Scholarlink

Federated Learning (FL) traditionally adopts unbiased client selection, which slows error convergence. Existing biased strategies often add communication overhead or fail under Non-Independent and Identically Distributed (Non-IID) data. We propose an efficient FL framework based on Multi-Armed Bandit (MAB) with two scoring methods—Upper Confidence Bound (UCB) and Thompson Sampling (TS)—to evaluate client contributions. In each round, top-ranked clients are trained with local Gradient Descent (GD), while others update with shuffled Stochastic Gradient Descent (SGD). Experiments show our method achieves comparable accuracy to baseline with fewer rounds and reduced time, and remains robust under heterogeneous data, supporting practical deployment.

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

Multi-Armed Bandit-Based Client Selection for Efficient Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dewei Ning,
  • Yong-Feng Ge,
  • Lili Sun,
  • Yunliang Jiang

摘要

Federated Learning (FL) traditionally adopts unbiased client selection, which slows error convergence. Existing biased strategies often add communication overhead or fail under Non-Independent and Identically Distributed (Non-IID) data. We propose an efficient FL framework based on Multi-Armed Bandit (MAB) with two scoring methods—Upper Confidence Bound (UCB) and Thompson Sampling (TS)—to evaluate client contributions. In each round, top-ranked clients are trained with local Gradient Descent (GD), while others update with shuffled Stochastic Gradient Descent (SGD). Experiments show our method achieves comparable accuracy to baseline with fewer rounds and reduced time, and remains robust under heterogeneous data, supporting practical deployment.