Generative Adversarial Networks-Based Channel Estimation for Intelligent Reflecting Surface Assisted mmWave MIMO Systems

Channel estimation is a challenging task in intelligent reflecting surface (IRS)-assisted communication systems due to the large amount of passive IRS elements. Recently, deep learning (DL) based channel estimation schemes for multiple-input multiple-output (MIMO) communication systems have achieved remarkable success. However, the performance of channel estimation algorithms still needs to be improved. Meanwhile, the loss functions in traditional DL-based methods are not well designed and investigated. In this paper, we propose a generative adversarial network (GAN) variant based channel estimation method to improve the channel estimation accuracy. Specifically, two DL networks are trained adversarially with the received signals as the conditional input to learn an adaptive loss function. Furthermore, the GAN variant can also learn the mapping from the received signals to the real channels. To improve the training stability of GANs, a loss function is proposed to ensure the correct optimization direction of training the generator. To further improve the estimation performance, we investigate the influence of the hyper-parameter of the loss function on the performance of our model. Our extensive simulation results show that the proposed method outperforms traditional DL-based methods and shows great robustness.