spectrum decision support system thesis

8/13/2019 Spectrum Decision Support System Thesis

1/122

Blekinge Institute of TechnologyLicentiate Dissertation Series No. 2012:04

School of Computing

A SPECTRUM DECISION SUPPORT SYSTEM

FOR COGNITIVE RADIO NETWORKS

Yong Yao


2/122

A Spectrum Decision Support Systemfor Cognitive Radio Networks

Yong Yao


3/122


4/122

A Spectrum Decision Support Systemfor Cognitive Radio Networks

Yong Yao

Licentiate Dissertation in

Telecommunication Systems

Blekinge Institute of Technology licentiate dissertation seriesNo 2012:04

School of ComputingBlekinge Institute of Technology

SWEDEN


5/122

2012 Yong YaoSchool of ComputingPublisher: Blekinge Institute of Technology,SE-371 79 Karlskrona, SwedenPrinted by Printfabriken, Karlskrona, Sweden 2012ISBN: 978-91-7295-231-7ISSN 1650-2140urn:nbn:se:bth-00527


6/122

To my family...


7/122


8/122

Abstract

Cognitive Radio Networks (CRNs) offer a promising capability of alleviating theproblem of spectrum insufficiency. In CRNs, the licensed spectrum channels areeither exclusively reserved for licensed users or temporarily used by unlicensedusers. The requirement for unlicensed users is to not harmfully impair thelicensed users transmissions. Because of this, the unlicensed users must solvethe task to decide which available channels should be selected. The selectionprocess is often referred to as spectrum decision, with the aim to optimize thetransmission performance of unlicensed users.

A support system for CRNs is introduced, which is called Spectrum Deci-sion Support System (SDSS). SDSS provides an intelligent spectrum decisionstrategy that integrates different decision making algorithms and takes into ac-count various channel characterization parameters. The objective is to developa scientic framework for decision making in CRNs, which includes theoreticalanalysis, simulation evaluation and practical implementation. Three importantcomponents of SDSS are discussed: 1) setting up an overlay decision maker, 2)prediction based spectrum decision strategy, and 3) queuing modeling of CRNs.The reported results indicate the feasibility of the suggested algorithms.

iii


9/122

iv


10/122

Acknowledgements

It is a pleasure to express my deep gratitude and appreciation to those who havecontributed to the conducting of this thesis. Foremost, I would like to thank mymain advisor Prof. Adrian Popescu for accepting me as a Ph.D student, and forhis guidance and immense knowledge that helped me grow up in research career.I also would like to thank my co-advisor Doctor David Erman for his invaluablesupervision, and Prof. Markus Fiedler for rmly supporting me along the way.

Further, my special thanks go to my fellow Ph.D students and colleaguesat Blekinge Institute of Technology (BTH) for their endless encouragement,

suggestion and help. In particular Said Rutabayiro Ngoga for his fruitful col-laboration, Alexandru Popescu for his insightful opinion, Junaid Shaikh for hisextensive motivation, and Selim Ickin for his interesting discussion.

Last but not least, I am deeply grateful to my wife Lu and daughter Yuxuanfor their innite understanding, comfort and accompany that make this thesispossible. I am also sincerely grateful to my parents Bingfeng Li and ZhishunYao for unconditionally supporting and assisting me spiritually throughout mylife.

Yong Yao

Karlskrona, May 2012

v


11/122


12/122

Contents

Page

1 Introduction 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Spectrum Decision Support System 9

2.1 Spectrum Decision . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2 Scientic Functionalities of SDSS . . . . . . . . . . . . . . . . . . 15

2.3 SDSS Features and Structure . . . . . . . . . . . . . . . . . . . . 17

2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Overlay Decision Making 21

3.1 Channel Characterization . . . . . . . . . . . . . . . . . . . . . . 21

3.2 Uniform Decision Criterion . . . . . . . . . . . . . . . . . . . . . 24

3.3 Group Decision Making . . . . . . . . . . . . . . . . . . . . . . . 27

3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 Channel Usage Prediction 33

vii


13/122

4.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.3 Channel Usage Prediction . . . . . . . . . . . . . . . . . . . . . . 40

4.4 Prediction Implementation . . . . . . . . . . . . . . . . . . . . . . 43

4.5 Spectrum Decision . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.6 Simulation Evaluation . . . . . . . . . . . . . . . . . . . . . . . . 50

4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

5 Competition-Based Channel Selection 55

5.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.2 Idle Time Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.3 Competition Among Secondary Users . . . . . . . . . . . . . . . 60

5.4 Channel Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.5 Fuzzication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.6 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . 67

5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6 Intra- and Inter-Handoff 73

6.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

6.2 System Model and Inter-handoff Schemes . . . . . . . . . . . . . 75

6.3 Queuing Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

6.4 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . 80

6.5 Determination of Inter-handoff Parameters . . . . . . . . . . . . 83

6.6 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . 85

6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

7 Conclusion 91

7.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

viii


14/122

A Complete publication list 93

Bibliography 95

ix


15/122

x


16/122

List of Figures

Figure Page

1.1 Vertical and horizontal properties of wireless communication sys-tem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2 Abstract illustration of SDSS. . . . . . . . . . . . . . . . . . . . . 6

2.1 Spectrum decision function. . . . . . . . . . . . . . . . . . . . . . 10

2.2 Classication of elements involved in spectrum decision. . . . . . 12

2.3 Three sets involved in spectrum decision. . . . . . . . . . . . . . 13

2.4 Information processing and overlay decision making. . . . . . . . 14

2.5 Abstract structure of SDSS. . . . . . . . . . . . . . . . . . . . . . 19

3.1 Modeling of channel characterization in interval [ t 0 , t H ]. . . . . . . 23

3.2 Three directed graphs under =1.00, 0.67, 0.50, respectively. . . 31

4.1 Example of infrastructure based CR network. . . . . . . . . . . . 354.2 CSMA/CA based protocol for SUs transmission. . . . . . . . . . 37

4.3 Two types of sensing errors: overlook, misidentication. . . . . . 38

4.4 Ten encoding steps. . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.5 The modeled digital tree. . . . . . . . . . . . . . . . . . . . . . . 45

4.6 Average throughput of SUs successful transmission, E [( H )]. . . 52

xi


17/122

4.7 Average wasted throughput due to SUs transmission collision, E [ X ( H )]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.8 Average wasted throughput due to SUs misidentication error, E [Y ( H )]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.1 SUs slot: sensing and receiving broadcast, information exchangeand data transmission are accomplished in the phases I, II andIII, respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

5.2 A binary sequence, which indicates an example of PU channeloccupancy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.3 Example of ve SUs competing for the use of the same channel. . 61

5.4 Two-Step Information-Exchange method for SU transmitters. . . 62

5.5 Membership functions of xn(t ) and ynm(t ) to FCA. . . . . . . . . . 67

5.6 Average dropping probability of SUs versus six scenarios . . . . . 69

5.7 Average blocking probability of SUs versus six scenarios . . . . . 70

5.8 Average success probability of SUs versus six scenarios . . . . . . 70

6.1 CR networks with inter-handoff. . . . . . . . . . . . . . . . . . . 75

6.2 System model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.3 State diagram for the modeled system, g > 0 , h > 0 . . . . . . . . 81

6.4 Inter-handoff calls (PU or SU) handed over by either a local cellor its neighboring cells. . . . . . . . . . . . . . . . . . . . . . . . . 83

6.5 Blocking probability of new SU calls P(1 )

bl versus p1 . . . . . . . . 86

6.6 Blocking probability of inter-handoff SU calls P(2 )

bl versus p1 . . . 87

6.7 Forced-termination probability of ongoing SU calls P f t versus p1 . 88

6.8 SU service-completion throughput Rse versus p1 . . . . . . . . . . 89

xii


18/122

List of Tables

Table Page

4.1 CUS history within 20 identical time slots . . . . . . . . . . . . 41

4.2 Decoded Contexts . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.3 All paths occurring under conditional events . . . . . . . . . . . . 46

4.4 Prediction computation for channel a . . . . . . . . . . . . . . . . 48

4.5 Parameter settings. . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5.1 Parameter Settings . . . . . . . . . . . . . . . . . . . . . . . . . . 68

6.1 The computed inter-handoff arrival rates p2 and s2 ; values (ob-tained from simulations) of SU inter-handoff throughput Rinter arein parentheses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.2 Numerical and simulation (in parentheses) results of blockingprobability of inter-handoff SU calls, P(

2 )bl . . . . . . . . . . . . . . 87

6.3 Numerical and simulation (in parentheses) results of forced-termination

probability of ongoing SU calls, P f t . . . . . . . . . . . . . . . . . 88

xiii


19/122

xiv


20/122

Acronyms

ACO Ant Colony OptimizationAI Articial IntelligenceAPI Application Programming In-

terfaceBER Bit Error RateBTH Blekinge Institute of Technol-

ogyCCC Common Control ChannelCDMA Code Division Multiple Ac-

cess

CI Computational IntelligenceCPU Central Processing UnitCR Cognitive RadioCTMC Continuous Time Markov

