Autoencoder-Based End-to-End Communication

Abstract

Wireless communication systems consist of a channel that has nondifferentiable components and parameters that cannot be estimated and modelled. If these channel parameters are estimated, the system's performance will drastically fall when applied in real-time, as it is not trained to handle conditions that cannot be estimated. Hence, an autoencoder-based end-to-end communication system using machine learning is implemented. It uses alternate training of transmitter and receiver. It eliminates the need to model the channel and its parameters. The system is also trained on the atmospheric effects that impact the channel during normal wireless communication. The proposed work proved to have better results than supervised learning and was able to adapt to the degradation caused by rain attenuation.