End-to-End Learning of Communications Systems

The fundamental problem of communication is that of reliably transmitting a message from a source to a destination over a channel by the use of a transmitter and a receiver. In order to come close to the theoretically optimal solution to this problem in practice, transmitter and receiver were subsequently divided into several processing blocks, each responsible for a specific subtask, e.g., source coding, channel coding, modulation, and equalization. Despite suboptimal, such an implementation has the advantage that each component can be individually analyzed and optimized, leading to the very efficient and stable systems that are available today. Although today's systems have been intensely optimized over the last decades and it seems difficult to compete with them performance-wise, we are attracted by the conceptual simplicity of a communication system that can learn to communicate over any type of channel without the need for prior mathematical modeling and analysis. The main contribution of this thesis is to demonstrate the practical potential and viability of such a system by developing a prototype consisting of two software-defined radios that learn to communicate over an actual wireless channel by means of a deep-learning alternating procedure that iterates between supervised training of the receiver and reinforcement learning-based training of the transmitter