面向DAG任务的分布式智能计算卸载和服务缓存联合优化  

Joint optimization of distributed intelligent computation offloading and service caching for DAG tasks

作　　者：李云[1,2] 南子煜 姚枝秀 夏士超 鲜永菊[1] LI Yun;NAN Ziyu;YAO Zhixiu;XIA Shichao;XIAN Yongju(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学软件工程学院,重庆400065

出　　处：《中山大学学报（自然科学版）（中英文）》2025年第1期71-82,共12页Acta Scientiarum Naturalium Universitatis Sunyatseni

基　　金：国家自然科学基金(62071077,62301099);中国博士后科学基金(2023MD734137);重庆市自然科学基金(2022NSCQ-LZX0191)。

摘　　要：建立了一种有向无环图(DAG,directed acyclic graph)任务卸载和资源优化问题,旨在应用最大可容忍时延等约束实现系统能耗最小化。考虑到网络中计算请求高度动态、完整的系统状态信息难以获取等因素,最后使用多智能体深度确定性策略梯度(MADDPG,multi-agent deep deterministic policy gradient)算法来探寻最优的策略。相比于现有的任务卸载算法,MADDPG算法能够降低14.2%至40.8%的系统平均能耗,并且本地缓存命中率提高3.7%至4.1%。A directed acyclic graph(DAG)was developed for task offloading and resource optimization,aiming to minimize system energy consumption under constraints such as maximum tolerable delay.Considering that computing requests are highly dynamic in the network and it is difficult to obtain complete system state information,the multi-agent deep deterministic policy gradient(MADDPG)algorithm is used to explore the optimal strategy.Compared to existing task offloading algorithms,the MADDPG algorithm can reduce the average system power consumption by 14.2%to 40.8%,and improve the local cache hit rate by 3.7%to 4.1%.

关 键 词：移动边缘计算 多智能体深度强化学习 计算卸载 资源分配 服务缓存 

分 类 号：TN92[电子电信—通信与信息系统]