MDL-AltMin: A Hybrid Precoding Scheme for mmWave Systems With Deep Learning and Alternate Optimization

Abstract

The hybrid precoding structure composed of analog and digital precoders has received increasing attention in mmWave massive <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">multiple-input multiple-output</i> (MIMO) systems because it can balance the energy consumption and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">spectral efficiency</i> (SE). However, it is challenging to obtain the optimal hybrid precoding scheme by joint optimization with lower computational complexity. This letter proposes a hybrid precoding scheme based on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">model-driven deep learning and alternate minimization</i> (MDL-AltMin), which is implemented by alternately solving analog precoder and digital precoder. During the alternation, we design an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">analog precoding network</i> (AP-Net) to solve the phase shift network in analog precoder with the goal of maximizing SE. The digital precoder is solved by the Lagrange multiplier method. In each alternate optimization process, the criteria for convergence is to minimize the error between the hybrid precoder and the optimal fully digital precoder. The simulation results show that the SE of our proposed scheme is very close to the fully digital precoding scheme based on singular value decomposition with lower computational complexity.