Passive Beamforming Design of IRS-Assisted MIMO Systems Based on Deep Learning

Abstract

In the intelligent reflecting surface (IRS)-assisted MIMO systems, optimizing the passive beamforming of the IRS to maximize spectral efficiency is crucial. However, due to the unit-modulus constraint of the IRS, the design of an optimal passive beamforming solution becomes a challenging task. The feature input of existing schemes often neglects to exploit channel state information (CSI), and all input data are treated equally in the network, which cannot effectively pay attention to the key information and features in the input. Also, these schemes usually have high complexity and computational cost. To address these issues, an effective three-channel data input structure is utilized, and an attention mechanism-assisted unsupervised learning scheme is proposed on this basis, which can better exploit CSI. It can also better exploit CSI by increasing the weight of key information in the input data to enhance the expression and generalization ability of the network. The simulation results show that compared with the existing schemes, the proposed scheme can effectively improve the spectrum efficiency, reduce the computational complexity, and converge quickly.