Channel Estimation Enhancement With Generative Adversarial Networks

Improving the accuracy of channel estimation is a significant topic in the context of wireless communications. For training-based channel estimations, increasing the length of a training sequence may improve the accuracy of channel estimation but causes a higher overhead. Nevertheless, this paper shows that benefiting from generative adversarial networks (GANs), which is an emerging deep learning framework, the accuracy of channel estimation can be improved without transmitting a longer training sequence. To this end, this paper proposes a GAN-based channel estimation enhancement algorithm, where GANs are trained online with receive sequence so as to obtain a longer mimic sequence and enhance channel estimation. In order to address the problem of improving the training stability and the learning ability of GANs, this paper proposes a novel framework by integrating a conditional GAN with an improved Wasserstein GAN. Furthermore, a strategy based on a lookup table is proposed to alleviate overfitting that may occur during the training of GANs. Simulation results indicate that the proposed GAN-based channel estimation enhancement algorithm can benefit the conventional training-based channel estimation, yielding lower relative error performance, especially in the low SNR regions.