IJDIWC

IJDIWC
International Journal of
ISSN 2225-658X (Online)
DIGITAL INFORMATION AND WIRELESS COMMUNICATIONS
Volume 3, Issue 1
2013
TABLE OF CONTENTS
Original Articles
PAPER TITLE
AUTHORS
A NEW METHOD TO CHANNEL ESTIMATION IN OFDM SYSTEMS Ali Asadi, Behzad Mozaffari
BASED ON WAVELET TRANSFORM
Tazehkand
PAGES
1
BICLUSTERING GENE EXPRESSION DATASET USING ENHANCED

Tamer Mohamed
BARYCENTER CROSSING MINIMIZATION
10
IMPROVING AND MEASURING

ALGORITHM IN RGB COLOR SPACE
19
COLOR
EDGE
DETECTION
Mohamed Almadhoun
COLLABORATIVE PROBLEM SOLVING USING PUBLIC SOCIAL

Melvin Ballera, Ismail Ateya
NETWORK MEDIA: ANALYZING STUDENT INTERACTION AND ITS
Lukandu, Abdalla Radwan
IMPACT TO LEARNING PROCESS
25
IMPACT OF A FORM OF ONLINE MATERIALS ON THE QUALITY OF Blanka Frydrychova Klimova,

EDUCATION A CASE STUDY
Petra Poulova
43
SEPARATION IN DATA MINING BASED ON FRACTAL NATURE OF

Marcel Jirina, Marcel Jirina Jr.
DATA
50
ADAPTIVITY OF A COLORING ALGORITHM TO UNRELIABLE

Ichrak Amdouni, Pascale Minet,
COMMUNICATIONS FOR DATA GATHERING IN WIRELESS SENSOR
Cedric Adjih
NETWORKS
Sabri Khssibi, Hanen Idoudi,
PRESENTATION AND ANALYSIS OF A NEW TECHNOLOGY FOR LOWAdrien Van Den Bossche,
POWER WIRELESS SENSOR NETWORK
Thierry Val, Leila Azzouz Saidane
Samira Chouikhi, Ines El Korbi,
A COMPARISON OF WIRELESS SENSOR NETWORKS COYacine Ghamri-Doudane, Leila
SIMULATION PLATFORMS FOR SMART GRID APPLICATIONS
Azouz Saidane
CMSA: A CENTRALIZED MAINTENANCE STRATEGY
ANTICIPATION FOR WIRELESS SENSOR NETWORKS
WITH Skander Azzaz, Leila Azouz

Saidane
67
81
93
103
ROMAN LAW AND CONTEMPORARY LEGISLATION

WIRELESS NETWORKS:
COUNTERMEASURES
DEVELOPMENTS,
THREATS
Wolfgang
Kahlig
Kahlig,
Eleonora
AND Mardiana Mohamad Noor, Wan

Haslina Hassan
112
125
International Journal of Digital Information and Wireless Communications (IJDIWC) 3(1): 1-9
The Society of Digital Information and Wireless Communications, 2013 (ISSN: 2225-658X)
A New Method to Channel Estimation in OFDM Systems Based on Wavelet

Transform
Ali Asadi1 and Behzad Mozaffari Tazehkand2
University of Tabriz
Faculty of Electrical and Computer Engineering University of Tabriz, Blvd. 29 Bahman, Tabriz, Iran
1
a_asadi89@ms.tabrizu.ac.ir, 2mozaffary@tabrizu.ac.ir
ABSTRACT
In OFDM systems, it is necessary to estimate the
channel to overcome the distortion caused by
channel fadings which can be induced by many
phenomena such as: delay spread, multipath effect,
mobility and Doppler shift. Most of the channel
estimation techniques are proposed in frequency
domain using the pilot symbols. One of them which
has less complicated is well-known as Least-Squares
(LS) method which is widely used in channel
estimation but it is more sensitive to noise respected
to the other reported techniques. In this paper, a new
threshold
based
method
using
wavelet
decomposition will be proposed which is based on
an initial LS estimation technique. The reported
simulation results show that the proposed method
has better performance compared to the other
methods such as Lee Method that has been published
recently.
KEYWORDS
OFDM signal, Channel Estimation, Channel State
(CS), Least Square (LS), Wavelet Transform, Cyclic
Prefix, Channel Impulse Response
1 INTRODUCTION
OFDM is a signaling technique that has been
applied widely in wireless communication
systems due to its ability to maintain effective
transmission and highly efficient bandwidth
utilization in the presence of various channel
impairments which one of them is frequencyselective fading. In OFDM systems the
available spectrum are divided into many
orthogonal
sub-channels,
which
are
instantaneously used to data transmission. Also,

in this technique the inter-symbol interference
(ISI) which is induced due to frequencyselective channels can be reduced by adding the
cyclic prefix (CP) [1].
In OFDM systems, channel estimation is
necessary to obtain the channel state
information (CSI), reducing the bit error rate
and also to achieve a distortion less output data.
There are various methods to channel estimation
such as: with or without a need to parametric
models, blind or pilot based methods, frequency
and/or time domain analysis, adaptive or non
adaptive techniques. Among these mentioned
methods, channel estimation in OFDM systems
is often done in frequency domain using pilot
symbols or training data [2]. The least square
and minimum mean-square error (MMSE) are
conventional
linear
channel
estimation
techniques which are based on pilot
arrangement. The LS method is less
complicated and simple respect to other
methods and consequently is used to channel
estimation, but it has a serious drawback which
is more sensitive to channel noise. MMSE
estimator has better performance than LS
method but suffers from a high computational
complexity because it requires knowledge of the
channel statistics and the signal-to-noise ratio
(SNR) [3]. Some different methods have been
developed to reduce the complexity and
improve the performance of the MMSE
estimation such as modified MMSE and
singular value decomposition (SVD) [4-5]. 1n
2006 Noh et al. proposed a method to
decomposing the covariance matrix to the
1
presented in section (3). The proposed method

will be described using wavelet transform in
section (4). In section (5), the simulation and
obtained results will be considered and finally in
section (6) conclusions will be qualified.
2 SYSTEM DESCRIPTIONS
2.1 OFDM Systems
OFDM technique converts a frequency
selective channel into a number of frequency
nonselective channels by dividing the available
spectrum into a number of overlapping and
orthogonal narrowband sub channels where
each of them sends own data using a subcarrier.
A block diagram of OFDM systems is shown in
figure (1).
Mapper
X(k)
S
/
P
Adding
Pilots
x(n)
x(n)
P
/
S
Adding
Cyclic
Prefix
IFFT
Channel
w(n)
+
X(k)
Demapper
simple and low order sub matrix so that they can

decrease the complexity of MMSE method [6].
Hsieh used a comb type pilot arrangement and
second order interpolation method to channel
estimation [7]. Coleri et al. compared the results
of many interpolation techniques to channel
estimation with Rayleigh fading such as linear,
second order, cubic, low pass filtering and
spline interpolation methods [8]. Edfors et al.
modified the MMSE and LS methods with
assumption that the channel model to be an FIR
in which the impulse response duration cant
greater than the Guard Interval of an OFDM
symbol [9]. Dowler assumed that if the
maximum delay caused by channel to be a
known parameter so estimation based on DFT
method can obtain better results [10]. Minn et
al. improved the results of Dowler method by
considering a sparse channel model [11]. In [12]
Kang et al. proposed a DFT based channel
estimation. In their method the effect of channel
noise in outside of maximum channel delay are
replaced with zero and finally a good estimation
were resulted. In 2009 Lee et al. obtained an
optimal threshold value based on wavelet
decomposition and therefore they could improve
the channel estimation [13].
In this paper, we propose a time-domain
approach to channel estimation using wavelet
decomposition. In this approach, initial channel
estimation is calculated by the LS estimator, and
then channel coefficients in time-domain are
obtained using IFFT transform. It is assumed
that the maximum delay caused by channel at
most is equal the length of cyclic prefix. In the
next step wavelet transform is applied to the
obtained coefficients in order to calculate a
threshold value [14]. Finally, the estimated
coefficients are de-noised and the good estimate
is obtained using the calculated threshold.
The outline of the remaining of this paper is
organized as follows. In section (2), the OFDM
system based on pilot channel estimation will be
considered. The wavelet transform will be
P
/
S
Channel &
Data
Estimation
x(n)
FFT
x(n)
Remove
Cyclic
Prefix
S
/
P
Figure 1. The OFDM System in baseband model
At first, in transmitter the binary inputs are

grouped to get an M-ary symbol. According to a
predefined baseband modulation such as QPSK
and MQAM, the obtained symbols are
modulated using a signal mapper subsystem. In
the next step, an S/P sub-block converts the
serial input symbols to a block data which can
be considered as a vector X=[X0, X1 , XN-1].
The vector size is N which determine the
number of subcarriers in OFDM signal. Any
subcarriers will be modulated by the obtained

symbols in data vector using IFFT technique
and consequently, the time domain of the
OFDM signal are calculated which can be
written as equation (1).
x(n)
1
N
LN 1
k 0
X Ke
2
kn
LN
0 n LN 1
(1)
Where L is an oversampled factor which can

be set to any number as: 2, 4, 8, 16 [15].
To prevent the effect of ISI in OFDM signals,
a guard time which well known as cyclic prefix,
must be add to the symbol. The equation (2) and
figure (2) shows the adding process.
x N n
x n
x n
n N c , N c 1, , 1
n 0,1, , N 1
(2)
Where Nc denotes the cyclic prefix length.

Nc
CP
OFDM Symbol
Figure 2. Cyclic Prefix adding process in OFDM symbol
Finally, the obtained OFDM signal is

converted to serial form and is transmitted to the
receiver side through a frequency selective
channel which is often considered as a Rayleigh
fading model with additive white Gaussian
noise (AWGN). The received signal in the
output of the channel can be modeled as
equation (3).
(3)
y n x n h n w n 0 n N '
Where w(n) is the channel noise which is
assumed to be AWGN, zero mean and its
variance equals 2 and h(n) is the impulse
response of the channel and N=N+Nc-1. It is
assumed that the channel is linear and time
invariant in period of an OFDM symbol. After
removing the cyclic prefix the above equation
can be rewritten as follow:
(4)
y n x n h n w n
0 n N 1
In the frequency domain we can obtain the

following equation.
Y k X k H k W k k 0, , N 1 (5)
Where H(k) and W(k) are the Fourier transform
of the h(n) and w(n), respectively. In this paper
it is assumed that there is no synchronization
error.
2.2 Channel Estimation
In any communication systems, channel
estimation is a most important and challenging
problem, especially in wireless communication
systems. Usually, the transmitted signal can be
degraded by many detrimental effects such as
mobility of transmitter or receiver, scattering
due to environmental objects, multipath and so
on. These effects cause the signal to be spread in
any transformed domains as time, frequency and
space. To reducing these effects anyone must
estimate the channel impulse response (CIR).
Channel estimation has a long history in single
carrier communication systems. In these
systems, CIR is modeled as an unknown FIR
filter whose coefficients are time varying and
need to be estimated [16]. There are many
channel estimation methods that can be used in
multicarrier communication systems but the
especial properties of multicarrier transmission
systems give an additional perspective which
forces to developing new techniques to channel
estimation in wireless communication systems.
In general, channel estimation methods based
on OFDM systems can be categorized into two
groups as blind and non-blind techniques. In the
former, all of the techniques use the statistical
behavior of the received signals and therefore,
to obtain the accurate CIR a large amount data
is required [17]. Finally, the complexity of
computations is very high. In the later, to obtain
a good estimation of channel, the transmitter
sends a collection of data aided as pilots whose
are previously known by the receiver. Often,
most OFDM based systems as IEEE 802.11a
3
and hyperLAN2 use pilots in frequency domain

in order to sampling the faded channel in
frequency domain.
Channel estimation based on pilot arrangement
which have been used in many application
systems especially wireless communication and
power line communication channels can be
divided in two main categories as block type
and comb type [8]. These arrangements are
shown in figures (3) and (4). In these figures
Tc and Bc parameters denote the time and
bandwidth coherency of the channel.
OFDM Symbol
pilot ,
Xm
data ,
Frequency
Tc
transmitting OFDM symbols contains data and

pilot subcarriers where the pilot spaces must be
equal to the frequency coherency which is
related to the delay spread caused by multipath
effects. In the comb type arrangement when the
density of pilots increase then the result of
channel estimation will be improve. There are
many techniques to channel estimation in both
arrangement types as LS, MMSE, DFT based,
modified DFT based, Decision Direct and
wavelet based methods.
In the comb-type arrangement, pilot symbols
are inserted and continuously transmitted over
specific pilot sub-channels in all OFDM
symbols according to the following equation:
m kP
otherwise
(6)
Where P denotes the pilot repetition rate or

pilot spaces in OFDM symbol which can be
calculate as P=N/Nc. Also Nc indicates the
pilot number.
Time
Figure 3. Block type arrangement of OFDM symbols
Frequency
OFDM Symbol
Bc
Time
Figure 4. Comb type arrangement of OFDM symbols
According to the type of channel fading, slow

or fast fading, anyone can uses the block or
comb type arrangements, respectively, to
channel estimation. In the block type, an OFDM
symbol which contains pilots in all subcarriers,
are transmitting periodically which equals the
time coherency of the channel which is related
to the Doppler effects and in other type the
2.3 Least Square Method

This method can be applied in both block and
comb type. In later arrangement, in frequency
domain, at first the channel output at pilot
locations is extracted. In the next step channel
estimation can be calculated using the extracted
subcarriers which are known to the receiver.
The corresponding equation can be written as
the following equation.
Y (k p )
H L S ( k p )
X (k p )
H ( k p ) W ( k p ), k p 1, 2,..., N
(7)
p
Where W(kp)=W(kp)/X(kp) is the noise

component at the estimated channel coefficients
in frequency domain and kp denotes a
subcarrier index at pth pilot. Then to obtaining
the channel estimation at the data subcarriers, an
interpolation technique is required. There are
some interpolation techniques in [8] but linear
interpolation is the simple one which can be
written as equation (8).
4
(k k p )
(k k p )
H ( k ) [1
]H (k p )
H (k p L ) .
L
L
(8)
Where L denotes distance between two

adjacent pilot subcarriers.
3 WAVELET DECOMPOSITIONS
Wavelet transforms provide a framework
where a signal is decomposed, with each level
corresponding to a coarser resolution, or lower
frequency band. There are two main groups of
transforms, continuous and discrete. Continuous
wavelet transform of the signal x(t) can be
expressed by (9):
1
tb
(9)
x(a, b)
x ( t ) * (
) dt a 0, b .
a
x (t )
j ,k
ck (t k )
(11)
(t ) h1 ( k ) 2 (2t k )
k
Where h0(k) and h1(k) are impulse response of

low-pass and high-pass filters respectively.
Based on these filters, each ideal signal can be
expanded as coefficients wavelet. Figure (5)
shows related decompositions in two levels.
HP
X0
X
LP
k j 0
j ,k
j ,k
(t )
(10)
( t ) 2 j / 2 (2 j t k ) .
Where k and j are translation and scale

parameters, Ck is scaling coefficient which is
used for representation of the low-resolution
approximation of the signal and dj,k is wavelet
coefficient which is used for giving detailed
information of the signal. Approximated and
detailed coefficients can be obtained based on
low-pass and high-pass filter respectively. These
filter coefficients are calculable according to
equation (11):
HP
LP
Where a is called the scaling factor and

also b denotes the translation factor and (t) is
wavelet function [18]. Although the continuous
wavelet transform is simple to describe
mathematically, both the signal and the wavelet
function must have closed forms, which make it
difficult or impractical to apply. The discrete
wavelet is used instead so that if we choose the
scale a and position b based on powers of
two, the analysis will be much more efficient
and accurate [19]. Any finite energy analog
signal x(t) can be decomposed into a coarse
approximation represented by scaling functions
(t) and details are represented by wavelet
functions (t) as follows:
(t ) h0 ( k ) 2 (2t k )
X1
X2
Figure 5. Wavelet decomposition in 2 levels
4 PROPOSED METHOD
Since LS estimator is much sensitive to noise,
we introduce a time-domain method to channel
estimation which its objective is improvement
the performance of LS algorithm. In this paper it
is assumed that the pilot spaces are related to the
maximum delay caused by the channel. The
impulse response of the channel can be
calculated using IFFT based on initial LS
estimated channel as following equation.
hLS ( n ) IDFT { H LS ( k )}
h1 ( n ) h ( n ) w '( n ) 0 n L 1
L n L 1
h2 ( n ) w '( n )
0
L n N 1
(12)
Where L can be controlled by the pilot spaces

and L can be determined by maximum channel
delay. The estimated channel impulse response
obtained by LS has most of its energy
concentrated on a few first samples because in
practice, the channel length is shorter than the
IFFT size which can be determined by the cyclic
prefix length. In this paper we will consider
threshold based channel estimation in wavelet
domain. Calculating of the threshold value is
very important to noise reduction in channel
5
estimation. According to [14], the calculated

threshold value using wavelet decomposition
can be obtained better results than other
threshold based techniques.
Since the noise-dominant components of the
estimated channel coefficients are in highfrequency part of wavelet decomposition, the
detail coefficients can be used to calculate the
threshold value [14] as the following equation.

median ( Di )
(13)
0.6745
Where Di denotes the detail coefficients.

By applying the calculated threshold value into
channel coefficients according to (14), the noise
in the channel will be removed effectively.
hLS ( n )
hLS ( n ),
hp (n)
0
(14)
hLS ( n )
Finally by taking the DFT of the obtained

channel impulse response, the estimated channel
frequency response will be resulted. The
proposed algorithm in more details is shown in
figure (6).
Y(k)
Di
Pilot
Extraction
H LS ( k )
IDFT
Wavelet
Decomposition
hLS ( n )
h2 (n)
h1 ( n )
Threshold
Computation
Denoising
Algorithm
h ( n )
Figure 6. Block diagram of the proposed method
5 SIMULATION RESULTS
In order to evaluate the proposed method it is
necessary that a simulation must be taken. The
simulation parameters are listed in table (1). In
all simulations the block coding and interleavers
are not used and also in order to better

evaluation of the proposed method the channel
equalization are not considered.
Table 1. Applied parameters to evaluate the proposed
method
Parameters
OFDM symbol size
CP size
Baseband modulation
Wavelet
Subcarriers Number
Channel length
Pilot spacing
Channel Model
Related Values
N=2048
Nc=512
QPSK , 16QAM
Db4
N=2048
L=10
15, 30
Rayleigh with AWGN
In the following we want to compare the

channel estimation results which are obtained
using the LS, Lee and proposed methods. The
channel model is considered as an FIR filter.
The corresponding filter taps is considered as a
complex random variable which has Rayleigh
distribution in order to model the Rayleigh
channel fading. Figures (7) and (8) show bit
error rate (BER) of estimation which is obtained
by LS, Lee and the proposed methods based on
16-QAM and QPSK modulations, respectively.
It is noted that the obtained result in the
proposed method is better than Lee and LS
methods in overall SNR conditions. It is shown
that the proposed method has improvement of
about 1-1.5dB compared to the Lee method.
Also mean square error (MSE) parameter for
more evaluation of proposed method compared
with the others is plotted in (9) and (10). In
these figures the pilot spacing is considered
equal to 30. Figures (11) and (12) are the BER
and MSE curves of the LS method, the Lee
method and the proposed method in case of 16QAM with pilot spacing 15 respectively. These
figures show that by reducing the number of
pilots, Lee method yields poor results compared
to the proposed method.
Figure 7. BER Performance comparison with 16QAM
Figure 8. BER performance comparison in case of QPSK
Figure 9. MSE performance comparison with 16QAM
Figure 10. MSE performance comparison with QPSK
Figure 11. BER performance comparison with 16QAM
Figure 12. MSE performance comparison with 16QAM
In the following we want to compare the

channel estimation results, in case of frequency
impulse response, which are obtained using the
LS, Lee and proposed methods. Again, the
channel model is considered as an FIR filter.
The filter taps is considered as a complex
random variable which has Rayleigh
distribution in order to model the Rayleigh
channel fading. Also the channel noise is
considered as AWGN with SNR=10dB. The
filter has 10 taps.
on wavelet decomposition is used in order to

reduce the noise effect. It is assumed that the
channel length is related to the maximum delay
spread caused by multipath channel. This
method is less sensitive to changes in
modulation parameters and pilot spacing
compared to Lee method that leads to
improvement in channel estimation. It is shown
that the proposed method has improvement of
about 1-1.5dB compared to the Lee method.
The obtained results are shown in the figure

(13), also in all figures for better comparisons
the frequency response of the ideal channel is
plotted. As shown in the figures the obtained
results in the proposed method are better rather
than other methods.
ACKNOWLEDGMENT
The authors would like thanks IRTC in order to
supporting on this project
7 REFERENCES
1.
Amplitude (dB)
-20
-30
-40
-50
0
Ideal
LS Method
50
100
150
Subcarrier Index
200
250
Amplitude (dB)
-20
-25
-30
Ideal
Proposed Method
-35
-40
0
50
100
150
Subcarrier Index
200
250
Amplitude (dB)
-20
-25
-30
Ideal
Lee Method
-35
-40
0
50
100
150
Subcarrier Index
200
250
Figure 13. The Estimated Impulse Response of The

