The annual award recognizes the scientific and practical relevance of a Ph.D. thesis in the fields of communications engineering, microsystems engineering, microwave engineering, or signal processing with particular focus on its relevance for future satellite communications systems within modern radio networks.
The Tesat-Spacecom GmbH & Co. KG (TESAT) located in Backnang, Germany, is a leading manufacturer of payload equipment for communication satellites.
The GOSATCOM conference represents a major event for experts on governmental and military communication solutions from industry, academia and government.
A highlight was put in this year’s conference on the development of multi-orbit solutions to provide secure and cost-efficient communication capabilities.
About the thesis:
Machine learning has become an important part of virtually all fields of science including communications engineering. In communications engineering, the majority of machine learning based approaches addresses networking aspects. Machine learning for the physical layer, i.e. for particular transmission methods, is less prominent. However, promising approaches are discussed e.g. in order to reduce signal processing complexity compared to conventional methods or in order to enable a practical implementation at all. Insights can particularly be expected for modern applications such as joint communications, sensing and control, where ultra low latency and high energy efficiency are required. Furthermore, the target can be solutions without explicit models or for scenarios where only very complex models exist. This may be the case e.g. for molecular communications or in highly non-linear transmission scenarios. As far as satellite communications is concerned, the approaches are interesting in order to reduce signal processing complexity in regenerative satellites or terminals, respectively, as well as for making low-cost hardware efficiently applicable in large satellite constellations, by enabling adaptive treatment of their (non-linear) characteristics.
The thesis by Maximilian Stark covers two different approaches and applications of machine learning in physical layer communications.
In the first part, he applies the information bottleneck method, a classification and clustering tool, in order to enable very coarsely quantized baseband signal processing in the receiver of a digital transmission system. The primary target here is the reduction of complexity, particularly of chip area and energy consumption while preserving excellent performance. The core idea is to preserve an adequately defined relevant information as good as possible in each stage of the signal processing chain despite coarse quantization. While coarse quantization causes significant performance degradation in conventional methods, an only marginal degradation compared to double precision implementations is obtained with a resolution of only three to four bits per sample.
The second part addresses the relatively new idea of end-to-end learning with a specific form of neural networks, so-called autoencoders. The concept of autoencoders is applied to digital transmission systems, where components of transmitter and receiver are learned jointly. The results deliver fundamental insights as well as innovative approaches for optimization of transmission schemes with non-linear or unknown channels.
