A Collaborative Cislunar Computing Architecture for Task Scheduling and Resource Coordination Via Multi-Agent Reinforcement Learning

With the increasing demand for lunar exploration missions and the surge in communication and computing requirements, satellite-based Non-Terrestrial Network (NTN) has emerged as critical infrastructure. However, existing computing architectures are challenged by the Moon's complex environment. Although extreme temperature and radiation pose hardware-layer threats, this study focuses on the algorithmic challenges of communication and energy constraints. Specifically, existing task scheduling paradigms are inefficient in leveraging computing power resources, and existing resource allocation algorithms exhibit high deployment costs and poor dynamic adaptability. This paper proposes a collaborative cislunar computing architecture that orchestrates heterogeneous computational resources spanning lunar surface devices, Lunar Orbit Satellites (LOS), and Earth Orbit Satellites (EOS), incorporating the Two-phase Proactive Selection Multi-Agent Group Relative Policy Optimization (TPS-MAGRPO) algorithm. At its core, the algorithm's Two-phase Proactive Selection (TPS) scheduler—a two-stage task scheduling mechanism—harnesses fragmented resources via initial monolithic scheduling and adaptive task decomposition, while integrating a lightweight multi-agent reinforcement learning algorithm to enable efficient task scheduling and computing power coordination. Simulation results demonstrate improvements in task completion rate and task completion efficiency compared with conventional algorithms, providing an effective solution for the efficient task computation requirements in future lunar exploration.