Towards Energy-efficient Resource Allocation for Federated Learning in Mobile Edge Computing

Federated learning has become a promising technology that enables edge devices to participate intelligent modeling without sharing data, and thus realizing edge intelligence in mobile edge computing (MEC). In this paper, we propose an energy-efficient resource allocation framework for federated learning in MEC. Different from existing works, we consider the heterogeneous and the dynamic nature (e.g., stragglers, diverging interests, and intermittent drop-out) of edge devices and their effects on the convergence and energy efficiency of federated learning. The proposed framework leverages multi-agent reinforcement learning to enable different devices to flexibly modify their federated learning policies based on the environment and their own status. The convergence and energy efficiency of federated learning can be further improved through collaborative decision-making and mutual compromise among devices. The numerical results showed that the proposed framework could greatly improve the convergence performance of federated learning model compared to baselines while achieving efficient and sustainable use of energy.