Machine Learning for Cognitive Radio Networks

Vera Stavroulaki, Aimilia Bantouna,

Yiouli Kritikou, Kostas Tsagkaris,

Panagiotis Demestichas,

Pol Blasco, Faouzi Bader, Mischa Dohler,

Daniel Denkovski, Vladimir Atanasovski,

Liljana Gavrilovska, and

Klaus Moessner

Learning mechanisms are essential for the attainment of

experience and knowledge in cognitive radio (CR) sys-

tems, exposed to high dynamics with often unpredictable

states [1]. These mechanisms can be associated with user

and device profiles, context, and decisions. The focus learning

user preferences is the dynamic inference and estimation of cur-

rent and future user preferences. The acquisition and learning of

context information encompasses mechanisms for the system to

perceive its current status and conditions in its present environ-

ment, as well as estimating (and forecasting) the capabilities of

available network configurations. Finally, learning related to

decisions addresses the building of knowledge with respect to

the efficiency of solutions that can be applied to specific situa-

tions encountered. Based on knowledge obtained through learn-

ing, decision-making mechanisms can become faster, since

the CR system can learn and immediately apply solutions

that have been identified as being efficient in the past.

Moreover, knowledge obtained through learning

mechanisms may be shared among nodes of a sys-

tem. Thus, more reliable and more optimal deci-

sions can be made by exploiting knowledge

obtained through learning mechanisms.

In this context, this article presents various

basic learning functionalities and relevant

requirements for the identification, collection,

and processing of information that can lead to

exploitable knowledge. The article also provides

an overview of potential implementation

approaches for these functionalities that

exploit diverse machine-learning techniques.

Scenarios

This section introduces a set of indicative

scenarios focusing on information acquisi-

tion and processing, learning, and knowl-

edge management. The target of these

scenarios is to provide an overview of CR

applications, where learning and knowl-

edge management play an important role.

Digital Object Identifier 10.1109/MVT.2012.2190196

Date of publication: 13 April 2012

©A

RT

VIL

LE,

LLC

.,IN

GR

AM

PU

BLI

SH

ING

JUNE 2012 | IEEE VEHICULAR TECHNOLOGY MAGAZINE 1556-6072/12/$31.00©2012IEEE ||| 91

In this respect, the particular scenarios have been selected

as they include different facets of information acquisition,

processing, and exploitation in the scope of advanced

coexistence technologies for radio optimization in licensed

and unlicensed spectrum.

Knowledge-Based Management of

Reconfigurable B3G Infrastructures

This scenario focuses on the operation of a cognitive net-

work management system (CNMS), which constantly moni-

tors the network context and is able to detect a problematic

situation [e.g., severe quality-of-service (QoS) degradation]

that has occurred in a network area (Figure 1).

The CNMS is able to handle these problematic situa-

tions by making the appropriate decisions for the reconfi-

guration of the managed infrastructure. Moreover, the

CNMS stores information on encountered contextual sit-

uations, including the solution that was applied for

handling them in a reference context repository. This

allows considering past network interactions to allow

faster and more efficient handling of problems. The on-

time reaction of the CNMS guarantees the smooth network

operation and the provision of services and applications

at the desired quality levels.

Self-Optimization of Cognitive User Devices

This scenario considers the behavior of user devices

that can comprise various awareness functionalities so

as to decide in an autonomous manner on the actions

they have to take to ensure that all of their running serv-

ices are constantly obtained at the best possible QoS

level. The need for action comes from the volatile con-

text of the network environment. There can be several

operators in the area, offering the same services but with

different quality and reliability. The user device, based

on the knowledge and experience developed over

time, when faced with similar contextual situations, can

autonomously make appropriate reconfiguration deci-

sions and continue to efficiently operate despite the

changing network conditions. A high-level view of the

Context Matching(k-NN)

ContextAcquisition

Registry

Request ID Solution ID

201202001 201202101

201202002 201202201

201202003 201202202

Registry

Request ID Solution ID

201202001 201202101

201202002 201202201

201202003 201202202

Context

1

2

Reference ContextRepository

Wireless Network Segment

ReconfigurationEnforcement

KnownSolution 3

Optimization Process(Decision Making)

2

Unknown

Context

New Context

CNMS

Profilesand

Policies6

ApplySolution

4

5Decision–Solution

Add New Contextand Solution

FIGURE 1 Knowledge-based management of reconfigurable B3G infrastructures [2].

LEARNING MECHANISMS ARE ESSENTIALFOR THE ATTAINMENT OF EXPERIENCE ANDKNOWLEDGE IN COGNITIVE RADIO SYSTEMS,EXPOSED TO HIGH DYNAMICS WITH OFTENUNPREDICTABLE STATES.

