Unsupervised Deep Learning-Based Hybrid Beamforming in Massive MISO Systems

Hybrid beamforming (HBF) is a promising approach for balancing the hardware cost, training overhead and system performance in massive MIMO systems. Optimizing the HBF through deep learning (DL) has gained considerable attention in recent years due to its potential in dealing with the nonconvex problems. However, existing DL-based HBF methods require wider or deeper neural networks to guarantee training performance, which not only leads to higher complexity in training and deploying, but also increases the risk of over-fitting. In this paper, we propose a low-complexity HBF method based on convolutional neural network (CNN) to solve the spectral efficiency (SE) maximization problem with constant modulus constraint for the analog phase shifters over the transmit power budget in a multiple-input single-output (MISO) system. An unsupervised learning strategy is derived for the constructed CNN to learn to generate feasible beamforming solutions adaptively and thus avoiding any label data when training them. Simulations show its advantages in both SE and complexity over other related algorithms.