分层智能体架构下的巨星座自适应管控研究  

A Hierarchical Agent Architecture for Adaptive Governance and Control of Mega-Constellations

作　　者：刘立祥[1,2] 孙楚雄 LIU Lixiang;SUN Chuxiong(Institute of Software Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100039,China)

机构地区：[1]中国科学院软件研究所,北京100190 [2]中国科学院大学,北京100039

出　　处：《航天技术与工程学报》2025年第1期93-101,共9页

摘　　要：针对数千甚至数万颗卫星组成的星座在任务规划、资源调度、干扰规避以及在轨维护方面面临的诸多挑战,提出了一种基于分层智能体的解决方案,通过结合大语言模型与强化学习技术,实现巨星座系统的自适应管理与协同优化。基于巨星座系统的高维复杂特性,研究设计了一个分层智能体架构。高层智能体采用大语言模型,以其多模态语义融合与高层次规划能力对任务进行全局性设计与资源分配,提供可解释的宏观调度策略。下层智能体利用强化学习算法,结合规则知识库,负责具体任务的执行、优化与动态适应。该架构通过上下层智能体间的连续反馈迭代与信息传递,实现了自适应协作,并提高了系统在复杂环境中的灵活性与稳健性。能够有效应对巨星座的高动态复杂环境,为未来大规模天基网络的智能化管理提供了具有潜力的技术路径与理论支持。In response to challenges of mission planning,resource scheduling,interference avoidance,and in orbit maintenance by Mega-Constellations,a hierarchical agent-based solution has been proposed,integrating LLM with reinforcement learning techniques to achieve adaptive management and collaborative optimization of megaconstellation systems.In order to study the high-dimensional complexity of mega-constellation systems,a hierarchical agent architecture was designed.The high-level agent utilizes LLM for global mission planning and resource allocation,leveraging their multimodal semantic integration and advanced planning capabilities to provide interpretable macrolevel scheduling strategies.The low-level agents employ reinforcement learning algorithms,combined with a rulebased knowledge base,to execute,optimize,and dynamically adapt to specific tasks.This architecture facilitates continuous iterative feedback and information exchange between layers,achieving adaptive collaboration and enhancing system flexibility and robustness in complex environments.This study highlights that a hierarchical a gent architecture effectively addresses the highly dynamic and complex environment of mega-constellations,providing a promising technical pathway and theoretical foundation for the intelligent management of future large-scale space-based networks.

关 键 词：巨星座 自主智能体 大语言模型(LLM) 强化学习(RL) 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]