Learning-Based Downlink Power Allocation in Cell-Free Massive MIMO Systems

TL;DR

This paper introduces deep learning-based algorithms for downlink power allocation in cell-free massive MIMO systems, achieving scalable and efficient performance with reduced computational complexity compared to traditional optimization methods.

Contribution

It proposes novel neural network models trained on heuristic schemes to enable distributed and scalable power allocation in cell-free massive MIMO systems.

Findings

Deep neural networks can effectively approximate optimal power allocation.

Distributed DNNs improve spectral efficiency using local information.

Clustered DNN models enhance performance with limited global data.

Abstract

This paper considers a cell-free massive multiple-input multiple-output (MIMO) system that consists of a large number of geographically distributed access points (APs) serving multiple users via coherent joint transmission. The downlink performance of the system is evaluated, with maximum ratio and regularized zero-forcing precoding, under two optimization objectives for power allocation: sum spectral efficiency (SE) maximization and proportional fairness. We present iterative centralized algorithms for solving these problems. Aiming at a less computationally complex and also distributed scalable solution, we train a deep neural network (DNN) to approximate the same network-wide power allocation. Instead of training our DNN to mimic the actual optimization procedure, we use a heuristic power allocation, based on large-scale fading (LSF) parameters, as the pre-processed input to the DNN. We train the DNN to refine the heuristic scheme, thereby providing higher SE, using only local information at each AP. Another distributed DNN that exploits side information assumed to be available at the central processing unit is designed for improved performance. Further, we develop a clustered DNN model where the LSF parameters of a small number of APs, forming a cluster within a relatively large network, are used to jointly approximate the power coefficients of the cluster.