
Mohammad Saroughi, Mohammad Fathi · Scientific Reports 2026 · 2026
DOI: 10.1038/s41598-026-72922-z
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Wireless federated learning (FL) has emerged as a promising solution to generate a global model from distributed local models on wireless smart clients. However, it faces challenges for heterogeneous clients with imbalanced and non-IID data distributions. These challenges lead to straggler clients, inefficient computation and communication resource allocation, and thereby resulting in unfavorable learning models. To address these issues, this paper formulates the long-term convergence of a wireless FL system in the presence of heterogeneous clients using a stochastic optimization problem. The Lyapunov drift-plus-penalty framework is adopted to decompose the long-term optimization into a sequence of tractable per-round subproblems. This decomposition allows for solving client scheduling, communication and computation resource allocation, and the number of local training epochs per client. Notably, an additive-increase multiplicative-decrease (AIMD) mechanism is presented to adaptively adjust the number of local training epochs per client at each global round. This mechanism mitigates the straggler clients effect caused by imbalanced data distributions, weak channel conditions, and limited computation capacity. Extensive experiments demonstrate that the proposed algorithm outperforms existing benchmarks in terms of convergence, learning accuracy, and long-term energy stability for non-IID data distributions.
No comments yet — start the discussion below.