ChainCTW Context Tree WeightingCUS Channel Usage StateDM Decision-MakingDSA Dynamic Spectrum AccessDTMC Discrete Time Markov ChainFCA Fuzzy Channel AvailabilityFCC

Federal CommunicationCommissionGSL GNU Scientic LibraryLAPI Lower Application Program-

ming InterfaceLZ78 Lempel-Ziv-78MAC Media Access Control

MDP Markov Decision ProcessesMHz MegahertzNS-2 Network Simulator 2OSA Opportunistic Spectrum Ac-

cessOS Operation SystemPOMDP Partially Observed Markov

Decision ProcessPPM Prediction by Partial MatchPST Probabilistic Suffix Tree

PU Primary UserQoS Quality of ServiceRSS Received Signal StrengthSDSS Spectrum Decision Support

SystemSOP Spectrum OpportunitySU Secondary UserSTL Standard Template LibraryTSIE Two-Step Information-

ExchangeUAPI Upper Application Program-

ming InterfaceUWB Ultra Wide BandUSRP Universal Software Radio Pe-

ripheralWLAN Wireless Local Area NetworkGB Gigabytes

xv


21/122

xvi


22/122

Chapter 1

Introduction

1.1 Introduction

As a result of static frequency allocation policy, todays licensed spectrum

is rarely fully utilized in the domains time, frequency and geographical loca-tion [13]. Spectral under-utilization is contrasting the increasing demand forradio resources (i.e., spectrum channels), which is due to the rapid growth of wireless services and applications such as sensor networks and smart phones.This conict reveals a serious problem of spectrum scarcity challenging inher-ently in 4G and beyond. To overcome the problem, attractive solutions likeDynamic Spectrum Access (DSA) have been advanced. The underlying idea isto open licensed spectrum to unlicensed users, thus allowing them to use so-called spectral white spaces [4, 5]. Meanwhile, a vital assurance raises as the

unlicensed users are not allowed to harmfully interfere with the licensed users(called primary users or PUs). So far, various strategies have been developedfor DSA. A detailed taxonomy of these strategies is reported in [6].

In parallel with investigations on DSA, Cognitive Radio (CR) has been ac-cepted as an emerging technology to improve the performance of radio commu-nication systems. CR was rst coined by Mitola in 1999 [7], and it is envisionedto act as a highly intelligent radio where transmission parameters like, e.g.,

1


23/122

CHAPTER 1. INTRODUCTION

frequency range, transmit power and modulation type can be altered by thelearning of radio environment. This novel communication paradigm leads toan enabling framework CR Networks rst suggested by Federal Communica-tion Commission (FCC). In general, DSA can be thought of as an importantapplication of CR [6], while the rise of CR networks allows for both academicand industry to realize DSA with the aim to obtain an efficient use of spectrumresources. In CR networks, the unlicensed users are equipped with CR devices,and they are called secondary users (SUs). By using CR technology, SUs arecapable of autonomous adaptation to radio environment changes.

Given the dynamic nature of PUs activity, one feasible DSA-based trans-mission model for SUs to use in CR networks is the Opportunistic SpectrumAccess (OSA) model [8]. In this model, if and when a channel is not used bythe PU, the channel becomes available for the use by SUs. These channels arealso known as spectrum holes . Further, it is incumbent for SUs to identify thespatio-temporally available channels. The identication can be done by eitherspectrum sensing or by database based solutions [911]. Furthermore, when thePU occupies a channel, the SU using the same channel must vacate this channel.

Otherwise, the PU transmission would be impaired.For OSA based CR networks, there may exist multiple available channels at

a time moment. Due to the above described PU protection mechanism, SUsneed to decide which channel should be selected for the use in the near future.The selection process is often referred to as spectrum decision . For a given SU,the goal of doing spectrum decision is to select an available channel, which couldoptimize the SUs transmission performance.

In this licentiate thesis, the research interest is focused on studying anddeveloping intelligent spectrum decision strategies for OSA based CR networks.In the following sections, we rst report related work on the investigation of spectrum decision strategies and the cognition approach to intelligent decisionmaking algorithms. Based on the observation of the related work, we thenpresent the motivation and contributions of our research work. After that, wegive the outlines of the thesis.

2


24/122

1.2. RELATED WORK

1.2 Related Work

1.2.1 Decision Strategies

To date, a large amount of spectrum decision strategies have been suggested. Forinstance, in [5], the authors characterize the channels by parameters like receivedsignal strength, interference, path loss, wireless link error, link layer delay. Basedon these parameters, the SUs can select the most available channels. However,

this strategy gives rise to a multiple-constraint based decision problem. Fuzzylogic is useful for solving this problem because of the capability of coping withvarious criteria for decision making purposes. For instance, in [12] the authorsadopt parameters like mobile speed and signal strength received at SU side torepresent the channel availability. Moreover, by introducing a set of fuzzy rules,the authors construct a fuzzy logic based decision system to help SUs in doingchannel selection.

Clearly, the above described spectrum decision strategies rely on the instantinformation. They are not exible for SUs to adapt to the radio environment

varying over time. Therefore, statistical information about communication en-vironment changes is desirable for SUs to do spectrum decision. Along withthis line, the two-state ON-OFF model is widely used to study the arrival anddeparture processes of both PUs and SUs [1315]. Here, the states ON andOFFrepresent PU being present and PU being absent in channels, respec-tively. In [13], the authors suggest an ON-OFF Markov chains based model totheoretically study the interaction between PUs and SUs. Based on a similarapproach, in [14] the authors suggest a forecasting strategy. This strategy isused to predict the spectrum usage by SUs with respect to power-level mea-

sured at SU side. The goal of the prediction is to prevent an overcrowded case,in which case a large number of SUs simultaneously use the same channels.

Alternatively, in [16] the authors suggest a hybrid ON-OF model, in whichthe duration of being OFF state is xed and the duration of being ON state isexponentially distributed. In [15], by applying the hybrid ON-OFF model, theauthors investigate both blocking and dropping probabilities of SUs. A blockevent means there are no channels available for SUs, and a drop event means a

3


25/122


SU is out from an available channel due to channel occupancy by a PU.

1.2.2 Cognition Approach

To be called cognitive, the SUs should have an ability of learning about commu-nication environment changes. To provide this ability, Computational Intelligence(CI) based algorithms are applied in spectrum decision strategies as indicatedin recent literature.

For instance, in [17], the authors suggest a biologically-inspired algorithmto achieve efficient spectrum sharing among SUs. To do this, they rst dene ahandoff event as that a SU vacates a channel and switches to another availablechannel. Then, the authors formulate the decision of performing a handoff as afunction of two parameters. The rst parameter is dened as the (total) transmitpower permitted by a channel. The second parameter is dened as the transmitpower required by a SU when the SU wants to access a channel. These twoparameters are respectively referred to as response threshold and stimuli inten-sity , which are introduced in adaptive task allocation model [18]1 . As a result,SUs can automatically access the most available channels in accordance withthe largest handoff-decision value. However, the biologically-inspired algorithmmay lead to a high handoff rate of SUs. To reduce the handoff rate, the authorsof [19] suggest an improved algorithm. In this algorithm, the values of two pa-rameters response threshold and stimuli intensity are derived with respect toan ON-OFF Markov process. Moreover, in [20,21], the authors apply the game theory in CR networks, so that SUs can behave in a socially constructive waywhen they want to select available channels.

1.3 Motivation

The above mentioned research works lay the ground to develop spectrum deci-sion. However, they do not consider the heterogeneity aspects and the interop-

1 This model is used to study the cooperation behavior in insect colonies [18].

4


26/122

1.3. MOTIVATION

!"#$% '()$%*

+$",-%. '()$%

/01 '()$%

2#)*34(5 '()$%

6 3 / 0

7 6 '

0 + 1 8

/ 0

1%-**9'()$%8$*3:;

63%$5$**


27/122


"#$%& '(()*+,-*#. /&,11*.0 2.-%&3,+%4 5"'/246

/&%7*+-*#.')0#&*-81

9-8%&')0#&*-814

9:%&),; %&?*1@),-*#.9AB%+-*:%4

C((%& '(()*+,-*#. /&,11*.0 2.-%&3,+% 5C'/246

D@EE;"#0*+

F%@&,)F%-$#&>

Figure 1.2: Abstract illustration of SDSS.

modied for CR based communications. For instance, solutions to IEEE 802.11and CDMA based CR networks are reported in [22,23] and [24], respectively.

The above described heterogeneity further leads to various channel charac-terizations. According to section 1.2, channel characterizations can either useinstant parameters like transmit power and bandwidth, or statistical parame-ters like idle time average. Since the SUs have a diversity in terms of differentQoS/QoE requirements, this gives rise to different channel selection criteria forSUs. For instance, a particular SU wants to use the available channels withthe largest bandwidth, while another SU wants to use the ones with the small-est delay. Moreover, different channel selection criteria may be associated withdifferent decision making algorithms. Therefore, to make SUs be adaptive tothe heterogeneity existent in CR networks, the spectrum decision maker needsan interoperability framework across various characterization parameters anddifferent decision making algorithms.

