Adaptive Sensor Scheduling for Federated Learning over 6G in-X Subnetworks

Abstract

In the upcoming 6G era, various end devices with data and computing resources can collaborate through federated learning (FL) to achieve the goal of ubiquitous intelligence. Inside any of those ‘X’ (i.e. everything) devices, however, a6G in-X subnetwork that wirelessly uploads training data from onboard sensors to an embedded access point (AP) may become a communication bottleneck during the process of FL. To support communication-efficient FL over in-X subnetworks, this paper proposes an adaptive sensor scheduling algorithm that jointly solves a queue stability problem and an age of information (AoI) optimization problem, based on Lyapunov optimization and MaxRatio scheduling methods. Simulations show that the proposed algorithm achieves higher FL accuracy, while guaranteeing queue stability at APs and better data freshness (quality) for data uploading over in-X subnetworks.