Channel estimation and equalisation using generative adversarial networks

Abstract

Channel State Information (CSI) estimators are used in everyday communication systems to estimate channel impairments that affect transmitted data, and to equalise the impaired data to a more accurate state. However, this can be a complex task as channels cause many non-linear impairments to transmitted data. In this study, we construct a simulated environment to investigate the effects of channel impairments on CSI data in Long Term Evolution (LTE) environments. Using the simulated environment, we also generate datasets with which we investigate the ability of several deep learning architectures to estimate CSI. We extend this investigation to adversarial training techniques that have had success on computer vision tasks that are similar to CSI estimation. These trained deep learning networks are evaluated in several wireless communication environments to investigate the effect of adversarial training on network performance. We start this analysis by investigating networks in the Single-In Single-Out (SISO) environment before moving to Multi-Antenna (MA) environments. In this process, we find that the performance of adversarially trained networks in an MA environment deviates from the expected performance indicated in the SISO training environment. Finally, we show that adversarial training has the potential to train better performing CSI estimators without increasing the computational complexity of the network when implemented in a wireless communications system.