92 ||| IEEE VEHICULAR TECHNOLOGY MAGAZINE | JUNE 2012

process of self-optimization of cognitive devices is pro-

vided in Figure 2.

Cognitive Systems and Opportunistic Networks

This scenario addresses the use of opportunistic net-

works (ONs) for enabling devices to communicate over

infrastructure networks even if there is no direct connec-

tion to an infrastructure network [3] (Figure 3).

The ONs are operator governed (through the provision

of resources and policies) and can be coordinated exten-

sions of the infrastructure for a particular time interval.

For the scope of this scenario, various learning and knowl-

edge mechanisms can be useful. These include mecha-

nisms for the identification and learning of context

information, such as identification of spectrum opportuni-

ties [4] for the setup of ONs, and candidate ON nodes such

as devices in the vicinity, identification, and learning of

user preferences.

Learning and Knowledge Management

Mechanisms

Figure 4 illustrates an overview of learning mechanisms

that can be derived from the analysis of the scenarios out-

lined in the previous section.

Acquiring and Learning User Information

The goal of this mechanism is to acquire, maintain, and

process information on user preferences and behavior as

well as the capabilities of the user device. Indicative infor-

mation includes: 1) the set of potential configurations

(e.g., the radio access technologies that the mobile device

is capable of operating with and the associated spectrum

and transmission power levels), 2) the set of services that

can be used and the sets of QoS levels associated with the

Acquiring andLearning User

Information

Acquiring and Learning Context

Information

AcquiringPolicies

Decision on OptimalDevice Configuration

Update of UserInformation and

Knowledge

Update of ContextInformation and

Knowledge Acquisition of Policies

Refinement of Userand Context

Information Under Policies

Selection of OptimalDevice Configuration

Policies

Cognitive User Device

User Information and Knowledge

Context Information and Knowledge

FIGURE 2 Self-optimization of cognitive user devices.

Macro BaseStation 1

Macro BaseStation 2

ON 1(Opportunistic Coverage Extension)

ON 2 (OpportunisticCapacity Extension)

FIGURE 3 Coverage and capacity extension with ONs.

JUNE 2012 | IEEE VEHICULAR TECHNOLOGY MAGAZINE ||| 93

use of a service, 3) the utility volume associated with the

use of a service at a particular quality level [quality of

experience (QoE)], and 4) the maximum price that the

user is willing to pay to use certain services at specific

QoS levels, i.e., for a certain QoE. Furthermore, this func-

tionality deals with learning elements of the user behavior.

The term utility is borrowed from economics, where utility

is used to express the satisfaction obtained from the con-

sumption of goods or services. In this scope, the utility

volume provides a ranking (by order of preference) of

service and QoS combinations, similar to the mean opin-

ion score. User preferences may vary depending on the

contextual situation and may change over time. Therefore,

the utility volume depends on a range of context-related

observable parameters. More specifically, the utility

volume, apart from the service and QoS level, may be

related to the location of the user, the time zone, the user’s

role, and the feedback obtained from the user.

Acquiring and Learning Context Information

On the network side, this mechanism enables collection of

the status of network elements (base stations), the status

of their environment as well as the user devices [5]. Essen-

tially, each element may use monitoring and discovery

procedures [6], which can be sensing based [7], and/or

pilot channel based [8]. Monitoring procedures provide,

for each network element of the segment and for a specific

time period, the traffic requirements, the mobility condi-

tions, the current configuration in terms of operating

parameters (e.g., spectrum assignment and power level)

and the QoS levels offered. Furthermore, it includes proce-

dures for capturing the channel state information, which

is reflected by the estimation of the mean signal-to-noise

ratio value for each subcarrier. Context information is

used by the system for the assessment of key performance

indicators, used for addressing potential problems and

triggering appropriate optimization procedures whenever

necessary. Diverse types of context information storage

may be applied depending on the aforementioned types of

information. The notion of reference context repository

outlined earlier enables the storage of formerly encoun-

tered contextual situations in addition to the correspond-

ing solution that was selected by an optimization process

for addressing these situations. The concept of radio envi-

ronment maps enables the storage of context and policy

information [9].

