Over-the-Air Computing Aided Federated Learning and Analytics via Belief Propagation Based Stochastic Bandits

Abstract

This paper considers a mobile edge cloud (MEC) system where several distributed users collaboratively learn a global model by exploiting the over-the-air computing (AirComp) aided federated learning (FL) mechanism. Previous works focus on the FL training process, ignoring the generalization ability of the trained global model. To overcome this, we propose a novel AirComp-aided FL and federated analytics (FL&FA) framework to improve the generalization ability by making full use of user data and opinions. We first formulate this problem as an online user selection problem. Then, we further model it as a stochastic multi-armed bandit (SMAB) framework, where arms are the decentralized users and rewards are user opinions in FA. To tackle the decentralized feature among users, we put forth a belief propagation-based upper confidence bound (BP-UCB) algorithm to solve this SMAB problem. In addition, we derive an upper regret bound for the BP-UCB algorithm, which increases logarithmically over time. Simulation results demonstrate that the proposed algorithm is close to the optimal solution by less than 3.0% and has a fast convergence rate among existing methods.