Federated Learning for Massive MIMO Power Allocation

This work proposes the use of federated learning (FL) for deep learning-based power allocation in massive MIMO systems. The conventional approach for learning-based power allocation builds on centralized training, where training data is collected at a single node which performs training. On the other hand, doing so requires extensive data exchange and processing at the central node, leading to high communication and computational overhead. In contrast, FL enables distributed training without the need to centralize training data, allowing for parallelization of the training load (hence speeding up training), as well as efficient use of the available communication resources. In this work we study FL for massive MIMO power allocation with heterogeneity-aware user sampling, and show that FL-based power allocation can achieve performance comparable to that of the centralized approach while significantly reducing communication overhead. Our experiments demonstrate the effectiveness of the proposed distributed learning strategies in terms of both convergence rate and spectral efficiency.