智能系统学报2024，Vol.19Issue(6)：1552-1561,10.DOI:10.11992/tis.202305054
基于数据质量评估的高效强化联邦学习节点动态采样优化
Client dynamic sampling optimization of efficient reinforcement federated learning based on data quality assessment
摘要
Abstract
Communication cost and efficiency are the key bottlenecks of federated learning due to the existence of sys-tem and statistical heterogeneities.Selecting only a subset of clients to perform model updates and aggregation can ef-fectively reduce communication costs among numerous participants.However,biased selection and uneven distribution of data quality across clients pose additional challenges to client sampling methods.Therefore,this paper proposes a method for client dynamic sampling optimization in efficient reinforcement federated learning based on data quality as-sessment(RQCS)to address the aforementioned issues.This method evaluates data quality on clients using a contribu-tion index based on the Shapley value and intelligently selects clients with high data quality for each round of federated learning.By leveraging reinforcement learning,the method aims to offset the bias introduced by uneven data quality dis-tribution,accelerate model convergence,and improve model accuracy.Experiments on the MNIST and CIFAR-10 data-sets show that the proposed algorithm not only reduces communication costs but also further accelerates convergence speed and achieves better performance in model accuracy compared to other algorithms.关键词
联邦学习/深度强化学习/客户端动态采样/贡献指数/数据质量/通信效率/沙普利值/模型精度Key words
federated learning/deep reinforcement learning/client dynamic sampling/contribution index/data quality/communication efficiency/Shapley value/model accuracy分类
计算机与自动化引用本文复制引用
赵泽华,梁美玉,薛哲,李昂,张珉..基于数据质量评估的高效强化联邦学习节点动态采样优化[J].智能系统学报,2024,19(6):1552-1561,10.基金项目
国家自然科学基金项目(62192784,U22B2038,62172056,62272058) （62192784,U22B2038,62172056,62272058）
中国人工智能学会-华为MindSpore学术奖励基金项目(CAAIXSJLJJ-2021-007B). （CAAIXSJLJJ-2021-007B）