1.4 Contributions

The main contributions of the licentiate thesis are as follows:

A system called Spectrum Decision Support System (SDSS) is suggested,

6


28/122

1.4. CONTRIBUTIONS

as shown in Fig. 1.2. To address the heterogeneity issue, we develop inSDSS the upper/lower Application Programming Interfaces (APIs). Theupper APIs allow SDSS users to embed SDSS in other tasks like, e.g., de-signing of CR networks based architectures [11], setting up of CR networksbased testbeds, queueing theory based analysis, cognitive MAC protocoldesign. The lower APIs allow SDSS users to modify the decision makingalgorithms in SDSS and integrate new ones in SDSS. To support interop-erability, SDSS adopts a new parameter Fuzzy Channel Availability (FCA)dened from fuzzy logic point of view. The goal of FCA is to transformdifferent types of parameter values in an uniform type with respect toFCA. Hence, SDSS is capable of jointly considering various character-ization parameters. Furthermore, we also build up an overlay decisionmaker in SDSS. This overlay decision maker is capable of coordinatingand employing different decision making algorithms.

A channel usage prediction based algorithm for spectrum decision is devel-oped. This algorithm is accomplished by using LeZi-update scheme andfuzzy logic. The channel usage prediction is based on the joint considera-

tion of sensing error, SUs competition and SUs transmission collision. Theprediction goal is to help SUs in knowing in advance the channel avail-ability, and thus the SUs can choose the most available channels. Theeffectiveness and performance of the developed algorithm is evaluated bysimulation experiments.

A fuzzy logic based decision making algorithm is developed for competition-based channel selection. The underlying decision criterion integrates boththe statistics of PUs channel occupancy and the competition level of SUs.

By using this algorithm, the SU competitors can achieve an efficient shar-ing of the available channels. Simulation results are reported to demon-strate the performance and effectiveness of the suggested algorithm.

A Markov chain based queueing model is developed to represent the OSAperformance in cellular CR networks. To do this, both SUs intra-cell andinter-cell spectrum handoff are considered. By assuming that multiple cellsshow identical statistics in steady state, we determine the values of arrival

7


29/122


rates of inter-cell handoff users. We also derive the corresponding perfor-mance metrics in terms of blocking and forced-termination probabilities of SUs, service-completion throughput of SUs and inter-handoff throughputof SUs. The numerical analysis is validated by simulation experiments.

1.5 Outline

The rest of the thesis is organized as follows. In Chapter 2: Spectrum DecisionSupport System, we discuss the spectrum decision function in CR networks andpresent the functionalities of SDSS. In Chapter 3: Overlay Decision Making,we develop an overlay decision maker for spectrum decision by using fuzzy logic.In Chapter 4: Channel Usage Prediction, we suggested a new spectrum deci-sion strategy based on channel usage prediction by using LeZi-update scheme.In Chapter 5: Competition-Based Channel Selection, we suggest a fuzzy logicbased hybrid decision making algorithm to alleviate the problem of competitionamong SUs. In Chapter 6: Intra- and Inter-Handoff, we analyze the OSAperformance in the presence of SUs intra- and inter-handoff. Finally, Chapter 7concludes the research and presents the future work.

8


30/122

Chapter 2

Spectrum DecisionSupport System

Spectrum decision has emerged as an important research issue in CR networks.This is because of its crucial functionality of determining the most available

channels for SUs to use. By using spectrum sensing and spectrum analysis ,SUs can obtain information about the channel availability. The information canbe either instant or statistical channel characterizations in terms of differentparameters. The parameter values can be collected by spectrum decision maker,which further performs spectrum decision on the basis of various decision makingalgorithms. Furthermore, the decision results guide SUs to the most availablechannels. The associated SUs activities refer to spectrum mobility and spectrum sharing 1 .

The spectrum decision is not independent from spectrum sensing, analysis,mobility and sharing. This raises the question of complexity of doing spectrumdecision in CR networks. In this chapter, we suggest a system Spectrum DecisionSupport System (SDSS) that makes spectrum decision simpler to perform.

1 Spectrum sharing indicates not only the sharing between PUs and SUs, but also thesharing among SUs themselves.

9


31/122

CHAPTER 2. SPECTRUM DECISION SUPPORT SYSTEM

2.1 Spectrum Decision

In this section, a mathematical model is advanced to represent the functionalityof spectrum decision in CR networks. We also build a graph-based model todetail the two tasks for doing spectrum decision in SDSS.

2.1.1 Mathematical Model

Fig. 2.1 shows the relationships among spectrum decision, spectrum sensing,spectrum analysis, spectrum mobility and spectrum sharing.

!"#$%&'(*#$+,+-.

!"#$%&'( /-0+1+%2!"#$%&'( !34&+.5

!

!

!

!"#$%&'( !#.,+.5!"#$%&'( 6.412,+,

!

!

! 7 3 4 & 4

$ % # & + 8 4

% + - . ,

! # 1 # $

% # 9 7 3

4 . . # 1

Figure 2.1: Spectrum decision function.

In this gure, the spectrum decision is represented as a black box that pro-vides a function in terms of input (characterization parameters), output (se-lected channels), and internal process (doing decision making).

Further, we let A (t ) denote the set of all parameter values at time t , and letc(t ) denote the index of a selected channel based on A (t ). Hence, the rule thatconnects A (t ) and c(t ) is given by one or more decision making algorithms. Welet f denote such rule and also call the rule as spectrum decision function .

According to rule f , if the spectrum decision maker can assign to A (t ) ex-actly one element, denoted by f (A (t )), we say that f (A (t )) is the value of f atA (t ). Furthermore, f (A (t )) indicates a decision making result, which leads tothe selected channel c(t ). Subsequently, we can represent the functionality of spectrum decision by an arrow diagram:

A (t ) f f (A (t )) c(t ) (2.1)

10


32/122

2.1. SPECTRUM DECISION

In the arrow diagram, f (A (t )) is a variable dependent upon A (t ), while A (t )is an independent variable compared with f (A (t )). However, given the selectedchannel c(t ), if a SU, denoted by s, starts using channel c(t ) at time t , the SUactivity in c(t ) may effect the channel characterization of channel c(t ) after timet . For instance, for other SUs, the available bandwidth of channel c(t ) may bereduced due to the use by SU s.

With respect to the arrow digram (2.1), the key objective that we deal within SDSS is spectrum decision function f . To implement f in SDSS, we focus on

two tasks, namely, the processing of channel characterizations and the overlaydecision making. To address the two tasks, we build up a graph-based model asfollows.

2.1.2 Graph-based Model

We rst consider the elements involved in doing spectrum decision. Here, byelement we mean an individual entity in CR networks, which can be like, e.g.,a PU, a SU, a channel, a CR network model, a decision making algorithm.Further, we classify various elements in three sets: active unit , information base , and algorithm factory , as shown in Fig. 2.2. The relationships amongthese three sets are shown in Fig. 2.3(a).

The active unit includes the PU, SU, radio channel, transmission schemeand CR coordinator. PUs, SUs and radio channels are essential elementsin CR networks. PUs and SUs can do transmission by using an ad-hoc oran infrastructure based manner. The corresponding transmission schemecan use either time-slotted basis or continuous basis. The channel avail-ability for SUs can be spatial invariant or be spatially varying [8]. CRcoordinator indicates the function of allocating available channels to SUs.CR coordinator can be deployed in either centralization or decentralizationway in CR networks. By centralization way, we mean the CR coordinatorcan be like a Support Node [11] or a Secondary Base Station [24]. By de-centralization way, we mean the function of CR coordinator is embeddedinto every SU.

11


33/122


Elements

Active unit

Coordinator Centralization based

Decentralization based

PUs, SUs Ad-hoc based

Infrastructure based

Channels Availability is spacial invariant

Availability is spatially varying

Transmission scheme Time-slotted basis

Continuous time basis

Information base

Information type Instant parameters

Statistical parameters

Collection of information

Spectrum sensing

Information exchange

Database

Algorithm factory

Articial Intelligence (AI)

Computational Intelligence (CI)

Markov decision process

Figure 2.2: Classication of elements involved in spectrum decision.

The information base consists of two different parts, namely, the informa-tion content and the way of collecting infromation The information content

12


34/122

2.1. SPECTRUM DECISION

!"#"$%"&$'())"#

*+##"$%,+) +.,).+/0(%,+)

*'(/($%"/,1(%,+)2

3$%,4"5),%

6).+/0(%,+) 7(2"

3#8+/,%'0 9($%+/:

(a) Triangle relationship

! " # $ #

% & ' $ ( ) #

& ( * + ,

- ' . ' %

& ' / ! "

# + + ' .

-&*%"#,&(%122$*#%"

34))5!*+&$*.

1+&!*.*+56+,2($'/

7#8'9"'*$5

:&"'$, ;;;

1.


