Enhancing Freshness and Energy Efficiency in Asynchronous Federated Learning: An AoI-Aware Online Approach
Jiebei Shi, Jie Chen, Ying Chen, Jiwei Huang · 2025
Federated Learning (FL) is an emerging distributed learning model that enables multiple devices to collaboratively train a shared model. However, in practical scenarios, device heterogeneity can cause outdated model parameters in synchronous FL. To address this issue, asynchronous updates are introduced to mitigate delays in parameter updates. However, asynchronous updates can lead to imbalanced updates, resulting in excessive energy consumption and reduced model accuracy. To tackle these challenges and ensure both parameter freshness and efficient resource utilization, this paper introduces Age of Information (AoI) as a metric to measure model freshness and optimizes AoI to ensure parameter freshness. Additionally, we consider Age of Peak (AoP) and impose a constraint on peak age to prevent prolonged periods of stale parameters, which lead to inaccurate models. Based on this, we formulate an optimization problem involving update decisions and resource allocation, aiming to optimize AoI and energy consumption while imposing a constraint on AoP. Since asynchronous updates and training delays are both stochastic and dynamic, we employ stochastic optimization techniques to decompose the problem into two subproblems and design a low-complexity algorithm using optimization theory. Extensive experimental results demonstrate the proposed framework's effectiveness in improving parameter freshness and energy efficiency.