Hybrid Precoding Scheme in Millimeter Wave Massive MIMO Based on Stochastic Gradient Descent

Abstract

As one of the key technologies of 5G, the precoding can be used to compensate for the path loss and increase the capacity of massive millimeter wave (mmWave) multiple-input-multiple-output (MIMO) systems. However, the traditional digital precoding has the disadvantages such as high computational complexity, high hardware cost and high power-loss due to the use of a large number of radio frequency (RF) chains. This paper proposes a hybrid precoding scheme based on the stochastic approximation with Gaussian (SAG) method of deep learning. By approximating the objective function with Gaussian smoothing, we can use the gradient descent scheme to obtain the required matrix. Compared with the optimal precoding, this method significantly reduces the computational complexity though it incurs slight loss of spectrum efficiency (SE) compared with the optimal precoding. Simulation results show that the proposed scheme outperforms the state of art when the number of data streams is slightly smaller than the number of RF chains.