MAB-Based Channel Scheduling for Asynchronous Federated Learning in Non-Stationary Environments

--Federated learning enables distributed model training across clients under central coordination without raw data exchange. However, in wireless implementations, frequent parameter updates between the server and clients create significant communication overhead. While existing research assumes either known channel state information (CSI) or that the channel follows a stationary distribution, practical wireless channels exhibit non-stationary characteristics due to channel fading, user mobility, and hostile attacks in telecommunication networks. The unavailability of both CSI and time-varying channel distribution can lead to unpredictable failures in parameter transmission, exacerbating clients staleness thus affecting model convergence. T o address these challenges, we propose an asynchronous federated learning scheduling framework for non-stationary channel environments, designed to reduce clients staleness while promoting both fair and efficient communication and aggregation. This framework considers two channel scenarios: extremely non-stationary and piecewise-stationary channels. Age of Information (AoI) serves as a metric to quantify client staleness under non-stationary conditions. Firstly, we perform a rigorous convergence analysis to explore the impact of AoI and per-round client participation on learning performance. The channel scheduling problem in the non-stationary scenario is addressed and formulated within the multi-armed bandit (MAB) framework and we derive the achievable theoretical lower bounds on the AoI regret. Based on this framework, we propose corresponding scheduling strategies for the two non-stationary channel scenarios that leverage the foundations of the GLR-CUCB and M-exp3 algorithms, along with derivations of their respective upper bounds on AoI regret. Additionally, to address the issue of imbalanced client updates in non-stationary channels, we introduce an adaptive matching strategy that incorporates considerations of marginal utility and fairness of clients. Simulation results demonstrate that the proposed algorithm achieves sub-linear growth in AoI regret, accelerates federated learning convergence, and promotes fairer aggregation. HE proliferation of Internet of Things (IoT) devices and the rise of edge computing have resulted in an increasingly decentralized distribution of data across end devices, such as smartphones and sensors. In traditional centralized machine learning approaches, data consolidation at a single location is required, which raises privacy concerns and incurs significant communication overhead.