Channel using the LS, Lee and Proposed Method,
SNR=10
6 CONCLUSIONS
In this paper, a new method to estimating the
faded channel based on wavelet decomposition
is presented. In this method, a threshold based
B. Muquet, Z. Wang, et al. Cyclic prefixing or zero padding

for wireless multicarrier transmissions? IEEE Tran.
Commun., 50(12): pp. 2136 2148, 2002.
2. YG Li, Pilot-symbol-aided channel estimation for OFDM
in wireless systems. IEEE Trans. Vehic. Technol. 49(4):
pp. 1207-1215(2000). doi:10.1109/25.875230
3. M. Sandell S. K. Wilson J. J. van de Beek, O. Edfors and P.
O. Borjesson. On channel estimation in OFDM systems. In
Proc. IEEE Vehicular Technology Conf., volume 2,
pp.815819, Chicago, IL., July 1995.
4. O. Edfors, M. Sandell, et al. OFDM channel estimation by
singular value decomposition. IEEE Tran. Commun., 46:
pp. 931 939, 1998.
5. M. R. Raghavendra and K. Giridhar Improving
channel estimation in OFDM systems for sparse
multipath channels, IEEE Trans on Signal Processing
Letters, vol.12,no.1, pp.52 55, 2005.
6. M. Noh, Y. Lee, and H. Park, A low Complexity LMMSE
Channel Estimation for OFDM, IEE Proc. Commun., vol.
153, no. 5, pp. 645-650, Oct. 2006.
7. M.Hsieh and C. Wei, Channel Estimation for OFDM
Systems based on Comb-Type Pilot Arrangement in
Frequency Selective Fading Channels, IEEE Trans.
Consumer Electron., vol. 44, no. 1, Feb. 1998.
8. S. Coleri, M. Ergen, A. Puri, and A. Bahai, Channel
Estimation Techniques Based on Pilot Arrangement in
OFDM Systems, IEEE Trans. On Broadcasting, vol. 48,
no. 3, pp. 223-229, sep. 2002.
9. O. Edfors, M. Sandell, J.J. van de Beek, S.K. Wilson and
P.O. Borjesson, Analysis of DFT based Channel
Estimators for OFDM," Wirel. Pres. Commun., vol. 12, no.
1, pp. 55-70, 2000.
10. A. Dowler, A. Doufexi and A. Nix, Performance
Evaluation of Channel Estimation Techniques for a Mobile
fourth Generation Wide Area OFDM System, IEEE
Vehicular Technology Conf., vol. 4, pp. 2036-2040, Sep.
2002.
11. H. Minn, and V.K. Bharagava, An Investigation into Time

domain Approach for OFDM Channel Esimation, IEEE
Trans. Bradcast., vol. 64, no. 4, pp. 240-248, Dec. 2000.
12. Y. Kang, K. Kim, and H. Park, Efficient DFT-based
channel estimation for OFDM systems on multipath
channels, IET Commun., vol. 1, No. 2, pp. 197202, 2007.
13. Y.-S. Lee et al. Channel estimation based on a time-domain
threshold for OFDM systems. IEEE Tran. Broadcast.,
55(3): pp. 656 662, 2009.
14. D. L. Donoho and I. M. Johnstone, Ideal spatial adaptation
by wavelet shrinkage, Biometrika, vol. 81, no. 3, pp. 425
455, 1994.
15. J. Tellado, Peak to Average Power Ratio Reduction for
Multicarrier Modulation, PhD thesis, University of
Stanford, Stanford, 1999.
16. H. Arslan and G. E. Bottomley, Channel Estimation in
Narrowband Wireless Communication Systems, Wireless
Commun. and Mobile Comp., vol. 1, no. 2, pp. 20119,
Apr. 2001.
17. H. Bolcskei, R. W. Heath Jr. and A. Paulraj, Blind
channel identification and equalization in OFDM-based
multiantenna systems, IEEE Trans. Signal Processing, vol.
50, No. 1, pp. 96-109, Jan. 2002.
18. Stphane Mallat, A Wavelet Tour of Signal Processing,
Academic Press, 1999.
19. C. S. Burrus, R. A. Gopinath, H. Guo, Introduction to
Wavelets and Wavelet Transforms, a primer, Prentice Hall
New jersey, 1998.
BICLUSTERING GENE EXPRESSION DATASET USING ENHANCED

BARYCENTER CROSSING MINIMIZATION
Tamer Mohamed
Helwan University
Faculty of Computers and Information
Biomedical Informatics Department
Sniper_fci_cs@yahoo.com
ABSTRACT
1. INTRODUCTION
There are two main categories of bi-clustering

approaches graph-based bi-clustering, and nongraph based bi-clustering. In graph-based biclustering algorithm the input dataset is converted
into bigraph such as bipartite graph, and apply
some heuristic local searching techniques to
minimize the number of crossings between edges in
the Bigraph such as BaryCenter (BC) used in
SPHier algorithm.
DNA is the secret of life. It is the blue print for

making new cells or even new organism. It
contains the genetic information that move from
generation to generation in what we called
inheritance.
The main problem of graph-based bi-clustering

algorithm is to extract the best bi-clusters, and this
leads to the ordering of gene expression dataset
before we apply biclustering algorithm. In bipartite
graph this is achieved through minimizing the
number of crossings in bipartite graph. As we
minimize the number of crossings in the bipartite
graph, the gene expression dataset becomes more
ordered, and this enhances the results of
biclustering algorithm.
The main goal of our proposed algorithm is the
enhancement of graph-based biclustering algorithm
by enhancing BaryCenter crossing minimization
heuristics of bipartite graph.
In the proposed algorithm we add the rank of each
node to the rank of its neighbors, and using the
position of each node in the calculations to give a
new rank to each node, and using this rank for
reordering the nodes of each layer.
KEYWORDS
BaryCenter Heuristics, Biclustering, Gene Expression,
Microarray.
DNA contains the genes that carry the phenotypes

that determine how the organism looks a like or the
physical properties of the organism, the hair is
black, and the eye is blue, and so on. Every gene
can carry phenotype such as the color of eye. The
phenotype property can be determined by two, or
more genes not just one gene [1]
In humans, genes are arranged into twenty three
pairs of chromosomes, or forty six chromosomes.
Twenty three chromosomes come from the father,
and the other twenty three come from the mother,
so the child inherits its properties from the two
parents. Everybody have two copies from each
gene one from father and one from mother [1].
DNA takes a form of helix of double strands. Each
strand connects to the other with a hydrogen bond.
Each strand is a long chain of nucleotides; every
nucleotide is made up of three components: base,
deoxyribose sugar, and phosphate.
These
nucleotides come up in pairs to make the DNA.
DNA consists of thousands of just four bases,
namely Adenine (A), Guanine (G), Thymine (T),
and Cytosine (C). Each nucleotide connects to
another one with phosphodiester bond. Phosphate
and sugar work as the backbone of the DNA like
stair and the bases like the steps of the stair.
10
RNA is more like DNA, it have some different
changes than DNA. RNA is a chain of nucleotides,
these nucleotides composed of ribose sugar instead
of deoxyribose in DNA, phosphate, and base. RNA
contains thousands of just four bases but with
Uracil in place of Thymine.
Proteins are the basic material that makes anybody.
It is responsible for all functions that occur in the
body, body structure, and all the physical
characteristic of the body. Proteins are the most
important molecule in the body, because any
mutations happened in the proteins may lead to the
protein losing its functionality, and hence its role in
the body, and hence diseases begin. Proteins are
long chains of Amino Acids (AA). There are only
twenty amino acids that made up all the proteins of
the body
Gene expression is the process of converting genes
into proteins in two basic steps: DNA is converted
into mRNA in a process called transcription, and
then mRNA converted to proteins in a process
called translation
DNA microarray is a technology that it is used to
measures the amount of gene expression in the cell.
It is used in the study of behavior of genes under
certain conditions, or at different time frames. This
technology enables scientists to study what happens
to genes in different disease conditions [2].
The result of microarray is a gene expression
dataset which is a tow dimensional array - data
table - where every row represents one gene and
every column represents one condition or sample.
Every cell in this table represent the expression
level Lij of gene i in condition j, as in table 1.
There is many Data mining techniques used for
analyzing gene expression datasets in order to
extract information and finding relations between
these data such as clustering and classification.
Clustering is used to assign data to groups, or
classes like classification, but the difference is that
in classification the classes are predefined, but in
clustering the classes are not determined before
applying the algorithm. Clustering algorithms can
be applied on gene expression datasets in order to
group genes that have the same expression together
and this is helping in the study of diseases such as
cancer. There are three ways of clustering:
1. Gene-based clustering in this approach genes

are treated as objects and samples are treated
as features.
2. Sample-based clustering is the opposite of
gene-based clustering, samples are treated as
objects and genes are treated as features.
3. Subspace clustering this technique combine
the previous two approaches the genes and
samples can be treated as objects and features
so that gene may be object or features and
samples are treated the same.
Biclustering or subspace clustering is the most
efficient technique in clustering because clustering
of genes depending on conditions or vice versa, and
this is not what happens exactly in the real life , a
subset of genes may have similar behavior on
subset of conditions, and here is the Biclustering
come to solve this problem.
Gene expression datasets can be represented in
bipartite graph where:
1. Genes are represented by the top layer.
2. Conditions are represented by the bottom
layer.
3. Edges that connect the genes in the top
layer with its corresponding conditions in
the bottom layer.
The edge weigh wij is the expression value of gene i
in condition j, where wij = 0 if there is no edge
between node i and j, and this means that the gene i
does not express in condition j.
We can represent data in table 1 in a bipartite
graph, as in figure 1,where top layer represents top
conditions and bottom layer represents genes, then
this Bipartite graph can be clustered directly, but
the results will be less accurate, and in order to
overcome this problem there is one more step
before biclustering, which is crossing minimization.
Crossing minimization aims to minimizing the
number of crossings between lines that connect the
two layers, the graph is reordered, and related nodes
11
grouped together, as in figure 2, and this enhances
the results of biclustering process.
Table 1. Gene Expression Data Matrix
Cond 1
Cond 2
Cond 3
Cond 4
Gene 1
Gene 2
Gene 3
Gene 4
Gene 5
2.1 CHENG AND CHURCH

Cheng and Church [6] define bicluster as a
submatrix for which the mean square residue ( )
of each bicluster is below than a predefined
threshold. The algorithm is run in two main phases:
first, looping through the gene expression matrix
removing rows and columns until ( ) is less than
the threshold; second, looping through the deleted
rows and columns and adding to bicluster as long as
( )is less than threshold. The ( )of an
element
in a submatrix
(
| || |
( )
Where
Sub Row Average
| |
( )
Sub Column Average

Figure 1. Bipartite graph representation of table 1
| |
( )
Sub Matrix Average

| || |
( )
2.2 SPHIER
Figure 2. Bipartite Graph after crossing minimization
2. RELATED WORK
Biclustering algorithms can be classified in two
main categories: graphical and non-graphical
biclustering. We focus our discussions on graphical
biclustering which is the main interest point of our
research for general discussion on biclustering
algorithms can be found in [3, 4]. There is another
good reference that focuses on Barycenter crossing
minimization which the main goal of our work is to
enhance Barycenter algorithm [5].
SPHier [7] extract clusters from dataset using

bipartite crossing minimization biclustering
techniques but instead of serial Bigraph crossing
minimization using parallel algorithm in order to
increase the performance of the algorithm beside
the enhancement made on Cheng and Church
biclustering algorithm to enable the local search for
clusters instead of global search because after the
Bigraph reordering the related values arranged
together in blocks and global search is useless.
2.3 TABU SEARCH
Ibrahim [8] using a Tabu search, which is a
metaheuristics technique depending on guide local
12
search technique to find global optimum solution.
He adapts the Tabu search in Marti [9] to solve the
crossing minimization problem and reach to better
order of dataset. Tabu search run in two main steps:
first, construction of initial solution and second,
iterative improvement of this solution until reaching
the optimum solution. Iterative improvement has
two steps: Intensification and Diversification and
each step of them has three steps: normal,
influential, and opposite and the algorithm moves
between these three steps as it insert or remove
nodes depending on their Barycenter and if this
move reduces the number of crossings in the graph
and if there is no enhancement it stops after a
specified number of iteration this process continue
running in this way until reaching the optimum
solution.
2.4 BIMAX
BiMax Algorithm was proposed by Prelic et al [10];
the main idea behind this algorithm is to
transforming the gene expression dataset inputs into
binary dataset (or 0s and 1s) by discretization
process, where 0s means no change in expression
level and 1,s from change in expression level.
is to make a fast biclustering method that would

produce statistically significant results.
It defines a bicluster as a subset of genes that jointly
respond across a subset of conditions. The gene is
considered as normal if its expression level
responding significantly in these conditions.
SMABA algorithm represents the gene expression
dataset as bipartite graph where genes represent one
layer of nodes, and conditions represents the other
layer. The nodes in tow layers are connected with
edges. The edges represent the weight (or
expression) level of each gene under each
condition. The weight of each subgraph (or
bicluster) is the sum of the weights of genecondition pairs in it, including edges and nonedges.
Hash table is created for the bipartite graph using
the weights of bipartite graph and a probabilistic
function. This hash table is used to find heavy
subgraph around each vertex using local search
depending on scoring function, and improves this
cluster until no score improvements are available.
2.5 DISCUSSION
A bicluster corresponds to a subset of genes that

jointly respond across subset of samples. The
algorithm uses divide-and-conquer mechanism. It
add row by row to the cluster, when it adds a row
the columns set is partitioned into CU - columns in
which the added row has ones, and CV - columns in
which the added row has ones.
Cheng and Church was one of the leading

algorithms of biclustering algorithms, but this
algorithm using global search heuristics to find
biclusters, and this technique consumes more time
in running the algorithm beside the effect of global
search that it gives results that it is not so accurate
and efficient.
The row set is split into GU- the rows that have only
ones in CU, GV - the rows that have ones in CV only,
and GW - the rows that have ones in both. If U and
V do not share any rows and columns of the matrix
E, i.e. GW is empty; the two matrices can be
processed independently from each other. If U and
V have a set GW of rows in common as shown in
Figure 3.6, special care is necessary to only
generate those biclusters in V, that share at least
one common column with CV.
SPHeir algorithm uses the same technique behind

Cheng and Church but using the bipartite graph
along with crossing minimization techniques for
solving the problem besides using local search
heuristics and that overcomes the problem of global
search heuristics, hence it gives more efficient and
accurate results.
2.1 SAMBA
SAMBA [11] stands for statistical algorithmic for
bicluster analysis. The main goal of this algorithm
Tabu search algorithm used the same idea as

SPHeir algorithm, as it used bipartite graph,
crossing minimization, and local search heuristics,
but the difference in this algorithm is the local
search heuristics for finding the biclusters. This
algorithm achieves better, more accurate and
efficient results, but this algorithm takes more time
for solving the problem, but as the processing
13
power of computers increases rapidly, so we think
the time is not factor as efficiency, because at the
end the target of biclustering is to get accurate
results.
SAMBA algorithm is one of famous algorithms that
achieves good statistical results, runs in a good
time, it also used the methodology of bipartite
graph, and used the weight of edges and non-edges
along with statically functions to give weight for
each subgraph, so it differs from the previous the
other algorithms that we discussed, and which using
bipartite graph in that it used the weight of edges
and non-edges in calculations, but other algorithm
depend on edges only in the calculations, and also
depend on local search heuristics to find biclusters.
BiMax algorithm looks like Cheng and Church in
that the two algorithms do not use graph
representation in solving biclustering problem,
BiMax algorithm depended on transforming the
dataset set into binary dataset in a process called
discretization, and then work on this dataset by
divide-and-conquer methodology by partitioning
the dataset to rows and columns, and add them to
clusters depending on the property they share, the
idea of discretization used in other algorithms such
as Tabu search algorithm to minimize the time
taken in crossing minimization process, this
technique decreases the running time.
We can conclude from this discussion that the most
significant algorithms of the algorithms we
discussed is Tabu search, and SMABA algorithm,
because they achieve the best results, and used
different valuable processing techniques.
3. PROPOSED ALGORITHM
There are many techniques used for Biclustering
one of these techniques is constructing a bipartite
graph from gene expression dataset, reordering
nodes in this graph using crossing minimization
technique then apply biclustering algorithm on the
reordered bipartite graph. Most of the algorithms
used to crossing minimization in bipartite graph
divided into three basic steps: the first step is the
construction of bipartite graph, the second step is
the initial ordering, and the last step is iterative
improvement.
we calculate the rank of each node in layer 0 using

Barycenter heuristics, as in equation 5, then
reordering these nodes depending on this rank
while keeping the layer 1 static then reordering
layer 0 by calculating the rank of each node in layer
1 using equation 5, and reordering nodes while
keeping the layer 0 static, and we will continue
iteratively in this manner until no more changes
occur in the order of nodes in the two layers.
(5)
Where:
Let vi represent the ith node in the non-static layer,
and set Ni represents the set of neighbors of vi, also
let rj represent the rank of jth member of the set Ni
[12].
We success to solve the problem of crossing
minimization in a way that achieves better results in
crossings number rather than Barycenter heuristics
[13]. We use the same way of Barycenter heuristics
used in [13] to minimize crossings in bipartite
graph, but we make some changes to the
calculations of weight of each node and using this
weight as a new rank for reordering nodes. We
calculate the Barycenter of each node as in
algorithm 2, then calculating the new weight by
adding the rank of each node to the sum of the
ranks of its neighbors as in algorithm 3, and taking
the position of each node into consideration of our
calculations as in algorithm 4, then using this rank
for reordering the nodes.
The algorithm of biclustering has three main basic
steps:
-
Construct bipartite graph

Bipartite graph crossing minimization as in
algorithm 1
Applying biclustering algorithm on bipartite
graph as in algorithm 5
14
Algorithm 1: Bipartite Crossing Minimization
1. Construct Bipartite graph BG
2. Compute the Barycenter for each
node in the two layers, as in
algorithm 2
3. Compute the new weight for each
node in the two layers, as in
algorithm 3
4. Reorder nodes in the two layers
5. Compute crossing minimization for
bipartite graph
6. Repeat steps from 2 to 5 until
the number of crossing of the
current bipartite graph is larger
than the number of crossings of
previous bipartite graph from
last iteration
minimization process groups related rows and

related columns together so the algorithm does not
need global search to find the biclusters.
The algorithm uses mean square residue score MSR - defined in equation 1 where MSR of each
biclusters is less than a predefined threshold
Algorithm 5: Biclustering Algorithm
Comment: Identify Biclusters in
Matrix representation of Bigraph BG
Reordered
Comment: Mean Squared Residue Score of Each

Bicluster is less than
Input:
Bipartite
minimization
graph
after
crossing
Output : Biclusters
startRow = 0;
startCol = 0;
Index = 0;
FOR i = 0 TO numCols DO
colSum[i] = 0;
blockSum[i] = 0;
Residue[i] = 0;
FOR r = startRow TO numRows DO
rowSum = 0;
FOR c = startCol TO numCols DO
colSum[c]=colSum[c]+MatrixBG[r][c];
rowSum = rowSum + MatrixBG[r][c];
blockSum[c]=blockSum[c-1]+colSum[c];
colCount = c - startCol + 1;
rowCount = r - startRow + 1;
blockCount = rowCount * colCount;
rowMean = rowSum / colCount;
colMean = colSum[c] / rowCount;
blockMean = blockSum[c]/blockCount;
Res = Matrix G[r][c] - rowMean colMean + blockMean
Residue[c] = Residue[c-1] + Res*Res;
msr = Residue[c]/ blockCount;
IF msr > THEN

Blocked = true;
stopCol = c;
BREAK;
Sum + = (|vi.Rank| - |neighborj.Rank|) *

(|vi.Position| - |neighborj.Position|)
END FOR
IF Blocked = TRUE THEN

Biclusterbix = Rows [startRow TO r],
Columns[startCol TO stopCol - 1];
Biclusters [Index] = bix;
Index = Index + 1;
startRow = r;
r = r - 1;
startCol = stopCol;
Algorithm 2: Computing the Barycenter for each node

in the two layers
For all i such that vi is a vertex in
top or bottom layer
Compute weighted mean for vi
IF Nodei.Rank != WeightedMean Then
Nodei.Rank = WeightedMean
PositionChanged = PositionChanged + 1
End if
End for
Algorithm 3: Compute the new weight for each node in

the two layers
FOR ALL i such that vi is a vertex in
top - or bottom - layer
vi.NewRank = | vi.Rank | + get vi
neighbors Rank as in algorithm 4
END FOR
Algorithm 4: Get neighbors Rank of vertex
FOR Each neighbors j of vertex i
After the process of crossing minimization we

perform the Biclustering process as in algorithm 5,
this algorithm find biclusters by performing local
search because the fact that the crossing
15
4. EXPERIMENTAL RESULTS
The proposed algorithm is written in C-sharp
language and the experiments are running on i5 2.3
GHz processor and 4 GB of RAM, and under
windows 7 operating system. We apply the
algorithm on cancer gene expression datasets:
lymphoma dataset [14], gastric cancer dataset [15],
and AML prognosis dataset [16] to test the
accuracy of the proposed algorithm relative to
Barycenter used in SPHier algorithm.
We apply the proposed algorithm and Barycenter
algorithm on the three cancer datasets, and compare
the results of the two algorithms according the
accuracy of crossing minimization.
After applying the proposed algorithm, and
Barycenter algorithm of crossing minimization on
the cancer datasets we get the results in table 2, that
it is represented in figure 3.
In this experiment we are focusing mainly on the
accuracy of the algorithm, because the most
important thing in the gene expression datasets
analysis is the accuracy of results more the time the
algorithm takes, because of the revolution in
computational power of processors solve many of
problems of computations time.
Figure 3 shows enhancements that our algorithm
made in the Barycenter algorithm, the proposed
algorithm achieves better results especially in large
size gene expression datasets with a small number
of conditions.
Because of the decrease in crossings number of
bipartite graph leads to grouping related node
together, the enhancement in crossings number
increases the efficiency of biclustering algorithm
[7].
Table 2. Crossings number of original dataset, Barycenter
algorithm and proposed algorithm
Original
Dataset
Barycenter
Algorithm
Proposed
Algorithm
Gastric
11,048,714,141
11,048,713,498
11,048,654,635
AML
Prognosis
131,451,569,724
131,451,501,126
131,450,516,864
Lymphoma
100,790,992,882
100,790,992,882
100,790,645,942
Standaderd
Deviation
300000
200000
BaryCenter
Algorithm
100000
Proposed
Algorithm
0
7129 12625 22283
Number of genes in thouthands
Figure 3. Comparison between standard deviation of
Barycenter algorithm and proposed algorithm from the
original crossings number
5. CONCLUSIONS AND FUTURE WORK

Many of researches used barycenter heuristics for
crossing minimization. In this research we used also
barycenter heuristics, but with a little modification
in way of calculations of the weight that it is given
to each node in the graph.
Bipartite graph crossing minimization algorithms
enhance the order of gene expression dataset, and
this leads to enhancement in the biclusters extracted
by the biclustering algorithm.
We find a way to enhance the barycenter heuristics
to give better arrangement of the gene expression
dataset, and this is help us to get better results of
biclustering algorithm.
In the proposed algorithm we add the rank of each
node to the rank of its neighbors, and using the
position of each node in the calculations, as we
discussed in the proposed algorithm, to give a new
rank to each node, which we depend on it to reorder
the nodes in each layer.
When we analyze the results shown in figure 3, and
by knowing that the number of genes, and
conditions of each dataset as stated in table 3, we
conclude that the best result of gene expression
datasets is the result of lymphoma dataset, which
has the minimum number of conditions, and
maximum number of genes. The efficiency of the
proposed algorithm affected by the number of
nodes in each layer, as we increase the difference
between then number of nodes in each layer, the
efficiency of the proposed algorithm increase.
As we stated in section 4, we didnt focus on the
performance of the algorithm, but generally the
16
proposed algorithm takes much time than
barycenter algorithm, because our algorithm depend
on counting the crossings number for every single
iteration of the algorithm, and this takes long time
to compute.
The area of biclustering is very complex and all
researches in this area is considered as a domain
specific solution, and in some cases its a problem
specific solution, we can find algorithms that find
good biclusters in some cases, and in other cases
these algorithms do not perform well, so there is no
perfect biclustering algorithm or magic algorithm
that can solve all the problems in a prefect way.
We want to state that the proposed algorithm is an
idea for finding the relations between numbers, and
its positions in the dataset, and how these positions
affect the process of crossing minimization. We
take this trend to make it our work in the research,
and we cannot say that it is the perfect solution of
crossing minimization problem, but it is a new
trend of research that we want to go through it.
2. Macgregor, P.F. and Squire, J.A. 2002. Application of

microarrays to the analysis of gene expression in cancer.
Clin.Chem. 48: 1170-1177.
3. Amos Tanay, Roded Sharany, and Ron Shamir,
Biclustering Algorithms: A Survey, 2004.
4. S. C. Madeira and A. L. Oliveira, Biclustering algorithms
for biological data analysis: A survey. IEEE/ACM
Transactions on Computational Biology and
Bioinformatics, 1, 2004.
5. A. Abdullah & A. Hussain, A new Biclustering technique
based on crossing minimization, Neurocomputing Journal
69, page 1882-1896, 2006.
6. Y. Cheng and G.M. Church. Biclustering of expression
data. In Proceedings of Intelligent Systems for Molecular
Biology. 2000.
7. Waseem Ahmad, Ashfaq Khokhar, SPHier: Scalable
Parallel Biclustering Using Weighted Bigraph Crossing
Minimization, 2007.
8.
I. M. El Henawy, Ahmed Hussain Kamal and Ibrahim