35/122


!"#"$%"&()*++"#

(,##"$%-,+ ,. /+.,01*%-,+

2$%-3"4+-%/+.,01*%-,+5*6"

/+.,01*%-,+70,$"66-+8

93"0#*:;"$-6-,+


36/122

2.2. SCIENTIFIC FUNCTIONALITIES OF SDSS

2.2 Scientic Functionalities of SDSS

To carry out information processing and overlay decision making , we suggestve scientic functionalities for SDSS, which are describe as follows.

2.2.1 Overlay Decision Maker

So far, most of previously suggested decision making algorithms work indepen-

dently from each other, and different algorithms are used for different researchgoals. On the other hand, the recent research on CR networks concentrates onsetting up practical CR network testbeds and architectures [11]. This raises twoimportant questions:

Development : do the researchers need to individually implement all sug-gested decision making algorithms, or instead to only suggest new decisionmaking algorithms?

Compatibility : if the newly suggested algorithms are evaluated to beuseful for spectrum decision, then how to easily make them applicable toother CR network testbeds or architectures?

These two questions motivate us to develop an overlay decision maker, with thegoal to manage and to coordinate different decision making algorithms to worktogether.

2.2.2 Learning and Prediction

To identify channel availability, SUs need to observe different parameter of chan-nels. The observation activity can be like, e.g., spectrum sensing. However, thecontinuous observation at SUs side is inefficient in terms of energy consumptionand hardware demanding. Hence, the observation usually is done on a periodictime basis, for instance, time-slot based spectrum sensing. However, by doingthis, the observation results may correspond to partial characterization aboutCR network environment.

15


37/122


To approach the full knowledge of channel availability, we suggest the SUs tolearn from historical observation results. Further, by using the learned knowl-edge, the decision maker is capable of estimating in advance the channel avail-ability in the near future. The corresponding estimation process is known asprediction.

2.2.3 Secondary Users Cooperation

In CR networks, the SUs can share the same information about the CR networkenvironment. When several SUs need the channel at the same time, a simpledecision maker may lead to the same available channels. In this case, the mul-tiple SUs would compete for the channel utilization over a single channel. If alarge number of SUs simultaneously use a particular channel, the channel maybecome overcrowded, and thus Quality of Service (QoS) performance for SUsdegrades. Therefore, it is necessary that SUs access different available channelsas much as possible.

Furthermore, to avoid SUs overcrowding in available channels, we suggestSUs to work in a cooperative manner. This means that the SUs exchange theinformation about channel utilization via the Common Control Channel (CCC)or other signaling protocols [25]. Based on this, the SUs can behave in self-organizing manner to efficiently share the available channels.

2.2.4 Queueing Modeling

The particular spectrum decision strategy may affect the overall performanceof the whole CR network system. Investigation of related performance char-acteristics is usually based upon theoretical analysis or simulation experiment.In SDSS, we employ queueing modeling to carry out theoretical analysis. Theconsidered queueing models are, e.g., m-server loss system, nite customer popu-lation system, nite customer population and nite storage system. In Chapter6, we develop a queueing model to study the performance of cognitive radiospectrum access with intra- and inter- spectrum handoff.

16


38/122

2.3. SDSS FEATURES AND STRUCTURE

2.2.5 Cognitive Radio Simulator

To validate the theoretical study on CR network performance, we develop a CRsimulator to conduct simulation based experiments. The simulator consists of two parts. The rst part is to congure the simulation parameters and scenar-ios. For instance, the simulation parameters related to ON-OFF model can belike the arrival rates of PUs and SUs. Further, a simulation scenario can becongured as an ad-hoc CR network, where the channel availability is spatiallyinvariant for all SUs. By this, we mean every SU can identify the same set of available channels at the same time.

The second part is to simulate the dynamic behaviors of both PUs and SUs,and to deal with the interactions among PUs and SUs. The interactions can belike, e.g., a PU accesses/releases a channel, a SU vacates a channel due to PUchannel occupancy.

2.3 SDSS Features and Structure

Based on above suggested ve functionalities, the development of SDSS has twogoals. The rst goal is to provide an intelligent spectrum decision strategy toSUs in CR networks. The second goal is to help SDSS users like, e.g., researchers,industries, to concentrate on other research issues regarding CR networks likespectrum sensing, spectrum mobility. To do these, we present the features andstructure of SDSS in the following.

2.3.1 SDSS Features

The main features of SDSS are as follows:

Setting up an uniform decision criterion. This is done by using Com-putational Intelligence (CI) techniques, such as Fuzzy Logic and NeuralNetwork. Fuzzy Logic is considered because of the capability of dealingwith various criteria for decision making purpose.

17


39/122


Providing several CR network models for theoretical analysis and simula-tion experiment. Based on a particular model, the SUs transmission per-formance can be evaluated with respect to specic metrics. For instance,in m-server loss system, the performance metrics can be the blocking prob-ability of SUs due to no available channel, the dropping probability of SUsdue to PU channel occupancy, SUs spectrum handoff probability becauseof channel switching, and SUs service-completion throughput.

Providing the Upper Application Programming Interface (UAPI). The

UAPIs are platform-independent interfaces to SDSS users 2 . In otherwords, SDSS users can conveniently embed SDSS into their own researchtasks called upper objectives . The upper objectives can be, e.g., simulationbased study of CR networks 3 , architecture setting up, testbed evaluation,algorithm development.

Providing the Lower Application Programming Interface (LAPI). TheLAPIs help SDSS users in updating the old algorithms existing in SDSSand adding new algorithms in SDSS. Once the new algorithms are addedinto SDSS, they are available for the use by all SDSS users. In particu-lar, the updated and added algorithms can be evaluated by experimentsor simulations on the basis of the specic upper objectives. Therefore,the functionalities of SDSS for spectrum decision can be improved andextended.

2.3.2 SDSS Structure

We abstractly illustrate the structure of SDSS in Fig 2.5. In the gure, the com-

ponents in the area of grey color compose the SDSS. We present the connectionsamong these components as follows:

The arrows of type (1) indicate that SDSS is applied to upper objectiveslike testbeds/architectures development, simulation experiments.

2 The SDSS users can be, e.g., academic/industrial researchers.3 For instance, a simulation experiment based on Network Simulator 2 (NS-2) [26, 27].

18


40/122

2.3. SDSS FEATURES AND STRUCTURE

!"# %&'( ') *+,-./01 2,-3*3'4!5# 64.,/47& )04-.3'4 ') 8988

!:# ;1# ?+27.34= 742 ;@7&07.3'4

;@7&07.3'4*

A.B,/ 1,.B'2*C D71, EB,'/FG D,4,.3- H&='/3.B1G I7-B34, J,7/434=G KKK

8988 L J'(,/ HM6*

M/,23-.3'4H&='/3.B1

H4. N'&'4FA+.313O7.3'4

A.B,/I'20&,*

A@,/&7F 9,-3*3'4 I7P,/%0OOF J'=3-

Q,0/7&Q,.('/P

8310&7.3'4A


41/122


The arrows of type (4) show the internal functions of SDSS. Fuzzy logicmodule is responsible for processing all input characterizations with re-spect to the uniform decision criterion. After doing this, the fuzzy logicmodule sends the processed information to the overlay decision maker.The overlay decision maker uses this information to do hybrid decisionmaking based on different decision making algorithms. Moreover, thecommunication between the fuzzy logic module and overlay decision makeris bidirectional. This implies that the decision making inside SDSS is aself-learning process.

2.4 Summary

In this chapter, we addressed the complexity of spectrum decision function. Weclassied the different elements, involved in CR network, in three sets. They areactive unit, information base and algorithm factory. Based on this classication,ve different functionalities of implementing the Spectrum Decision SupportSystem (SDSS) are presented. The ve functionalities are: overlay decisionmaker, learning and prediction, SUs cooperation, queueing modeling approach,CR simulator. We further discussed the four features of SDSS, and illustratedthe structure of SDSS as well as corresponding descriptions.

20


42/122

Chapter 3

Overlay Decision Making

In CR networks, different decision making algorithms may have different de-cision criteria. These decision criteria may give different constraints to spec-trum decision maker. Therefore, if multiple different decision making algo-rithms are jointly taken into account, the spectrum decision becomes a multiple-constraint based decision making problem. Towards solving this problem, athree-dimension model is used to represent channel characterization, and anoverlay decision maker is constructed by using fuzzy logic and graph theory.

3.1 Channel Characterization

We consider a CR network with M radio channels denoted by c1 , c2 ,... , cm,... , c M .

We assume that every channel is characterized by a number, N , of parameters,each with label of e1 , e2 ,...,e N , respectively. Further, we assume that these N parameters are independent from each other.

Let amen (t ) denote the observed value of a specic characterization parameteren on channel cm at time t , where n { 1 , 2 ,..., N } and m { 1 , 2 ,..., M }. At timet , the characterization of M channels under N parameters can be represented bya matrix, denoted by A (t ):

21


43/122

CHAPTER 3. OVERLAY DECISION MAKING

A (t ) =

a 1e1 (t ) a1e2 (t ) ... a

1e N (t )

a 2e1 (t ) a2e2 (t ) ... a

2e N (t )

......

