PhD defence by Jakob Thrane

Title: Optimisation of future mobile communication systems using deep learning


Principal supervisor: Associate Professor Henrik Lehrmann Christiansen, DTU Fotonics, Denmark
Co-supervisor: Associate Professor Darko Zibar, DTU Fotonics, Denmark

Evaluation Board
Professor Olav Breinbjerg, DTU Elektro, Denmark
Laurent Roullet, Nokia Bell Labs, France
Professor Preben Mogensen, Aalborg University, Denmark

Master of the Ceremony
Associate Professor Michael Stübert Berger, DTU Fotonics

Mobile communication networks are complex systems consisting of many incredible engineering achievements. The current and future mobile communication systems are, and in the future even more so, complex to manage and optimize. Sub-optimal mobile networks result in cost-ineffective deployments causing poor customer experience and increased operational and capital expenses for operators. Deep Learning (DL) have in the recent years provided with impressive results in complex tasks such as speech recognition and computer vision. The complex reality of mobile communication systems is expected to increase, and Deep Learning has been hailed as a necessary component co overcome such challenges. The content of this dissertation is the exploration of novel methodologies available in Deep Learning toolbox. This dissertation has resulted in several advancements for applying Deep Learning to complex tasks in mobile communication systems.

Accurately, three novel methods are presented.
1) Accurate signal quality predictions in unseen locations with low data complexity using geographical images. 2) Significant improvements to channel estimation applied on sparse reference signals in uplink and 3) An adaptive reinforcement learning algorithm capable of avoiding contamination in the radio environment. In addition to this, several study items are presented. Most noticeably the outcome can be summarized as, 1) Complex ray-tracing methods show little to no performance gain compared to simple empirical models for mobile communication propagation modelling. 2) Current deep-indoor propagation models show poor generalized performance and require novel solutions.


Fri 03 Jul 20
13:00 - 16:00


Lyngby Campus, building 341, auditorium 23


Via Zoom. Write to to sign up