Hamed, Using Tabu Search as a Crossing Minimization
Technique toward Biclustering Gene Expression
Datasets. In: Forth International Conference on
Intelligent Computing and Information Systems, Cairo,
Egypt, 2009.
9.
Marti R. (1998), A Tabu Search Algorithm for the

Bipartite Drawing Problem, European Journal of
Operational Research, Vol. 106, pp. 558-569.
Table 3 number of genes and conditions of gene expression

Datasets used in the experiment
Number of genes
Number of conditions
Gastric
7,129
30
AML Prognosis
12,625
58
Lymphoma
22,283
29
We summarize the Future works for the proposed

algorithm in the following directions: firstly, it can
be enhanced to increase the efficiency of crossing
minimization algorithm for minimum size gene
expression datasets or for datasets with large
number of conditions, and secondly, the proposed
algorithm can be enhanced in a way that enables to
implements it in a parallel way, finally, we can
extended the proposed algorithm to be applied to
different types of datasets other than gene
expression datasets.
6. REFERENCES
1. Peter Clote, Rolf Backofen, Computational Molecular
Biology, an Introduction, John Wiley & Sons Ltd, 2000.
10. A. Prelic, S. Bleuler, P. Zimmermann, A. Wille, et al. A

systematic comparison and evaluation of biclustering
methods for gene expression data, Bioinformatics 22
(2006), 1122-1129.
11. Arifa Nisar, Waseem Ahmad, Wei-keng Liao, Alok N.
Choudhary: High Performance Parallel/Distributed
Biclustering Using Barycenter Heuristic. SDM 2009:
1050-1062
12. A. Tanay, R. Sharan and R. Shamir, Discovering
statistically significant biclusters in gene expression
data. Bioinformatics, 18 (2002), 136-144.
13. Ahmad, W., Khokhar, A.: chawk: A highly efficient

biclustering algorithm using Bigraph crossing
minimization. In: Second International Workshop on
Data Mining and Bioinformatics, VDMB 2007, Vienna,
Austria (Held in Conjunction with VLDB2007). (2007).
17
14. Raetz EA, Perkins SL, Bhojwani D, Smock K et al. Gene
expression profiling reveals intrinsic differences between
T-cell acute lymphoblastic leukemia and T-cell
lymphoblastic lymphoma. Pediatr Blood Cancer 2006
Aug; 47(2):130-40.
15. Hippo Y, Taniguchi H, Tsutsumi S, Machida N et al.
Global gene expression analysis of gastric cancer by
oligonucleotide microarrays. Cancer Res 2002 Jan 1;
62(1):233-40.
16. Yagi T, Morimoto A, Eguchi M, Hibi S et al.
Identification of a gene expression signature associated
with pediatric AML prognosis. Blood 2003 Sep 1;
102(5):1849-56.
18
IMPROVING AND MEASURING COLOR EDGE DETECTION ALGORITHM

IN RGB COLOR SPACE
Mohamed D Almadhoun
University College of Applied Sciences
Palestine, Gaza
mdmadhoun@ucas.edu.ps
ABSTRACT
Edge detection process has a widespread usage in
computer vision applications. But it has a different
output when its input image changes from color to
grayscale. This changeability of results make us add
modification on edge detection process procedures to
correctly detect all those edges in color images that
cant be detected in gray ones. This research reviews
the proposed solution of Dutta and Chaudhuri on color
edge detection algorithm that works using RGB color
space, detects problem of a huge set of undetected
edges by their proposed algorithm and find solutions
for that, and applies complexity and performance
analysis and experiments to compare the proposed
algorithm with Canny and Sobel edge detection
algorithms.
KEYWORDS
Computer vision, Color edge detection, Enhancement
on edge detection, Complexity analysis, Performance
Comparison.
INTRODUCTION
Edges are sudden variation in the gray level or

color of image pixels [2], edge detection is an
important operation for reducing processed data
with preserving useful and informative object
boundaries [1]. Original color edge detection uses
some specific procedure consisting of classical
operators to combine the RGB colors of an image
to shape the original edges of image [3], then
applies a traditional edge detection algorithms
which passes through three steps, starts by
reducing noise which acts like fake edges, then
edge enhancement to strengthen edged and
suppress elsewhere which is expressed as high

pass filter, and finally edge localization to apply
threshold of deciding which pixels are edges and
which are not [2]. Research has shown that 90% of
edges in color images can be found their
corresponding grayscale images, and the
remaining 10% of edges may not be detected in
intensity images due to change in color, which
may cause a fail of vital computer vision tasks [1]
[4].
Dutta and Chaudhuri in [1] proposed an algorithm
for edge detection of color images which starts by
median filter to suppress unwanted noise in the
image, then maximum directional differences of
sum of gray values (Red+Green+Blue) are
calculated for each pixel, then a single value autothreshold is applied, and finally an edge thinning
technique is applied to generate edge map [1].
A problem of RGB transformation procedure was
detected in Dutta and Chaudhuri algorithm which
makes a huge set of couple of different colors are
transformed to a single value by the RGB
transformation procedure. This problem means
there are lots of edge in the color image will not be
detected.
By this research, an enhancement on Dutta and
Chaudhuri algorithm will be proposed. In addition,
a complexity analysis, and performance
measurement operations in terms of execution
time and number of basic operations will be shown
with comparison to Canny and Sobel edge
detection algorithms.
19
2
ENHANCEMENT ON ALGORITHM
The second step of algorithm that was proposed by

Dutta and Chaudhuri calculates a directional color
difference. This step applied a transformation on
the RGB pixel values to make it one value instead
of three to reduce the computational overhead of
color vector [1]. This step used Equation 1
Pixel(i,j)=2*red(i,j)+3*green(i,j)+4*blue(i,j)
modification that can make algorithm detect this

edge
Pixel(i,j)=2*red(i,j)+3*green(i,j)+4*blue(i,j)+2*Hue
(2)
By entering image of figure 1 (a) to the algorithm

with modified calculation, we got out resulting
image in Figure 2 with correct detected edges
(1)
But this equation will fail in differentiating

between lots of colors like the two colors
(RGB=150,50,250)
and
(RGB=10,210,200),
because the value of pixel(i,j) will be the same
Pixel(i,j)=2*150+3*50+4*250 = 1450
Figure 2. Output image after modification of calculating
color difference
Pixel(i,j)=2*10+3*210+4*200 = 1450
So it will not consider the divider between two
areas of those two colors as a boundary.
Figure 1 and following images show how applying
algorithm did draw the edge between the two
colors
COMPLEXITY ANALYSIS
Basic operation is that operation which contributes

most towards the running time of the algorithm.
To analyze time efficiency of an algorithm, we
count the number of repetitions of the basic
operation as a function of input size [7].
Herewith, we can analyze Dutta and Chaudhuri
algorithm by passing through its steps. Their
algorithm consists of four main steps
1- Smoothing by Adaptive Median filter
Number of comparisons in this step is at
(a)
(b)
Figure 1. (a) Original input image. (b) Output image.
most=number of window comparisons x

n=(9+16+25+36+49+64+72)xn = 271 x n
Where n is number of image pixels
To solve this problem we have to consider a forth

item in Error! Reference source not found., this item
should fulfill this differentiation. By experiment
and from studying other color modes, it was found
that Hue in the HSB is the most suitable item to
finish this problem. Equation 2 shows the
And in the best case it will be 9xn, when

always med>min and med<max
So complexity is (n)
2- Directional Color Difference Calculation
20
This step runs two different loops, number
Its clear that both algorithms run in a cost of
of assignments in the first loop is n, where
(n),
n=number of image pixels, and number of
perspective
assignments in the second loop is at most
comparing number of basic operations repetitions,
4xn
Canny repeats operations less than the proposed
So, its (n).
algorithm repetitions and this makes Canny
3- Threshold Technique
which
of
means
time
they
are
complexity,
equal
from
but
when
finishes his job faster especially when applying
Number of times of executing basic
algorithms on multi images like processing video.
operation in this step is n, where n=number
As for Sobel edge detection algorithm it is based
of image pixels, so its (n).
on 3x3 neighbourhood based gradient operator to
4- Edge thinning
This operation has 2xn of assignments and
complexity of (n) when n=number of
image pixels.
find gradient magnitude at each point of input

image. It used four different convolution kernels to
find edge map from all directions [5]. It follows
these steps:
As for Canny edge detection algorithm, it is multistage starts by detecting edges with miximizing
probability of edge points and minimizing nonedge points, then localization and resulting
resulting one real edge [6], by passing through the
algorithm it follows these steps:
1- Convert to gray: costs n iterations if n is
the number of image pixels
2- Normalize contrast: runs n+255+n, so its
complexity is (n), where n is number of
image pixels.
3- Compute gradients: costs at most is 3 x (k
x n)+ n iterations for the four loops, so
its (n). [k is the number of kernel
elements]
4- Perform hysteresis: costs at most n, so its
(n).
5- Threshold edges: costs n iterations, so its
complexity if (n)
1- ConvertToGray: costs n iterations where

no for vertical is the number of image
pixels
2- Sobel convolution: costs (6+6)xn, So its
complexity is (n)
4
PERFORMANCE MEASUREMENT AND

ANALYSIS
Comparing different edge detection algorithms

makes us do the best choice of using one of them
in some application. Performance measurement
experiment is needed to compare performance of
the three edge detection algorithms: Canny, Sobel,
and the proposed by Dutta and Chaudhuri. Ten
images were downloaded from Cambridge image
database, image resolution was 816x616, with jpeg
format, and RGB color.
In each of the three edge detection algorithms, a
counter for counting the number of basic
operations were added inside its code, and another
statement for getting the execution time of each
algorithm.
21
Performance measurement experiment applied the
three algorithms on one image at first, then on two
images, on three, until reaching to apply 10 images
at a time for each algorithm. Applying algorithm
in a loop on more than one image discovers the
difference between them when using algorithm in
video processing.
process and the clock time at the end of process

[8].
Figure 3 shows the curves of execution time
values for processing multi-images in each
algorithm. This experiment was applied using PC
computer with Dual core CPU of 1.73 GHz, 2 GB
RAM, and 32-bit windows OS. Its clear that the
angle of canny curve is the smallest, and execution
time values in all ten stages of canny experiment
are the lowest, that means canny edge detection
algorithm is the most suitable for video processing
operation.
Since we interested in how quickly program is

executed, performance metric of computer
program is the time required to execute it and
finish its job completely. The basic technique to
measure time of execution is by reading the
difference value between clock time at the start of
Execution Time (In MilliSeconds)

18000
16000
Time in MilliSeconds
14000
12000
10000
8000
6000
4000
2000
0
# of processed images
Sobel
10
1064.4699 2036.1599 2908.2182 3861.8333 4763.6225 5937.2051 6830.3131 7679.9482 8578.4878 9605.5161
Proposed 2408.2262 3463.8285 5182.9545 6720.1038 8367.0729 9976.6211 11724.613 13379.057 15010.007 16549.499
Canny
758.33158 1345.6685 1750.5082 2319.3174 3066.8976 3335.9226 4047.0674 4639.1153 5192.539 5696.5223
Figure 3. Curves of execution time values for processing multi-images in each algorithm
As for counting number of basic operations, next

figure shows that canny has the smallest number
of basic operations, but the contrary thing that
number of basic operations of the proposed
algorithm is smaller than Sobel despite of larger

execution time for the proposed algorithm.
22
The main reason here is that basic operations of
the proposed algorithms need more execution time
than those in Sobel.
Basic Operation Running Times

100000000
90000000
Number of Basic Operations
80000000
70000000
60000000
50000000
40000000
30000000
20000000
10000000
0
# of processed images
Sobel
10
8967024 17934048 26901072 35868096 44835120 53802144 62769168 71736192 80703216 89670240
Proposed 6531668 13063336 19595004 26126672 32658340 39190008 45721676 52253344 58785012 65316680
Canny
2944842 5889684 8834526 11779368 14724210 17669052 20613894 23558736 26503578 29448420
Figure 4. Curves of number of basic operations for processing multi-images in each algorithm
CONCLUSION
By this research, problem of failed detection of

edges between a set of colors in Dutta and
Chaudhuri color edge detection algorithm in RGB
color space was defined and resolved by adding
Hue value from HSB color space to the procedure
of transformation from RGB to single
distinguishing value. In addition, a complexity
analysis was applied on Dutta and Chaudhuri
proposed algorithm, and compared to Canny and
Sobel edge detection algorithms with respect to
time execution and number of basic operations.
Comparison results show that Canny has the best
performance properties.
2.
3.
4.
5.
6.
REFERENCES
1.
Dutta, S., Chaudhuri, B.: A Color Edge Detection

Algorithm in RGB Color Space. In: Proc.
7.
International Conference on Advances in Recent

Technologies in Communication and Computing.
pp. 337 - 340, (2009).
Efford, N.: Digital Image Processing - A Practical
Introduction Using Java. Pearson Education
Limited, ISBN 0-201-59623-7, (2000).
Zhang, L., Mao, X., Zhou, C.: Edge Detection of
Color Image Based on HI*S* Color Space. In: Proc.
International
Conference
on
Information
Engineering and Applications. Lecture Notes in
Electrical Engineering Volume 154, 2012, pp 15291534, (2011).
Chen, H.: A novel color edge detection algorithm in
RGB color space. In: Proc. IEEE 10th International
Conference on Signal Processing (ICSP), (2010).
Arpitha, D., Arakeri, M., Reddy, R.: An Approach
for Color Edge Detection with Automatic Threshold
Detection. In: Proc. International Conference,
ADCONS 2011. Lecture Notes in Computer
Science Volume 7135, 2012, pp 117-124, (2011).
Cheng, Y.: An improved Canny Edge Detection
Algorithm.
Lecture
Notes
in
Electrical
Engineering Volume 126, pp 551-558, (2012).
Levitin, A.:
Introduction to the Design and
Analysis of Algorithms (2nd Edition). Addison
23
Wesley, ISBN-10: 0321358287 | ISBN-13: 9780321358288, (2006).
8. Lilja, D.: Measuring Computer Performance: A
Practitioner's Guide. Cambridge University Press,
New York, NY, ISBN 0-521-64105-5, (2000).
24
COLLABORATIVE PROBLEM SOLVING USING PUBLIC SOCIAL NETWORK

MEDIA: ANALYZING STUDENT INTERACTION AND ITS IMPACT TO
LEARNING PROCESS
Melvin Ballera
Faculty of Information Technology
Strathmore University Kenya
maballera@yahoo.com.ph
Ismail Ateya Lukandu

Faculty of Information Technology
Strathmore University Kenya
iateya@strathmore.edu
ABSTRACT
This paper examines the use of social network media at
three aspects in African and Libyan perspective.
Firstly, to use social network media as an open network
learning environment that
provide service for
interaction necessary for learners to support
socialization and collaboration during problem solving.
Secondly, to use social media as a tool to support
blended learning in e-learning system and encourage
non-native English students to express their ideas and
fill the gap of communication problems. Thirdly, to
analyze the interaction of the learner in social media
threaded messages and its relation to group and
individual performance using different social schema
and social network analyses. Quasi-experimental
results indicate that there is an increase on the
cognitive level of students at different level while
qualitative results reveal that it helps deepen learning,
memorable, have freedom to express opinions and
lessen pressure and increase communication and
socialization.
KEYWORDS
blended learning, collaboration, e-learning, network
learning, socialization, social network media
1 INTRODUCTION
Learning English is not mandatory in Libya, few
years back the language has never been integrated
in the curriculum and English was static and never
been practice outside the school. Thus, foreign
lecturers who were tasked to deliver computer
science and information technology education
suffered communication gap among learners and
used different method to lessen the problem. One
method is to employ e-learning systems that
allows students to learn freely at their home and at
school whenever possible. Given the emerging
Abdalla Radwan
Computer Science Department
Sirte University Libya
radwan2004@hotmail.com
technology, from static e-learning (absence of

multimedia and purely textual) becomes dynamic
as evidence by several researches in e-learning
development. The use of multiple agents [1],[2]
and social media in e-learning system have been
used [3][4], genetic algorithm and personalization
[5][6] interactive and socially capable agent [7]
[8]and socially deceiving agent [9] have been
studied. With those papers, its been very obvious
that socialization becomes a major component of
e-learning systems to cater the needs of the
students. Through this paper, the students are
encourage to use social network media and
collaborate, solve problem using FaceBook,
Yahoo Messenger, Window Live Messenger,
Skype or other network technologies capable to
support networked learning. Further, we have
investigated the use of social network media as
part of the blended e-learning system by analyzing
the threaded messages of the students by studying
its impact on the performance both in group and
individual level during collaborative problem
solving.
Collaborative problem solving receives a
significant attention for its potential to increase
problem solving skills [10],[11],[12],[13], improve
critical thinking [14],[15],[16] and knowledge
acquisition [17],[18] and academic achievements
[19]]among learners. Collaboration describe social
interaction within a group or a team, when
students actively talk, share their cognitive
resources and to produce a single outcome [20].
Students work in teams and act as one and
confronting problems as they occur [21]. Although
they get insufficient information, students must
settle on the best possible way the problem
presented to them [22]. In problem solving, a
number of alternative solutions must be examined
and analyzed to meet the goal. However, meeting
25
the goal is not that easy, since it will entail proper
communication and conversation among members
in the group. One should not feel superior or
inferior among others; otherwise a catastrophic
effect in the learning process will occur.
In the field of e-learning, collaboration usually
takes place using pre-programmed agent or
animated character agent by virtually deceiving
the learner that someone is helping him in the
learning process, but in reality it was preprogrammed based on learners prior knowledge
and personal profile. Many researchers have been
trying to develop collaborative software and
integrate to e-learning module but none of which
have surpassed the power of human socialization.
For example, early e-learning system with
animated agent cannot joke, greet and show facial
expressions because of its limited domain,
complexities, issues and constraints [23]. These
simple gestures can build rapport and develop
personal and affective relationships among
members and somehow affect the overall
performance of the learning group. Thus, instead
of pre-defined and pre-programmed socialization;
social networking has been adopted for
collaboration.
The growth of social networking has created new
opportunities for collaboration in problem solving
[13],[24]. In 2009, University of Cambridge
posted a problem in mathematics using blogs and
solved within six weeks known as Polymath I [25].
Social networking media let people rendezvous,
connect or collaborate, support network of people,
share content and services that are more adaptable
and responsive to changing needs and goals [26].
Social media have already led to widespread
adoption of portfolios for learners bringing
together learning from different context and
providing an on-going record of lifelong learning,
capable of expression in different forms.
The rapid diffusion and public acceptability of
social network media such as Facebook, Yahoo
Messenger, Windows Media Live and Skype have
enable users to connect with people more than
ever before. Student used social media at school
for various purposes such as socialization, sharing
experiences and exchange information and vice

versa [27]. While many instructional strategist or
educators are concerned with how they should
treat social media in order to prevent classroom
disruptions, social media provide affordable
resources that can build social learning
environment in a way was not possible before and
collaborative problem solving. Recent research
shows that the educational use of social media
have significant potential in collaborative problem
solving [1],[24],[28] and blended learning in elearning systems.
In this paper, our main objective is to experiment
whether public social media can help students in
collaborative problem solving, improve learning
delivery by employing such technology and one
way to encourage non-native English student to
express their idea. The researcher believed that
students suppressed their sharing ability and teambased attitude due to communication difficulty. In
the experiment we will try to prove and encourage
e-learning stakeholder to consider implementing elearning with blended environment using social
media. The researcher believed that we can
establish the benefits of employing social media in
the learning process such as knowledge transfer,
skill acquisitions, teambuilding, collaboration,
socialization and their perceptions towards the
used of this technology.
The paper is organized according to the following;
discussion of study and methodologies, findings
and discussions, and conclusions
2 THE STUDY
2.1 Participants and Materials
Participants were enrolled in Design and Analysis
of Algorithms, one of the core computer courses
that requires mathematical analysis and
algorithmic program. The study has been
conducted for two semesters at university with 48
students, the first semester composed of 20
students with 12 females and 8 males and divided
into 4 groups consisting of 5 members while the
second semester composed of 28 students, which
is composed of 21 females and 7 males and
26
divided into 5 groups consisting of 5 or 6
members; overall, there were 9 groups in the
study. Group members were randomly selected as
suggested by [29],[30].
The topics included in the course Algorithm in elearning module have been selected/driven by
either by the problems practical importance or by
some specific characteristic making the problem
an interesting research subject. The following
topics have been included in the module such as
sorting techniques, searching algorithms, string
processing, graph problems, combinatorial
problems. These topics are all suitable for
collaborative problem solving.
like button signifying that sessions is relevant to

the problem. The participants are also permitted
and welcomed to use the open forum for both
casual conversation and information sharing
(Facebook), invite to conference (Yahoo
Messenger) and request for remote assistance
(Windows live Messenger). All activities are
conducted on group page, which are setup for use
by small study groups.
During face-to-face session, students are required
to submit their progress report including the
printout of their sessions, program code or pseudocode and computation of time complexity. They
are also required to present and share to other
groups their solution to problems.
2.2 Program Overview

