Over-the-Air Federated Learning in User-Centric Cell-Free Networks

This letter presents a new over-the-air federated learning (OTA-FL) system supported by a user-centric cell-free (UCCF) network. We propose a two-level hierarchical deep reinforcement learning (HDRL) framework that minimizes mean squared error (MSE) derived from convergence analysis by jointly optimizing AP-device association (ADA) and power control (PC). Specifically, a soft actor-critic (SAC) with customized data pre-processing is designed for addressing ADA, and a multi-actor-attention-critic (MAAC) with tailored preprocessing and policy networks is designed for handling PC with complex state-action space. Simulations show that our method improves MSE by at least 31% and achieves better OTA-FL convergence than its benchmarks.