Adaptive Federated Learning in Heterogeneous Wireless Networks with Independent Sampling

• DOI: 10.48550/arxiv.2402.10097
• Publication Date: 2024-02-15
• Authors: Jiaxiang Geng; Yanzhao Hou; Xiaofeng Tao; Juncheng Wang; Bing Luo
• Keywords: Federated Learning; Adaptive sampling

Federated Learning (FL) algorithms commonly sample a random subset of clients to address the straggler issue and improve communication efficiency. While recent works have proposed various client sampling methods, they limitations in joint system data heterogeneity design, which may not align with practical heterogeneous wireless networks. In this work, we advocate new independent strategy minimize wall-clock training time FL, while considering both computation. We first derive convergence bound for non-convex loss functions then propose an adaptive bandwidth allocation scheme. Furthermore, efficient algorithm based on upper bounds rounds expected per-round time, heterogeneity. Experimental results under network settings real-world prototype demonstrate that scheme substantially outperforms current best schemes models datasets.