Reinforcement Learning-Based Device Scheduling for Renewable Energy-Powered Federated Learning
Yangguang Cui, Kun Cao, Tongquan Wei
Abstract
Due to its unique privacy protection advantages, emerging federated learning (FL) is regarded as a significant technique to enable Industry 4.0. However, the industrial deployment of FL encounters the primary obstacles of limited device energy and system communication resources. Nowadays, renewable energy-powered devices have been deployed in various industrial fields to tackle the challenges of unsustainable and limited energy of battery-powered devices. Inspired by this, this article proposes a novel FL protocol to groundbreakingly improve the performance of renewable energy-powered FL systems. Specifically, with the underlying theory of FL as the guide, the proposed protocol features a reinforcement learning-based device scheduling solution to adapt to intermittent renewable energy supply. Following this device scheduling solution, an integer linear programming-based bandwidth management scheme is introduced to optimize communication efficiency. Experimental results on two representative data distribution situations demonstrate that compared with the state-of-the-art schemes, our FL protocol can boost up to 36.63% and 50.99% accuracy, respectively.