Generative Adversarial Networks for Pseudo-Radio-Signal Synthesis
For many wireless communication applications, traffic pattern modeling of radio signals combined with channel effects is much needed. While analytical models are used to capture these phenomena, real world non-linear effects (e.g. device responses, interferences, distortions, noise) and especially the combination of such effects can be difficult to capture by these models. This is simply due to their complexity and degrees of freedom which can be hard to explicitize in compact expressions. In this paper, we propose a more model-free approach to jointly approximate an end-to-end black-boxed wireless communication scenario using software-defined radio platforms and optimize for an efficient synthesis of subsequently similar 'pseudo-radio-signals'. More precisely, we implement a generative adversarial network based solution that automatically learns radio properties from recorded prototypes in specific scenarios. This allows for a high degree of expressive freedom. Numerical results show that the prototypes' traffic patterns jointly with channel effects are learned without the introduction of assumptions about the scenario or the simplification to a parametric model.