基于联邦学习与DQN的缓存策略  

Caching strategy based on federated learning and DQN

作　　者：桂易琪 王鹏程 王威 李鹏海 张乐君 GUI Yiqi;WANG Pengcheng;WANG Wei;LI Penghai;ZHANG Lejun(School of Information Engineering,Yangzhou University,Yangzhou 225127,China;Research and Development Center for E-Learning,Ministry of Education,Beijing 100039,China;Cyberspace Institute Advanced Technology,Guangzhou University,Guangzhou 510006,China)

机构地区：[1]扬州大学信息工程学院,江苏扬州225127 [2]教育部电子学习研究与开发中心,北京100039 [3]广州大学高级技术网络空间研究所,广州510006

出　　处：《扬州大学学报(自然科学版)》2025年第2期45-53,共9页Journal of Yangzhou University:Natural Science Edition

基　　金：国家自然科学基金资助项目(62172353);未来网络科研基金资助项目(FNSRFP-2021-YB-48)。

摘　　要：命名数据网络(named data networking, NDN)的“泛在缓存”特性引发数据副本率过高、缓存空间不能充分利用等问题。为了解决上述问题,从内容流行度和缓存节点角度出发,以最大化缓存命中率作为目标,提出了一种基于联邦学习与深度Q网络(deep Q-network, DQN)相结合的缓存节点选择算法。通过自编码器(auto encoder, AE)模型预测内容流行度,得到所需缓存的内容;使用DQN模型对缓存节点选择的决策过程进行建模,根据节点的状态获得缓存对象的最佳缓存节点;采用主动缓存替换算法和被动缓存替换算法,减少了传输路径上的缓存冗余,提高了节点上数据包的利用率。结果表明,该方案在缓存命中率、延迟和链路负载等缓存方面较现有策略表现更为出色。The characteristic of“ubiquitous caching”of Named Data Networking(NDN)has led to problems such as excessively high rates of data replication and the inefficiency of cache space utilization.To address these issues,a caching node selection algorithm which combines federated learning with the Deep Q-Network(DQN)model is proposed in this paper,with the goal of maximizing the cache hit rate.The auto-encoder model is used to predict the content popularity and obtain the content needed to be cached.DQN model is used to model the decision-making process of cache node selection and the best caching node is obtained according to the state of the nodes.The active and passive cache replacement algorithms are employed to reduce the cache redundancy on the transmission path and improve the utilization rate of data packets on the node.Experimental results show that the proposed solution exhibits superior performance over existing strategies in terms of cache hit rate,latency,and link load.

关 键 词：命名数据网络 缓存策略 深度学习 联邦学习 

分 类 号：TP393.09[自动化与计算机技术—计算机应用技术]