Deep Learning based Hybrid Beamforming Design for IRS-Aided MIMO Communication

In this letter, we study an intelligent reflecting surface (IRS)-aided multiple-input multiple-output (MIMO) communication system in which a transmitter equipped with hybrid precoding architecture communicates to a receiver with hybrid combining architecture via an IRS. We propose a deep learning (DL) based approach to jointly learn and predict the features of the IRS reflection matrix and the hybrid beamformers while maximizing the spectral efficiency. To accomplish this, we first propose the alternating direction method of multipliers (ADMM) algorithm to design the IRS matrix. This matrix is then used to form an effective channel that serves as input to a two-stage neural network. The neural networks are trained sequentially to estimate the IRS matrix, as well as the hybrid precoding and combining matrices. Simulation results verify that our proposed model achieves a performance that is close to the full-digital based scheme; achieves comparable performance to some state-of-the-art algorithms while with reduced computational complexity; and outperforms the scheme implemented without an IRS.