Enhancing Adaptive Behavior in Federated Learning: The AdaptiveMesh Algorithm for Interactive Mobile and Bandwidth-Limited Resource-Constrained Wireless Environments

Abstract

Federated learning (FL) enables decentralized model training while preserving data privacy, but its application in resource-constrained environments faces challenges, including computational load, slowing down CPU speed, and network bandwidth limitations. This paper introduces and evaluates the AdaptiveMesh (Adaptive Federated Learning for Wireless Mesh Environments) algorithm, an adaptive approach designed to optimize resource usage in FL scenarios, with a focus on resource-constrained Internet of Things (IoT) devices, mobile applications, and commercial cloud systems. We compare the performance of AdaptiveMesh with a naive non-adaptive algorithm, focusing on key metrics such as CPU utilization, device temperature, accuracy, training time, and network bandwidth. Our study examines these algorithms’ effectiveness on both physical devices, including Raspberry Pis, mini PCs, and laptops, as well as virtual machine instances in the Google Cloud Platform. The results demonstrate that the AdaptiveMesh algorithm dynamically adjusts training parameters in response to real-time device metrics, thereby optimizing resource usage, preventing device overload, and reducing training time under constrained conditions, whereas the naive approach often leads to device inefficiencies and potential temporary shutdowns, as devices attempt to manage heat or workload by throttling performance. This comparison underscores the importance of adaptive mechanisms in FL, especially in resource-limited environments.