scheme is validated by extensive simulation experiments.