This mechanism enables user devices to obtain infor-

mation about the network context in a radio environment

with various coexisting technologies. This network

context information should include data about available

access technologies in a given area and their corres-

ponding status (e.g., used frequencies, available resour-

ces, and coverage), information about the device status

(e.g., coverage at the current location, power available,

and technology capabilities), information about the status

of other devices in the area (e.g., activity and the ability to

cooperate). It should be noted that such information can

be directly obtained at the mobile device (e.g., received

power level) as well as at other nodes (e.g., load level in a

given access point). Therefore, appropriate mechanisms

are required, which enable the exchange of context

information/knowledge between nodes in a cognitive net-

work. Learning context information on devices aims to

estimate the most likely capabilities (in terms of QoS

parameters, such as bit rate) that

can be achieved by a given con-

figuration, in a given location, and

time zone. Such estimations can

be exploited to increase reliability

of decisions. Functionality for learn-

ing context information also ena-

bles estimation of situations that

are likely to be encountered in

future time points, given a certain

situation in the current time zone,

thus enabling proactive handling of

certain circumstances.

Acquiring, Deriving,

and Maintaining

Policy Information

This mechanism is responsible for

the acquisition, derivation, and

management of information related

to policies. Policies specify rules or

constraints that should be taken

into account for the selection of the

Acquiring andLearning User

Information

Acquiring andLearning Context

Information

Acquiring,Deriving and Maintaining

Policies

Decision Making(Knowledge-Based

Optimization)

ConfigurationActions

Knowledge(e.g., Encountered

Contextual Situations, Past Decisions/Actions,

Policies)

FIGURE 4 High-level view of learning functionalities in the scope of self-management in CR

systems.

94 ||| IEEE VEHICULAR TECHNOLOGY MAGAZINE | JUNE 2012

optimal configuration of a service area, network element,

or user device. In this sense, policies refine the set of data

comprised in context information.

Policies are usually stored in a database. A policy

database may be used as a central storage point for

the policies and policy users in the network. Such a

database enables keeping track of active users and

policies in the network as well as user-policies associa-

tions. Policy information should form a set of database

tables so that fast and efficient usage of the stored infor-

mation is enabled.

A key feature related to policies in a CR system is the

derivation of spectrum-sharing policies on the spectrum

vacancies (currently not utilized by primary users) for the

secondary users. The policy derivation requirements can

be divided based on the following:

n input parameters (sensing based and nonsensing based)

n type of derivation (reactive, history/predictive derivation)

n type of policies (based on location, functionalities,

connectivity and spectrum).

The derivation of spectrum policies may be realized in

a static or dynamic manner. The static approach relies on

long-term environment measurements, allows learning of

certain repetitive behaviors in the network and leads to

derivation of static spectrum sharing policies. The

dynamic approach relies on shorter real-time measure-

ments that enable policy derivation coping with the dyna-

mism of the wireless environment on a shorter time scale.

Learning Related to the Efficiency of Decisions

The goal of this mechanism is to build knowledge related to

decisions regarding the system configuration and to exploit

this knowledge to decrease the time required to reach a deci-

sion. More specifically, this functionality comprises mecha-

nisms that allow storing information about problems that

occur (e.g., QoS degradation) as well as their solutions and

to identify whether a problem currently addressed is similar

to an older one for which a solution is already available. This

allows for the overall decision-making process to speed up,

because it becomes less complex due to the limited number

of choices that should be checked. Thus, required system

adaptations can be performed faster.

This mechanism should enable

recording and storing of informa-

tion on situations encountered,

actions that were taken in reaction

to these situations, and, potentially,

information on the effectiveness/

efficiency of decisions taken. A

simple way of implementing this is

with the use of an appropriate

registry table. Implicitly, some of

the requirements for this function-

ality are related to necessities for

context acquisition.

A key requirement for this functionality is related to

the ability of the system to identify similar contextual sit-

uations so that it can be derived whether a previously

selected and enforced action should be directly applied

without the need of executing anew the decision-making

process. The direct application of known (good) solutions

facilitates the reduction of the time required to handle

contextual situations.

Docitive Networks

Despite cognition and learning having received a consider-

able interest, the process of knowledge transfer, i.e.,

teaching, over the wireless medium has received fairly lit-

tle attention to date. To this aim, in [10], a new and largely

unexploited paradigm referred to as docitive radios has

been introduced. The term docitive comes from ‘‘docere’’

= ‘‘to teach’’ in Latin, which relates to radios (or general

networking entities) and transfer their prior knowledge to

other radios. These radios are not only supposed to teach

end results (e.g., in form of ‘‘I sense the spectrum to be

occupied’’) but rather elements of the methods of getting

there. It capitalizes on the advantages but, most impor-

tantly, mitigates major parts of the aforementioned draw-

backs of purely CRs.

