联邦学习网络中基于分层博弈的边缘关联和资源分配  

Hierarchical game-based edge association and resourceallocation in federated learning networks

作　　者：范露露 倪郑威 李职杜 FAN Lulu;NI Zhengwei;LI Zhidu(School of Information and Electronic Engineering,Zhejiang Gongshang University,Hangzhou,310018,P.R.China;School of Communications and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China)

机构地区：[1]浙江工商大学信息与电子工程学院,杭州310018 [2]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《重庆邮电大学学报(自然科学版)》2025年第2期261-272,共12页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：浙江省自然科学基金项目(LQ22F010008);重庆市教委科学技术研究项目(KJQN202300646)。

摘　　要：在分层联邦学习网络中,引入边缘服务器需要面对边缘服务器与终端设备间交互时延和边缘聚合模型质量两大挑战。针对上述问题,利用分层博弈模型对带宽资源分配和边缘关联机制进行分析。在下层,利用演化博弈思想,提出低时延的边缘关联算法,建立终端设备自适应选择边缘服务器的机制;在上层,利用非合作博弈思想,提出平衡模型质量与带宽消耗的资源分配算法,建立边缘服务器自适应调节带宽的机制。仿真结果表明,提出的方法能够智能调节联邦学习网络中的节点,使终端设备达到低延时的边缘关联状态,边缘服务器权衡模型质量与资源消耗达到最佳带宽分配状态,可以为联邦学习网络解决边缘关联和资源分配问题提供一种新的方法。In hierarchical federated learning networks,introducing edge servers needs to consider the edge server-end device interaction latency problem and the edge aggregation model quality problem.To address these above issues,an analysis of the bandwidth resource allocation and edge association mechanism was conducted through a hierarchical game.In the lower layer,a low-latency edge association algorithm was proposed using the evolutionary game theory to establish a mechanism for end devices to select edge servers adaptively.In the upper layer,a resource allocation algorithm balancing the model quality and the bandwidth consumption was proposed using the non-cooperative game theory to establish a mechanism for the edge servers to adjust the bandwidth adaptively.Simulation results show that the proposed method could regulate the nodes intelligently in the federated learning network so that the end devices reached a low-latency edge-associated state,and the edge servers balanced the model quality and resource consumption to reach an optimal bandwidth allocation state,providing a new approach for federated learning networks to solve edge association and resource allocation problems.

关 键 词：分层联邦学习 演化博弈 非合作博弈 边缘关联 资源分配 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]