JultikaUniversity of Oulu repository

Abstract

Federated learning (FL) is a promising decentralized training method for on-device machine learning. Yet achieving a performance close to a centralized training via FL is hindered by the client-server communication. In this work, a novel joint client scheduling and resource block (RB) allocation policy is proposed to minimize the loss of accuracy in FL over a wireless system with imperfect channel state information (CSI) compared to a centralized training-based solution. First, the accuracy loss minimization problem is cast as a stochastic optimization problem over a predefined training duration. In order to learn and track the wireless channel under imperfect CSI, a Gaussian process regression (GPR)-based channel prediction method is leveraged and incorporated into the scheduling decision. Next, the client scheduling and RB allocation policy is derived by solving the aforementioned stochastic optimization problem using the Lyapunov optimization framework. Then, the aforementioned solution is extended for scenarios with perfect CSI. Finally, the proposed scheduling policies for both perfect and imperfect CSI are evaluated via numerical simulations. Results show that the proposed method reduces the accuracy loss up to 25.8% compared to FL client scheduling and RB allocation policies in the existing literature.