Project

Back to overview

Machine Learning for Communications

English title Machine Learning for Communications
Applicant Burg Andreas Peter
Number 182621
Funding scheme Project funding (Div. I-III)
Research institution Laboratoire de circuits pour télécommunications EPFL - STI - IEL - TCL
Institution of higher education EPF Lausanne - EPFL
Main discipline Electrical Engineering
Start/End 01.11.2018 - 31.10.2022
Approved amount 467'692.00
Show all

All Disciplines (2)

Discipline
Electrical Engineering
Information Technology

Keywords (7)

Full-Duplex; Wireless; Sensing; Machine Learning; Communications; Neural Networks; Coding and Modulation

Lay Summary (German)

Lead
Algorithmen für maschinelles Lernen (ML), bzw. künstliche Intelligenz, erleben seit kurzem eine Renaissance in verschiedensten Anwendungsgebieten. Obwohl die Kernidee, die dem menschlichen Gehirn nachempfunden ist, bereits seit vielen Jahrzehnten existiert werden solche Methoden erst seit kurzem in grossem Umfang eingesetzt da sie oft einen grossen Rechenaufwand erfordern. Insbesondere der einmalige Trainingsprozess ist dabei mit einem grossen Aufwand verbunden und erfordert eine grosse Menge an Daten. ML Methoden werden heute vor allem in der Bild- und Spracherkennung eingesetzt, sind aber im Prinzip auch für viele weitere Anwendungen und Probleme geeignet. Ein interessantes, neues Anwendungsgebiet für diese Art von Algorithmen ist die drahtlose Datenkommunikation.
Lay summary

Im Gegensatz zur Bild- und Sprachverarbeitung setzt die drahtlose Datenkommunikation fast ausschliesslich auf klassische Signalverarbeitung die auf klar formulierten Signalmodellen basiert die analytisch gut nachvollziehbar sind. In diesem Sinne erscheint der Einsatz von ML wie ein Wiederspruch oder sogar ein Rückschritt. Bedenkt man aber dass herkömmliche Kommunikationsmodelle oft stark vereinfacht sind und viele komplexe Zusammenhänge vernachlässigen, wird klar, dass neue Methoden, die stark vereinfachte Modelle ersetzen können grosse Leistungsgewinne versprechen.
Ziel dieses Projektes ist es den Einsatz von Methoden für maschinelles Lernen im Bereich der drahtlosen Datenkommunikation zu untersuchen. Da für viele Probleme bereits stark optimierte konventionelle Algorithmen existieren, ist es wichtig genau diejenigen Probleme zu identifizieren bei denen vereinfachte Signalmodelle die Performance begrenzen. Drei Bereiche sind dabei von besonderem Interesse: i) der Übergang von der Demodulation komplexer Symbole zur Dekodierung Fehler-korrigierender Codes, ii) die Unterdrückung von Selbstinterferenz in Full-Duplex Systemen, und iii) die Verwendung von Kommunikationssignalen um Rückschlüsse auf die Umgebung von Sender und Empfänger zu erhalten. Die Forschung in diesem Projekt beschäftigt sich dabei nicht ausschliesslich mit algorithmischen Aspekten, sondern bezieht auch Aspekte der Implementierung und Komplexität mit ein. Wir hoffen somit existierende Algorithmen durch einfachere und leistungsfähigere ML Methoden zu ersetzen und so die Effizienz der drahtlosen Kommunikation in vielen Bereichen zu verbessern.

Direct link to Lay Summary Last update: 20.02.2019

Responsible applicant and co-applicants

Employees

Associated projects

Number Title Start Funding scheme
146753 Full-duplex Radio, Theory and Experiments 01.04.2013 Project funding (Div. I-III)

Abstract

The field of machine learning has seen tremendous advances in the past few years, largely due to the abundant processing power and the availability of vast amounts of data that enable effective training of deep networks for complex models. The main motivation for using machine learning usually comes from that fact that in some areas, such as image recognition and natural language processing, constructing hand-crafted models that are elegant, tractable, and practically useful is nearly impossible. The field of communications, however, is built on precise, but often simple mathematical models that are well understood and have been shown to work exceptionally well for many practical applications. Unfortunately, the ever-increasing demands have forced communications systems designers to push the boundaries to such an extent that in many applications conventional mathematical models and signal processing techniques are no longer sufficient to accurately describe the encountered complex scenarios. Specifically, there is an increasing number of cases where rigorous mathematical models are either not known or impractical from an analytical, or even computational perspective. At this point, machine learning methods may come to the rescue as they do not require rigid pre-defined models and can extract relevant and meaningful structure from large amounts of data to provide useful results. In fact, contrary to many other machine learning applications, in communications it is relatively simple to collect such labeled data, either from simulations or from testbeds or the massive amount of every-day communication, making the training of machine learning algorithms for numerous applications comparatively straightforward. The two aforementioned observations render machine learning an ideal tool that carries the potential to augment and revolutionize signal processing for communications and even exploit communications for further applications. Some well-knownresearch groups have recently started to venture into this area, but the idea is still very much in its infancy and there is a number of pressing issues that need to be addressed, such as:1. Applications: Machine learning algorithms can, in principle, learn to perform any task from the data they are given. However, since in the field of communications we already know how to solve many problems optimally, it is crucial to carefully identify the applications where machine learning is actually beneficial and not a re-invention of a (needlessly complicated) wheel.2. Architectures: Different applications require vastly different machine learning approaches and architectures. Thus, it is important to identify the best architecture for each scenario and to develop novel architectures and approaches whenever necessary.3. Adaptability: In most communications applications, and in particular in wireless communications, the environment can change rapidly over time. Machine learning algorithms need to be able to adapt, either through appropriate offline training or through low-complexity and fast-converging online training.The main goal of this project is to exploit expert knowledge from the communications field to put machine learning to service where conventional signal processing struggles to achieve the expected quality of results. To this end, we propose to examining a number of concrete practical applications. The first application is to employ neural networks in order to perform digital self-interference cancellation in full-duplex radios with better performance and lower complexity than existing approaches. The second application has to do with using machine learning techniques in order to perform blind detection of channel codes, as well as joint detection and decoding of coded non-linear modulation schemes. As a final application, we propose to employ machine learning techniques that exploit the residual self-interference signal in full-duplex radios to build context-aware radios that can make inferences about the environment in which they are operating.
-