. . . ...

a me1 (t ) ame2 (t ) ... a

me N (t )

......

. . . ...

a M e1 (t ) a M e2 (t ) ... a

M e N (t )

(3.1)

The matrix A (t ) is called the channel characterization at time t . Further-more, we assume that the observation of characterization parameters is doneon a periodic time basis with uniform duration . We also assume that theobservation starts at time point t 0 and ends at time point t H , where t H and t 0satisfy the equation:

(t H t 0 ) = H (3.2)

Thus, we can obtain a set of channel characterization matrixes at discrete timepoints {t 0 , t 1 ,..., t H }, which is expressed as:

{A (t 0 ), A (t 1 ),..., A (t H )} (3.3)

For a particular parameter, we further assume that every channel holds thesame value during the time interval [ t h , t h+ 1 ], where h = 0 , 1 , 2 ,..., H 1 . Subse-quently, the characterization of M channels during the time period [ t 0 , t H ] can berepresented by a three-dimension model. The three dimensions are associatedwith three domains, namely, time, frequency (i.e., channels) and parameter. Anexample of modeling of ve channels under multiple parameters such as PU

channel occupancy, delay and bandwidth is illustrated in Fig. 3.1.

3.1.1 Suggested Decision Making Strategies

Based on channel characterization model, the spectrum decision maker dealswith two main problems: learning from historical information and doing decisionmaking according to multiple constraints.

22


44/122

3.1. CHANNEL CHARACTERIZATION

!" !!" "

!" #!" $

!" % !

!!

""

#$

%&'

"

""

!!

($

)&'

!

!"

"!

#$

(&'

!

!"

!!

($

*&'

#$

+&'

($

,&'

*$

*&'

#$

)&'

$%&'()'*+',-.

/0)0&'1')2

! 3! " 4! 3! #4

! 3! $%# 4! 3! $ 4

" " "

" " "

" " "

" " "

" " "

" " "

" " "

" " "

" " "

51"') /0)0&'1')2

/6 5--+70,-.

8'90.

:0,;


45/122


3.2 Uniform Decision Criterion

Since different parameters may vary in distinct metrics and measures, this givesrise to a multiple-constraint based decision problem of nding the most avail-able channel for SUs. Fuzzy-logic is suggested to solve this problem. We rstintroduce a parameter named as Fuzzy Channel Availability (FCA).

Proposition : Fuzzy Channel Availability is a fuzzy-logic based parameter torepresent three different levels of channel availability for SUs. The three levels

are respectively formalized as three fuzzy sets, namely, high-level, medium-level and low-level channel availabilities.

The use of FCA is to map different types of parameter values to an uniformtype, i.e., fuzzy membership degree.

3.2.1 Fuzzy Membership Degree

Let en denote the set of all possible values by observing parameter en on allchannels during a long time period. For the mth channel at time t , we havea men (t ) en . By considering fuzzy set theory, we introduce a function gen actingas characteristic function of set en . Specically, gen is generalized to a mem-bership function in such a way that, for every amen (t ) en , we assign gen (a

men (t ))

a value from the unit interval [ 0 .0 , 1 .0 ].

For example, at time t , we dene a fuzzy set B that is determined by the setof pairs:

B = {(a men (t ), gen (a men (t ))}, amen (t ) en , gen (a men (t )) [0 , 1 ] (3.4)

where gen (a men (t )) is called the fuzzy membership degree that quanties the gradeof membership of amen (t ) to B. For instance, the case of gen (a

men (t )) = 0 means

that amen (t ) is not a member of B, whereas the case of gen (amen (t )) = 1 means that

a men (t ) is a full member of B. In addition, the case of 0 < gen (amen (t )) < 1 implies

that amen (t ) belongs to B only partially.

24


46/122

3.2. UNIFORM DECISION CRITERION

3.2.2 FCA Based Channel Availability

We adopt the notations en , en and en to denote three fuzzy sets high-level,medium-level and low-level under parameter en , respectively. Accordingly,their fuzzy membership functions are denoted by g en , g

en , and g

en , respectively.

For a given channel cm at time t , we have:

[FCA |en , m, t ] = {[ en |m, t ], [ en |m, t ], [ en |m, t ]} (3.5)

where

[ en |m, t ] = {(a men (t ), g en (a men (t ))}



(3.6)

and amen (t ) en , g en (a

men (t )) , g

en (a men (t )) , g

en (a men (t )) [0 .0 , 1 .0 ].

In equation (3.6), g en (amen (t )), g

en (a men (t )) and g

en (a men (t )) are called fuzzy mem-

bership degrees of amen (t ) to fuzzy sets en , en , and en , respectively. The threefuzzy membership degrees further form a vector:

V men (t ) = g en (a

men (t )) , g

en (a

men (t )) , g

en (a

men (t )) (3.7)

We call V men (t ) the FCA-based characterization of observed parameter en in chan-nel cm at time t . Since V (t ) is a three-coordinate vector, it is not convenientto carry out the numerical computing regarding decision making. This hasprompted the development of method to compound three coordinates into a joint value referred to as the channel availability.

3.2.3 Fuzzy-Comparison

To obtain the above mentioned joint value, we adopt a fuzzy-comparison basedalgorithm developed by Saaty [28]. The algorithm is based on using a paired-comparison of three fuzzy sets importance in deciding on which channel is mostavailable.

25


47/122


Let , + , and denote the importance of high-level, medium-level,and low-level, respectively. For example:

Since high-level has strong importance over low-level , we assign

with

5.

Since high-level and medium-level respectively have weaker importancethan medium-level and low-level , we assign both

+

and +

with 3.

Since high-level , medium-level , or low-level has equal importance over it-self, we have

= + +

=

= 1 .

Similar to [29], we can therefore obtain a fuzzy-comparison matrix, denoted by , as:

=

/ + / /

/ + + / + / +

/ + / /

=

1 3 5

1 / 3 1 3

1 / 5 1 / 3 1

(3.8)

The matrix is used to determine the numerical values of , + , and ,respectively. Given the eigen value and eigen vector of matrix , theysatisfy the eigen equation and characteristic equation as:

=

det ( I ) = 0(3.9)

where I is an unit matrix. The largest real eigen value corresponds to an eigenvector, denoted by and it is given by:

= { , + , } {0.94

,0.31

,0.19

} (3.10)where the three coordinates , + and are associated with , + and ,respectively.

Consequently, three coordinates of V (t ) can be composed in the term of alinear combination:

men (t ) = g en (a

men (t )) + g

en (a men (t )) + + g

en (a men (t )) (3.11)

26


48/122

3.3. GROUP DECISION MAKING

We call men (t ) the FCA-based decision factor of parameter en on channel cm attime t . We can further compute different FCA-based decision factors me1 (t ),

me2 (t ),

..., me N (t ), which correspond to parameters e1 , e2 ,..., e N , respectively.

3.2.4 Hybrid Decision Making

Although the computed FCA-based decision factors of different parameters havethe same value type, their respective weights for doing decision making still need

to be congured. Assume that we can assign every parameter en { e1 , e2 ,..., e N }with a weight value wn . We have a weight vector , denoted by W that is givenby:

W = ( w1 , w2 ,..., wn ,..., w N ) (3.12)

where n = 1 , 2 ,..., N .

Moreover, let m(t ) denote the numerical channel availability (for SUs) of channel cm at time t . If the decision maker jointly takes into account the multipleparameters e1 , e2 ,..., en , m(t ) can be computed as:

m(t ) = me1 (t )w1 + me2 (t )w2 + ... + men (t )wn + ... + me N (t )w N (3.13)

Therefore, the most available channel in this particular model is based the de-cision with the largest value of m(t ), which is given by:

m(t ) = max{ 1 (t ), 2 (t ),..., M (t )} (3.14)

where m = 1 , 2 ,..., M .

3.3 Group Decision Making

In subsection 3.2.4, we assumed that the weights of all FCA-based decisionfactors can be prescribed. However, in some particular cases, it is possible thatwe can not pre-assign a FCA-based decision factor with an adequate weightvalue. Therefore, we introduce another algorithm called group decision to dospectrum decision. The group decision is proposed by Blin [30], and it is also

27


49/122


based on fuzzy logic. For instance, the author of [31] adopts the group decisionto make the choice of the optimal medicine.

In our work, we dene E and C as the sets of N parameters and M channels,respectively, i.e.:

E = {e1 , e2 ,..., en ,..., e N } (3.15)

andC = {c1 , c2 ,..., cm,..., c M } (3.16)

where n = 1 , 2 ,..., N and m = 1 , 2 ,..., M .

Different parameters may lead to respective decision criteria of estimatingchannel availability. Take two parameters delay and bandwidth as example.In general, the larger the bandwidth of a channel is, the higher the channelavailability becomes. Oppositely, large transmission delay in a channel implieslow channel availability.

Given the decision criterion of parameter en , if the availability of channelcm is higher than the one of channel cm , this comparison result is denoted by

cmen

cm

. In the following subsections, we consider an example to present thedetailed algorithm of doing group decision.

