Scholarlink

With growing concerns over data privacy, federated learning (FL) enables collaborative model training without exposing raw data. However, FL faces persistent challenges in efficient client selection and effective resource utilization, especially under heterogeneous and Non-IID conditions. This paper presents a reinforcement learning-based client selection strategy leveraging a Dueling Double Deep Q-Network (DD-DQN) to enable adaptive and informed decision-making. Additionally, a novel aggregation mechanism is proposed, which assigns dynamic weights based on normalized Q-values, local training loss, and client data volume to improve aggregation accuracy. Experiments conducted on CIFAR-10, CIFAR-100, and MNIST under Non-IID partitions demonstrate that the proposed method improves global model accuracy by approximately \(1.0\%\) on average compared to FedAvg and FLASH-RL, while also achieving more stable convergence and higher system efficiency. These results validate the effectiveness of the proposed approach in enhancing federated learning performance under challenging data distributions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Reinforcement Learning-Based Client Selection Strategy for Efficient Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mingzhi Jiang,
  • Ziqi Song,
  • Chen Cui,
  • Shen Wang

摘要

With growing concerns over data privacy, federated learning (FL) enables collaborative model training without exposing raw data. However, FL faces persistent challenges in efficient client selection and effective resource utilization, especially under heterogeneous and Non-IID conditions. This paper presents a reinforcement learning-based client selection strategy leveraging a Dueling Double Deep Q-Network (DD-DQN) to enable adaptive and informed decision-making. Additionally, a novel aggregation mechanism is proposed, which assigns dynamic weights based on normalized Q-values, local training loss, and client data volume to improve aggregation accuracy. Experiments conducted on CIFAR-10, CIFAR-100, and MNIST under Non-IID partitions demonstrate that the proposed method improves global model accuracy by approximately \(1.0\%\) on average compared to FedAvg and FLASH-RL, while also achieving more stable convergence and higher system efficiency. These results validate the effectiveness of the proposed approach in enhancing federated learning performance under challenging data distributions.