2.3 Data Collection
The study was created as a blended learning
approach by combining on-site studies, face-toface guidance and collaborative problem solving
using social media such as Facebook, Windows
Live Messenger and Yahoo Messenger. Each
group freely chooses social media on their own
with 5 Facebook, 3 Yahoo Messenger and 1
Windows Live Messenger.
During on-site study, the students get familiarize
with the topics e.g sorting problems. The elearning module discussed how time complexities
will be computed and expressed in mathematical
notations. There are many possible solutions of the
sorting problem and 2 of which will be presented
using simulations, with program code, and time
complexities computation. Others will be left for
collaborative problem solving using social media.
The students are required to discuss their problem
and share their idea in the social networking site
until they agree and arrive with the final solution
of their given problem. The final output of the
students will include computation of the time
complexity and its algorithmic program
implementation.
During collaboration, facilitator who is in-charge
with the course is a member of all the groups and
usually view and see the communication
threading. The facilitator will never comment and
just read the threaded messages but can press the
The data used in this study are both quantitative

and qualitative. The quantitative data used survey
to collect the demographic profile of the learner
such as gender, age, year level and grade point
average. After reading the on-line materials of
Design and Analysis of Algorithm, the students
were asked to answer set of questionnaires about
the course. The same set of questionnaires were
administered after collaboration and compared
their performance both for group and individual
using Blooms Cognitive Test [31]. In analyzing
the threaded messages of the group, four
approaches were used; counting in and outmessages, coding the learning process using
Veldhuis-Diermanse Schema Technique, tutoring
and communication process analysis using
Anderson Schema and social network connectivity
(SNA) using NodeXL software to analyze density
of
collaboration
and
social
centralities
[32],[33],[34].
The qualitative data on the other hand were
collected via students description and opinions and
then combined to note emerging pattern to gain
understanding of the learner experience in the
study. Data were re-analyzed until saturation had
been reached.
27
40% is out-messages. In-messages are the number
of messages contributed by all groups related to
the problem at hand while the out-messages are
interjected social interaction. Though outmessages are threaded communication that are not
related to problem, it becomes integral in the
overall collaboration by providing informal
communication. Informal communication entice
the collaboration by making jokes, creating
relationship, encourage to talk and share opinions,
thus, allowing learner to increase their level of
collaboration. Among the 9 groups studied, Group
9 emerge with 101 messages (66 in-messages and
30 out-messages), Group 1 contributed
82
messages (62 in-messages and 20 out-messages)
while Group 6 contributed 82 total messages (34
in-messages and 48 out-messages). It appears from
the table that Group 9 is the most serious group
during collaboration as indicated by 66 inmessages. Group 6, 4 and 2 are the most least
sociable group with 5,10,15 out-messages reported
respectively. This is perhaps attributed to random
selection of participants. Group 8 has a 51 outmessages making it the most sociable group as
indicated by many times used of social interjection
like greetings to one another, asking how is family
life, asking about weather, news and others,
evident to Arab learners everyday life.
3 FINDINGS AND DISCUSSIONS

For the purpose of epitomizing the results and
analysis, some groups were selected for
interpretation. We are also assuming that the
learning process are actually represented by the
expression we are coding drawn from the schema.
The limitation of the coding process is that
linguistic expression varies accordingly based
upon the intentions and motivation of the learner.
Nevertheless, the codings suggest some clear
pattern in the study.
3.1 Quantitative Analysis
The quantitative data regarding the demographic
profile of the learner shows that there are 15 males
and 33 females with age mean of 22 and standard
deviation of 2.0. Year level of all the learners
belong to 4th year with a grade point average 63.5
and a standard deviation of 10.1
3.1.1 In and Out Messages
Table 1 shows the in and out-messages of all the
group with 573 total messages encoded during the
collaboration. Out of 573 messages, 344 or 60% of
the total encoded is an in-messages while 229 or
Table 1. In and Out-Messages of Different Groups During Collaboration

Type of Massages
(No. of Members)
In-Messages (344)
Out-Massages (229)
Total Messages (573)
Group 1
(5 )
Group 2
(5)
Group 3
(5)
Group 4
(5)
Group 5
(6)
Group 6
(6)
Group 7
(5)
Group 8
(5)
Group 9
(6)
62
20
82
24
15
39
26
25
51
60
10
70
34
48
82
28
5
33
26
20
46
18
51
69
66
35
101
3.1.2 Coding Schema

Analyzing the threaded messages, two coding
schema have been adapted for analysis. The first
coding schema is developed by VeldhuisDiermanse Schema [34], was used to code units of
meaning focusing on learning processes that
includes the three main categories: cognitive
activities - used to process learning content and to
attain the learning goals; metacognitive knowledge
and metacognitive skills - used to regulate the
cognitive activities; and affective activities - used

to cope with feeling occurring during learning.
The second schema is the Anderson [35], was
used to attempt to reveal the ways in which the
participants were facilitating and regulating each
others learning, while undertaking the
collaborative problem solving. This schema
involves three activities: design and organization,
facilitation of discourse and direct instruction
activities. Capturing these activities using strict
syntactic rules analysis is difficult and not possible
28
due to elaborative nature of discussion
nevertheless it was coded using the guided
example provided by the two coding schema.
procedural patterns, and concepts that serve in the

development of intellectual abilities and skills.
In the affective category, the same set of group
emerge, Group 9 produce 18 messages, Group 1
with 17 and Group 4 with 15 messages. These
means that these groups value the feelings of other
learner by carefully criticizing their opinions and
their suggestions. It includes manner to deal with
emotions,
feelings,
values,
appreciation,
enthusiasms, motivations, and attitudes. Group 2,
5 and 6 are the least affective group. This could be
attributed to feeling of strangeness within the
group. Looking at the meta-cognitive learning
category, Group 9 has 21 messages, Group 4 has
20 messages and Group 1 has 15 messages.
Usually, this category viewed learners as more on
planning, clarifying matters, asking for more
suggestions and monitoring other learners. This
means that students collaborated and discussed
solutions, the learner with the most convincing
possible solutions will be adapted and become
basis in the threading. The learners are trying to
contribute and help to achieve the goal.
Veldhuis-Diermanse Schema
Table 2 summarized the coded learning processes
using Veldhuis-Diermanse Schema where type of
learning has been divided into four subgroups;
Cognitive, Affective, Meta-cognitive and Others.
Other category is the number of out-messages.
Among the 344 in-messages from Table 1, it was
further breakdown to Cognitive category with 155
or 45% of the total in-messages , Affective with 91
or 26%, and Meta-cognitive with 98 or 29%. For
cognitive category, Group 1 has 30 messages,
Group 9 with 27 messages and Group 4 with 25
messages, they were ranked 1, 2 and 3
respectively. In this category, member of the
groups debated and discussed, used external
information and experiences and linked internal
information or concepts found. It involves
knowledge and the development of intellectual
skills by recall or recognition of specific facts,
Table 2. Units of Meaning Coded for Learning Process(Veldhuis-Diermanse Schema)

Type of Learning Process
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Group 7
Group 8
Group 9
30
17
15
62
12
4
8
24
10
10
6
26
25
15
20
60
18
6
10
34
13
7
8
28
14
8
4
26
6
6
6
18
27
18
21
66
Cognitive (155)
Affective (91)
Metacognitive (98)
Total (344)
Anderson Schema
messages, contributing 22% in their group

respectively. This is not surprising given that the
group was engaged in collaborative activity that
drew own members individual resources and
other materials to which they were directed before
the activity commenced. Members were hesitant at
first how to present their ideas and questions to the
group. Due to communication gap and
estrangement, students are not reluctant to share
knowledge unless being asked to do so, not to
mention their personality issues and poor
understanding of the problem at hand.
Table 3 summarized the Anderson schema of how

communication and tutoring takes place in the
social network media. This schema is divided into
three subgroups; direct instruction, facilitation and
instructional design. Direct instruction receives the
lowest messages in the schema throughout the
entire activity of all group with a total of 70 coded
messages or 20% out of 344 in-messages. Group 9
has 16 messages, contributing 24% , Group 1 has
15, contributing 24% and Group 4 has 13
Table 3. Units of Meaning Coded for Tutoring and Communication Process (Anderson Schema)
Type of Tutoring Process
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Group 7
Group 8
Group 9
Direct Instruction (70)

Facilitation (120)
Instructional Design (157)
Total (344)
15
20
27
62
7
7
10
24
2
14
10
26
13
18
29
60
8
12
14
34
8
9
11
28
7
9
10
26
3
5
10
18
16
20
30
66
29
Facilitation has 120 total messages out of 344 inmessages or 35% have been coded into this
category. Group 1 and Group 9 coded with 20
messages respectively followed by Group 4 with
18 messages. In this category, indicators show that
student making agreement/disagreement, seeking
understanding/consensus, reinforcing student
contributions and assessing the efficacy of the
collaboration process.
The learning and tutoring pattern that have

emerged from this coding analysis provide some
insights into the dynamics of individuals and
group behavior in social network environment.
Collaborative learning is dependent on individual
contributions. But the benefits of doing
collaboration have a big impact on individual
learning process; transforming learners to become
independent.
Instructional Design composed of 157 messages or

45% of the total 344 in-messages. Examining the
performance of each group in this category shows
that instructional design category receives the
highest number of messages in the schema. From
the table, Group 9 has a total of 30 messages
followed by Group 4 with 29 messages and Group
1 with 27 messages. This is not surprising because
this is where the students are starting to set their
goals and how to deal with the problem. Learners
are currently negotiating how to deal with the
problem, establishing parameters and deadlines,
and assigning task to members. An indication that
collaboration will take place effectively.
3.1.4 Blooms Cognitive Level and T-test

To determine the benefits and impact of
collaborative problem solving, an exam was
conducted. Table 4.1 and 4.2 shows the
comparative cognitive level based on pre and postexam results by the groups. In the exam, a total of
18 items have been asked with 3 items in each
category. The two tables show the six
classification based on Bloom Taxonomy of
Cognitive learning; Evaluation, Synthesis,
Analysis, Application, Understanding and
Knowledge.
Table
4.1
show
the
results/performance of each group upon reading
the e-learning module while Table 4.2 was taken
after the collaborative problem solving.
Table 4.1 Blooms Cognitive Level Coding of Different Groups (after reading e-learning module)
Cognitive Level
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Group 7
Group 8
Group 9
Evaluation
Synthesis
Analysis
Application
Understanding
Knowledge
1
0
1
1
1
2
0
0
0
0
1
1
0
0
1
0
0
1
1
1
0
1
2
1
0
0
0
0
1
1
0
0
0
1
1
1
0
0
0
0
0
1
0
0
0
0
1
0
1
1
1
1
2
3
Cognitive Level
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Group 7
Group 8
Group 9
Evaluation
Synthesis
Analysis
Application
Understanding
Knowledge
2
2
3
3
3
3
0
1
1
1
2
2
0
1
1
2
3
3
3
1
2
2
3
3
1
1
1
2
3
3
1
1
1
1
2
3
0
0
1
1
2
2
1
0
1
1
2
1
1
2
3
2
3
3
Table 4.2 Blooms Cognitive Level Coding of Different Groups (after collaborative problem solving)
Based on group performance, Group 1 and Group

4 were taken for analyses. For group 1 there is a
significant increased in all the categories;
Evaluation increased by .33, Synthesis, Analysis,
Application, and Understanding are all increased
by .66, while Knowledge increased by .33. For

Group 4, it shows a dramatic increase for
Evaluation by .66, Synthesis remain at .33,
Analysis increase by .66, Application increase by
.33, Understanding increased by .33 and
30
Knowledge remain at 1; a perfect score for pre and
post-exam. For knowledge category, students
exhibit memory of previously learned materials by
recalling the content of the e-learning module, so it
is not surprising that in the pretest, these categories
have high result.
Understanding category also increased due to prior
knowledge taken by the students from the elearning module, students were able to organize,
compare and interpret data given a change of the
input of the algorithmic problems of the course
presented to them during collaboration. Initially,
Synthesis and Evaluation have the lowest value
but it is where the cognitive level significantly
increased. The highest cognitive level in Blooms
taxonomy is Evaluation where student are
expected to demonstrate the ability to validate data
based on predefined criteria. Student for example,
were able to prove that the time complexity of
sorting technique varies accordingly as input
changes. Another criteria in analyzing the
algorithm is generality, where certain inputs can
be changed into different format and re-evaluate
their coded program to fit the new input. In the
Synthesis, students can change the program from
one data types to another, e.g. from integer, to
character, or string or another data types in their
coded program. These changes
were not
specifically mentioned in the e-learning module

but it was observed and illustrated during
collaborative problem solving.
In the study, we want to show that working by
group collaboratively in the area of problem
solving
can
increase
critical
thinking,
understanding, and other cognitive domain, that
when applied individually it will lead to an
increase of individual performance credited from
collaboration. Table 4.3 shows the average
individual cognitive difficulty level of Group 9.
The result of the pretest (after the e-learning
module) and the posttest (after the collaboration)
show that there is a significant increase of
individual learners in different cognitive level;
Knowledge from .60 to .98, an increase of .38,
Understanding from .28 to .93, an increase of .64,
Application from .55 to .77, an increase of .22,
Analysis from .06 to .60, shows a dramatic
increase of .54, Synthesis from .38 to .55, an
increase of .17 and Evaluation from .11 to .50. an
increase of .39. Further, the table reveals that from
Analysis up to Evaluation level, majority of the
learner failed to answer correctly during the
pretest
but
increase
significantly
after
collaboration. This is quite considerable since
students are still relying on their individual
knowledge and effort.
Table 4.3 Blooms Cognitive Level

(Individual Indicators - Before and After Collaboration)
Group 9
Knowledge
Understanding
Application
Analysis
Synthesis
Evaluation
Before
After
Before
After
Before
After
Before
After
Before
After
Before
After
Member 1
0.33
0.98
0.33
0.66
0.33
0.66
0.00
0.33
0.33
0.33
0.00
0.33
Member 2
0.33
0.98
0.33
0.98
0.00
0.66
0.00
0.33
0.33
0.66
0.00
0.66
Member 3
0.66
0.98
0.33
0.98
0.33
0.66
0.00
0.66
0.33
0.66
0.00
0.66
Member 4
0.98
0.98
0.33
0.98
0.66
0.99
0.00
0.98
0.33
0.66
0.33
0.33
Member 5
0.66
0.98
0.00
0.98
0.98
0.99
0.33
0.98
0.00
0.66
0.33
0.66
Member 6
0.66
0.98
0.33
0.98
0.98
0.66
0.00
0.33
0.33
0.33
0.00
0.33
Average
0.60
0.98
0.28
0.93
0.55
0.77
0.06
0.60
0.28
0.55
0.11
0.50
31
Table 4.4 Comparative Group Performance of All the Group in 2nd Semester
( T-test Before and After Collaboration)
Table 4.4 shows students pretest and posttest

result analyses which include the mean, standard
deviation, maximum score, minimum score, range
and t-test. Among the 5 groups, Group 9 emerged
with a highest posttest 12.83 mean with an
increase of 8, followed by Group 8 with 12 with an
increase of 7.2, Group 7 with10.8 with an increase
of 7, Group 6 with an increase of 6.6 and Group 5
with an increase of 6.94. The standard deviation
before and after increased at all groups while
minimum and maximum before and after
collaboration shows a dramatic increased at all
levels, signifying a knowledge transfer has
occurred during collaboration. To test the
difference between the mean of the two exams if
statistically significant,
t-test evaluation was
conducted. Group 6 received the highest t-value
with .000742, followed by Group 8 with .000254,
Group 9 with .000235, Group 7 with .000215 and
Group 5 with .000118 respectively. All the Tvalue is accepted with .05 cut-off criterion. This
mean that the difference is not likely to happen by
chance, and therefore, statistically significant.
Overall, there is a general indication that there is a
positive effect of using public social media as a
collaboration tool. Further concluded that a
knowledge transfer occurred and skill acquisition
acquired.
3.1.5 Social Network Analysis

Social network analysis (SNA) is the mapping and
measuring of relationships and information flows
between members in the groups. By conducting
SNA several factors and essential benefits can be
revealed and study how the process of
collaboration somehow affect the performance of
individuals. Some students become leader then
gradually faded while other become active and
later inactive and many others. Through this, we
will be able to unmask properties of socialization
and collaboration and reveal the centralities. To
understand networks and their participants, we
evaluate the location of actors in the network.
These measures give us insights into the various
roles and groupings in a network who are the
connectors, leaders, bridges, isolates, where are
the clusters and who is in them, who is the core of
the network, and who is the periphery.
Starting with SNA analysis of Figure 1 and 2, it is
clear that participation over time is dynamically
changing as all groups becoming more
interconnected and communicated to solve
problem. Initially, the overall tendency is to act as
a group and getting familiar to each other. There
are central and peripheral participants, but the
interaction are not centralized around a few
dominant participants. Based on the preliminary
collaboration phase, Eman dominates Group 5,
32
Sadiya for Group 6, Mona for Group 7, Fathiya
and Saber for Group 8 and 9 respectively.
Participants tend to communicate to the person
whom they perceived intelligent or have
exceptional characters among the groups.
As deadline approaches, collaboration tend to
change and communication effort become more
dense. Figure 2 shows how interaction seems to
change more drastically throughout the ending
part. The central participant in the preliminary
collaboration phase have established their position
except for Group 9 where initially it was Saber but
was taken gradually by Mosbah. It is very
interesting to note that other central players of
other groups seem to collaborate with Mosbah,
making him the most active and the most
influential node in the network. He become the

connector or a hub in the network making the
connections to other groups vulnerable. If Mosbah
node is damaged or removed, the network quickly
fragments into unconnected sub-networks. A
highly central node can become a single point of
failure. Also noted that during preliminary
collaboration, many nodes have been isolated such
as Hadi, Aisha, Mariam, Amira and Naja2 but
gradually become active and communicate among
members in the group. Central node or the most
connected node in each group network seems
given an authority by all member to communicate
outside, a common practice of an organized team
learning.
Figure 1. Preliminary Density Group Collaboration
33
Figure 2. Post Group Collaboration
Taking deeply the SNA analysis of Group 9,

Figure 3 and Figure 4 show the connections and its
collaboration network. This network visualizes all
the connections each participant has with other
members in the group and how communication
evolves over time as they work collaboratively on
the learning task. Also, in these figures the number
associated with the network represent the volume
of communication between participants. Figure 3
was taken in the first week of collaboration with a
total of 50 in and out messages while Figure 4 was
taken on the third week of the collaboration with
101 in and out- messages. At the beginning, Saber
has initiated the most of the discussion with 16
messages. He acts as a central member of the
group as depicted in the graph while Mosbah the
most intelligent as revealed in the individual
cognitive performance level contributed only 7
messages. In the network analysis we cannot
identify which one contributes the best in the
problem solving since we are only showing the
threaded messages to whom participants
communicated with. At the end of the
collaboration interaction pattern seem to change.
Figure 4 revealed now that Mosbah increased his

collaboration by overtaking Saber and become the
new central player of the group with 25 messages
while Saber contribute 18 messages. There was no
evidence what triggered Mosbah to increase his
collaborative effort but examining the pattern of
this student, he is slow at the beginning and
suddenly climbing at the end of the collaboration,
other students follow the same pattern.
Collaboration effort also increased with Ali and
Saed both 8 to 15 messages, while Mufta from 5 to
16 and Ahmed from 9 to 14 messages.
In this study, students were asked to participate in
the social network media at their convenience to
exercise freedom and flexibility, but were given a
deadline to meet the requirements. This study want
result to be real as possible. Interestingly, students
increase their individual contributions as deadline
comes. Another interesting feature we can read
from this graph is how tightly knit the
collaboration effort this learning group is. All
participants get responses from almost all the
members. Although they have different levels of
34
contribution, everybody is engaged. . For this, in
the beginning student gets to know each other and
slowly
increases
the
collaboration
and
communication. As time goes by and each
become relax to deal one another. Surely, there
was an increased of learning since no one among
the groups lose interest during their collaborative
problem solving. The interaction pattern clearly
showed transformation of membership as learners
Figure 3. Social Network Analysis

(Density of Collaboration for the 1st Week of Group 9)
Measures of Centrality
There are many important properties within a
social network that relate to the identification of
the power of relationships in the network structure.
It was revealed that certain participants in
collaborative online learning communities have a
greater power within the community than others.
This was evident in the study when comparing the
fact that some participants attracted a number of
responses to their queries while others attracted no
responses and quickly become isolated. Taking
into account the preliminary group collaboration
of Figure 1 and the post group collaboration of
Figure 2, the following centrality has been
computed Table 5.1 and 5.2 for degree centrality,
Table 5.3 and 5.4 for betweenness centrality and
Table 5.5 and 5.6 for closeness centrality.
Degree Centrality refers to the number of
connections that a node contains and indicates the
level of activity of a node within a group. The
gradually move towards the center of the network.

