Low-complexity deep learning-based beamforming in MISO systems

Abstract

This study proposes a low-complexity deep learning-based beamforming neural network (BFNN) for massive multiple-input single-output (MISO) systems. We adopt an unsupervised learning-based convolutional neural network (CNN) model. The network is trained to obtain an analog phase shifters (PSs)-based beamforming vector of a given user by maximizing the system spectral efficiency (SE) while maintaining the transmitted power constraint. The channel state information (CSI) for millimeter wave (mmWave) channel and signal-to-noise-ratio (SNR) are used as inputs to the network. We also proposed a novel input feeding arrangement to the network and assessed its performance by using different input data representations. Simulation results show that the CNN-BFNN has the lowest complexity compared to a fully connected neural network (FCNN) and the existing conventional algorithms. Furthermore, the CNN model with fast Fourier transform (FFT) input provides the highest SE performance among all other input data representations.