Machine Learning-Based Spectrum Efficiency Hybrid Precoding With Lens Array and Low-Resolution ADCs

Hybrid precoding is an important issue in millimeter wave (mmWave) massive multi-input and multi-output (MIMO) system. Specially, energy-saving hybrid precoding architectures and efficient hybrid precoding schemes provide ideas for solving this issue. In this paper, we propose a hybrid precoding/combining architecture that is low-cost and easy to implement. Specifically, a hybrid precoding architecture is realized by the lens sub-arrays at the base station (BS). Moreover, a hybrid combining architecture applies the low-resolution analog-to-digital converters (ADCs) at the front end of the radio frequency (RF) chains at the receiving terminal. Based on the hybrid precoding/combining architecture, the hybrid precoder and combiner are jointly optimized to maximize the spectrum efficiency (SE) in the downlink systems, which is a combinatorial optimization problem due to hardware constraints. The cross-entropy (CE) approach in machine learning (ML) is a simple way to solve the combinatorial optimization problem benefiting from its adaptive update procedure. Therefore, we propose an adaptive hybrid precoder/combiner design scheme (AHDS), in which a hybrid precoding algorithm based on the improved CE (ICE) inspired by ML is adopted to design the optimal hybrid precoder, and an approximate optimization method (AOM) is suggested when designing the hybrid combiner. In general, compared with the existing hybrid design schemes, the proposed AHDS is demonstrated to have significant advantage in SE with low computational complexity.