3.3.1 Example

Let us consider six parameters, i.e., E = {e1 , e2 , e3 , e4 , e5 , e6 }, and three availablechannels, i.e., C = {c1 , c2 , c3 }. We adopt the denotation to stand for either

e1,

e2,

e3,

e4, or

e5.

With respect to six parameters, aassume that we can obtain the followingdecision results for channel selection.

According to either e1 , e3 , or e4 , the decision maker judges that c1 c2 c3 .

According to e2 , the decision maker judges that c2 c1 c3 .

According to either e5 or c6 , the decision maker judges that c2 c3 c1 .

28


50/122


Based upon these different decision results, the goal of group decision is to ndthe best channel.

3.3.2 Group Order

To do group decision, we rst introduce a parameter called group order . Thegroup order is denoted by R , which is a 3 3 matrix and it is dened as:

R =

R(c1 , c1 ) R(c1 , c2 ) R(c1 , c3 )

R(c2 , c1 ) R(c2 , c2 ) R(c2 , c3 )

R(c3 , c1 ) R(c3 , c2 ) R(c3 , c3 )

(3.17)

In matrix R , the coordinate R(ci, c j) is given by:

R(ci, c j) = | i, j|

| E | (3.18)

where

i

, j = {en : decides that c i c j}

i, j = 1 , 2 , 3

n = 1 , 2 , 3 , 4 , 5 , 6

(3.19)

Based on equation (3.18), we can compute all coordinates of R as:

R(c1 , c2 ) = |{e1 , e3 , e4 }|

|{e1 , e2 , e3 , e4 , e5 , e6 }| =

3

6 = 0 .50

R(c1 , c3 ) = |{e1 , e2 , e3 , e4 }|

|{e1 , e2 , e3 , e4 , e5 , e6 }| = 4

6 0 .67

R(c2 , c1 ) = |{e2 , e5 , e6 }|

|{e1 , e2 , e3 , e4 , e5 , e6 }| =

3

6 = 0 .50

R(c2 , c3 ) = |{e1 , e2 ,e3 , e4 , e5 , e6 }||{e1 , e2 , e3 , e4 , e5 , e6 }|

= 6

6 = 1 .00

29


51/122


R(c3 , c1 ) = |{e5 , e6 }|

|{e1 , e2 , e3 , e4 , e5 , e6 }| =

2

6 0 .33

R(c3 , c2 ) = 0

|{e1 , e2 , e3 , e4 , e5 , e6 }| =

0

6 0 .0

R(c1 , c1 ) = R(c2 , c2 ) = R(c3 , c3 ) = 0 .0

Finally, we can obtain R as:

R =

R(c1 , c1 ) R(c1 , c2 ) R(c1 , c3 )

R(c2 , c1 ) R(c2 , c2 ) R(c2 , c3 )

R(c3 , c1 ) R(c3 , c2 ) R(c3 , c3 )

=

0 .0 0 .50 0 .67

0 .50 0 .0 1 .00

0 .33 0 .0 0 .0

(3.20)

3.3.3 Graph Based Decision

Based on the above dened group order R , we introduce a variable equal toone of coordinates of R , i.e., { 1 .00 , 0 .67 , 0 .5 , 0 .33 , 0 }. Further, let R denotethe cut set of R . By this, we mean that the coordinates in R are from Rand they must not be less than . Namely, in our case, we have ve cut sets:

R1 .00 = { R(c2 , c3 )} = {1 .00 }

R0 .67 = { R(c2 , c3 ), R(c1 , c3 )} = {1 .00 , 0 .67 }

R0 .50 = { R(c2 , c3 ), R(c1 , c3 ), R(c1 , c2 ), R(c2 , c1 )}

= {1 .00 , 0 .67 , 0 .5 , 0 .5 } R0 .33 = { R(c2 , c3 ), R(c1 , c3 ), R(c1 , c2 ), R(c2 , c1 ), R(c2 , c1 )}

= {1 .00 , 0 .67 , 0 .5 , 0 .5 , 0 .33 }

R0 .0 = { R(c2 , c3 ), R(c1 , c3 ), R(c1 , c2 ), R(c2 , c1 ), R(c1 , c1 ),

R(c2 , c2 ), R(c3 , c2 ), R(c3 , c3 )}

= {1 .00 , 0 .67 , 0 .5 , 0 .5 , 0 .0 , 0 .0 , 0 .0 , 0 .0 }

30


52/122


!"

#" $"

!"

#" $"

!"

#" $"

% & "

Figure 3.2: Three directed graphs under =1.00, 0.67, 0.50, respectively.

Based on each of the above ve cut values, we further build a directedgraph, in which three channel denotations, i.e., c1 , c2 , c3 are taken into con-sideration as three vertexes. In such a graph, we also dene a directed arrowci c j corresponding to the pair ( ci, c j) referred to as an edge in the graph. Fora particular directed graph, if every pair of vertexes is connected by a directedarrow, the graph is called total graph . Furthermore, if a vertex in a total graphhas the highest member of arrows, the vertex corresponds to the best channel.

In group decision algorithm, the key process is to nd a total graph withregard to a specic cut set. In our example, we start with doing group deci-sion by considering the rst cut set R1 .00 = { R(c2 , c3 )}. The correspondingdirected graph is shown in Fig. 3.2-a. However, the gure shows that the graphis not total as there are no edges between c1 and either c2 or c3 .

Regarding cut set R0 .67 = { R(c2 , c3 ), R(c1 , c3 )}, we show the directedgraph in Fig. 3.2-b. This graph is not total either for the same reason with thegraph under R0 .67 .

Regarding cut set R0 .50 = { R(c2 , c3 ), R(c1 , c3 ), R(c1 , c2 ), R(c2 , c1 )}, weshow the directed graph in Fig. 3.2-c. In the graph, since all three vertexesare connected by directed arrows, the graph is a total graph. We nd that: 1)both vertexes c1 and c2 are connected by three arrows, but 2) the vertex c3 is

31


53/122


connected by two arrows. As a result, we can make an overlay decision makingas: c1 = c2 c3 . In other words, both channels of c1 and c2 have the samechannel availability, and they have higher channel availabilities than the one of c3 .

3.4 Summary

In this chapter, we presented two spectrum decision strategies for doing spec-trum decision. In the rst strategy, we introduced a parameter called FuzzyChannel Availability (FCA) from fuzzy logic point of view. By using FCA,different types of characterization parameter values are transformed into anuniform type. We further integrated the transformed values together. The in-tegrated value stands for the numerical channel availability. Thus, the mostavailable channel in this particular case is determined by the largest integratedvalue. The second strategy refers to a group decision algorithm developed onthe basis of graph theory and fuzzy logic theory.

32


54/122

Chapter 4

Channel Usage Prediction

This chapter is about a new strategy suggested for spectrum decision in CR net-works. By jointly considering sensing error, SUs competition and SUs transmis-sion collision, a new parameter called Channel Usage State (CUS) is introduced.For a particular channel, we predict the respective probabilities of occurrence of CUS states by using the LeZi-update scheme. We also adopt a fuzzy comparisonalgorithm to combine the prediction results as a joint value. The largest jointvalue is associated with the most available channel for access by SUs in the nearfuture. By comparing with random channel access, the suggested strategy canimprove the SUs transmission throughput. This is demonstrated by simulationevaluations.

4.1 Motivation

Due to channel availability varying over time, one approach for optimal spectrumdecision is to learn from channel occupancy statistics of PUs. For example,the authors of [13] suggest a continuous-time Markov chains based modelingapproach to study the arrival and departure processes of both PUs and SUs.However, there are three problems existing in statistics based spectrum decision.

33


55/122

CHAPTER 4. CHANNEL USAGE PREDICTION

They are sensing error, SUs competition and SUs transmission collision, asdescribed in the following. First, the performing of spectrum sensing may beimperfect in practical implementation. The imperfect sensing may create errorsin terms of overlook and misidentication of channel availability [8]. Second,multiple SUs may select the same channel with respect to the same statisticalinformation. As a result, they are likely to compete for the channel utilization inthe selected channel. Third, SUs competition can be alleviated by using CarrierSense Multiple Access with Collision Avoidance (CSMA/CA) based protocol[22, 23]. Under such protocol, one or more SUs may have the same smallestbackoff time in the same selected channel. In this case, their transmissionscollide with each other when backoff time expires.

The above described problems are widely reported in recent studies. How-ever, in most of them, the effects on spectrum decision are separately investi-gated. For instance, the authors of [8] suggest a decentralized MAC protocol forSUs in the presence of sensing errors. While, this protocol does not consider theSUs competition and transmission collision. In [32], the authors report on theimpact of sensing errors and SUs transmission collision, but they do not give

the detailed solutions to overcome the impacts. In [14], under an assumption of perfect sensing, the authors report on an overcrowded case that a large numberof SUs simultaneously use the same channels.