As illustrated in Figure 5, the canonical cognitive cycle

is advantageously extended by docition. It is realized by

means of an entity that facilitates knowledge dissemina-

tion and propagation with the nontrivial aim to help

learning. With this valuable knowledge at hand, a

newcomer CR is able to reduce the exploration phase and

increase its performance. In [11], docition is applied in the

context of femtocell networks. Femtocells use Q-learning,

an algorithm that originates from reinforcement learning,

Action

Information Acquisition

and Learning

Cognitive Radio

Docitive RadioDocition

DecisionDecision

FIGURE 5 Canonical cycle and its extension through docition.

LEARNING RELATED TO DECISIONSADDRESSES THE BUILDING OF KNOWLEDGEWITH RESPECT TO THE EFFICIENCY OFSOLUTIONS THAT CAN BE APPLIED TOSPECIFIC SITUATIONS ENCOUNTERED.

JUNE 2012 | IEEE VEHICULAR TECHNOLOGY MAGAZINE ||| 95

to allocate the optimal power and frequency to transmit.

In this context, the most expert femtocells share their

Q-tables with the newcomer femtocells that join the net-

work. Within this extended cognitive cycle, newcomer

femtocells capitalize on docition to use the knowledge

of expert radios and avoid learning from scratch. The

benefits of docition are gauged against regular CR

networks, showing better speed of convergence and

higher accuracy.

Docition is inspired by the so-far-successful problem-

based learning (PBL) concept used at schools. In PBL,

teachers are encouraged to be coaches and not informa-

tion providers, with the aim to have pupils work as a team

using critical thinking to synthesize and apply knowledge,

apprehend through dialog, questioning, reciprocal

teaching, and mentoring. Translated back to the wireless

setting, this implies a distributed approach where

nodes share potentially differing amounts of intelligence

acquired on the run. This, in turn, is expected to

sharpen and speed up the exploration process. Any

achieved gains, however, need to be gauged against the

overhead incurred as a result of the exchange of doci-

tive information.

The potential impact of docitive networks into next-

generation high-capacity wireless networks is ensured by

means of latest European Telecommunications Standards

Institute Broadband Radio Access Networks standardiza-

tion activities [12].

An Overview of Potential

Implementation Approaches

Learning and knowledge management mechanisms for CR

applications may be implemented by exploiting a variety

of machine-learning and pattern-recognition approaches-

such as Bayesian networks, neural networks, and fuzzy

cognitive maps [13]. This section provides an overview of

potential (indicative) approaches for the implementation

of learning and knowledge management for CR applica-

tions, focusing on a subset of machine learning techniques.

Learning User Preferences with the Use of

Bayesian Statistics

This section presents an approach for dynamically

learning user preferences regarding the perceived QoS

level per service/application [14] with the use of Bayesian

statistics concepts. The aim is to estimate the probability

of the level of user satisfaction for a specific service

and perceived QoS level, given a certain location and

time zone.

The user profile can be modeled as a collection of

parameters that can be classified into two main groups:

observable and output parameters. Observable parame-

ters include the currently running services/applications,

corresponding QoS levels, and associated QoS parameters

such as location, time zone, and provided user feedback.

User feedback is obtained in the following manner. The

user initiates a specific service. At the initial stages, it is

considered that the user is indifferent between service

provision choices. Every time the user obtains a service, a

rating facility, embedded in the learning mechanism,

allows the user to rate how much he/she liked the particu-

lar service provision. Different rating options are pro-

vided, also enabling nonexperts to provide the system

with feedback on their preferences. Users are also given

the choice to decline by providing a rating. The value of

output parameters is dynamically updated over time

based on the value of observable parameters. The focus

here is on one output parameter, namely, the utility

volume. The utility volume provides a ranking, by order of

preference, of service and QoS combinations. User prefer-

ences may vary depending on the contextual situation and

may change over time. Conditional probabilities for the

utility volume are calculated. These yield the most likely

user preferences per QoS level, which in turn can be used

as input in a decision-making mechanism for deciding, for

instance, on the most appropriate configuration of a

user’s device.

Learning Context Information with the Use

of Bayesian Networks

In this case, Bayesian statistics are used to estimate the

future behavior of candidate networks in terms of QoS

capabilities based on collected measurements.

The set of candidate networks is determined by the

context of operation, the device profile, and the policies of

the network operator (NO). Initially, the context in which

the user and device are found and the capabilities of the

device provide the networks that are available in the area.

The set of candidate networks is a subset of the available

networks, as it comprises only those available networks

that are compliant with the policies of the NO. Compliance

with the policies means that the NO allows the selection of

the networks for the particular user and device.

