基于D2D网络的去中心化学习系统关键技术研究  

Research on Key Technologies of Decentralized Learning System Based on D2D Network

作　　者：孙凯 张亚楠 孙贞 刘胜利 SUN Kai;ZHANG Yanan;SUN Zhen;LIU Shengli(State Grid Liaocheng Power Supply Company,Liaocheng Shandong 252400,China;State Grid Dong’e Power Supply Company,Liaocheng Shandong 252200,China;Shanghai University,Shanghai 200444,China)

机构地区：[1]国网聊城市供电公司,山东聊城252400 [2]国网聊城市东阿县供电公司,山东聊城252200 [3]上海大学,上海200444

出　　处：《通信技术》2024年第11期1153-1157,共5页Communications Technology

基　　金：国家自然科学基金(62301477)。

摘　　要：终端直通(Device-to-Device,D2D)网络承载的去中心化学习系统可以在不依赖中心服务器的情况下,利用分布式的数据与算力完成常见的神经网络模型训练,以支撑如车联网、工业互联网等场景的边缘智能应用。然而,系统中终端算力、数据和信道等资源的异构性及D2D网络通信资源的局限性给模型训练的性能带来较大的挑战。为此,从中继选择、网络拓扑优化、簇头选择及广播速率优化等方面阐述了可以解决上述问题的关键技术,并对未来的研究方向进行了概述,如新型的模型汇聚算法及动态资源管理算法等。最后,总结指出需要在上述的研究方向上进一步探索,以提高基于D2D网络的去中心化学习的效率。The decentralized learning system carried by the D2D(Device-to-Device)network can complete common neural network model training by using distributed data and computing power without relying on a central server,so as to support edge intelligent applications in scenarios such as the Internet of Vehicles and the Industrial Internet.However,the heterogeneity of computing power,data,and channel resources in the system as well as the limitations of D2D network communication resources poses great challenges to the performance of model training.For this reason,the key technologies that can solve the above problems are explained from the aspects of relay selection,network topology optimization,cluster head selection,and broadcast rate optimization,and future research directions,such as novel model convergence algorithms and dynamic resource management algorithms,are outlined.Finally,the conclusion of the paper points out that further exploration is needed in the above research directions to improve the efficiency of decentralized learning based on D2D networks.

关 键 词：D2D网络 去中心化学习 中继选择 网络拓扑优化 

分 类 号：TN929.5[电子电信—通信与信息系统]