Page d'accueil // Université // Actualités // Événements // PhD Defence: Active Learning in Cognitive Radio Networks

PhD Defence: Active Learning in Cognitive Radio Networks

twitter linkedin facebook google+ email this page
Add to calendar
Conférencier : Anestis Tsakmalis
Date de l'événement : mardi 18 juillet 2017 14:00 - 16:00
Lieu : Room E004, JFK Building
29, Avenue J.F. Kennedy

Members of the defense committee:

Dr Radu State, chairman
Prof. Dr. Ana Isabel Perez-Neira, vice-chairman
Dr Symeon Chatzinotas, supervisor
Prof. Dr. Björn Ottersten, Université du Luxembourg, member
A-Prof. Dr Antonio G. Marques, member

Abstract: In this thesis, numerous Machine Learning (ML) applications for Cognitive Radios Networks (CRNs) are developed and presented which facilitate the efficient spectral coexistence of a legacy system, the Primary Users (PUs), and a CRN, the Secondary Users (SUs). One way to better exploit the capacity of the legacy system frequency band is to consider a coexistence scenario using underlay Cognitive Radio (CR) techniques, where SUs may transmit in the frequency band of the PU system as long as the induced to the PU interference is under a certain limit and thus does not harmfully affect the legacy system operability. This thesis starts with an overview of the ML literature for CRNs in Chapter 2 and continues with the contributions which are divided into the three following chapters.

In Chapter 3, we propose a ML approach for detecting the Modulation and Coding scheme (MCS) of a PU. This Spectrum Sensing (SS) task considers Higher Order Statistical (HOS) features of the sensed PU signal and an efficient ML classifier, the Support Vector Machine (SVM), to identify the modulation scheme used by the PU. The outcome of this identification is combined with the log-likelihood ratios (LLRs) of the PU signal code syndromes to find the PU MCS. This process is called Modulation and Coding Classification (MCC) and it will play an important role in the next part of this thesis.

In Chapter 4, we take advantage of the MCC process in order to bypass the absence of communication between the PU and the SU systems. Due to lack of cooperation between the two systems, the CRN may exploit this multilevel MCC sensing feedback as implicit channel state information (CSI) of the PU link in order to constantly observe the impact of the aggregated interference it causes. Changes in the PU MCS because of the CRN induced interference are considered as PU reactions following the PU Adaptive Coding and Modulation (ACM) protocol. In the examined case study, this MCC feedback allows the CRN to sequentially adjust the SU transmit power levels in order to jointly tackle maximizing the CRN throughput, a Power Control (PC) optimization objective, and learning the interference channel gains which basically constitute the PU interference constraint of the aforementioned optimization problem. Ideal candidate learning approaches for this problem setting with high convergence rate are the Cutting Plane Methods (CPMs). The work of this part laid the foundation of the Active Learning (AL) thesis perspective enabled by PU pieces of feedback.

In Chapter 5, we aim solely at learning the interference channel gains by sequentially probing the PU system. Here, we no longer consider the MCC feedback but the ACK/NACK binary packet which is acquired by eavesdropping the reverse PU link and indicates whether the probing-induced interference is harmful or not. This rudimentary piece of feedback is chosen in order to focus on developing sophisticated probing design techniques for learning the PU interference constraint, since we have already demonstrated the benefits of using the multilevel MCC feedback. We adopt an approach related with AL, a ML field where a learning algorithm sequentially chooses unlabeled data and requests its label in order to reach to a learning solution with as less as possible label queries. This means that the unlabeled data chosen in each step must be selected or designed intelligently so that it delivers more information about the learning solution. This process is clearly correlated with designing the SU transmit power levels, the probing, in order to render it as more informative as possible. Additionally, we incorporate the probability of each binary feedback being correct into this intelligent probing mechanism by developing multivariate Bayesian AL methods inspired by the Probabilistic Bisection Algorithm (PBA) and the CPMs.

Finally, in Chapter 6 we summarize the conclusions and discuss the promising research applications of AL in the CR framework.