We have also seen that it is not necessarily the
case that the most active members always regulate
and dominate the discussion, in the case of Saber
he communicated very well from the beginning to
open the discussion then gradually taken by
Mosbah. Saber is more socially engaged while
Mosbah is more concentrated in getting the goal.
Figure 4. Social Network Analysis

(Density of Collaboration for the 3rd Week of Group 9)
general idea is that the greater a nodes degree,

the more potential influence it has on the network
and the more potential influence the network has
on it. Table 5.1 and 5.2 shows the degree centrality
among the five groups. For Group 5, Eman is the
most active in the network and manage to maintain
till the end of the collaboration but what is so
interesting to note is that, as the collaboration
takes place, its original value of .32 becomes .28,
a decrease of .04. This means that during the
process of solving the problem, Eman decreased
its degree centrality while other member increased.
Other members actively participated in the social
network threading. In Group 6, the reverse
happened, Sadiya from .33 to .4 and increased of
.07. It means that she was able to maintain her
leadership and become the connector or hub of the
group. Group 7 performance likewise shows a
dramatic change where all member decreased their
degree of centrality except for Wadian, whom she
increased her participation in the threading.
Almost all members of Group 8 decreased their
35
degree centrality an indication that all groups were
actively involved and working very close while
Group 9 shows a very good relationship to each
other as their degree centrality becomes all equal
at the end of the collaboration. It means that there
is no competition among the members as they are
all involved in solving problem and just focused to

deliver the output as one group activity. Overall
the minimum and maximum degree, average and
median degree decreased, a general conclusion
that the students work as team in solving the
problem.
Table 5.1 Degrees of Centrality of Preliminary Group Collaboration
Table 5.2 Degrees of Centrality of Post Group Collaboration
Betweenness centrality refers to the fraction of the

number of shortest paths that flow through a node.
It is an indicators how information flows through
graph. Nodes that occur on many shortest path will
have a higher value of betweenness than those that
do not. Table 5.3 and Table 5.4 shows the value of
preliminary and post collaboration betweenness
centrality taken from Figure 1 and Figure 2
respectively. In the preliminary collaboration
stage, Eman have 108 for Group 5, Sadiya have 90

for Group 6, Mona have 88.33 for Group 7,
Fathiya have 104 for Group 8 and Mosbah have
200 for Group 9. They become the most influential
and have the power to control the information flow
in their respective network. However having said
that, there is an interesting observation which
happened in the post collaboration betweenness
centrality value, there is general and conclusive
36
observation that collaboration takes place and
students gradually contributing to problem solving
as the maximum betweenness and average
decreased drastically. It means that students
actively participated in problem solving.
Analyzing Group 9, preliminary shows that
Mosbah dominates the discussion but gradually all

members equally participated in the process. If one
of the nodes will be damaged or destroyed, the
information flow will not be affected since all
information are readily available among the
members as indicated by 0 betweenness value.
Table 5.3 Betweenness Centrality of Preliminary Group Collaboration
Table 5.4 Betweenness Centrality of Post Group Collaboration
Closeness refer to the geodesic distance of a given

node to all other nodes in the graph. The closeness
of a node indicates how easily a node can be
reached. In general, a node with a relatively high
level of closeness can be more easily reached and
receive information more quickly. Table 5.5 and
Table 5.6 shows all the value of closeness
centrality taken from the network of Figure 1 and
Figure 2. According to the table for both
preliminary and post collaboration stage, Eman

have the highest in Group 5 with .019 to .167,
Sadiya have the highest in Group 6 with .017 to
.250, Mona have the highest in Group 7 with .018
to .018 and Fathiya have the highest in Group 8
with .02 to .250 and Mosbah has the highest in
Group 9 with .023 to .2. They are the most nodes
that can efficiently obtain information in the
network and into other groups. They have the
37
shortest paths to all others they are all closed to
anyone else and they are in excellent position to
monitor the information flow of the network.
Group 9 has all equal value at post collaboration
closeness centrality, thus allowing each member to

have equal access to all nodes in the network more
quickly.
Table 5.5 Closeness Centrality of Preliminary Group Collaboration
Table 5.5 Closeness Centrality of Post Group Collaboration
3.2 Qualitative Analysis

Another way to study the impact of collaborative
problem solving using social media is taking event
recall technique and interviewing the learner after
collaboration. The qualitative data were collected
via written students description, observation and
opinions based on the learner experience in the

study and upon showing them the table results
from the previous discussion. Their responses have
been re-analyzed until saturation has been reached.
The four themes have been identified in Table 5.
38
Table 5. Qualitative Themes
Theme
-Learning is fun, better, better, better! Thinking more!
-Memorable! Because of video, I remember because it is interactive and can repeat the simulation.
-We are free, we can say what we want, and get helped, No pressure.
-Support socialization, allowed us to chat, practiced English
Learning is fun, better, better, better! Thinking

more! Participants agree that e-learning provides
them an opportunity to study anywhere,
whenever scenario is fun. During collaboration
students can communicate with friends while
working on their problem. This is demonstrated by
out-messages where students greet one another
several time as a daily customary way in the Arab
world. The out-messages ware not taken as
negative in the threaded communications since it
builds rapport among learners and making their
collaborative more effective and relaxing.
Mabrouka said that Asalamalaikum, How
are you! The study is great. If my husband
will forbid us to come, we can still study at
home, in that case, we can continue
studying while supporting our family,I
am happy that I passed the exam, I learn
a lot from e-learning and love chatting with
my friends.
Rokiya agree that we should have asked
our classmates, especially the smart ones to
help us more in our study to pass our
courses, making our study better and
better, Oh I am happy that it shows I am
learning, I hope it will continue.
Mosbah on the other hand tell us that We
have more time to think, have more time to
solve problems than in the presence of our
instructor, through this, we will not ask
again and again. I am very hesitant from
the beginning to collaborate because I
dont know the other guy, but he became
my good friends and then I collaborate and
we solve the problem, I am very happy
with the study
These excerpts illustrate how the participants

enjoy the learning process. It can also be noted
that the exchange of messages increases from all
the participants signifying their interest, finish
their activity and submit on-time their output thus,
collaboration increases as shown in the social
network diagram. The Blooms taxonomy
collaborates the outcome of the study. Having fun

while learning increases significantly the cognitive
skills of the learner.
Memorable! Because of video, I remember
because it is interactive and can repeat the
simulation. Majority of the participants in each
group agreed that using interactive simulation and
videos
make
the
lessons
memorable.
Transforming this memorability of the lesson
makes the collaboration effective and continue
since students have basis to attack the problem at
hand. Without such understanding, group
performance during collaboration suffer.Sample
excerpts from the students;
Muftah says that The interactive
simulation somehow fill the gap of
communication problems., since it gives us
more time to process data rather than
human instructor. The video is very
good, while I am in chatting I am thinking
what I saw and I am sharing it with my
groupmates.
Entesor mentioned that It gives us idea to
solve the problem because of the video, it
provides us guide to solve the problem step
by step. I am happy with the results, my
group mates asked me to see the
computation of time complexity but I dont
have it so I asked in FB, I managed it.
Hamed commented further that I can
play several times and then perform by
myself without the instructor supervision
because I remember the step. During
collaboration, I was able to respond to
some of my groupmates because I
remember the step from simulation.
Thus, simulations and visualizations tools make it

possible for student to bridge experience and
remember better the lessons presented to them. It
has been observed that students remember the
concepts make them actively participate in the
39
collaboration. The inclusion of interactive
simulations in the e-learning improve the quality
and outcomes of the study. By knowing the
concepts, learners actively collaborate and
negotiate among themselves as shown in the social
network analysis.
We are free, we can say what we want, and get
helped, No pressure.- The pouring of in and outmessages during collaboration relate this, they
have freedom to express their opinions in the
social media without limitations. Many students
try to communicate with other learners in and out
of their circle to seek help. Students can
collaborate without the pressure and watchful eye
of their instructor. This is illustrated through the
following excerpts:
Fatma claimed that Libya is free so are
we to express our opinion, I studied in my
house and chat with my friends, it make
sense since we can tell to our parents that
we are studying instead of just searching
and chatting non-sense, I learned a lot so
it is not surprising that I have a good mark
in the exam.
Rowida also said that We can say what
we want, seek help to our friends, and
collaborate with the groups. I get some
inputs and explanation from the other
group, and it helps me a lot. Sometimes we
have a good output if no pressure is given
to us, just ample time to solve the
problem.
With this line of reasoning, students viewed the

social media as environment to express their
opinions and seek help with other learners. The
increase of coded messages could be attributed
with this thinking. Although it is good for the time
being, this attitude could lead to spoon feeding
type of learning, where learner could just request
from others and instantly provides an answer to
problems. An indications we need to watch and
address in the next papers.
Support socialization, allowed us to chat and
practice English Adapting social media
definitely increase socialization. Socialization is a
learning perspective that needs to be address and
include in any on-line education suitable and
essential in the area of collaboration. Construction

of knowledge takes place in a social context such
as this study where students were asked to solve
problem collaboratively, thereby increasing their
cognitive and affective skills. In the case of Saber,
the most sociable keeps the group communicate
while Mosbah is constructing or organizing the
solution .
Mosbah said Saber can communicate well
because he knows how to express in English
like me, but lack the ability to start solving
the problem. Luckily, he helps me keeps the
participants to continue the communication
process. He communicates and socializes
well. FB is very powerful, it help us a lot to
communicate.
Saber said Mosbah is good and very
serious, so I open topics to other while we
are waiting for his comments and suggestion.
But there is a serious problem in
communication for others, They should take
more time to chat and practice English.
Aisha mentioned that We have to practice
more English to communicate with other
learners, members who are fluent in English
tends to dominates discussion and we need
to deal with that. Perhaps I need to chat more
with Nadia.
The more in and out messages the more

communication takes place and this can be viewed
as a way of practicing English. In the study, we
have slightly relate the impact of threaded social
media as a medium of practicing English language
for computer science students, students are strictly
instructed to use English as medium of
communication. Many messages have been
discarded because it was written in Arabic. The
impact of social media allows students to
socialize, a factor that can never be removed in
ideal educational setting. The power of
socialization in the collaboration allows student to
negotiate, criticize their own and others students
contributions, ask for explanations and counter
arguments if necessary and in doing so, learners
modify and develop their own learning process.
40
4 CONCLUSIONS
This paper has presented the results of an approach
to content analysis of messages exchanged during
collaborative problem solving of approximately
three weeks duration. This analysis has enable the
tentative identification of patters of individual and
group learning during the activity. The students
interactions have been analyzed using the
Veldhius-Diermanse schema and Anderson
schema and relate the results to group and
individual
performance.
The
cognitive
performance level of students increased
dramatically after collaboration.
The social network analysis likewise shows the
density of collaboration among learners. This
provides little insights into a key aspect of the
individual and group processes and argued that
there is a need to perform more analysis to
understand fully the richness of these learning
interaction. It also concluded that the more
threaded communication the better, but it doesnt
guarantee that such quantity is transformed into
effective collaboration. The out-messages for
example are communication but not related to the
problem at hand and yet, it was concluded that this
provides informal communication and support
socialization.
In the interview, four themes have been selected
based on the written recall of events, experiences
and observed during the study. This recall has the
potential to access aspects of learning that are not
directly available in discussion or transcripts. But,
having many messages, performing saturation or
filtering is difficult. In summary, the study
successfully reported the used of event recall
among participants, the used of several content
coding analysis and the used social network
analysis.
2.
3.
4.
5.
6.
7.
8.
9.
5 REFERENCES
10.
1. Ballera, Melvin and Aziza Ehmaid Omar. ( 2012).
Exploring Social Networking Technology and
Multiple Pedagogical Agents: How, When and to
What Extent they Facilitate Learning in E-Learning
System. Proceedings of 2012 International
11.
Conference on Information Communication

Technologies in Education (ICICTE 2012). Rhodes
Island, Greece. July 2-5, 2012.
Veletsianos, G (2012). How do learners respond to
pedagogical agents that deliver social-oriented nontask messages? Impact on student learning,
perceptions, and experiences. Computers in Human
Behavior 28, 275-283.
Sanjaya S. and S. Pramsane (2008). Providing the
eLearning Services for the Blog, Special Issue of
the International Journal of the Computer, the
Internet and Management (IJCIM), Vol. 16
No.SP3, Assumption University of Thailand,
December, 2008.
Reategui, E., Polonia, E., Roland. (2007). The role
of animated pedagogical agents in scenario-based
language elearning: a case study. International
Conference on Computer Aided Learning (ICL
2007) . September 26-28, Villach, Austria.
Ballera, M., Musa. A. (2011). Personlized
eLearning System using Three Parameters and
Genetic Algorithms Proceedings of Society for
Information Technology & Teacher Education
International Conference, -March 7, 2011 (pp 569574) Nashville, Tennessee, USA.
Huang M-J., M-Y. Chen and S-C. Lee, (2007).
Integrating data mining with case-based reasoning
for chronic diseases prognosis and diagnosis,
Expert Systems with Applications, 32, 856867.
Ballera, M., Elssaedi, M. M. (2012). Incorporating
Social Oriented Agent and Interactive Simulation
in E-learning: Impact on Learning, Perceptions,
Experiences to non-Native English Students. In
Proceedings of World Conference on E-Learning
in Corporate, Healthcare and Education 2012 (pp.
495-503). Chesapeake, VA: AACE.
Doering, A., Veletsianos, G., & Yerasimou, T.
(2008). Conversational agents and their
longitudinal affordances on communication and
interaction. Journal of Interactive Learning
Research, 19(2), 251270.
Ballera. M. Elssaedi, M.M., Zhody, A.K. (2013).
Collaborative e-Learning with Multiple Imaginary
co-Learner: Design Issues and Implementation.
Proceedings of International Conference on
Information and Technology(ACIT2013), Zurich,
Switzerland. January 14-15.
Brindley, J., Walti, C., Blaschke, L. M. (2009).
Creating Effective Collaborative Learning Groups
in an Online Environment. International Review
of Research in Open and Distance Learning.
Meister, J.C., & Willyerd, K. (2010). The 2020
workplace: How innovative companies attract,
41
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
develop, and keep tomorrows employees today.

New York: HarperCollins Publishers.
Shaw, S. (2006). New reality: Workplace
collaboration is crucial. Eedo Knowledgeware
Whitepaper.
Lavonen, Jari; Autio, Ossi; Meisalo, Veijo (2004).
Creative and Collaborative Problem Solving in
Technology Education: A case Study in Primary
School Teacher Education. The Journal of
Technology Studies. 107115.
Chapman, C. Ramondt, L., &Smiley, G. (2005).
Strong Community, Deep Learning: Exploring the
Link. Innovations in Education and Teaching
International, 47(3), 217-230.
Brookfield, S. D. (1995). Becoming Critically
Reflective Teacher. San Francisco: Jossey-Bass.
Gaytan, J., & McEwen, B. (2007). Effective
online instructional and assessment strategies.
American Jour-nal of Distance Education, 21(3),
117-132.
Palloff, R. M., Pratt, K. (2005). Collaborative
Online: Learning Together in Community. San
Francisco: Jossey-Bass.
Chen, P., Gonyea, R., Kuh, G. (2008). Learning at
a distance: Engaged or not? Innovate, 4(3).
Springer, L., Stanne, M.E., & Donovan, S.S.
(1999). Effects of small-group learning on
undergraduates
in
science,
mathematics,
engineering, and technology: A meta-analysis.
Review of Educational Research, 69(1), 21-51.
Hennesy, S., Murphy, P. (1999). The potential for
collaborative problem solving in design and
technology. International Journal of Technology
and Design Education, 9(1), 1-36.
Smith, G.G., Sorensen, C., Gump, A., Heindel,
A.J., Caris, M., & Martinez, C.D. (2011).
Overcoming stu-dent resistance to group work:
Online versus face-to-face. The Internet and Higher
Education, 14(2), 121-128.
Finegold, A., & Cooke, L. (2006). Exploring the
attitudes, experiences and dynamics of interaction
in online groups. The Internet and Higher
Education, 9(3), 201-215.
Valetsianos, G. (2010). Contextually relevant
pedagogical
agents:
Visual
appearance,
stereotypes, and first impressions and their impact
on learning. Computer and Education, 55(2), 576283.
Fan, Sandra. (2010). Roles in Online Collaborative
Problem Solving. IEEE Symposium on Visual
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
Languages and Human-Centric Computing, 265266.

Gowers, T. (2009). Can Polymath be Scaled up?
Atwell, Graham. (2007). The Personal learning
Environments the future of eLearning? eLearning
Papers, vol. 2. ISSN 1887-1542.
Hue, K. (2011). Students and teachers use of
Facebook., Computers in Human Behavior, 27,
662-676.
Junco, R. (2012). The relationship between
frequency of Facebook use, participation in
Facebook activities, and student engagement.
Computer and Education, 58(1), 162-171.
Huxham, M., & Land, R. (2000). Assigning
students in group work projects. Can we do better
than ran-dom? Innovations in Education &
Training International, 37(1), 17-22.
Roberts, T.S., & McInnerney, J.M. (2007). Seven
problems of online group learning (and their
solutions). Educational Technology & Society,
10(4), 257-268.
Whitton, Diana (2000). Revised to Blooms
Taxonomy. Retrieved June 10, 2012, from
http://www.nwlink.com/~donclark/hrd/bloom.html.
Freeman, L. 2000, Visualizing Social Networks.
Journal of Social Structure', vol. 1, no. 1.
http://www.cmu.edu/joss/content/articles/volume1/
Freeman.html
Newman, M. J. (2005), A Measure of Betweenness
Centrality Based on Random Walks, 27, pp. 3954.
Opsahl, T., Agneessens, F., Skvoretz, J. (2010).
Node
centrality
in
weighted
networks:
Generalizing Degree and Shortest Paths. Social
Networks 32 (3): 245.
Veldhuis-Diermanse, A. E. (2002).CSClearning?
Participation, learning activities and knowledge
construction in computer-supported collaborative
learning in higher education. Wagenigen: Grafisch
Service Centrum Van Gils.
Anderson, T., Rourke, L., Garrison, D.R., Archer,
W. (2001). Assessing teaching presence in a
computer conference context. Journal of
Asynchronous Learning Networks, 5(2), 1-17.
42
Impact of a Form of Online Materials on the Quality of Education

A Case Study
Blanka Frydrychova Klimova and Petra Poulova
University of Hradec Kralove
Rokitanskeho 62, Hradec Kralove, 500 03, Czech Republic
Blanka.Klimova@uhk.cz, Petra.Poulova@uhk.cz
ABSTRACT
With a massive arrival of information
technologies in the past quarter century there
has been a change of the set processes in the
society. These changes concerns also
education. Online courses are nowadays
well established forms of learning and
teaching all over the world. They are used as
supporting courses of traditional classes, as
complementary courses of hybrid or blended
courses or pure online courses. Although
they cannot fully replace a teacher, they are
now becoming a powerful tool in the
educational process, both formal and
informal. Nevertheless, in all cases they
must be designed well to attract and
motivate students to work and study the
uploaded information. Therefore, this article
focuses on the analysis of online materials
and their impact on the process of learning.
KEYWORDS
eLearning, Internet, Learning Management
System, online courses, online material,
learning, survey.
1 INTRODUCTION
In the Czech Republic there are
nowadays 26 public universities, two
state universities and more than 40
private universities. The Faculty of
Informatics and Management (FIM) is
one of the five faculties of the University
of Hradec Kralove educating specialists
in the area of informatics, economics,
management and management of

tourism. Likewise at the other
universities, also at FIM in the past 20
years there has been a increased
influence
of
information
and
communication technology (ICT) on
different areas of faculty life. FIM
presents itself with the help of its web
and facebook pages; it uses an
information system accessible from the
internet for the support of study
administration; it publishes an electronic
newsletter for undergraduate and
graduate students; besides classic books
and journals students have an access to
electronic information sources; or it uses
an electronic information system for the
support of economic processes. Faculty
has been intensely involved in the
application of ICT in form of eLearning
since 1997 as the teachers see this as a
way of improving the quality of their
teaching. They also discover that
eLearning
contributes
towards
increasing the
effectiveness and
efficiency of the educational process.
Moreover,
it
enhances
learners
autonomy. This was a gradual process.
In the first phases only a few teachers
used electronic courses for the courses of
further education. Since 2002 more and
more teachers have got involved into the
creation of eLearning courses and they
have been designing online courses for
part-time and full-time students in a
virtual learning environment (Learning
43
Management System) called Blackboard
Learn .
At present, more than 220 e-courses are
offered, with more than 45 of these
being foreign language courses. Several
of them can be taught completely online.
However, most of them are blended as
there are usually a few tutorials. At
tutorials (face-to-face teaching) students
usually discuss the problems they come
across when doing different types of
tasks or writing assignments. Moreover,
sometimes there are regular classes and
e-courses are used as reference courses
(i.e. students can once again read the
information obtained in class and do
some additional exercises to practise
their knowledge) for further self-study or
revision of the lecture. For more
information see Frydrychov Klmov &
Poulov [1].
The creation of online courses is, of
course, very demanding for teachers and
therefore a number of trainings focused
on the use of ICT, course design in the
virtual learning environment, and also on
the methodology of online materials
design from the pedagogical and didactic
point of view were prepared for them in
the past years. The eLearning course
should provide students with information
needed for their successful study in the
form of study materials. Furthermore, it
should enable students to verify the
acquired knowledge with the help of a
number of tests, provide them with a
simple tool for communication with a
teacher and other students and a
possibility to submit and possibly revise
their assignments or term projects. In
addition, it should enable teachers to run
the educational process in an effective
way and help with record keeping.
Figure 1. Blackboard Learn at FIM
In the online courses students especially

exploit
study
materials
which
undoubtedly play a crucial role in the
acquisition of their knowledge and skills.
Moreover, not much hands-on research
has been done in this field yet. However,
if the online courses are developed well,
they definitely motivate students in their
self-study.
Semradova [2], for example, outlines the
following nine key prerequisites of
effective online course materials:
The learning material has to be clear
and accessible.
It is important to collaborate with
professionals.
One should be aware of the dangers
and pitfalls of selective perception,
selective processing and selective
memory.
It is necessary to think carefully
about the content and organization
of materials.
One should keep in mind the issue
of "cognitive dissonance" - the
source of which is a tension between
a wide choice, the need to deal with
what makes us tired, bored,
44
burdensome, and it is difficult to
understand, respectively, on the
contrary, it seems trivial and
obvious.
In the preparation for teaching some
of the less frequent motifs, based on
the teachers need to be curious, to
discover, to be playful and
constructive, the need for security,
comfort, contact, the desire for
autonomy, and the need to manage
time ...
The learning texts should be
effective; they should include
headlines and subheading in order to
be well arranged, and clearly
structured.
Study materials should respect the
hermeneutic approach.
Courses and learning materials
should be an intellectual challenge.
For a detailed description of the creation
of any online study materials see [3] or
[4].
2 RESEARCH AND ITS FINDINGS
In order to find the latest attitude of FIM
students to the study materials, the
authors made a survey in the winter
semester of 2011. The survey
concentrated on the eLearning form of
study with respect to the study materials.
A form of electronic questionnaire was
chosen for this research.
Overall, 484 students participated in the
survey. 66% were men and 34% women
(Fig. 2), which in fact reflects the nature
of the faculty since most respondents
were from the field of informatics:
applied informatics (AI), information
management (IM) and information and
knowledge management (IZM) - 360,
other respondents studied financial
management (FM 27), sport
management
(SM
17)
and
management of tourism (MCR 56)
see Fig. 3.
Figure 2. Respondents sex
Figure 3. Respondents field of study
Altogether 2400 students are enrolled at

the faculty. That means that more than
20% of students participated in the
research.
The structure of respondents group
followed the structure of students
according to their fields of study. There
was just a slight majority of students in
the field of information management.
This higher ration was probably caused
by the topic of the survey and the chosen
form of questionnaire. Just this very
group of future informatics uses the
online study materials most often.
An average age of the respondents was
22 years 316 students (see Fig. 4). In
accordance with the students age
structure, the respondents age ranged
45
from 19 to 41. More than 85% of
respondents were under 25 years old.
would rather have the online classes only

(see Fig. 6).
Figure 4. Respondents age
Figure 6. Preferred classes
Students usually have enough experience

in working with e-subjects because 191
of them admitted studying 2-5 e-courses
and 104 of them even worked in more
than 10 e-courses. Only 20% of
respondents studied just one subject with
online materials. (See Fig. 5.)
The survey confirmed that the most

exploited eLearning tools are study
materials (459 respondents), followed by
assignments, quizzes, and self-tests (180
respondents), then by a syllabus and
study objectives (144 respondents) and
the last group formed communication
tools (91 respondents). See Fig. 7.
Figure 5. Experience in working with e-subjects

