GAN-Based Channel Estimation for IRS-Aided Communication Systems

This paper proposes a generative adversarial network (GAN) based channel estimation scheme for intelligent reflecting surface (IRS)-aided single-input multiple-output (SIMO) communication systems. The proposed novel GAN-based deep learning technique is efficient to estimate channels in IRS-aided wireless communication systems with high accuracy. The generator of GAN can reproduce data whose distributions are similar to the actual underlying channel. Consequently, the proposed approach does not require the statistical distribution of the underlying channel to be known in advance. Simulation results prove that the proposed GAN-based channel estimation approach outperforms the conventional least square estimation (LSE) approach significantly in terms of estimation accuracy as well as provides better performance than a fully connected deep neural network (DNN) and convolutional neural network (CNN)-based methods.

I. Introduction

In order to support the rising demand for ubiquitous wireless connectivity anywhere in the upcoming Internet-of-Everything (IoE) era, along with the soaring data-hungry applications development, fifth-generation (5G) cellular networks may not be adequately efficient to meet the demands in terms of capacity. This observation leads the researchers to carry out cutting-edge research to explore newer dimensions in upcoming sixth-generation (6G) cellular technology. Intelligent reflecting surface (IRS) is a key enabler of data transmission technology, with a vision to be deployed in 6G cellular communication systems to significantly enhance spectral efficiency. IRS is the advanced version of massive multiple-input multi-output (mMIMO) data transmission system [1], which is the prime transmission technology in 5G cellular networks. IRS is a controllable metasurface comprised of a large number of passive reflecting elements (PREs) that use very little power to control the phase and/or amplitude changes of incident signals to the IRS [2], [3], [4], [5].