分布式星群中的协同计算卸载与资源分配  

Cooperative Computing Offloading and Resource Allocation in Distributed Satellite Clusters

作　　者：刘津宇 姜兴龙 胡海鹰[1,3] 梁广 LIU Jinyu;JIANG Xinglong;HU Haiying;LIANG Guang(Innovation Academy for Microsatellites,Chinese Academy of Sciences,Shanghai 201203,China;University of Chinese Academy of Sciences,Beijing 100049,China;Shanghai Engineering Center for Microsatellites,Shanghai 201203,China)

机构地区：[1]中国科学院微小卫星创新研究院,上海201203 [2]中国科学院大学,北京100049 [3]上海微小卫星工程中心,上海201203

出　　处：《宇航学报》2024年第6期970-982,共13页Journal of Astronautics

基　　金：国家自然科学基金(U21A20443)。

摘　　要：在分布式星群在轨建造、维修的全天基场景中,星群动态性大,时延敏感类(A类)与时延容忍类(B类)任务由资源受限的终端卫星并发产生,而传统的天地协同卸载场景中不考虑卫星动态性和任务类型的多样性。针对全天基场景的需求,构建了一种分布式星群边缘计算架构,提出了一种差分自适应奖励系统DDPG(DARS-DDPG)算法。通过将差分自适应奖励系统引入传统DDPG算法,使算法在学习过程中能够区分出不同类型任务的重要性,并自适应调整两类任务的惩罚系数,使两类任务的完成率达到最高。仿真表明,DARS-DDPG学习出的策略相较于基线策略以及传统DDPG学习出的策略,在任务时延以及A、B类任务完成率上都有大幅提升。In the all-space-based scenario of on-orbit construction and maintenance of distributed clusters,the clusters are highly dynamic,and the tasks of delay-sensitive(class A)and delay-tolerant(class B)are generated concurrently by resource-constrained terminal satellites,while the traditional space-ground cooperative offloading scenario does not consider the satellite dynamics and the diversity of task types.Aiming at the needs of all-space-based scenarios,a distributed star clusters edge computing architecture is constructed,and a differential adaptive reward system-deep deterministic policy gradient(DARS-DDPG)algorithm is proposed.By introducing the differential adaptive reward system into the traditional DDPG algorithm,the algorithm can distinguish the importance of different types of tasks in the learning process and adaptively adjust the penalty coefficients of the two types of tasks.The completion rate of the two types of tasks is the highest.Simulation results show that compared with the baseline strategy and the strategy learned by the traditional DDPG,the strategy learned by DARS-DDPG has a significant improvement in task delay and the completion rate of A and B tasks.

关 键 词：计算卸载 资源分配 卫星集群 深度强化学习 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] TN919.2[自动化与计算机技术—计算机科学与技术] V19[电子电信—通信与信息系统]