DPC-CNN Algorithm for Multiuser Hybrid Precoding With Dynamic Structure

Abstract

This paper presents a dynamic partially connected (DPC) structure-based convolutional neural network (CNN) hybrid precoding with multi-user optimization algorithm. In the proposed algorithm, a multi-output CNN framework is constructed to simultaneously optimize the phase shifter and switch precoders, including custom ‘Out’ layer, deep neural network (DNN)-based analog phase shifter subnetwork, namely DNN-Fps, and DNN-based switch subnetwork, called DNN-Fs. Specifically, the DNN-Fps is designed to obtain the vectorized phase shifter precoder with constant modulus constraint. The DNN-Fs is utilized to output the vectorized switch precoder with the binary constraint. The ‘Out’ layer is defined to obtain the vectorized analog precoder with constant modulus and binary constraints. Moreover, to further improve the real-time performance of hybrid precoding, a dynamic pruning technique is applied to remove the redundant parameters for the DPC-CNN model. Finally, the DPC-CNN is trained using the loss function with the residual between the vectorized analog precoders of the fully connected (FC) and DPC structures. Theoretical analyses and simulation experiments show that compared to the FC and partially connected structures, the proposed DPC-CNN hybrid precoding algorithm can achieve a balance between spectral efficiency and energy efficiency with less execution time.