Motivation:

Motivated by the fact that the innovative machine learning based autoencoder design of communication systems requires a well-known system model, model-free approaches for deployment in adaptive and real-time applications are of interest. Some recent ideas comprise channel estimation using Variational Generative Adversarial Networks and reinforcement learning for training of a communication system.

Goal:

The aim of this thesis is to investigate one approach to overcome the mentioned limitation and to implement it in a modern deep learning library. In the end, the performance should be evaluated in practice with real data on a software defined radio platform.

Requirements:

In order to work on this thesis, a successful participation in the lectures Wireless Communications and Communication Technologies is required. Also, a solid understanding in linear algebra, optimization theory and stochastics is needed. Programming skills in Python are advantageous but not necessary.