摘要
针对节点数据分布差异给联邦学习算法性能带来不良影响的问题,提出了一个基于标签量信息的节点选择算法。算法设计了一个关于节点标签量信息的优化目标,考虑在一定时耗限制下选择标签分布尽可能均衡的节点组合优化问题。根据节点组合的综合标签分布与模型收敛的相关性,新算法降低了全局模型的权重偏移上界以改善算法的收敛稳定性。仿真验证了新算法与现有的节点选择算法相比拥有更高的收敛效率。
Aiming at the problem that the difference of node data distribution has adverse effect on the performance of federated learning algorithm,a node selection algorithm based on label quantity information was proposed.An optimization objective based on the label quantity information of nodes was designed,considering the optimization problem of selecting the nodes with balanced label distribution under a certain time consumption limit.According to the correlation between the aggregated label distribution of selected nodes and the convergence of the global model,the upper bound of the weight divergence of the global model was reduced to improve the convergence stability of the algorithm.Simulation results shows that the new algorithm had higher convergence efficiency than the existing node selection algorithm.
作者
马嘉华
孙兴华
夏文超
王玺钧
谭洪舟
朱洪波
MA Jiahua;SUN Xinghua;XIA Wenchao;WANG Xijun;TAN Hongzhou;ZHU Hongbo(Sun Yat-sen University,Guangzhou 510006,China;Nanjing University of Posts and Telecommunications,Nanjing 210023,China)
出处
《物联网学报》
2021年第4期46-53,共8页
Chinese Journal on Internet of Things
基金
国家重点研发计划(No.2019YFE0114000)
国家自然科学基金资助项目(No.92067201)
江苏省自然科学基金资助项目(No.BK20212001)
广东省基础与应用基础研究基金资助项目(No.2021A1515012631,No.2019A1515011906)。
关键词
联邦学习
节点选择
通信时延
federated learning
node selection
communication delay
