REAFL: Accelerating federated learning with personalized iteration allocation and peer learning

Abstract

In recent years, federated learning (FL) has made significant promise for wireless communication edge computing (WCEC) environments. However, challenges like device heterogeneity, limited edge resources, and Non-IID data disrupt symmetry in data distribution and model updates, which is crucial FL efficiency and performance. To address these issues, we propose Resource-Efficient Accelerating Federated Learning (REAFL), which combines Particle Swarm Optimization (PSO) and Momentum Gradient Descent (MGD) to enhance model training efficiency while maximizing resource utilization in WCEC environments. Our approach innovatively leverages PSO’s inherent self-learning and peer-learning capabilities to facilitate robust local model training on individual devices, subsequently refined by MGD optimization. Through a clear toy example, we demonstrate the importance of dynamically adjusting the local iteration quantities for heterogeneous devices during FL training. Theoretical analysis of REAFL’s convergence under a fixed time budget reveals the relationship between local iteration quantities and the optimal global model. Building on this, we propose a novel FL framework to dynamically adjust the local iteration quantities for devices. Extensive experiments show that REAFL outperforms existing benchmarks, offering improvements in accuracy, resource efficiency, and resilience to Non-IID data distributions.