Input, i.e., measurements from candidate networks,

may be collected for offline or online analysis. In the cur-

rent approach, online analysis is applied, and its objective

is to estimate the future state using the observations up to

the current time. This estimation is modeled as a probabil-

ity density function. The advantage of this approach is

that no past observations need to be explicitly stored.

THE ACQUISITION AND LEARNING OF CONTEXTINFORMATION ENCOMPASSES MECHANISMSFOR THE SYSTEM TO PERCEIVE ITS CURRENTSTATUS AND CONDITIONS IN ITS PRESENTENVIRONMENT, AS WELL AS ESTIMATING (ANDFORECASTING) THE CAPABILITIES OFAVAILABLE NETWORK CONFIGURATIONS.

96 ||| IEEE VEHICULAR TECHNOLOGY MAGAZINE | JUNE 2012

QoS levels for networks are considered as a combina-

tion of QoS parameters. Each parameter can refer to a spe-

cific aspect, e.g., bit rate and delay. The QoS capabilities of

the candidate networks are determined by the context of

operation (e.g., interference conditions) and by the poli-

cies of the operator.

In summary, the overall learning process evolves as fol-

lows. Measurements are collected from the candidate net-

works, which are in line with operator policies. Based on

these measurements, the conditional probabilities are

updated, which provide an estimation of how probable it

is that a specific QoS parameter will reach a certain value,

given a certain network. The next step is the update of the

probability density function values, which offers a more

aggregate estimation regarding the probability to achieve

a certain combination of QoS parameters that corre-

sponds to a QoS level, given a certain network. These

probabilities can be exploited by knowledge-based net-

work selection schemes [15].

Learning Context Information with the Use

of Self-Organizing Maps

This section focuses on an implementation approach for

learning context information with the use of self-organiz-

ing maps (SOMs). The elements that serve as inputs for

discovering the network capacity are parameters that are

obtained, given a configuration, e.g., received signal

strength identifier and number of input bytes. The variable

that expresses the network capacity is the bit rate.

According to this approach, SOM is used for mapping

multidimensional data in a two-dimensional (2-D) map. To

do so, SOM requires a training process where the data are

converted in data samples and, finally, in vectors that are

mapped with respect to their resemblance. Thus, the cre-

ated map depicts the clustering of data based on the

resemblance of the corresponding vectors.

At this point, it is essential to clarify that the term data

sample differs from the term data as follows. A data sam-

ple consists of more than one data. In fact, each data sam-

ple is a combination of values, each of which refers to a

different observed parameter.

When applying the SOM approach, as soon as a pat-

tern has been identified, a new data sample can be

mapped on it using the same process. The identification

of patterns in data is based on the Euclidean distance

(selected due to its straightforward applicability)

between vectors of data samples. Data samples that are

similar to each other will be grouped into the same

cluster. This means that the bit rate observed at the

same time with the variables that formulate one data

sample of the cluster is expected to be the same with

the respective bit rate of the other data samples of the

cluster. Thus, for inferring the network capacity, the

cluster in which a newly obtained entry belongs is iden-

tified. Further information about the SOM technique

and its application to learning context information can

be found in [16].

Enabling Learning Related to the Efficiency of

Decisions with the Use of k-Nearest Neighbor(s)

Learning related to the efficiency of decisions requires the

identification of the similarity of contextual situations so

as to enable appropriate updating of information. For the

identification of the similarity of contextual situations an

approach based on the k-nearest neighbor(s) (k-NN) algo-

rithm [17] can be applied. Elements of the current context

information are compared with past network conditions

(reference contexts) based on several threshold compari-

sons regarding the total number of users, the Euclidean

and profile distance among users (the Euclidean distance

is a typical metric used in k-NN algorithms). More specifi-

cally, given the current and a set of reference contexts, the

algorithm checks each reference context to match with

the current context according to an overall distance. The

overall distance is based on 1) the total number of ses-

sions distance, 2) the total number of sessions per service

distance, and 3) the sessions distribution distance. Taking

into account these subdistances, the overall distance for

each reference context from the current context is calcu-

lated as the sum of the distances calculated in each phase.

The outcome is the reference context with the minimum

overall distance from the current context, if such exists in

the reference context repository. If no such context

exists, then a new record is created in the repository for

the current context. If it exists, then the decision that

was recorded for this solution is derived. If it is rated as a

good decision, it can be applied without the need of

running an optimization process. If the decision applied

in the past has a bad rating, the optimization process

can be triggered and the previously recorded decision can

be replaced in the repository. In all cases, after the imple-

mentation of the decision, the rate on the efficiency of

the decision corresponding to the identified context is

updated accordingly.

