摘要
博弈强化学习作为人工智能领域的新兴范式,是当前解决低轨星座边缘计算问题的主流方法。融入博弈论的多智能体深度强化学习方法为复杂、动态、不确定性的星座边缘计算问题提供了新思路。通过梳理总结卫星组网、任务卸载以及资源调度3种卫星边缘计算主要研究方向,详细阐述了博弈强化学习范式基础,并从博弈模型、深度Q网络、深度确定性策略梯度以及近端策略优化等方面分别阐述了3种研究方向上的典型应用现状,最后对该领域的前沿挑战进行分析,期望为博弈强化学习范式与低轨星座边缘计算领域的交叉融合研究提供参考。
As a new paradigm in the field of artificial intelligence,game reinforcement learning is an advanced mainstream method to solve the edge computing problem of low-orbit constellation.The multi-agent deep reinforcement learning integrated into the game perspective provides a new idea for dynamic,complex and uncertain constellation edge computing problems.By summarizing the three main research directions of satellite edge computing,namely satellite networking,task unloading and resource scheduling,the basis of game reinforcement learning paradigm is elaborated,and the typical applications in the three research directions are described respectively from the methods of game model,deep Q network,deep deterministic strategy gradient and near-end strategy optimization.In the end,the paper looks forward to the frontier challenges in this field,expected to provide a reference for the cross-fusion research of game reinforcement learning paradigm and low-orbit constellation edge computing.
作者
谷学强
张万鹏
谭思雨
罗俊仁
周棪忠
GU Xueqiang;ZHANG Wanpeng;TAN Siyu;LUO Junren;ZHOU Yanzhong(College of Intelligence Science and Technology,National University of Defense Technology,Changsha 410073,China;Hunan Institute of Advanced Technology,Changsha 410205,China)
出处
《智能科学与技术学报》
CSCD
2024年第3期301-318,共18页
Chinese Journal of Intelligent Science and Technology
基金
国家自然科学基金项目(No.92271108,No.62173336)。
关键词
低轨星座
边缘计算
博弈论
多智能体强化学习
low-orbit constellation
edge computing
game theory
multi-agent reinforcement learning
