Beamforming design via machine learning in intelligent reflecting surface-aided wireless communication

Beamforming design is a pivotal issue in intelligent reflecting surface (IRS) assisted wireless communication. The capacity of the classic regular IRS-based schemes with a few numbers of elements is not convincing. In order to deal with this issue and gain spatial degrees of freedom, we offer an irregular IRS architecture and investigate a weighted sum rate (WSR) maximization problem so as to enhance the system capacity. WSR maximization subject to the transmit power is a nonconvex problem and confronting with this issue is arduous. Despite some existing approaches exhibit proper results, several defects such as computational complexity, acquiring local optimal solutions and so on are still controversial. In this paper, unlike these conventional techniques, a machine learning (ML) inspired beamforming design is presented. In the offered method, the goal is to employ a deep learning (DL) model which, via utilizing only omni or quasi-omni beam patterns, learns how to predict the precoding vectors. In order to improve the support of this system, instead of hiring position information, uplink received signal are used for beamforming prediction. In addition, a joint optimization method was considered in order to iteratively handle the optimization problem. Moreover, other fruitful advantages such as negligible training overhead and no need for training before deployment are attained. Simulation results, based on accurate ray tracing, affirm that the offered method access premiere performance compared with conventional beamforming approaches.