基于自适应聚合时间的半同步联邦资源优化算法  

Semi-synchronous Federated Resource Optimization Algorithm Based on Adaptive Aggregation Time

作　　者：李铁 庄琲 林尚静 韩志博 LI Tie;ZHUANG Bei;LIN Shangjing;HAN Zhibo(Beijing Key Laboratory of Work Safety Intelligent Monitoring,School of Electronic Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学电子工程学院安全生产智能监控北京市重点实验室,北京100876

出　　处：《无线电通信技术》2024年第4期696-703,共8页Radio Communications Technology

基　　金：湖北省工程研究中心2024年开放课题。

摘　　要：联邦学习是一种高效的分布式机器学习方法,其中多个设备使用自己的本地数据进行分布式模型训练,不需要交换本地数据,仅通过交换模型参数来构建共享的全局模型,从而保护用户的隐私。为了平衡模型性能和通信延迟,在半同步联邦学习场景下,使用权重参数建立了一个最小化模型性能和聚合时间的加权和的优化问题。优化变量包括进行全局模型更新的聚合时间、用户调度以及参与上传的用户的带宽和发射功率,通过使用交替方向乘子法(Alternating Direction Method of Multipliers, ADMM)将所提混合整数非线性规划(Mixed Integer Non-Linear Programming, MINLP)问题分解为两个子问题进行求解。仿真实验证明,所提算法能够以牺牲4%模型性能为代价,降低73%的聚合时间,显著提高了通信效率。Federated learning is an efficient distributed machine learning method,in which multiple devices use their own local data for distributed model training.There is no need to exchange local data.Instead,it only need to build a shared global model by exchanging model parameters,thereby protecting the user’s privacy.In order to balance model performance and communication delay,in the semi-synchronous federated learning scenario,an optimization problem that minimizes the weighted sum of model performance and aggregation time is established using weight parameters.Optimization variables include the aggregation time for global model updates,user scheduling,and the bandwidth and transmit power of participating upload users.The proposed Mixed Integer Non-Linear Programming(MINLP)problem is decomposed into two sub-problems to solve using Alternating Direction Method of Multipliers(ADMM).Simulation experiments prove that the proposed algorithm can reduce the aggregation time by 73%at the expense of 4%model performance,and significantly improve communication efficiency.

关 键 词：物联网 半同步联邦学习 用户调度 资源优化 

分 类 号：TN92[电子电信—通信与信息系统]