Approximating the Void: Learning Stochastic Channel Models from Observation with Variational Generative Adversarial Networks

Abstract

Channel modeling is a critical topic when considering accurately designing or evaluating the performance of a communications system. Most prior work in designing or learning new modulation schemes has focused on using simplified analytic channel models such as additive white Gaussian noise (AWGN), Rayleigh fading channels or other similar compact parametric models. In this paper, we extend recent work training generative adversarial networks (GANs) to approximate wireless channel responses to more accurately reflect the probability distribution functions (PDFs) of stochastic channel behaviors. We introduce the use of variational GANs to provide appropriate architecture and loss functions which accurately capture these stochastic behaviors. Finally, we illustrate why prior GAN-based methods failed to accurately capture these behaviors and share results illustrating the performance of such as system over a range of complex realistic channel effects.