Online Client Association and Adaptive Modulation in Hierarchical Federated Learning

Federated Learning (FL) enables privacy-preserving training, but it often suffers from substantial communication delay. Hierarchical FL (HFL) mitigates this by inserting edge servers (ESs) between clients and cloud servers. Prior HFL work typically optimizes global accuracy under idealized assumptions of reliable links and perfect knowledge of communication and computation capabilities. These assumptions overlook noisy links, where symbol errors corrupt model updates, and the coupling between client association and modulation in online settings. This work aims to minimize the number of training rounds via joint, online control of (i) client-ES association and (ii) perclient adaptive modulation-mode selection. The central challenge is twofold. First, to make effective association decisions under unknown, time-varying computation and channel conditions. Second, to balance upload speed against reliability. We propose BiBandit-HFL, an online algorithm that casts the problem as coupled bandits: a multi-player multi-armed bandits for client-ES association and a per-client multi-armed bandits for modulationmode selection. Simulation results show that BiBandit-HFL has faster convergence compared with online baselines.