Figure 7. eLearning tools and their exploitation
95% of respondents were composed of

full-time students, 5% part-time a 65%
were students of bachelor study
programmes, 30% students Master study
programmes and 5% were doctoral
students.
Moreover, some students added their

comments to this question. Their
opinions on the exploitation of the
eLearning tools were both positive and
negative.
Although the students of informatics

fields were prevalent in the survey, the
survey revealed that most respondents
(58%) preferred the blended classes to
the traditional, face-to-face classes
(33%) and only 9% of respondents
Positive:
I am (generally) satisfied.
It would be good to introduce more
e-subjects.
It depends, but generally, I am
satisfied.
46
I am satisfied although it is not true

for a few e-subjects.
I use them for my exam preparation
and so far it has been enough.
Negative:
The Blackboard is not working
properly at the moment.
Some
e-materials
could
be
structured in a better way.
Study materials are rarely balanced.
Everyone works with the electronic
materials differently, both students
and teachers. The structure of
lectures is sometimes not clear,
interesting, following only the
printed version. As for the
presentations, rules of a good
presentation are often broken.
Moreover, some teachers do not
know how to work with the tools
available to them, which is again at
the expense of quality, not
mentioning the outdated material.
The quality of individual lectures of
individual courses sometimes differs
but most of them are OK.
I would welcome more examples or
step-by-step tutorials.
There could be fewer learning
materials but it is more important to
get students involved and motivated
in these materials.
In most cases students evaluate the
online
materials
positively,
but
sometimes with some reservations, such
as:
Yes and no some materials are too
concise, other are too extensive. Only in
a few cases the study materials are
balanced. But in most cases I am
satisfied.
An overwhelming majority of students
(96%) found the electronic texts easy to
read in comparison with a dissatisfied

minority (4%). See Fig. 8.
Figure 8. Is it easy for you to read the online

texts?
The main reasons of respondents

dissatisfaction can be listed as follows:
complexity of study materials;
more information in bullets is
needed;
often a chaotic storehouse of
documents;
the texts often contain abbreviations
which are not explained;
the texts are sometimes more
difficult to understand.
Furthermore, 93% of respondents were
satisfied with the content structure of
their electronic materials and only 7% of
respondents were not. (See Fig. 9.)
Figure 9. Satisfaction with the content structure

of electronic materials
287 students also answered question

What do you like in your written
electronic materials most? Majority of
47
replies (almost 50%) concerned their
accessibility. Students really appreciate
accessing the study materials any time
and anywhere, most preferably, from the
cosiness of their homes. Other positive
factors of the electronic materials are as
follows:
an easy access to study materials;
a
possibility
of
copying,
downloading and printing;
low/no costs of learning materials;
ecological approach;
well-written and structured;
a chance of self-study;
individual pace of study;
a chance to return to study materials
and complete ones learning`
self-tests, a chance to check ones
knowledge;
interaction;
no need to take notes during the
face-to-face lectures.
3 CONCLUSION
From the above described results of the
survey, it is clear that students welcome
an opportunity to work online although,
predominantly,
in
the
so-called
combined/blended/hybrid
form
of
learning. The reason is that they can
complete their knowledge and thus
finish their learning process, which they
did not manage to do during the face-toface classes. They are not forced to do it
immediately at school or in the library
because they can access the online
course from the cosiness of their homes
any time they feel like that. In addition,
if they did not understand anything
during the lecture, they can contact their
teacher online and ask him/her.
In the online course students mostly
exploit the study materials. Therefore,
tutors should pay careful attention to
their creation. As both theory and

practice (survey) reveal:
Study materials should have a clear,
concise, logical and simple structure
(information in bullets is preferred).
They should be well-balanced (i.e.
there should be an adequate amount
of
relevant
teaching
matter
including learning objectives and
exercises/assignments/self-tests).
The
materials
should
be
comprehensible and up-to-date.
They should be easily navigated.
They should be interactive with
appropriate multimedia components.
They should be linked to other
suitable materials and relevant
websites.
Thus, Frydrychova Klimova [5] provides
the following simple framework for the
creation of any topic-based study
materials, be they online or traditional,
i.e. textbooks:
Maximally a two page document
consisting of the following items should
be developed:
topic (a concise sentence or a phrase
of the lesson content);
learning goal (a short statement
motivating the participants to study
the particular lesson); prerequisites
(previous knowledge required to
master the lesson);
skills (a description of the
knowledge/ skills to be gained in the
particular lesson);
explanation of the basic concept and
ideas of the teaching matter
discussed in the lesson (in the form
of text and questions);
conclusion with self-tests, tasks,
quizzes (with keys), or an
assignment; and
bibliographical sources and or links
to them.
48
If such a material is developed, then both
teaching and learning is a challenge.
Moreover, such written educational
materials might increase students
interest in online studies, respectively
his/her motivation for self-study.
4 ACKNOWLEDGEMENT
The paper is supported by the project
Excellence, no. 2216: The ICT
reflection within the cognitive processes
development .
5 REFERENCES
1.
2.
3.
4.
5.
Frydrychova Klimova, B. & Poulova, P.

(2012). Reflection on the development of
eLearning in the Czech Republic.
Proceedings of the 16th WSEAS
International Conference on Recent
Researches in Communications and
Computers. Greece: Corfu, 433-437.
Semradova, I. et al. 2010. Reflections on the
exploitation of a virtual study environment.
Hradec
Kralove: MILOS VOGNAR
Publishing House.
Branch, R.M., Dohun, K. & Koenecke, L.
(1999). A checklist for evaluating online
materials. Retrieved March 26, 2011, from
http://www.libraryinstruction.com/evaluatin
g.html.
Wright, R.C. Criteria for evaluating the
quality of
online courses. Retrieved
November
22,
2012,
from
http://elearning.typepad.com/thelearnedman/
ID/
evaluatingcourses.pdf.
Frydrychova Klimova, B. (2012). Teaching
formal written English. UHK: Gaudeamus.
49
SEPARATION IN DATA MINING BASED ON FRACTAL NATURE

OF DATA
Marcel Ji ina1 and Marcel Ji ina, Jr.2
1
Institute of Computer Science AS CR

Pod Vodrenskou v 2,
Prague 8, 18207 Czech Republic
marcel@cs.cas.cz
Faculty of Biomedical Engineering, Czech Technical University in Prague

Nm. Stn 3105, 272 01, Kladno,
Czech Republic
jirina@fbmi.cvut.cz
ABSTRACT
The separation of the searched data from the
rest is an important task in data mining. Three
separation/classification
methods
are
presented. We use a singularity exponent in
classifiers that are based on distances of
patterns to a given (classified) pattern. The
approximation of so called probability
distribution mapping function of the
distribution of points from the viewpoint of
distances from a given point in the form of a
scaling exponent power of a distance is
presented together with a way how to state it.
Considering data as points in a metric space,
three methods are based on transformed
distances of neighbors of a given point in a
multidimensional space via functions that use
different estimates of scaling exponent.
Classifiers data separators utilizing
knowledge about explored data distribution in
a space and suggested expressions of the
scaling exponent are presented. Experimental
results on both synthetic and real-life data
show interesting behavior (classification
accuracy) of classifiers in comparison with
other well-known approaches.
KEYWORDS
Nearest neighbor; fractal set; multifractal; IINC
method; correlation dimension
1 INTRODUCTION
One task is that we have a large amount
of data after extensive search, but we are
interested only in their small portion and
we wish to separate them from the rest
with a minimal error. These tasks are also
rather common in particle physics where,
during measurements, an extensive
amount of data (events) is registered but
only their small portion contains
information which has not been
previously known. This separation task
can be generalized in such a way that
there are several classes of data in which
we need to sort them.
Often, in classification problems, the only
known fact is the learning set, i.e. the set
of points each of known class. The
problem is how to estimate the probability
to which class a query point x of the data
space belongs. The different approaches
to the classification can be divided into
parametric and nonparametric methods.
Parametric methods include neural
networks of different kinds [6], decision
trees or forests [2], also known as the
CART method, the support vector
machine SVM [1] and many more.
50
Nonparametric methods are mostly based
on the Bayesian approach [6], [10], and
the k nearest neighbors (k-NN) method
[5], [6], [24].
Comparing
different
classification
methods applied to different data sets, one
can find a very interesting fact that
sometimes a simple approach outperforms
the other methods including those which
are very sophisticated. This is the case of
the 1-NN method and the k-NN method,
for example. These methods are
thoroughly
elaborated since Cover
and Hart [5], and many variants have
appeared up to now, see e.g. [8], [24].
Simple methods like 1-NN can be
improved by some preprocessing, in fact,
by some modification of the learning set.
Best known are bagging and bootstrap
approaches. Other approaches set some
weights to learning set samples or to
individual
features
(coordinates),
eventually modify features of individual
samples (move samples) to get better
results [8], [24]. Here we describe new
algorithms in a pure form without any
preprocessing.
Classifiers directly or indirectly take the
distribution of data points around a given
query point into account. To express the
distribution of points from the viewpoint
of distances from a given point, the
probability distribution mapping function
was introduced [16]. The approximation
of this function in the form of a suitable
power of the distance is presented. How
to state this power the distribution
mapping exponent is described. This
exponent is used for probability density
estimation in high-dimensional spaces
and for classification. It will be seen that a
core notion in the transformation
mentioned above is a slightly redefined
singularity or scaling exponent to fit the
notion of distance between points. The
scaling considered here is related to
distances between pairs of points in a

multivariate space. Thus, it is closer to the
correlation dimension by Grassberger and
Procaccia [11] than to box-counting or
other fractal or multifractal dimension
definitions [19], [26].
2 MULTIDIMENSIONAL DATA
2.1 Data Space
Let the learning set U of total N samples
be given. Each sample xt={xt1, xt2, xtn};
t = 1, 2, ... N , xtk R ; k = 1, 2, ..., n
corresponds to a point in n-dimensional
metric space Mn, where n is the sample
space dimension. For each xt U, class
function T: Rn {1, 2, ... C}: T(xt) = c is
introduced. With the class function the
learning set U is decomposed into disjoint
classes Uc = {xt U | T(xt) = c}; c {1,
2, ..., C}, Cc=1 U c , Uc Ud = ; c, d
{1, 2, ..., C}; c d. Let the cardinality of
C
set Uc be Nc ; c =1 N c = N .
2.2 Indexing Data
As we need to express which sample is
closer or further from some given point x
(the query point), we can rank points of
the learning set according to distance ri of
point xi from point x. Therefore, let points
of U be indexed (ranked) so that for any
two points xi, xj U there is i < j if
ri < rj; i, j = 1, 2, ... N, and class
Uc = {xi U | T(xi) = c}. Of course,
ranking depends on point x and eventually
metrics of Mn. We use Euclidean (L2) and
absolute (Manhattan, L1) metrics here. In
the following, indexing by i means the
ranking just introduced.
51
3 BASIC TYPES OF CLASSIFIERS
Classifiers are based on several
principles:
A cut of distributions of individual
features (variables) of a pattern; this is
rather close to the Bayes classifier,
used in physics.
Bayes classifier. For each variable
empirical distribution densities (histograms) are estimated for each class.
For a query pattern with unknown
class, the corresponding density is
estimated from empirical or empiricbased histogram for each feature.
Then, for each class a product of these
densities is computed and the largest
product corresponds to the class to
which the pattern belongs.
Distance based classifiers. These
classifiers need the pattern space be a
metric space. Therefore,
some
transformation is often used and
metrics introduced. The simplest is
the well-known 1-NN (nearest
neighbor) classifier that associates the
class of its nearest neighbor to the
query pattern. Several (k) nearest
neighbors (k-NN) can be used;
sophisticated variants use a special,
even adaptive metrics.
Cluster-based classifiers try finding
clusters according to given criteria,
and in a cluster, a threshold is used for
classification as in Random Forests.
Kernel-based classifiers first identify
parameters of kernels and then
estimate a class according to common
densities.
GMDH classifiers and approximators
that are special forms of neural
networks where neurons have two
inputs only and each neurons
response is a full quadratic
polynomial of two input signals [22].
A distance-based kind of classifiers (a

metrics distance is introduced) that
can also be considered as kernel
classifiers;
Use two observations:
1. Multidimensional data are fractal,
even multifractal (nonfractal data
are totally random data in the data
space or in its subspace).
2. Distance-based fractal measure is
the correlation dimension.
4 CORRELATION DIMENSION
The correlation integral, in fact, a
distribution function of distances between
all pairs of points in a set of points in a
space with a distance, was introduced in
[11]. Correlation integral CI(r) is defined
by
CI (r ) = limN
1
{number of pairs (i, j) : X i X j < r}
N2
.
In a more comprehensive form one can
write
C I ( r ) = Pr( X i X j < r ) .
Grassberger and Procaccia [11] have
shown that for small r the CI(r) grows like
power C I ( r ) ~ r and that correlation
exponent
can be taken as the most
useful measure of the local structure of
the strange attractor. This measure allows
distinguishing between deterministic
chaos and random noise.
The correlation integral can be rewritten
in the form
CI (r ) = limN
2
h(r X j X i ) ,
N ( N 1) 1i< j N
(1)
where h(.) is Heaviside step function, and

considering all N(N-1)/2 pairs of points of
set of N points. From it
= lim r
ln C I ( r )
ln r
Our class of classifiers is:

52
Thus, if the correlation integral as a
function of r is depicted in log log
coordinates, there appears a nearly
straight line with a slope equal to
correlation dimension v. This was shown
in [11] and the method based on this fact
is
called
Grassberger-Procaccias
estimator of v. There are other methods
for estimation of correlation dimension .
One of the most cited is Takenss (viz lit.
odkaz 27) estimator [27].
5 DISTRIBUTION MAPPING
FUNCTION
Here we single out
three notions
introduced or used in [14]-[15]. The
(DMF) is a mapping of the probability
distribution of points in n-dimensional
space to the distribution of points in onedimensional space of the distances. The
distribution density mapping function
(DDMF) is a one-dimensional analogy to
the probability density function. The
power approximation of the probability
distribution mapping function in the form
of (distance)q is introduced, where we call
the exponent q the distribution mapping
exponent (DME).
These notions are local, i.e. are related to
a particular (query) point. We show that
the distribution mapping exponent q is
something like a local value of the
correlation dimension according to
Grassberger and Procaccia, [11]. It can be
also viewed as the local dimension of the
attractor by Froehling [9] or singularity
eventually scaling exponent (exponent)
in the sense of Stanley and Melkin, [26].
To be more exact, let us introduce two
definitions
to study a probability
distribution of points (patterns) in the
neighborhood of a query point x in
n-dimensional Euclidean space En.
Definition 1. Let the probability

distribution mapping function D(x, r) of
the query point x in En be
function D( x, r ) = p( z )dz , where p(z) is
B ( x,r )
the probability density of the points at z; r

is the distance from the query point x and
B(x, r) is the ball with center x and radius
r.
Definition 2. Let the distribution
density mapping function d(x, r) of the
query point x in En be function
d ( x, r ) =
D ( x, r ) , where D(x, r) is a
r

of the query point x with radius r.
Let us try to transform the true
distribution of points so that the
distribution density mapping function is
constant, at least in the neighborhood of
the query point.
Definition 3. The power approximation
of the probability distribution mapping
function D(x, r) is the function rq such
that
D ( x, r )
rq
const
for r
0+ . The
exponent q is the distribution mapping

exponent.
These definitions were published and
used elsewhere [14]-[15].
Intuitively, Fig. 1 gives illustration of
mapping functions; Pure means
theoretical dependence, True means
real data. From top to bottom:
In Fig. 1a the dependence of order of near
point on its distance from the query point
is shown. Note that the (first) nearest
neighbor is rather far from the query point
and that distances of farther points from
the query point differ little. Positions of
furthest several points are influenced by
boundary effect.
In Fig. 1b the same dependence is shown
in logarithmic scale; linear dependence
53
appears here. Note the slope that is
q = 2.5.
Fig. 1. Illustration of mapping functions; Pure

means theoretical dependence, True means real
data. From top to bottom:
a. Dependence of order of near point on its
distance from the query point.
b. The same dependence in logarithmic scale.
c. Dependence of percentage of points (instead
of count) on distance to the q-th power.
d. Density (histogram) of variable z = rq
confirms its uniform distribution.
In Fig. 1c the dependence of percentage

of points (instead of count) on distance to
the q-th power is shown. In fact, this
graph is the distribution function of
variable z = rq. Again, linear dependence
appears. It means a uniform distribution
(with the exception of several farthest
points). It also appears that radii ri ,
i = 1, 2, grow proportionally to the q-th
q
root of index i ri i .
In Fig. 1d the density (histogram) of
variable z = rq confirms its uniform
distribution.
Exponent q is called the distribution
mapping
exponent
according
to
Definition 3 or referring to fractal
nature of data the scaling exponent as it
scales distances according to scaling law
z = rq .
We use this exponent in a classification
method.
54
6 DECOMPOSITION OF THE
CORRELATION INTEGRAL TO
LOCAL FUNCTIONS
The distribution mapping exponent
(DME) reminds one of the so-called
correlation dimension by Grassberger and
Procaccia [11], and corresponds to the
generally used definitions of power
scaling laws especially to singularity
exponent. It can be seen that the
correlation integral is a distribution
function of distances between all pairs of
points of the data points given. The
is a distribution function of the distances
from one fixed point x. In the case of
finite number of points N, there are N(N 1)/2 distances between pairs of points and
from them one can construct an empirical
correlation integral. Similarly, for each
point there are N - 1 distances and from
these N - 1 distances one can construct an
empirical
probability
distribution
mapping function. There are exactly N
such functions and the mean of these
functions gives the correlation integral.
This is also valid for N going to infinity.
Let us return to Eq. (1). In this section we
show that the correlation integral is the
mean of distribution mapping functions.
Theorem. [15] Let the correlation
integral, i.e. the probability distribution of
distances lij between all pairs of points
from the learning set, be CI(r) and let
D(xi, r), where lik is the distance of k-th
neighbor from point xi, be the distribution
mapping function corresponding to point
xi. Then, CI(r) is a mean value of D(xi, r):
N
C I ( r ) = lim
1
N N
D ( xi , r ) .
(2)
i =1
Proof.
Let h(x) be Heaviside step function. Then,
the correlation integral is
N N 1
CI (r ) = lim
1
N N ( N 1)
h(r lij )
i =1 j =1
and also
N 1
C I ( r ) = lim
1
N N
i =1
1
N 1
h ( r lij )
j =1
Comparing the last formula with (1), we

get directly (2).
Note the inner sum. It is the distribution
mapping function corresponding to the ith point.
7 THE EXPONENT IN A CLASSIFICATION METHOD
Informally, consider partial influences of
individual points to the probability that
point x is of class c. Each point of class c
in the neighborhood of point x adds a
little to the probability that point x is of
class c, where c = {0, 1} is the class mark.
This influence is the larger the closer to
point x the point considered is, and vice
versa.
From another point of view, let
Pr (T ( x) = c | T ( xi ) = c ) be the probability
that the given point x is of class c if
neighbor point number i is of the same
class as point x. Note that points of
learning set U are indexed so that for any
two points xi, xj U there is i < j if
ri < rj; i, j = 1, 2, ... N as mentioned in
Chap. 2.2. In the following, K is a
constant that is used to normalize the
probability that point x belongs to a class
to 1:
For the first (nearest) point i = 1
1
Pr (T ( x) = c | T ( x1 ) = c ) = K q ,
r1
for the second point i = 2
1
Pr (T ( x) = c | T ( x 2 ) = c ) = K q ,
r2
and so on, generally for point No. i
55
Pr (T ( x) = c | T ( xi ) = c ) = K
1
.
ri q
A question arises where do these
considerations come from? The answer is
simple:
Let us consider a (fixed) query point x.
Construct a ball B(x, r) with center x and
radius r. Let the volume of this ball be
V(x, r) and it holds V ( x, r ) = S n r n . Sn is a
constant
dependent
on
space
dimensionality n. Let there be i points in
B(x, ri). We are interested in the density
of points in the neighborhood of point x.
i
i
It
is
p ( x, ri ) =
=
.
V ( x, ri ) Kri n
Considering this density constant (in the
sense that we get the same density for
different values of i, i.e. for different radii
ri), radii ri would grow proportionally to
the n-th root of i, ri n i . But as
shown above the space can have
(locally
or
globally)
effective
dimensionality q lesser than n. Thus, we
should compute density of points using
the volume of q-dimensional ball using
i
i
formula p( x, ri ) =
=
, where
Vq ( x, ri ) S q ri q
Sq is a constant. Taking one isolated point

only in distance ri, i = 1, 2,..., we get
formulae for the first, second,... point as
above.
Individual points are independent; and
then we can sum up probabilities above.
Thus, we add the partial influences of k
individual points together by summing up
p (c | x) =
xi U c
Pr (T ( x) = c | T ( xi ) = c ) = K
1 / ri q
xi U c
.
(3)
Note that the sum goes over indices i for
which the corresponding samples of the
learning set are of class c, c = 1, 2, C,
where C is the number of classes. It can
be seen that any change of distance ri of
any point i from point x will influence the

probability that point x is of class c. It can
also be considered as a kernel method
with kernels ( xi , ri , c) = 1 / ri q located at
all points xi, i= 1, 2,.. in distances ri.
We can rewrite (1) in a more suitable
form for practical computation.
N
1 / ri q
p ( x | c) =
xi U c
N
i =1
1 / ri
(4)
The upper sum goes over indices i for

which the corresponding samples of the
learning set are of class c. At the same
time all N points of the learning set are
used instead of number k. Moreover, we
often do not use the nearest point (i = 1).
It can be found that its influence is more
negative than positive on the probability
estimate.
8 A THEOREM
Theorem 1
Let the task of classification into two
classes be a given. Let the size of the
learning set be N and let both classes have
the same number of samples. Let
, 1 < < n be the correlation dimension,
let i be the index of the i-th nearest
neighbor of point x (without respect to
class), and ri > 0 its distance from point x.
Then,
N
p (c | x) = lim
1 / ri
xi U c
N
i =1
(4)
1 / ri
is probability that point x belongs to class

c.
Proof [18]:
For each query point x one can state the
56
D(x, ri, c). We set this function so that it

holds (C is a constant)
D ( x, ri , c) = Cri
q
1 / ri q
lim
p (c | x ) =
x U c
N
lim
in the neighborhood of point x. Using

q
derivation, according to variable z = ri , we

get d ( x, ri , c) = C . By the use of z = ri , the
space is mapped (distorted) so that the
distribution density mapping function is
constant in the neighborhood of point x for
any particular distribution. The particular
distribution is characterized by the particular
value of the distribution mapping exponent q
at point x. In this mapping, the distribution of
the points of class c is uniform.
q
i =1
1 / ri q
and due to the same limit transition in the

numerator and in the denominator we can
rewrite it in form (5).
N
1
q
xi U c ri
is faster the larger DME q is. Usually, for
multivariate real-life data the DME is also
large (and the correlation dimension as
well).
Note that the convergence of S c =
Let us consider sum

i =1
d ( x, ri q , c) / ri q . For
this sum we have

N
lim
i =1
9 THREE METHODS
d ( x, ri q , c) / ri q = p(c | x) lim
N
i =1
1 / ri q
(6)
because d(x, ri , c) = d(x, z, c) = p(c|x) for
all i (uniform distribution has a constant
density).
Given the learning set, we have the space
around point x sampled by the
individual points of the learning set. Let
pc(ri) be an a-posteriori probability that
point i at distance ri from the query point
x is of the class c. Then, pc(ri) is equal to
1 if point xi is of class c and pc(ri) is
equal to zero, if not, i.e. if the point is of
the other class. Then, the particular
q
realization
of
p (c | x )
i =1
1 / ri q
is
sum
1 / ri q . Using this sum, we can
xi U c
rewrite (6) into the form

