Clustered Federated Learning for Massive MIMO Power Allocation

Power allocation in cell-based massive MIMO networks has been studied extensively in recent years. One major bottleneck for traditional optimization techniques for massive MIMO power allocation is the processing time. Recently, federated learning (FL) has been leveraged to address this challenge, by training a machine learning model for optimal power allocation, where the model is trained using the channel dynamics across distributed parties. However, current FL mechanisms train a single model to serve all users, which often fails to perform consistently across all scenarios, particularly when the training dataset exhibits high heterogeneity. To address the demands of real-time systems under heterogeneous network conditions, we propose clustered FL for massive MIMO power allocation, where the training mechanism is aimed at achieving max-min fairness of downlink spectral efficiency across heterogeneous network dynamics. To do so, personalized models are trained for different clusters of user equipments (UEs) with similar channel dynamics. A sample-efficient contextual multi-armed bandit (CMAB) mechanism is then implemented to accelerate the clustered FL training process. Our experiments demonstrate that the proposed distributed learning framework improves the convergence rate significantly compared to heterogeneity-agnostic FL mechanisms, while enhancing robustness to noisy datasets.