Conclusions

The article presented various basic learning functional-

ities for the identification and processing of information

that can lead to exploitable knowledge in CR networks.

The article also provided an overview of potential imple-

mentation approaches for these functionalities that

exploit diverse machine-learning techniques.

POLICIES SPECIFY RULES OR CONSTRAINTSTHAT SHOULD BE TAKEN INTO ACCOUNT FORTHE SELECTION OF THE OPTIMALCONFIGURATION OF A SERVICE AREA,NETWORK ELEMENT, OR USER DEVICE.

JUNE 2012 | IEEE VEHICULAR TECHNOLOGY MAGAZINE ||| 97

Acknowledgment

This work supports training activities in the context of

advanced coexistence technologies for Radio Optimiza-

tion in Licensed and Unlicensed Spectrum -Network of

Excellence (ACROPOLIS) project (http://www.ict-acropolis.

eu). This article reflects only the authors’ views, and the

community is not liable for any use that may be made of

the information contained therein.

Author Information

Vera Stavroulaki (veras@unipi.gr) received her diploma

degree in informatics from Athens University of Econom-

ics and Business in 2000 and Ph.D. degree in electrical

and computer engineering from the National Technical

University of Athens in 2004. She is currently an assist-

ant professor at the Department of Digital Systems,

University of Piraeus, Greece. She has ten years of

experience in European and national research and devel-

opment projects. Her current interests include cognitive

management functionality for autonomous, reconfigur-

able devices and networks, service governance and

virtualization, cloud computing, Internet of things, and

future Internet.

Aimilia Bantouna (abantoun@unipi.gr) received her

diploma degree from the School of Applied Mathematics

and Physical Sciences of the National Technical Univer-

sity of Athens in 2007. She obtained her M.Sc. degree in

digital communications and networks from the Depart-

ment of Digital Systems, University of Piraeus, Greece, in

2010. Since June 2011, she has been pursuing her Ph.D.

degree in machine-learning techniques of future net-

works and is a research engineer at the University of

Piraeus Research Center. She is also involved in various

European Union (EU)-funded projects.

Yiouli Kritikou (kritikou@unipi.gr) received her

diploma degree in 2003 and her Ph.D. degree in 2009

from the Department of Digital Systems at the University

of Piraeus. Since September 2003, she has been a

research engineer at the University of Piraeus, Labora-

tory of Telecommunication Networks and Integrated

Services. She also conducted a postdoc research in

advanced services in the context of the future Internet.

She has participated in numerous national and interna-

tional projects. Her research interests include design,

specification, and development of services for wireless

networks, concentrating on user profile, ubiquitous

service delivery, and technoeconomic aspects of serv-

ices delivery.

Kostas Tsagkaris (ktsagk@unipi.gr) received his

diploma and Ph.D. degree (Ericsson’s Awards of Excel-

lence in Telecommunications) from the School of Electri-

cal and Computer Engineering at the National Technical

University of Athens, Greece. Currently, he works as a

senior research engineer and adjunct lecturer at the

Department of Digital Systems in University of Piraeus,

Greece. He has been involved in many research projects

and published more that 100 papers in the areas of

design, management, and optimization of wireless auto-

nomic/cognitive networks.

Panagiotis Demestichas (pdemest@unipi.gr) is a

professor at the Department of Digital Systems, Univer-

sity of Piraeus, Greece, where he is also the head of the

department and leader of the Laboratory of Telecommu-

nication Networks and Integrated Services. He received

his diploma and Ph.D. degrees in electrical engineering

from the National Technical University of Athens. He has

been involved in numerous international research and

development programs. His research interests include

design and performance evaluation of wireless and wire-

line, broadband networks, software engineering, service

and network management, algorithms and complexity

theory, technoeconomic management of networks, and

services. He has several publications in international

journals and conferences.

Pol Blasco (pol.blasco@cttc.es) graduated in tele-

communications engineering in 2009 at the Technical

University of Catalonia [Universitat Politecnica de Cata-

lunya (UPC), Barcelona]. He graduated in electrical and

information theory engineering in 2008 at the Technical

University of Darmstadt (Germany). In 2011, he obtained

his M.Sc. degree for the Signal Theory Department of

UPC (Barcelona). Between 2007 and 2009, he pursued

research in several international research centers and

agencies. In 2009, he joined the Technical Telecommuni-

cation Centre of Catalonia (CTTC) to pursue his Ph.D.

degree in the area of intelligent energy.

