Intelligent reflecting surface (IRS) can improve a communicationenvironment by controlling the reflection phase of theradio waves. There are some works attempting to optimize theperformance of the IRS-aided MISO system. In [1] and [4], theyproposed two approaches to joint design beamforming (BF) at thebase station (BS) and phase shifts (PS) with perfect channel stateinformation (CSI). In [2] and [3], they proposed two deep learning(DL) based methods for only predicting the phase shifts vector atIRS. Unlike the reported works, in this report, we propose a DLbased method without CSI. We formulate the joint design problemto maximize the achievable rate (Rate) of the user and develop atwo-step deep neural network (DNN). We estimate the angle ofdeparture (AoD) of IRS in the first step, predict CSI by the AoDand design the BF and PS vector in the second step. Simulationresults show that comparing with two benchmarks, the proposedmethod has a better rate performance.
展开▼