基于独立采样的6G无线异构联邦学习优化方法  

Optimization for Wireless Heterogeneous Federated Learning on Independent Sampling

作　　者：耿嘉翔 李博宇 侯延昭[1,2] 崔琪楣 陶小峰[1,2] GENG Jiaxiang;LI Boyu;HOU Yanzhao;CUI Qimei;TAO Xiaofeng(Beijing University of Posts and Telecommunications,Beijing 100876,China;Peng Cheng Laboratory,Shenzhen 518055,China)

机构地区：[1]北京邮电大学,北京100876 [2]鹏城实验室,广东深圳518055

出　　处：《移动通信》2024年第8期90-95,共6页Mobile Communications

基　　金：北京市自然基金海淀原始创新联合基金(L232001)“面向6G‘通信-计算-控制’融合的确定性工业互联网架构与关键技术研究”;国家自然科学基金区域创新发展联合基金“面向工业应用的超低时延高可靠智能无线组网机理与关键技术研究”(U21A20449);中央高校基本科研业务费专项资金资助“面向6G确定性服务的智能资源预留研究”(2242022k60006)。

摘　　要：联邦学习作为新兴的分布式机器学习框架,能够为6G网络提供高效、安全的分布式数据资源使用方案。旨在研究最小化无线异构联邦学习系统的收敛挂钟时间。首先,给出了独立采样下联邦学习的收敛界限;进一步,提出带宽自适应分配方法以解决“掉队者”问题;最后,构建并求解最小化系统总挂钟时间的优化问题,得到最优采样概率。实验结果表明,该方法显著减少了联邦学习的收敛挂钟时间,验证了其实用性和优越性。Federated learning(FL),as an emerging distributed machine learning framework,provides an efficient and secure solution for utilizing distributed data resources in 6G networks.This paper aims to minimize the convergence clock time of heterogeneous wireless federated learning systems.Firstly,a convergence bound of FL is provided under independent sampling.Then,a bandwidth-adaptive allocation method is proposed to address the"straggler"issue.Finally,an optimization problem is formulated to minimize the total system clock time to obtain the optimal sampling probability.Experimental results demonstrate that the proposed method significantly reduces the convergence clock time for federated learning,validating its practicality and superiority.

关 键 词：联邦学习 客户端采样 无线异构网络 

分 类 号：TN929.5[电子电信—通信与信息系统]