Machine learning based power control in cellular and cell-free massive MIMO systems

TL;DR

This paper explores how machine learning can improve power control in large wireless networks, comparing it to traditional methods.

Contribution

The paper introduces a new metric to evaluate machine learning performance in power control across different network architectures.

Findings

DNN-based power control can match or outperform traditional methods in spectral efficiency.

The proposed metric effectively captures performance differences under varying network conditions.

Machine learning offers lower latency in dense and real-time wireless networks.

Abstract

Effective power control (PC) is essential for optimizing performance in large-scale multiple-input multiple-output (mMIMO) networks. Traditional methods such as the weighted minimum mean square error (WMMSE) algorithm offer reliable estimates but require substantial computational overhead This study examines PC in mMIMO systems, focusing on aggregate spectral efficiency (sum SE) and the per-user SE cumulative distribution function (CDF). This investigation explores the impact of factors such as the number of UEs, access points/base stations (APs/BSs), and deep neural network (DNN)-based PC implementations in both cellular (CL) and cell-free (CF) architectures. We introduce a new metric (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$\:\varDelta\:\mathrm{A}\mathrm{U}\mathrm{C}$$\end{document} ) the area between the per-user SE CDFs of the DNN-based PC and the WMMSE baseline - as a compact and interpretable measure of ML versus optimization performance under deployment scaling. To the best of our knowledge, this is the first paper to systematically apply this metric across both cellular and cell-free mMIMO architectures while varying AP/BS count, antenna count, user density, and dataset size. By combining this metric with RMSE, sum-rate change, and execution-time analysis (Figs. 1, 2, 3, 4, 5 and 6, Table 6), we provide prescriptive guidance on when DNN-based PC not only matches but also outperforms WMMSE in both performance and real-time latency, enabling practical deployment in dense and low-latency networks.

Figures