Scholarlink

<p>With the growing volume of data collected on devices and the growing importance of training models directly on these devices, Federated Learning (FL) was introduced. FL enables devices to train a shared model without exchanging raw data while respecting the privacy of data. However, FL faces some challenges including bandwidth limitations and device heterogeneity, hindering simultaneous updates on all devices. In this paper, a lightweight dual-reinforcement learning-based method is proposed for the optimal selection of participating devices in the learning process, as well as determining their level of participation. The proposed mechanism selects the most appropriate devices during communication rounds to improve validation accuracy and minimize rounds. Also, it determines the participation level of each selected device in every round based on their characteristics. According to the experiments, the proposed method can decrease latency compared to FedAvg and FLASH algorithms, while also elevating accuracy.</p>

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

AdaptFedDS: adaptive federated learning edge devices selection by using a light-dual reinforcement learning approach

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Fazeleh Tavassolian,
  • Mahdi Abbasi,
  • Abbas Ramezani,
  • Atefeh Salimi,
  • Amir Taherkordi,
  • M. Reza Khosravi

摘要

With the growing volume of data collected on devices and the growing importance of training models directly on these devices, Federated Learning (FL) was introduced. FL enables devices to train a shared model without exchanging raw data while respecting the privacy of data. However, FL faces some challenges including bandwidth limitations and device heterogeneity, hindering simultaneous updates on all devices. In this paper, a lightweight dual-reinforcement learning-based method is proposed for the optimal selection of participating devices in the learning process, as well as determining their level of participation. The proposed mechanism selects the most appropriate devices during communication rounds to improve validation accuracy and minimize rounds. Also, it determines the participation level of each selected device in every round based on their characteristics. According to the experiments, the proposed method can decrease latency compared to FedAvg and FLASH algorithms, while also elevating accuracy.