Modeling Time-Varying and Frequency-Selective Channels with Generative Adversarial Networks

Modeling realistic time-varying and frequency selective channels is critical for designing the wireless systems that are robust to fading channels. Modeling a wireless channel is a time consuming process which involves measurement campaigns in various environments and then fitting the right statistical model for each measurement scenario. In this paper, we propose a method to generate channel based on the generative adversarial network. This approach is helpful in generating time-varying and frequency-selective channels for the scenarios which do not have a known statistical model. Typically, this method is helpful in early stages of characterization of the channels for new bands or mediums. For example, millimeter wave, tera-hertz bands or new mediums like molecular communications. Furthermore, this method generates the channel trace by capturing the time-varying nature of the channel efficiently from the channel measurements itself and uses this learned correlation model in generating new such channels very efficiently. The accuracy of the channels generated by the proposed method is verified through conditional cumulative distribution function for the third generation partnership project (3GPP) defined spatial channel models.