Dynamic AP Clustering and Power Allocation for CF-mMIMO-Enabled Federated Learning Using Multi-Agent DRL

Federated learning (FL) is recognized as a pivotal paradigm for 6G, offering decentralized model training without compromising data privacy. Recent works have proposed deploying FL in cell-free massive MIMO (CF-mMIMO) networks for reliable model transmission between FL clients and the server. Nevertheless, the problem of simultaneous access point (AP) clustering (i.e., dynamically forming AP groups to facilitate client-server communication) and transmit power allocation has not been thoroughly investigated. Furthermore, most existing solutions do not simultaneously consider the fast decision-making requirements brought by user mobility and the scalability of solutions in large-scale networks. To address this gap, we propose DACPA, a multi-agent deep reinforcement learning (DRL)-based scheme that accounts for client mobility (walking speed) and heterogeneous computing capabilities. DACPA strategically assigns each client a customized AP cluster and corresponding transmit power configuration, thereby optimizing model update latency. Extensive simulation results demonstrate the superior performance of DACPA in terms of convergence stability, spectral efficiency, global model update latency, and average energy consumption.