Deep Neural Network-based Phase Reconfiguration for Spectral Efficient IRS-enabled mmWave Communication System

Intelligent Reflecting Surface (IRS) is emerging as an exciting new technology enabling communication beyond $\mathbf{5 G}$. IRS can maneuver the direction of radio signal propagation as per requirement, introducing intelligent beamforming in the channel. Jointly acquiring the channel state information (CSI) and optimal phase matrix of the IRS has been one of the challenges in IRS-based communication. Acquisition of the phase matrix of the IRS for the estimated CSI leads to a complicated optimization problem with a large signaling overhead. State-of-the-art optimization techniques to resolve this problem are found to be highly complex. This work proposes a deep learning-based approach like deep neural networks (DNNs) to find the phase matrix concerning IRS to optimize the spectral efficiency (SE) of the IRS-enabled millimeter-wave (mmWave) system. However, the CSI of the system is obtained by a sparse Bayesian learning (SBL) scheme. The DNN predicts the optimal phase matrix for the estimated CSI. An intelligent grouping of the IRS cells is also proposed to reduce the training overhead. It is further shown that an effective grouping of cells will reduce the computational complexity compared to non-grouping without significant SE loss.