Deep Learning Based Beam Allocation in Switched-Beam Multiuser Massive MIMO Systems

Abstract

This paper purposes beam allocation problem for a massive MIMO system using Deep Neural Network (DNN). We use supervised Machine Learning (ML) algorithm for Beam-Selection and Switching (BSS) that provides best communication performance to the receiver. These beams are narrow and highly directed that are produced by Butler method to achieve high gain. The brute-force and suboptimal search methods are used to allocate beams that have run time complexity but the proposed ML algorithm has training complexity that achieves nearly the same average sum data rate and on the run time it takes only query data to predict beams. Interpretation of BSS is achieved by multiclass classification that is to compare with the labeled training data of beams for prediction. We estimate channel equation according to the position of the user for training and use different KPIs for labeling that helps in the selection of optimal beam for the user. Simulation results show that the proposed scheme is able to predict and allocate beams with the accuracy of 91.6 to 97.7% that provides best communication performance to the receivers. We further explore the effect on the system performance in terms of average sum data rate by increasing or decreasing the number of users compared to beams.