Deep Learning to Predict the Number of Antennas in a Massive MIMO Setup Based on Channel Characteristics

Abstract

Deep learning (DL) solutions learn patterns from data and exploit knowledge gained in learning to generate optimum case-specific solutions that outperform pre-defined generalized heuristics. With an increase in computational capabilities and availability of data, such solutions are being adopted in a wide array of fields, including wireless communications. Massive MIMO is expected to be a major catalyst in enabling 5G wireless access technology. The fundamental requirement is to equip base stations with arrays of many antennas, which are used to serve many users simultaneously. Mutual orthogonality between user channels in multiple-input multiple-output (MIMO) systems is highly desired to facilitate effective detection of user signals sent during uplink. In this paper, we present potential deep learning applications in massive MIMO networks. In theory, an infinite number of antennas at the base station ensures mutual orthogonality between each user’s channel state information (CSI). We propose the use of artificial neural networks (ANN) to predict the practical number of antennas required for mutual orthogonality given the variances of the user channels. We then present an analysis to obtain the practical value of antennas required for convergence of the signal-to-interference-noise ratio (SINR) to its limiting value, for the case of perfect CSI. Further, we train a deep learning model to predict the required number of antennas for the SINR to converge to its limiting value, given the variances of the channels. We then extend the study to show the convergence of SINR for the case of imperfect CSI.