Faouzi Bader (cbader@cttc.es) received his Ph.D.

degree (with honors) in telecommunications at the Uni-

versidad Polit�ecnica de Madrid-Spain in 2002. Then, he

joined CTTC in Barcelona-Spain as a research associate,

in which since 2006, he was a senior researcher. He is a

Senior Member of the IEEE. His research interests

include broadband networks, radio resource manage-

ment (RRM) in cognitive radio, and relaying scenarios.

Mischa Dohler (mischa.dohler@cttc.es) is leading

the Intelligent Energy group at CTTC in Barcelona, with

a focus on smart grids and green radios. He is working

on machine-to-machine, wireless sensor, femto, coopera-

tive, cognitive, and docitive networks. He has published

around 150 technical journal and conference papers at a

citation h-index of 26 and citation g-index of 55, holds a

BASED ON KNOWLEDGE OBTAINED THROUGHLEARNING, DECISION-MAKING MECHANISMSCAN BECOME FASTER, SINCE THE CR SYSTEMCAN LEARN AND IMMEDIATELY APPLYSOLUTIONS THAT HAVE BEEN IDENTIFIED ASBEING EFFICIENT IN THE PAST.

98 ||| IEEE VEHICULAR TECHNOLOGY MAGAZINE | JUNE 2012

dozen patents, authored, coedited, and contributed to

19 books, has given 25 international short courses, and

participated in various standardization activities. He is

the editor-in-chief of ETT and has been holding various

editorial positions for numerous IEEE and non-IEEE jour-

nals. He is a Senior Member of the IEEE and Distin-

guished Lecturer of IEEE ComSoc.

Daniel Denkovski (danield@feit.ukim.edu.mk)

received his Dipl.-Ing. and M.Sc. degrees in Telecommu-

nications at Ss. Cyril and Methodius University (UKIM)

in Skopje, in 2008 and 2010, respectively. He is currently

a Ph.D. student in electrical engineering and information

technologies at UKIM. Since 2009, he has been a re-

searcher and a member of the Wireless Networking

Group at the Faculty of Electrical Engineering and Infor-

mation Technologies. His research work in the areas of

cognitive wireless networking and policing, spectrum

sensing, cooperative communications, and RRM has

resulted in more than 20 publications in relevant confer-

ences and journals.

Vladimir Atanasovski (vladimir@feit.ukim.edu.mk)

currently holds the position of assistant professor at the

Institute of Telecommunications at the Faculty of Electri-

cal Engineering and Information Technologies, UKIM in

Skopje. He received his B.Sc., M.Sc., and Ph.D. degrees

from UKIM in Skopje, in 2004, 2006, and 2010, res-

pectively. He is the author/coauthor of more than 70

research journal and conference publications and

technical papers. His major research interests lie in

the areas of cognitive radio networks, resource man-

agement for heterogeneous wireless networks, traffic

analysis and modeling, cross-layer optimizations, and

ad-hoc networking.

Liljana Gavrilovska (liljana@feit.ukim.edu.mk) cur-

rently holds the positions of full professor, head of the

Institute of Telecommunications and the Center for Wire-

less and Mobile Communications at the Faculty of

Electrical Engineering and Information Technologies, in

Skopje. She is the author/coauthor of more than 150

research journal and conference publications and

technical papers, a coauthor of the book Ad Hoc Net-

working Towards Seamless Communications (Springer,

2006) and coeditor of the book Application and Multidisci-

plinary Aspects of Wireless Sensor Networks (Springer,

2010). Her major research interest is concentrated on

cognitive radio networks, future mobile systems, wireless

and personal area networks, cross-layer optimizations,

ad-hoc networking, traffic analysis, and heterogeneous

wireless networks.

Klaus Moessner (K.Moessner@surrey.ac.uk) is a

professor in the Centre for Communication Systems

Research at the University of Surrey, United Kingdom.

He earned his Dipl-Ing (FH) degree at the University of

Applied Sciences in Offenburg, Germany, an M.Sc. degree

from Brunel University, United Kingdom, and his Ph.D.

degree from the University of Surrey. His research inter-

ests include reconfigurability of the different system

levels, including reconfiguration management, service

platforms, service-oriented architectures, and IMS, as

well as scheduling in wireless networks and adaptability

of multimodal user interfaces.

References[1] J. Mitola, III and G. Q. Maguire, ‘‘Cognitive radio: Making software

radios more personal,’’ IEEE Pers. Commun., vol. 6, no. 4, pp. 13–18,Aug. 1999.

[2] A. Saatsakis and P. Demestichas, ‘‘Context matching for realizingcognitive wireless network segments,’’ Wirel. Pers. Commun., vol. 55,no. 3, pp. 407–440, Nov. 2010.