Given a realistic CR networks, all three problems, i.e., sensing errors, SUscompetition and SUs transmission collision, may not be avoided. Therefore,they need to be jointly taken into consideration in developing practical yeteffective spectrum decision strategy. To our best knowledge, there has been littlestudies so far along with this line. Therefore, in this chapter an infrastructurebased CR network is considered, where multiple SUs employ CSMA/CA based

protocol to share multiple channels. The issue of transmission throughput of SUs is addressed by jointly considering sensing errors, SUs competition, and SUstransmission collision. Based on this, a new parameter called Channel UsageState (CUS) is suggested to characterize the channel utilization by PUs andSUs. By using the LeZi-update scheme [33], the uncertainty associated withfuture CUS is predicted based on CUS history. By using a fuzzy comparisonprocedure, the joint value of prediction results is computed, and it indicates the

34


56/122

4.2. SYSTEM MODEL

numerical channel availability. The corresponding system model is described inthe following section.

4.2 System Model

4.2.1 Network Model

We consider a CR network consisting of M radio channels, each with labelc1 , c2 ,..., c M . Every channel has an identical amount of bandwidth, which isdenoted by B. At any specic time moment, every channel can be only used byone user, namely, a PU or a SU.

!"#$

!%&

!%'

Figure 4.1: Example of infrastructure based CR network.

The activity of PUs is assumed to use a synchronous time-slotted basis.Namely, in every slot, PU is either present or absent in a channel during thewhole slot duration. The length of every slot identically equals in time do-main. Moreover, the channel occupancy by PUs is assumed to follow a two-stateBUSY-FREE Markov process. State BUSY means the event that the PU is oc-cupying a channel for one or more consecutive slots. Similarly, state FREEmeans the event that there is no PU in a channel for one or more consecutiveslots. The time periods of two states are integer times of . They are alsoassumed to be exponentially distributed with mean values 1 / p and 1 / p for

35


57/122


states BUSY and FREE, respectively.

In the network, there exist N SUs labeled by s1 , s2 ,..., s N . The channel avail-abilities are assumed to be spatially invariant for all N SUs. In other words, if SUs can do perfect sensing on a channel within a PU slot, the sensing result,i.e., PU being present or absent, is the same at every SU. We also assume thatthere is a Secondary-Access Point (S-AP) in the network. S-AP can be like abase station or a support node [11]. In addition, each of the N SUs alwayswants to transmit data to S-AP via an idle channel, as shown in Fig. 4.1. By

an idle channel, we mean the channel is not being used by any PU or SU at atime moment.

4.2.2 SU Activity Model

By communicating with S-AP, the activity of SUs can be synchronized withPUs. As a result, all SUs operate in a time-slotted basis, which has an uniformslot length same with PUs. At the beginning of each slot, each SU performsspectrum sensing to identify channel availabilities. The sensing duration hasan uniform value equal to , where < . Further, the sensing is assumed tobe sequentially performed on M channels one by one 1 . The sensing result isassumed to be imperfect with probability , where (0 , 1 ). The imperfectionmay be due to factors like sensing-duration limitation and hardware sensitivity.

Once SU identies a set of available channels, it can randomly select oneof them for data transmission. Since multiple SUs may want to use the samechannel in the same slot, this may lead to competition among them. To solvethis problem, a CSMA/CA based protocol is suggested for SUs. This protocol

is briey described in the following (Fig. 4.2).In a slot, after nishing the spectrum sensing, SU rst selects out an available

channel if it exists. Then, SU generates a random backoff time , the value of which is uniformly selected in {0 , w, 2 w,..., Cw}. C is an integer, w is called unitsize of contention window, and the value of Cw is less than ( ). Further,

1 For small number of channels, the sequential sensing is a major approach in recent studies[34]

36


58/122

4.2. SYSTEM MODEL

!"#$%&& '()*+), -.")*/+**+%) 012

Figure 4.2: CSMA/CA based protocol for SUs transmission.

when the backoff time expires, SU detects the existence of other SUs signal 2 . If no other SU has started transmission in the selected channel, the particular SUtransmits data to S-AP in the the remaining slot duration ( ). Otherwise,SU keeps silence until next slot. Moreover, a successful transmission can beacknowledged by S-AP at the end of the slot. For the simplicity of analysis,the duration of receiving acknowledgment message from S-AP is assumed to bezero.

4.2.3 Sensing Errors and Transmission Collision

Due to imperfect spectrum sensing, two different types of errors may occur insensing results: 1) a free channel is sensed to be busy (so-called overlook), and2) a busy channel is sensed to be free (so-called misidentication), as shown inFig. 4.3. Overlook error leads to the overlooking of transmission opportunityin available channels. When misidentication error occurs, SU may select a

busy channel. Further, by performing signal detection after backoff time, SUmay know the existence of PU signal in the selected channel, and thus it keepssilence until next slot. Since SU may be not able to differentiate PU signal fromother SU signals, it may obtain incorrect information that another SU is usingthe selected channel.

2 As described in [22], the detection usually operates at a much higher signal strength thanspectrum sensing, and thus the result is assumed to be perfect without any error.

37


59/122


!"##$%&'

!"##$%&'

!

()*++#, &.*.%&/+ "#*,/.'

!"##$%&' !"##$%&'

! + "#&%,.&

12#",113 4/&/0#+./5/(*./1+

Figure 4.3: Two types of sensing errors: overlook, misidentication.

On the other hand, the CSMA/CA based protocol may lead to transmissioncollision among SUs. That is, two or more SUs may select the same free channelin a slot. If these SUs generate the same backoff time, they can start transmittingat the same time when backoff time expires. Subsequently, transmission collisionamong them happens, and thus they experience unsuccessful transmission by

missing acknowledgment message from S-AP.By jointly taking into consideration sensing errors, SUs competition and SUs

transmission collision, a channel usage model is presented in the next subsection.

4.2.4 Channel Usage Model

Consider that the PU activity changes at discrete time points {t 0 , t 0 + , t 0 +2 ,..., t 0 + H }. Let hi denote the time interval [ t 0 + ( i 1 ) , t 0 + i ], where

i = 1

,2,

3,..., H . Let Am(i) denote the availability of channel cm in slot hi, wherecm { c1 , c2 ,..., c M }. Am(i) is given by:

Am(i) =1 , cm is actually free 0 , cm is actually busy

(4.1)

Let m,n(i) denote the possible backoff time used by SU sn in channel cm withinslot hi, where sn { s1 , s2 ,..., s N }. If channel cm is selected by SU sn in slot hi,

38


60/122

4.2. SYSTEM MODEL

m,n(i) is in value interval [ 0 ,Cw ]. Otherwise, m,n(i) = + .

Dm,n(i) =

1 , [m,n(i) = min{m,1 (i), m,2 (i),..., m, N (i)}| Am(i) = 0 , m,n(i) = + ]

0 , [m,n(i) > min{m,1 (i), m,2 (i),..., m, N (i)}| Am(i) = 0 , m,n(i) = + ]

or [ Am(i) = 1 , m,n(i) = + ](4.2)

Further, let Dm,n(i) denote the decision on whether or not SU sn can transmitdata in the selected channel cm within slot hi. The decision result is givenby using CSMA/CA based protocol. Dm,n(i) = 1 means that SU sn can starttransmission when its backoff time expires. Oppositely, Dm,n(i) = 0 means thatSU sn should keep silence after backoff time expires. The numerical expressionof Dm,n(i) is shown by equation (4.2).

Furthermore, let T m,n(i) denote the transmission status of SU sn in channelcm within slot hi, which is:

T m,n(i) =1 , successful 0 , unsuccessful due to collision

(4.3)

For the time interval [ t 0 , t i + H ], let ( H ) denote the total throughput success-fully used by SUs, and let X ( H ) denote the total throughput wasted by SUsdue to transmission collision among them. ( H ) and X ( H ) are computed byequations (4.4) and (4.5), respectively. Moreover, according to subsection 4.2.2,the misidentication error on channel availability may make a SU select a busy

channel. Although SU can keep silence when it nds the PU signal, this raisesa wasting case that this SU may miss the transmission opportunity in anotheravailable channel. Let Y ( H ) denote the total throughput wasted by SUs due tomisidentication error. It is computed by equation (4.6).

Compared with SUs random channel selection, our goal is to develop a newspectrum decision strategy that can optimize SUs throughput ( H ) in the pres-ence of sensing errors, SUs competition and SUs transmission collision. This

39


61/122


( H ) = H

i= 1

N

n= 1

[ B( m,n(i)) Dm,n(i)T m,n(i)|m,n(i) = + ] (4.4)

X ( H ) = H

i= 1

N

n= 1

[ B( m,n(i)) Dm,n(i)( 1 T m,n(i)) |m,n(i) = + ] (4.5)

Y ( H ) = H

i= 1

N

n= 1

[ B( m,n(i))( 1 Dm,n(i)) Am(i)|m,n(i) = + ] (4.6)

strategy is based on the channel usage prediction, which is presented in thefollowing sections.