N
p(c | x) lim
i =1
1 / ri q = lim
1 / ri q
xi U c
Dividing this equation by the limit of the

sum on the left hand side, we get
Modifying considerations above, one can

find that there are three methods
according to what kind of data we
consider and the way we deal with ratio
1/rq.
1. If data are multifractal, then variable
q differs for different places of data
space: It is necessary to compute q for
each query point extra. The apparent
advantage is a more exact estimate
than in the global approach below
that uses, in fact, the mean of all qs
over the whole learning set.
2. For monofractal data, there is q = v; q
has a constant value in the data space
and equals to the correlation
dimension. The advantage is that we
need
to
compute
correlation
dimension v only once. There are
many methods leading to a rather
exact estimate of the correlation
dimension.
3. IINC (Inverted Indices of Neighbors
Classifier) is based on the observation
explained below. The method needs
no q or v estimation, on the other
hand, it uses sorting.
57
9.2.1 Sensitivity of classification error

to error in CD estimation
For error sensitivity to the value of
correlation dimension no particular
threshold was used. Instead, we use a
more general classification quality
measure here, the size of area under the
ROC (Receiver operating characteristics
[7]) curve (the AUC) of dependence of
sensitivity, i.e. the acceptance of class 1
samples (often called signal) on
specificity, i.e. on the suppression of
class 0 samples (often called background,
i.e. background error). It holds that the
larger the area under the curve (AUC,
classification efficiency) the better
classification in a general sense. The ideal
case is unit area, i.e. ROC curve going
through point (0, 1), which means 100 %
sensitivity (signal efficiency) and 100 %
specificity, i.e. zero background error.
0.5 4
0.5 2
0.5
A UC
9.1 Multifractal Data Classifier

This classifier was described in Chap. 7.
Its advantage that more exact estimate of
the distribution mapping exponent q may
appear as an disadvantage not only for
repeated computation of q for each query
point separately, but also for large error in
its estimate. It is given by a relatively
small amount of data especially in cases
of small learning data sets like in e.g. Iris
data [21] where there is the learning set of
total 150 samples. Fortunately, it can be
shown that the method is not too sensitive
to exactness of stating the q. In any case
one must check the estimated value of q
for cases where q is too close to data
space dimensionality n or even larger than
n, or too close to zero. In these cases it is
advisable to use q estimated for another
point already computed and near to the
query point or to compute q for several
near points, exclude cases now discussed
and take a mean. Many similar
approaches can be designed including
estimate of correlation dimension for
small subset of the learning set consisting
of points near to the query point.
0.4 8
0.4 6
0.4 4
9.2 Monofractal Data Classifier

Beside the advantage that we need to
compute correlation dimension v only
once, there are two problems related to it.
First, the correlation dimension estimation
can be time-consuming for large learning
data sets because one must compute all
N(N-1)/2 distances between points of all
pairs of points of the learning set. Second,
do we know if our data are monofractal?
To answer these questions we show here
the sensitivity of classification error to
error in correlation dimension estimation,
and the fact that the spread of the
distribution
mapping
exponent
is
relatively small.
0.4 2
0.4
0
10
15
20
25
Estimat ed v alue q o f th e correlatio n dim ension v
Fig. 2. General classification efficiency AUC as

the function of the estimated value q of
correlation dimension for data Higgs. Data
dimensionality is 23, note that 6.5 is value of the
correlation dimension, i.e. the optimal value of q.
In Fig. 2 the classification efficiency as

the function of the value q of estimated
correlation dimension for data Higgs
is shown. These data have estimated
correlation dimension = 6.5. For this
value of q, the minimal error is 0.472, the
value of error is 0.481 for q = n = 23. This
58
is rather a small difference, only 1.93 %,
showing that error in correlation
dimension estimate need not be critical.
9.2.2 The correlation dimension and

spread of the DME
In Table 1 and Fig. 4 features of six
different data sets and corresponding
distribution-mapping
exponent
are
summarized. Data sets originate from UCI
Machine Learning Repository [21]. Note
that mean distribution-mapping exponent
is, in fact, the correlation dimension. It
can be seen that
Mean DME (in fact, an estimate of
correlation dimension) is much smaller

than dimension for all data varying from
a little more than 6.2 % (data Ionosphere)
to nearly 49.5 % (data RKB).
DME of a data set lies in a rather narrow
band; normalized mean squared variation,
sigma/mean, / varies from 7.357 % to
less than 19 %.
Note that lines for Heart, German, and
Higgs data look suspiciously similar but
these data come from very different
independent sources.
Table 1.
Parameters of DME distribution for different data sets
and color notation for Fig. 3. Data sets are from UCI
MLR.
mean
Dimen mean sigma/ DME/
sion
DME Mean dim.
Data
Color
Higgs
Red
23
1.75 0.098 0.076
German
Heart
Adult
RKB
Ionosphere
Aquam.
Violet
Green
Blue
Coral
20
13
14
10
33
2.71
2.42
5.28
4.94
2.06
0.079
0.074
0.158
0.185
0.139
0.136
0.186
0.377
0.495
0.062
Fig. 3. Histograms of distribution mapping exponent for

six different data sets. Histograms are normalized by
mean value of the DME and have a unit area. For legend
see Table 1.
Here we can conclude that data are

generally multifractals as the scaling
exponent, DME varies from point to point
of the set. On the other hand, these
variations usually lie in a rather narrow
band and thus mean DME, i.e. the
correlation dimension, may suffice for
characterization of the fractal nature of
data.
9.3
IINC Inverted Indices of

Neighbors Classifier
This is a modification of the first
(multifractal) method above; based on
observation that the distribution mapping
function grows linearly in log(count)log(distance) coordinates with slope q and
thus linearly in coordinates count vs.
(distance)q , as shown in Fig. 1c.
So, there the index of the neighbor i is
proportional to riq, i~ riq. Then, 1/riq ~ 1/i.
From this follows that we can use the
neighbors index i instead of its distance
to the q-th (or v-th) power, riq . Rewriting
(3), we get
p (c | x ) =
xi U c
Pr (T ( x ) = c | T ( x i ) = c ) = K
p (c | x ) = K
1 / ri q
xi U c
1/ i
xU c
59
This approach is very simple and very
effective; no estimate of the singularity
(fractal) exponent is needed.
9.3.1 Kernel method

All three approaches, and especially this
last one, are suggestive of kernel methods
but with a rather strange kernel. A kernel
[6], [13] is a non-negative real-valued
integrable function K satisfying
K (u )du = 1 and K(-u) = K(u) for all u.
(7)
If x1, x2, xN are independent and
identically distributed samples of a
random variable with probability density
function
f,
then kernel
density
approximation of f is
f h ( x) =
1
Nh
K
i =1
x xi
h
where K is some kernel function (often

standard Gaussian) and h is the bandwidth
or smoothing parameter. A kernel
function is also the inner product
k(x, x)=( (x). (x)), where data x are
implicitly mapped into the feature space
F: x (x) [13].
The use of kernels for classification and
regression is beyond any discussion;
support vector machines are subject of
many papers and books. Here we point
out several basic facts. First, there is a
kernel trick that allows computation of
(relatively simple) kernel function instead
of search for representing features,
computation a nonlinear transform
from samples (pattern) space into feature
(kernel) space, and then a scalar product
evaluation. As a basis of the support
vector machine, the learning error is
minimized via quadratic programming
problem. To get a unique solution, there is
a condition of positively definite Gram
matrix, and then kernel should be
positively definite. This is a nearly
standard demand. Generally, the kernel
need not be necessary positively definite.

Second, the kernel need not be finite, as
according to (7). In [25] it is shown nicely
that perceptron is also a kernel machine
and sigmoid kernel is nowhere positively
definite. Moreover, integral (7) is infinite
in this case. In other words, we want to
say that in the context of the method we
present here these usually very basic and
prohibitive conditions need not be
studied.
9.3.2 Zipfian distribution (Zipf's law)
The Zipfian distribution (Zipf'
s law)
[20], [28] predicts that out of a population
of N elements, the frequency of elements
of rank i, f(i;s,N), is given by probability
mass function
1 is
f (i; s, N ) = N
,
(8)
s
1t
t =1
where N is the number of elements, i is

their rank, s is the value of the exponent
characterizing the distribution. The law
may also be written as:
1
f (i; s, N ) = s
,
i H N ,s
where HN,s is the Nth generalized
harmonic number.
The simplest case of Zipf'
s law is a "1/f
function" arising when s = 1. Given a set
of Zipfian distributed frequencies of
occurrence of some objects, sorted from
the most common to the least common,
the second most common frequency will
occur as often as the first. The third
most common frequency will occur 1/3 as
often as the first, and so on. Over fairly
wide ranges, and to a fairly good
approximation, many natural phenomena
obey Zipf'
s law. Note that in the case of a
"1/f function", i.e. s = 1, N must be finite
and its denominator is equal to HN, the socalled harmonic number, i.e. the sum of
truncated harmonic series; otherwise the
60
denominator is a sum of harmonic series,

which is divergent. This is not true if
exponent s exceeds 1, s > 1, then the
series is convergent,
1
(s) =
< ,
s
t =1 t
where is Riemann'
s zeta function.
10 GENERALIZATION
Formulas (4), and (8) need some

generalization since there are several
classes C and a different number of
samples Nc, c = 1, 2, ..., C of each class in
the learning set. General formulas are as
follows:
1
1 / ri q
N c xiU c
p (c | x) = C
1
1 ri q
k =1 N k xi U k
(Eventually q = v),
1
1/ i
N c xiU c
.
p (c | x) = C
1
1i
k =1 N k xi U k
We can see the introduction of relative
representation of different numbers of
samples Nc of each class, i.e. introducing
a priori probabilities.
11 DISTRIBUTION MAPPING
EXPONENT ESTIMATION
In this section, we suggest a procedure

how to determine the distribution
mapping exponent for a classifier, which
classifies into two classes. The extension
for many classes will be straightforward.
As before (Chap. 2), let the learning set U
of total N samples be given. Let U be
composed of points (patterns, samples)
xcs, where c = {0, 1} is the class mark and
s = 1, 2, , Nc is the index of the point
within class c; Nc is the number of points

in class c and let N = N0 + N1 be the
learning set size. Points xcs of one class
are ordered so that index i = 1
corresponds to the nearest neighbor, index
i = 2 to the second nearest neighbor, etc.
In Euclidean metrics, ri = ||x - xci || is the
distance of the i-th nearest neighbor of
class c from point x. xi is the i-th nearest
neighbor of point x.
To estimate the distribution mapping
exponent q we use a similar approach to
the approach of Grassberger and
Procaccia [11], for the correlation
dimension estimation.
We look for exponent q so that ri q is
proportional to index i, i.e.
ri q = ki , i = 1, 2, ..., Nc,
c = 0 or 1,
where k is a proportionality constant,

which will be eliminated later, so we need
not bother with it. Using a logarithm, we
get
q ln(ri ) = ln(k ) + ln(i ) , i = 1, 2, ..., Nc
(10)
This is a task of estimating the slope of a
straight line linearly approximating the
graph of the dependence of the neighbors
index s as a function of distance in loglog scale. This is the same problem as in
the correlation dimension estimation
where equations of the same form as (9)
and (10) arise. Grassberger and Procaccia,
proposed a solution by linear regression.
Other
authors
proposed
different
modifications and heuristics later. Many
of these approaches can be used for the
distribution mapping exponent estimation,
e.g. the use of Nv < Nc nearest neighbors
instead of Nc eliminates the influence of a
limited number of the points of the
learning set. Nv may be equal e.g. to one
half or the square root of Nc. The accuracy
of the distribution mapping exponent
61
estimation is the same problem as the

accuracy of the correlation dimension
estimation. On the other hand, one can
find that a small change of q does not
essentially influence the classification
results, as discussed already (Chap.
9.2.1).
We solve the system of Nv equations (10)
with respect to an unknown q by the use
of standard linear regression for both
classes. Thus, for two classes we get two
values of q, q0 and q1. To get a single
value of q we use the arithmetic mean,
q = (q0 + q1)/2. For more classes, the
arithmetic mean of the qs for the
individual classes is used.
12 EXPERIMENTS AND RESULTS
12.1 Synthetic Data
Synthetic data [24] are two-dimensional
and consist of three two-dimensional
normal distributions with identical a priori
probabilities. If denotes the vector of
the means and Cm is the covariance
matrix, there is
Class A: = (2, 0.5)t, Cm = (1, 0; 0, 1)
(identity matrix)
Class B: = (0, 2)t, Cm = (1, 0.5; 0.5, 1)
Class C: = (0, -1)t, Cm = (1, -0.5; -0.5,
1).
In this experiment, we employed a simple
strategy of using a half of the total
number of samples in the learning set,
which were nearest to the query point.
Fig. 4 shows the
Fig. 4. Comparison of classification errors of the synthetic

data for different approaches. In
the legend, 1-NN (L2) means the
1-NN method with Euclidean
metrics; CW, PW, and CPW are
three variants of the method by
Paredes and Vidal; points are
estimated from the reference 24,
Bayes means the Bayes limit.
QCregre means the method
presented here.
Legend:
results obtained by different methods for

different learning sets sizes from 8 to 256
samples and a testing set of 5000 samples,
all from the same distributions and
mutually independent. Each point was
obtained by averaging over 100 different
runs.
In our method QCregre, we employed a
simple strategy of using a half of the total
number of samples in the learning set,
which were nearest to the query point in
this experiment. For other methods, i.e. 1NN method with L2 metrics and variants
of the LWM method [24], the values were
estimated from the literature cited.
In Fig. 4, it is seen that the use of class
probability estimation with the method
presented in this synthetic experiment
outperforms all other methods shown in
Fig. 4; and for a large number of samples,
it quickly approaches the Bayes limit.
62
12.2 Data from the Machine Learning
Repository
Data sets intended just for running with a
classifier were prepared by Paredes and
Vidal and are available on the Internet.
We used all data sets of this corpus. Each
task consists of 50 pairs of training and
testing sets corresponding to 50fold cross
validation. For DNA data, Letter data
(Letter recognition), and Satimage
(Statlog Landsat Satellite) [21], single
partitioning into a training and testing set
according to the specification in [21] was
used. We also added the popular Iris data
set with tenfold cross validation. The
results obtained by the QCregre approach,
in comparison with data published in [24],
are summarized in Table 2. Each row of
the table corresponds to one task from
[21]. For tasks where the data are not
available, only the results for 1-NN
method with L2 metrics were amended. In
the QCregre method, we used a rather
complex strategy of robust modification
of linear regression, as described above.
The interesting point is the experiment
with the simplest strategy of using a half
of the samples nearest to the query point.
For some tasks we obtained very good
results.
Table 2. Classification error rates for different

data sets and different NN-based approaches by
[18] and LWM1. Empty cells denote data are
not available.
In Table 3, the results for tasks Heart,

Ionosphere, and Iris are shown
together with the results for other
methods published in [24]. We can briefly
characterize these data sets as follows:
The task Heart indicates the absence or
presence of heart disease for a patient.
For the task Ionosphere, the targets
were free electrons in the ionosphere.
Good radar returns are those showing
evidence of some type of structure in the
ionosphere. Bad returns are those that do
not return anything; their signals pass
through the ionosphere.
The task Iris is to determine whether an
iris flower is of class Versicolor or
Virginica. The third class, Setoza is
deleted, as it is linearly separable from the
other two. 100 samples, four parameters
and tenfold cross validation were used.
We do not describe these tasks in detail
here as all of them can be found in the
descriptions of the individual tasks of the
Repository; and also the same approach to
testing and evaluation was used.
Especially, we used splitting the data set
63
into two disjoint subsets, the learning set

and the testing set, and we used cross
validation. We also checked some
standard
Table 3. Classification errors for three different tasks

shown for the different methods presented in the
Machine Learning Repository. The note [fri] means
the results according to the report by Friedman [8].
The results computed by authors are shown in bold.
methods for comparison:

1-NN standard nearest neighbor
method
Sqrt-NN the k-NN method with k
equal to the square root of the number
of samples of the learning set
Bay1 the nave Bayes method using
ten bins histograms
LWM1 the learning weighted
metrics by Paredes and Vidal [24].
For k-NN, Bayes, LWM and our method,
the discriminant thresholds were tuned
accordingly.
All
procedures
are
deterministic (even Bayes algorithm) and
thus no repeated runs were needed.
estimate of correlation dimension v

for all queries.
The local method that supposes
multifractal data; it needs an estimate
of scaling exponent q for each query a
new. Exactness depends on the quality
of estimate of q (GrassbergerProcaccias algorithm for a
neighborhood, Takenss method,
estimating the slope of DMF or other
methods of correlation dimension
estimation can be adopted [3], [4],
[12], [27].)
The IINC method, in spite of its
simplicity, is based on the assumption
of multifractal data, as shown above;
the estimate of q is, in fact, inherent.
The first two methods need no sorting

(perhaps in v or q estimation only), the
IINC needs sorting according to the
distance of points from the query point.
In any case, complexity is proportional to
the data dimension, the size of the
learning set; in the IINC method times
log(size of the learning set) counts more.
There is no exponential growth of
complexity with problem dimension. (We
project a multidimensional space into
one-dimensional space of distances.)
Software for the methods described here
is available on the net, see [14].
14 ACKNOWLEDGEMENT
The work was supported by the

Technology Agency CR under project of
series ALFA No. TA01010490. The
project was also supported by the Faculty
of Biomedical Engineering of the Czech
Technical University in Prague, RVO:
68407700, Czech Republic.
13 CONCLUSION
Three methods were presented here:

The global method that supposes
mono-fractal data; it needs one
15 REFERENCES
1.
Abe, S., Support Vector Machines for Pattern

Classification
(Advances
in
Pattern
64
2.
3.
4.
5.
Recognition). 2005: Springer-Verlag New

York, Inc.
Breiman, L.: Random Forests, Machine
Learning Vol. 45, No. 1, pp. 5-32. (2001)
Camastra, F.: Data dimensionality estimation
methods: a survey. Pattern Recognition 6,
2945 2954 (2003)
Camastra, P., Vinciarelli, A.: Intrinsic
Dimension Estimation of Data: An Approach
based on Grassberger-Procaccias Algorithm.
Neural Processing Letters 14(1), 27 34
(2001)
Cover, T.M., Hart, P.E.: Nearest Neighbor
Pattern Classification. IEEE Transactions on
Information Theory IT-13(1), 21 27 (1967)

Duda, R.O., Hart, P.E., Stork, D.G.: Pattern
classification, 2nd edn. John Wiley and Sons,
Inc., New York (2000)
7. Fawcett, T.: ROC Graphs: Notes and Practical
Considerations for Researchers. Technical
report, Palo Alto, USA: HP Laboratories,
38pp. (2004).
8. Friedmann, J.H.: Flexible Metric Nearest
Neighbor Classification. Technical Report,
Dept. of Statistics, Stanford University (1994)
9. Froehling, H. et al.: On determining the
dimension of chaotic flow. Physica 3 D
(1981), pp. 605-617.
10. Gama, J.: Iterative Bayes. Theoretical
6.
Computer Science 292, 417 430 (2003)

11. Grassberger, P., Procaccia, I.: Measuring the
strangeness of strange attractors. Physica 9D,
189 208 (1983)
12. Guerrero, A., Smith, L.A.: Towards coherent
estimation of correlation dimension. Physics
letters A 318, 373 379 (2003)
13. Herbrich, R.: Learning Kernel Classifiers.
Theory and Algorithms. The MIT Press,
Cambridge, Mass., London, England, 2002.
14. Ji ina, M. and Ji ina, Jr., M.:
IINC Classifier Based on Inverted Indexes of
Neighbors
Software.
On-line
http://www.marceljirina
.cz/files/IINC.zip
(2013)
15. Ji ina, M. and Ji ina, Jr., M.: Classification by
the Use of Decomposition of Correlation
Integral. /Foundations of Computational
Intelligence/. Berlin : Springer, 2009
(Abraham, A.; Hassanien, A.; Sn el, V.) S.