[3] (July 2010–Dec. 2012). FP7/ICT project OneFIT (Opportunistic net-works and cognitive management systems for efficient applicationprovision in the Future InterneT, ICT-2009-257385) [Online]. Available:http://www.ict-onefit.eu

[4] F. Bouali, O. Sallent, J. P�erez-Romero, and R. Agust�ı, ‘‘A framework

based on a fittingness factor to enable efficient exploitation ofspectrum opportunities in cognitive radio networks,’’ in Proc. 14th Int.Symp. Wireless Personal Multimedia Communications, 2011 (WPMC),Brest, France, pp. 1–5.

[5] M. J. Van Sinderen, A. T. Van Halteren, M. Wegdam, H. B. Meeuwis-sen, and E. Henk Eertink, ‘‘Supporting context-aware mobile applica-tions,’’ IEEE Commun. Mag., vol. 44, no. 9, pp. 96–104, 2006.

[6] K. Tsagkaris, A. Katidiotis, and P. Demestichas, ‘‘Neural networkbased learning schemes for cognitive radio systems,’’ Comput. Com-mun., vol. 31, no. 14, pp. 3394–3404, 2008.

[7] H. Kim and K. G. Shin, ‘‘Efficient discovery of spectrum opportunitieswith MAC-layer sensing in cognitive radio networks,’’ IEEE Trans.Mobile Comput., vol. 7, no. 5, pp. 533–545, 2008.

[8] J. Perez-Romero, O. Sallent, R. Agusti, and L. Giupponi, ‘‘A novel on-demand cognitive pilot channel enabling dynamic spectrum alloca-tion,’’ in Proc. 2nd Int. Symp. New Frontiers in Dynamic SpectrumAccess Networks 2007 (DySPAN), Dublin, Ireland, pp. 46–54.

[9] Y. Zhao, B. Le, and J. Reed, ‘‘Network support: The radio environ-

ment map,’’ Cognitive Radio Technology, 2nd ed. New York, Elsevier,Apr. 2009. ch. 11, pp. 325–336.

[10] L. Giupponi, A. M. Galindo, P. Blasco, and M. Dohler, ‘‘Docitive net-work—An emerging paradigm for dynamic spectrum Management,’’IEEE Wirel. Commun Mag., vol. 17, no. 4, pp. 47–54, 2010.

[11] A. Galindo-Serrano, L. Giupponi, and M. Dohler, ‘‘Cognition and doci-tion in OFDMA-based femtocell networks,’’ in Proc. IEEE Global Commu-nications Conf. (IEEE GLOBECOM), Miami, FL, Dec. 6–10, 2010, pp. 1–6.

[12] ETSI BRAN, ‘‘Very high capacity density BWA networks: System

architecture, economic model and technical requirements,’’ Euro-pean Telecommun. Standards Inst. Broadband Radio Access Net-works Tech. Committee, Sophia Antipolis, France, Tech. Rep. 101534 v0.0.3, Dec. 2011.

[13] B. Kosko, ‘‘Fuzzy cognitive maps,’’ Int. J. Man-Mach. Stud., vol. 24,no. 1, pp. 65–75, 1986.

[14] V. Stavroulaki, Y. Kritikou, and P. Demestichas, ‘‘Acquiring andlearning user information in the context of cognitive device manage-ment,’’ in Proc. IEEE ICC, 2009, Dresden, Germany, pp. 1–5.

[15] P. Demestichas, A. Katidiotis, K. Tsagkaris, E. Adamopoulou, andK. Demestichas, ‘‘Enhancing channel estimation in cognitive radiosystems by means of Bayesian networks,’’ Wirel. Pers. Commun.,vol. 49, no. 1, pp. 87–105, Apr. 2009.

[16] A. Bantouna, K. Tsagkaris, and P. Demestichas, ‘‘Self-organizingmaps for improving the channel estimation and predictive modellingphase of cognitive radio systems,’’ in Proc. 20th Int. Conf. ArtificialNeural Networks, (ICANN), Thessaloniki, Greece, Sept. 15–18, 2010.

[17] H. Same, ‘‘K-Nearest neighbor finding using MaxNearestDist,’’ IEEETrans. Pattern Anal. Mach. Intell., vol. 30, no. 2, pp. 243–252, Feb.2008.

THE FOCUS LEARNING USER PREFERENCES ISTHE DYNAMIC INFERENCE AND ESTIMATION OFCURRENT AND FUTURE USER PREFERENCES.

JUNE 2012 | IEEE VEHICULAR TECHNOLOGY MAGAZINE ||| 99