4.3 Channel Usage Prediction

4.3.1 Channel Usage State

To predict the channel usage, a new parameter is introduced, which is calledChannel Usage State (CUS). CUS is used to represent the channel utilizationobserved at SU side. It has three states, denoted by three symbols , S , and P,respectively. For a SU sn in channel cm within slot hi, the meanings of states ,S , and P are as follows.

State means a situation when no other active users (PU or SU) are iden-tied by SU sn . It includes two different cases:

SU sn senses channel cm to be free, but sn does not select cm to transmitdata.

SU sn accomplishes a successful transmission in channel cm within slot hi,i.e., Am(i) = Dm,n(i) = T m,n(i) = 1 .

State S means a situation when SU sn has detected the signal from otheractive users. It includes three different cases:

40


62/122

4.3. CHANNEL USAGE PREDICTION

SU sn senses channel cm to be free, but it is actually occupied by PU, i.e., Am(i) = Dm,n(i) = 0 .

Channel cm is actually free, but SU sns backoff time is not the smallestone in slot hi, i.e., Am(i) = 1 and Dm,n(i) = 0 .

SU sns transmission collides with other SUs, i.e., Am(i) = Dm,n(i) = 1 andT m,n(i) = 0 .

State P means a situation when SU sn senses channel cm to be busy. Thus,channel cm is not a candidate for channel selection by SU sn .

4.3.2 CUS History

Consider that the CUS observation on a channel is carried out by a SU in con-secutive slots {h1 , h2 ,..., h L}. The observation results constitute a CUS history,where L is called history length. Taking an example of L = 20 , means that SUobtains 20 observation results, as shown in the example in Table 4.1.

Table 4.1: CUS history within 20 identical time slots

Slot h1 h2 h3 h4 h5 h6 h7 h8 h9 h10

CUS P S P P

Slot h11 h12 h13 h14 h15 h16 h17 h18 h19 h20

CUS P S S P P

In the table, we are interested in two different types of statistics. The rstone is about the respective occurrence probabilities of three states in the 20slots. Since state P occurs 6 times and state S occurs 3 times, the occurrenceprobabilities of states S and P can be simply computed as 6 / 20 = 30% and3 / 20 = 15% , respectively. The second type of statistics is about the state changeof CUS within two consecutive slots. For instance, in slots h5 and h6 , the states

41


63/122


of CUS are S and P, respectively. The state change of CUS from slot h5 to sloth6 is denoted by S P . Clearly, the state change of S P is different from theone of P S , because their state changes have opposite orders in time domain.

By learning from CUS history, SUs are able to know in advance the futurechannel availabilities. However, it is difficult to precisely predict in which statethe CUS on a channel is expected to be. Thus, we instead focus on predictingthe probability of every CUS state occurring in the near future.

4.3.3 Predictor Selection

The Markov family of predictors is considered to conduct prediction. In [33,35, 36], the authors report on studies done on different Markov process basedpredictors like, e.g., Prediction by Partial Match (PPM), Lempel-Ziv-78 (LZ78),Probabilistic Suffix Tree (PST), Context Tree Weighting (CTW). For thesepredictors, the prediction is based on the historical information over a specictime period. This information is commonly called history while the numbers of recording information is called storage length .

In [35], the authors report on a comparison of the performances of the abovementioned predictors. Various metrics are considered, e.g., storage length, num-ber of alternative symbols, Markov class (variable or xed), computation effi-ciency and computation complexity. The study shows that LZ78 is robust whena few alternative symbols exist in the history and it is preferable for predic-tion computation under moderate storage length. Moreover, the authors of [33]report on the study of the advanced algorithm of LZ78 (called LeZi-update

scheme) for mobility purposes. By using it, the uncertainty associated withthe location of a mobile can be predicted with a high degree of accuracy givenits movement history. The movement history considers both the occurrence of visited areas and the transitions of different areas over time.

In our work, the CUS history covers only three symbols , S , P and the cor-responding building has a similar manner with the movement history presentedin [33]. Subsequently, we use the LeZi-update scheme to do CUS prediction.

42


64/122

4.4. PREDICTION IMPLEMENTATION

4.4 Prediction Implementation

The procedure of doing LeZi-update based prediction consists of two parts:parsing history and prediction computing. The rst part is to parse the CUShistory into different phrases and contexts . A phrase or a context is the sequenceof one or more symbols from the set { , S , P}. The phrases and the contexts areobtained by using encoding and decoding algorithms, respectively. The secondpart is to model a digital tree based upon the parsed contexts, and thus theprediction computing is performed.

4.4.1 Parsing CUS History

P SP P PSS PP .............................. 1

P SP P PSS PP .............................2

, P SP P PSS PP .........................

3

, P, SP P PSS PP .........................4

, P, , S P P PSS PP ..................... 5

, P, , S , P P PSS PP 6

, P, , S , P , P PSS PP 7

................

, P, , S , P , P , P SS PP 8

...............

, P, , S , P , P , P , SS PP 9 .......... , P, , S , P , P , P , SS , PP 10

.......... , P, , S , P , P , P , SS , PP ,

...............

Figure 4.4: Ten encoding steps.

Regarding CUS history P SP P PSS PP , the encoding algo-rithm is started by considering the rst symbol . Since there is no phrasebefore this consideration, is determined as the rst phrase. There are tworules after a phrase has been determined:

43


65/122


rule-1 : consider the next symbol that is following this phrase in the his-tory;

rule-2 : if this phrase has been determined before, consider the combina-tion of it and the next symbol.

According to the two rules, we show in Fig 4.4 the ten steps of determiningdifferent phrases. Consequently, P SP P PSS PP is encoded as , P, , S , P , P , P, SS , PP , .

The decoding algorithm is to acquire the different contexts that have oc-curred in the encoded phrases. Taking a phrase PP for example, we saythat its contexts are , P, P , PP and PP . By decoding all encodedphrases, we show all decoded contexts in Table 4.2.

Table 4.2: Decoded Contexts

P S P P P SS PP PP

4.4.2 Prediction Computing

First, we use the decoded contexts to model a digital tree . The root of this treeis called level-0 node with the meaning of null context . We let the notation denote the root. The symbols that are the rst symbol in the decoded contexts,are un-repeatedly modeled as the leaves of the root. The leaves of the root arecalled level-1 nodes. The symbols that are the second symbols in the decoded

contexts are un-repeatedly modeled as the leaves of every level-1 node. Theleaves of level-1 nodes are called level-2 nodes. It is clear that different level-1nodes may have different leaves. Similarly, we can obtain the leaves of level-2 nodes, which are called level-3 nodes. Due to the null property of , we consider as a leaf of any node in any level of the tree. We show the modeled digitaltree in Fig 4.5.

By observing the digital tree, we need to nd out all different tree branches,

44


66/122

4.4. PREDICTION IMPLEMENTATION

!"#$

!

!"%$&"'$

!"#$&"#$

& ! !

!

"##$

!"($ "($

"#$

")$

"#$

!"#"!$% !"#$%

!"#"!$' !"#$%

!"#"!$( !"#$%

&""'

Figure 4.5: The modeled digital tree.

which are arising from the the root, e.g., two different branches ( -) and(-) . Then, we count the occurrence time of every branch in all decodedcontexts. For example, in context , the branch( -) occurs three times,while the branch ( -) it occurs twice. Moreover, for a particular branch,we additionally mark the number of its occurrence time in the end node of thebranch. Since the branch PP occurs once in all decoded contexts, we markthe number 1 with the level-3 node P, as shown in Fig. 4.5.

By using the modeled digital tree, we do the prediction computing. Thecomputation is based on a blending strategy, which is used in the prediction by

partial match (PPM) scheme [37]. Further, the computation relies on the pre-determined order of Markov model. In [38], the authors studied various orderMarkov models by comparing prediction accuracy, coverage, model complexityand model size. Similar to [33], in our work we adopt the order-2 Markov modelfor prediction computing. The detailed computation process is as follows.

Regarding CUS history P SP P PSS PP , we consider asthe order-0 conditional event for prediction computing. We also consider the

45


67/122


Table 4.3: All paths occurring under conditional events

Channel a : three conditional events

under under under

(1) (3) P(1) (5) S (2) P(4)

(3) (1) PP (1) (3) P (1) P (1)

(5 ) PP (2) SS (1) (1)

PP (1) P (1) (1)

last symbol and two consecutive symbols of suffix as order-1 andorder-2 conditional events, respectively. Further, by observing the digital treein Fig. 4.5, we need to nd out all different paths occurring under each of threeconditional events.

Here, a path can be thought of as a special branch that is arising from eitherthe root or a leaf in the digital tree. The paths indicate the contexts that mayoccur in the near future. If a path occurs times at different places in the tree,we denote it as path ( ). We report all found paths and their occurrence timesin Table 4.3.

We then compute the occurrence probability of each found path by usingthe formula:

pr ( path ) = j2 z2

+ z2

j1 z1

+ z1

j0 z0

(4.7)

where

z0 , z1 , z2 : the frequencies of all possible paths (including ) occurringunder order-0, order-1 and order-2 events, respectively.

j0 , j1

spectrum decision support system thesis

Documents