39-55. ISBN 978-3-642-01535-9. - (Studies in
Computational Intelligence. 205)
16. Ji ina, M. and Ji ina, Jr., M.: Classifiers Based

on Inverted Distances. In: K. Funatsu, K.
Hasegawa: New Fundametal Technologies in
Data Mining. (Chap. 19) INTECHWEB.org. ,
2011. Also available at:
.
http://www.intechopen.com/books/show/title/
new-fundamental-technologies-in-datamining (2011)
17. Ji ina, M. and Ji ina, Jr., M.: Using
Singularity Exponent in Distance Based
Classifier. Proceedings of the 10th
International Conference on Intelligent
Systems
Design
and
Applications
(ISDA2010) November 29 - December 1,
2010 Cairo Egypt, paper No. 1011, pp. 220224, ISBN: 978-1-4244-8135-4.
18. Ji ina, M. and Ji ina, Jr., M.: Utilization of
Singularity Exponent in Nearest Neighbor
Based Classifier. Journal of Classification
(Springer) 2012, in print
19. Mandelbrot, B.B.: The Fractal Theory of
Nature. W. H. Freeman and Co., New York,
1982.
20. Maslov, V.P.: On a General Theorem of Set
Theory Leading to the Gibbs, BoseEinstein,
and Pareto Distributions as well as to the
ZipfMandelbrot Law for the Stock Market.
Mathematical Notes, vol. 78, no. 6, 2005, pp.
807813.
21. Merz, C.J., Murphy, P.M., Aha, D.W.: UCI
Repository of Machine Learning Databases.
Dept. of Information and Computer Science,
Univ. of California, Irvine (1997),
http://www.ics.uci.edu/~mlearn/MLSummary.
html
22. Oh, S.K., Pedrycz, W.: The Design of Selforganizing Polynomial Neural Networks.
Information Sciences (Elsevier), Vol. 141,
Apr. 2002, No. 3-4, pp 237-258.
23. Osborne, A.R., Provenzale, A.: Finite
correlation dimension for stochastic systems
with power-law spectra. Physica D 35, 357
381 (1989)
24. Paredes, R., Vidal, E.: Learning Weighted
Metrics to Minimize Nearest Neighbor
Classification Error. IEEE Transactions on
Pattern Analysis and Machine Intelligence
20(7), 1100 1110 (2006)
25. Schlkopf, B., Smola, A.J.: Learning with
Kernels.
Support
Vector
Machines,
Regularization, Optimization, and Beyond.
The MIT Press, Cambridge, Mass., London,
England, 2002.
26. Stanley, H.E., Melkin, P.: Multifractal
phenomena in physics and chemistry.
65
(Review) Nature Vol. 335, 29 Sept. 1988, pp.
405-409.
27. Takens, F.: On the Numerical Determination
of the Dimension of the Attractor. In:
Dynamical Systems and Bifurcations. Lecture
Notes in Mathematics, vol. 1125, pp. 99 106.
Springer, Berlin (1985)
28. Zipf, G.K.: The Psycho-Biology of Language.
An Introduction to Dynamic Philology. The
MIT
Press,
1968.
(Eventually:
http://en.wikipedia.org/wiki/Zipf'
s_law)
66
Adaptivity of a Coloring Algorithm to Unreliable

Communications for Data Gathering in Wireless
Sensor Networks
Ichrak Amdouni, Pascale Minet, Cedric Adjih
Inria Rocquencourt, 78153 Le Chesnay cedex, France
Email: {ichrak.amdouni, pascale.minet, cedric.adjih}@inria.fr
ABSTRACT
KEYWORDS
Wireless sensor networks (WSNs) are prone

to node/link failures, message losses, and
dynamic node joins and departures. For
instance, in data gathering applications that
constitute a common type of applications in
WSNs, links between nodes and their parent
in the data gathering tree may be broken.
Protocols supporting such applications
should adapt their behaviour to guarantee
reliable wireless communications while
keeping a low overhead. In particular, this
paper focuses on the optimization of a
known coloring algorithm called SERENA
(SchEdule RoutEr Node Activity).
SERENA assigns colors to nodes such that
no two interfering nodes share the same
color. Each color is mapped to a time slot
during which nodes having the associated
color can transmit data. To ensure collision
free communications, SERENA should be
aware about the set of interfering nodes.
However, in case of topology changes, this
set may vary. Consequently, SERENA
should adapt to this. Our solutions
proactively select one or more parent
backups and guarantee that the coloring
remains valid if a parent is replaced by its
backup. Simulation results show that
reliability is obtained at the price of a small
increase in the number of colors used to
color the network.
Wireless sensor networks, data gathering

applications, topology changes, unreliability,
graph coloring, SERENA.
1. CONTEXT AND STATE OF

THE ART
1.1. Wireless Sensor Networks
The integration of sensors in buildings,
industrial process, environment or
vehicles, coupled with the efficient
delivery of sensed information, provides
tremendous benefits. Potential benefits
include: enhanced homeland security,
improved manufacturing productivity,
conservation of natural resources and
improved emergency response [2], [3].
In addition to these promising
applications, the size and cost of
wireless sensor nodes continue to
decrease. Both contribute to a
widespread use.
1.2. Energy Constraint in WSNs
The major problem in deploying WSNs
is their dependence on battery power
which is limited. Indeed, WSNs are
expected to operate for months if not
years on small inexpensive batteries with
limited lifetimes. Therefore, energy
efficiency is typically the primary goal
in these networks. Such stringent design
requirement can be met by a plethora of
approaches like energy efficient routing
[4], [5], [6], reduction of the number of
67
data transmissions [7], [8] and node

activity scheduling [1], [9], [10].
1.3. Node Activity Scheduling
As the energy consumed in sleeping
state is smaller than the energy
consumed in any other state, keeping
sensor nodes in the sleep state is a good
approach to save energy. However, to
prevent network partition and message
loss when some nodes are in the sleep
state, node activity scheduling is needed.
Depending on their vision of the network
and the stack layer they focus on, these
methods approach the scheduling
problem at high or low granularity:
1) High granularity: The network is seen
as a set of nodes cooperating to monitor
a given area like in a surveillance
application for instance. Among these
nodes, there are nodes that are qualified
redundant from the application point of
view. The goal of these protocols is to
build a set of active nodes, such that
only nodes belonging to this set must be
active, redundant nodes can sleep to save
their energy. In [11], Cardei and Du
introduce a new energy saving method
where sensor nodes are rganized into a
maximum number of disjoint sets. The
sets are activated successively to
monitor the whole region while nodes in
the other sets are in a low energy sleep
mode. This work has been extended in
[12], where a node can belong to several
sets to improve network lifetime. GAF
[13] is an energy-aware geographical
routing protocol. GAF distributes nodes
into a virtual grid such that nodes in one
cell are equivalent from routing point of
view. Nodes in one cell are activated
successively. At any time, only one node
is active in one cell while its neighbors
turn off their radio.
2) Low granularity: These solutions deal
with each non redundant node
individually to schedule its transmissions

and receptions and allow it to sleep as
soon as possible. They depend on the
MAC layer and are in general integrated
in this layer. Methods based on
CSMA/CA aim at reducing the overhead
induced by the RTS/CTS exchange.
For instance, S-MAC [14] is based on
sleep/listen scheme. Each node prepares
its activity schedule (the time of its next
sleep) and exchanges it with its 1-hop
neighbors. After receipt of such a
schedule from its 1-hop neighbors, each
node deduces when these neighbors will
be awake, it updates its schedule such
that it wakes up and listens to them in
order to receive data destined to it. Many
other variations of S-MAC have arisen
such as T-MAC [15] with an adaptive
duration of the active period, D-MAC
[16] that reduces network latency, OMAC [17] that improves the throughput.
TDMA based protocols allocate time
slots to each node and schedule medium
access based on these slots. In general,
energy efficiency is achieved by
allowing nodes to sleep when they are
neither receiving nor transmitting and by
avoiding collisions. Examples of these
methods are TRAMA [18], FLAMA
[19] and FlexiTP [9]. For instance,
FlexiTP is proposed for data gathering
applications. Any slot is assigned to one
transmitter and one receiver. All other
nodes can sleep during this slot. A node
selects the first available slot in its
neighborhood up to 2 hops and
advertises its schedule. The exchange of
this schedule allows any node to
determine the forbidden slots and to
select slot(s) to transmit and slot(s) to
receive data. FlexiTP supports link
failure and node appearance in the
network by updating the slot schedule
locally and disseminating the updated
schedule to nodes in interference range.
68
This solution does not support

immediate acknowledgement. We
speak
about
immediate
acknowledgement when the receiver
uses the slot of the sender to transmit its
acknowledgement. Among the TDMAbased algorithms, there are algorithms
that rely on graph coloring to achieve
energy efficient scheduling.
1.4. Graph Coloring
The network is modeled as a graph and
the vertices of this graph are colored.
Graph coloring consists in assigning a
color to each vertex such that two
interfering vertices have not the same
color and the number of colors used is
minimized. At the MAC layer, this
coloring is exploited as follows: a color
is mapped to a time slot during which all
nodes having this color can transmit.
Examples of these algorithms are
TDMA-ASAP [10] and SERENA [1].
For instance, TDMA-ASAP [10] is
designed for data gathering applications
but
does
not
support
neither
communications with broadcast nor
immediate acknowledgement. It consists
in a centralized level by-level node
coloring algorithm and slot stealing
algorithm. It performs level by level
coloring starting from the leaves such
that any child is scheduled before its
parent. Moreover, any node listens for
unused slots owned by any of its onehop neighbors that has the same parent.
It can then steal any slot if the owner of
this slot is not using it. Contention is
used to prevent collision between
multiple stealers. Another example of
graph coloring methods is the SERENA
[1] algorithm that we have developed to
allow sensors to save energy. Ensuring
the adaptivity of SERENA to unreliable
wireless communications is the goal of
this paper.
1.5. Problem Statement

In this paper, we focus on data gathering
applications supported by WSNs. In
these WSNs, a data gathering tree rooted
at the sink is used to collect the data sent
by sensor nodes. In addition to the
energy constraint, WSNs are prone to
interferences, message losses and
node/link failures. In particular, in a data
gathering tree, when one node can no
longer reach its parent because the
parent or the link with this parent has
failed, the tree must be repaired. In
general, this is the task of the routing
protocol. However, route repair is
beyond the scope of this paper. We focus
rather on the adaptivity of SERENA to
the scenario of parent update in the data
gathering tree. The issue is that
assuming
broadcast
and
unicast
communications between a node and its
parent, a node updating its parent may
gain new interfering nodes.
Consequently, since these nodes were
not considered as interfering in the initial
coloring, color conflicts may occur. By
color conflict, we mean the fact that two
interfering nodes share the same color
and hence their packet transmissions
may collide. One possible solution is to
recolor the network. Our aim is to avoid
this overhead and to enhance the
robustness of SERENA.
1.6. Methodology
Our methodology consists in proactively
determining potential substitute(s) of the
parent in the data gathering tree, called
parent backup and integrating it in the
coloring. In other words, we color the
network such that when the parent is
replaced by a predetermined backup, no
color conflict occurs and the initial
coloring remains valid. We propose two
solutions based on SERENA and we
compare them with SERENA. The
69
benefit of these solutions is the

enhancement of the reliability and the
efficiency of wireless communications.
We will show by simulation that the
tradeoff is a reasonable increase of the
data delivery delays and the energy
consumption.
1.7. Organization of the Paper
The remainder of this paper is organized
as follows. In Section 2, we present the
general principles of SERENA and we
explain how these principles are adapted
to data gathering applications. We end
this section by discussing the robustness
of SERENA regarding the unreliable
nature of tree links. In Section 3, we will
present the two proposed solutions.
Section 4 presents a comparative study
of SERENA with the designed solutions.
We conclude in Section 5.
2. PRESENTATION OF SERENA
SERENA aims at coloring sensor nodes
such that no two interfering nodes share
the same color and the number of colors
is minimized. We say that the coloring is
valid if it ensures that any two
interfering nodes have different colors.
The general principles of SERENA are
presented in Section 2.1, while in
Section 2.2 we adapt these rules to data
gathering applications.
2.1. General Principles of SERENA
The SERENA coloring algorithm is
distributed and localized. Any node u
proceeds as follows:
1) u determines the set of its conflicting
nodes, that is the nodes that cannot share
the same color as it, denoted N(u).
2) u computes its priority, denoted
prio(u).
3) u colors itself according to the priority
order in N(u), that is when all conflicting
nodes having a higher priority than itself

are already colored.
4) u selects the smallest color unused in
N(u).
The benefits of SERENA are:
1) No bandwidth is lost in collisions, the
overhearing and the interferences are
reduced. The use of the same color by
several nodes ensures spatial reuse.
2) At the energy level, there are many
benefits: (1) No energy wasted in
collisions, (2) Sensors can sleep to save
energy without losing messages sent to
them, (3) Sensors are awake only when
needed: there is no idle listening.
3) Minimizing the number of colors
needed to color the network reduces the
total number of slots used to gather data
to the sink. Hence, end-to-end data
gathering delays are reduced. Also, as
we will explain in Section 2.2, SERENA
is adapted to data gathering applications
to ensure data delivery to the sink in a
single cycle.
As previously mentioned, SERENA
colors are associated to slots. This
implies a medium access scheme where
time is divided into cycles composed of
slots. Many of these medium access
methods like IEEE 802.15.4 [26]
consider a cycle composed of an activity
duration and an inactivity duration. Used
within such a cycle, the number of colors
computed by SERENA represents the
activity duration of the network.
Consequently, the smaller the number of
colors is, the longer the inactivity period
of sensors is. Also, reducing the number
of colors reduces the end-to-end data
gathering delays. That is why, for
SERENA, reducing the number of
colors is the main goal.
70
2.2. Rules of SERENA for a Data

Gathering Application
SERENA adopts a cross-layering
approach to fit the requirements of the
application
while
keeping
good
performance. For instance, in [20], we
adapted SERENA to data gathering
applications. This adaptation includes
the following steps:
1) Node priority definition:
In case of tree communications, the
priority of any node u is equal to
Descendants(u), that is the number of its
descendants in the data gathering tree.
This choice tends to alleviate the color
constraints imposed to the descendants
of the node and thus to reduce the
number of colors used. Indeed, let us
consider a data gathering tree where the
root has n children. Each child is a root
of a subtree. Since any two children
cannot have the same color, each one
will have a different color. Let u be the
child with the largest number of
descendants. If u is the last node to be
colored, it will be assigned the color n.
And since the color of any node is higher
than the color of its parent (see Rule R2
at the end of this section), descendants of
u will have colors higher than n.
Moreover, the biggest subtree is likely to
use the highest number of colors, so
there will be colors used only in this
subtree. Consequently, the total number
of colors will increase. This is justified
by simulation results published in [1],
[21], [22].
To avoid this, SERENA starts by
coloring the node u with the highest
number of descendants. Ties are broken
by the smallest node identifier.
2) The set of conflicting nodes:
Determining accurately the set of
conflicting nodes of any node is crucial
because: (1) If it includes less nodes than
necessary, the coloring would allow

conflicting nodes to have the same color
and hence collisions may occur, (2) If it
includes more nodes than necessary, a
color will be forbidden to a number of
nodes higher than necessary and hence
the total number of colors will increase.
Consequently, the set of conflicting
nodes must be finely defined taking into
account all the communication types
supported (e.g. broadcast, unicast with
immediate acknowledgements). In [20],
we have studied SERENA in the case
broadcast
and
immediate
acknowledgements are supported (the
immediate acknowledgement is sent at
the time slot of data reception). We have
identified the different sets of conflicting
nodes per node. Figure 1 illustrates the
determination of one of these collision
cases.
Fig. 1. Collision case in the case broadcast and

immediate acknowledgement are supported.
Figure 1 depicts a data gathering tree

where arrows link a node and its parent
in this tree and dotted lines are radio
links that do not belong to the tree. We
apply 2-hop coloring that is a coloring
that prevents any 1-hop or 2-hop
neighbors from having the same color.
Consequently, nodes D and G that are 3
hops away share the same color (the
colors of other nodes are not presented
71
here). Consider the following scenario.

First, node D sends its data to node A
while node G broadcasts a packet to
nodes S and F. Second, A acknowledges
the packet of D in the same time slot
during which F receives data from G.
Hence, this immediate acknowledgement
collides with data sent by G to node F.
To avoid this scenario, G which is a
parent of 1-hop neighbor of the parent of
D must have different color than D.
Similarly, we study all collision cases in
SERENA [20]. The set of conflicting
nodes of any node u must contain:
a) Set 1: its one-hop and two-hop
neighbors;
b) Set 2: the children of the one-hop
neighbors
c) Set 3: the parents of the one-hop
neighbors of its children;
d) Set 4: the children of the one-hop
neighbors of its children;
e) Set 5: the parents of the one-hop
neighbors of its parent (as in Figure
1).
Notice that in SERENA, the knowledge
of the 1-hop and 2-hop neighbors is
usually provided by the routing protocol.
3) Color selection:
a) Rule R1: any node u colors itself if

and only if all nodes in N(u) having a
higher priority than u are already
colored. It takes the smallest color
unused in N(u).
b) Rule R2: the color of u must be
higher than the color of its parent.
Rule R2 is added to reduce data
gathering delays. To illustrate this,
consider the example in Figure 2. The
colors produced by SERENA are
presented by integers next to the circle.
Based on these colors, scheduling
medium access according to the
decreasing order of colors enables each
node to aggregate the information
received from its children before
transmitting it to its parent. It is the case
of the node C for instance that accesses
the medium once all its descendants
have
transmitted
their
packets.
Consequently, information from all
nodes can reach the sink in a single
cycle.
Remark 1: Notice that this approach
implies that each node needs to buffer
the packets received from its children.
We assume that in our data gathering
application, sensors generate a periodic
traffic of small size and that aggregated
data can be transmitted in a single time
slot. This assumption is realistic
especially when intermediate nodes
perform some operations on the received
data like computing the average value or
the maximum value of the received data.
However, there are cases where
aggregation is not possible due to the
size of the gathered data. This context is
beyond the scope of this paper.
Fig. 2. Example of SERENA coloring applied

to a tree with the corresponding slot assignment.
Any node follows the two following
rules:
72
2.3. SERENA Robustness in Case of

Link Failure
Topology changes may invalidate
SERENA for the following reason. We
illustrate this via the example of Figure
3.
Fig. 3. Node D updates its parent B and gains a

new conflicting node G.
In Figure 3, arrows are tree links while

dotted lines are radio links that do not
belong to the tree. We suppose this tree
is colored. Initially, D has as a parent the
node B, D is not conflicting with G so D
and G share the same color. However, D
can no longer reach its parent node B, so
it replaces it by node A. Consequently,
while assuming tree communications
where node D communicates with A
rather than B, node G becomes a new
conflicting node to D (G belongs to Set 5
of D).
More generally, after any node u updates
its unreachable parent by a 1-hop
neighbor, Sets 1, 3, 4 remain unchanged,
while Sets 2 and 5 are updated.
Notice also that even if the coloring
remains valid, it may be no longer
optimized. Indeed, Rule R2 may be not
satisfied (see Section 2.2), hence data
gathering delays are increased.
To cope with these potential color
conflicts, the trivial solution is to recolor
the network. However, the drawback is
the induced overhead. Our aim is to
avoid this overhead.
3. IMPROVING
RELIABILITY OF TREE
COMMUNICATIONS
To realize our goal, we keep the same
principles of this solution and propose
two extensions called Sol1 and Sol2. Our
key idea consists in (1) determining
a priori one or several backups per
parent and (2) taking into account these
potential parents in the coloring
SERENA. Indeed, we extend the set of
conflicting nodes with the backup and
any other node that may belong to one of
the 5 sets previously defined if the
parent was replaced by the backup.
Consequently, the parent replacement by
a predetermined backup does not induce
any color conflict and the initial coloring
remains valid. Depending on the number
of backups, we propose two solutions.
3.1. First Solution (Sol1): Tree Coloring
with One Backup per Parent
Our first solution Sol1 tolerates a single

failure per parent, this failure can be the
parent failure or the failure of the link
with this parent. Any node u has a single
parent backup. This parent backup is a 1hop neighbor that belongs to a smaller
tree level than the node considered (the
level of a node is its distance to the root
in number of hops, assuming the root
level is 0). This parent backup is denoted
uncle. If the node that is unable to reach
its parent is unable to find an uncle, it
will select a brother, that is a 1-hop
neighbor that belongs to the same tree
level and has a smaller identifier than u.
Remark 2: Notice that with these
constraints, any node has 0 or 1 backup.
Moreover, the tree repaired by the
choice of either the uncle or brother does
not induce routing loops.
73
As illustrated in Figure 4, the set of

conflicting nodes must contain Set 1, Set
2, Set 3, Set 4 and Set 5 and also:
Set 6: the children of the 1-hop
neighbors of the parent backup of u.
Set 7: the parent of the 1-hop
neighbors of the parent backup.
(a) Initial conflicting nodes (S2,S3,S4,S5)
We apply Rule R2 and maintain an order

between the color of a node and the
color of its parent. As we will explain
hereafter, two variants of Sol1 are
compared: (1) in the first one the color
of a node is greater than the color of its
parent backup, (2) the second variant
omits this condition in order to
maximize the probability of a node to
have a parent backup.
3.2. Second
Solution (Sol2): Tree

Coloring with Several Backups per
Parent
(b) Conflicting nodes added with backup

(S6,S7)
Fig. 4. The set of conflicting nodes of node A.
The second solution Sol2 consists in 3hop coloring where the set of conflicting
nodes of each node contains its
neighbors up to 3-hop. Consequently, a
node has as potential parent backup all
its 1-hop neighbors. The rule R2 is taken
into account with the initial parent.
However, it is impossible to define an
order between the color of a node and its
parent backups which are in fact its 1hop neighbors. So, Sol2 implies that
there is no order between the color of a
node and the color of its parent backup.
Remark 3: Notice that applying this
solution, routing loops are possible.
However, we address this issue by one
of the two following hints:
We assume that the routing protocol
handles the loops with a detection
and resolution technique.
Or, when a node chooses a parent
backup, it chooses only a node that
has a smaller identifier.
74
Remark 4: Notice that both Sol1 and

Sol2 can be easily integrated with
routing protocols that are based on
redundant paths. These redundant paths
will constitute the backup nodes. In fact,
most of the existing multipath routing
protocols were mainly developed to
provide fault tolerance at the network
layer. In this category we can cite for
instance DBR2P [23] and AODV-BR
[24]. For WSNs, we can cite Braided
Multipath Routing Protocol [25].
4.1. The Trade-off Reliability, Data

Delivery Delays and Energy
Consumption
Table I compares the performances of
The aforementioned solutions denoted
respectively Sol1 and Sol2. For Sol1 two
variants are evaluated:
With order: where the color of a
node is greater than the color of its
parent backup.
Without order: when this order is not
required.
4. COMPARATIVE STUDY
In this section, we present a table that
compares
qualitatively
the
aforementioned two solutions with the
basic solution SERENA. We then
present the simulation results.
From Table I, we can draw the following

conclusions.
1) Sol2 guarantees the highest reliability
whereas Sol1 tolerates only one failure
per parent. Moreover, as we will see, the
number of colors is the highest because
all the 3-hop neighbors of a node cannot
share its color.
2) The integration of Rule R2 allows to
have optimal end-to-end delays as long
as the link with the parent is operational.
75
Otherwise, each time there is an

inversion between the color of a node
and the color of its new parent, an
additional cycle is required, since in this
case the child accesses the medium after
its new parent.
3) Concerning the overhead, Sol2
induces higher storage overhead because
a node stores information about all its
neighbors up to 3 hops. This results also
in a higher number of messages
exchanged as we will show through the
simulation results.
To conclude, we notice that the
preference of a solution to another is a
trade-off between the tree reliability
from one hand and the delays and energy
efficiency from another hand.
4.2. Simulation Results
1) Simulation setup:
We developed our Java simulator.
Wireless nodes are randomly deployed
in the network. The parameters we vary
in our simulations are the number of
nodes and the density. The network area
size depends on both the number of
nodes and the density considered. We
build a tree of minimum path length to
the sink. For a given number of nodes
and density value, we generate 20
random topologies. Hence, each result is
the average of 20 simulation runs.
2) Impact of the rule R2 on the number

of colors
(a) Sol1
(b) Sol2
Fig. 5. Number of colors in cases density=10
with/without Rule R2.
We proved that integrating the Rule R2

in SERENA has the benefit of ensuring
data delivery in a single cycle. However,
this rule prevents a node from selecting a
color even if it is not already taken by
any of its conflicting nodes only because
it is smaller than the color of its parent.
Consequently, this would increase the
total number of colors used. We want to
evaluate the cost of this rule in terms of
number of colors for both Sol1 and Sol2.
Figure 5 shows that the maximum
number of added colors is about 6 colors
for Sol1 and Sol2 for 100 nodes. We
notice also that this number increases
76
with the number of nodes. For instance,

for 100 nodes and density=10, Sol2
requires about 17% additional colors for
the backup support. This overhead is
light regarding the benefit in terms of
delays. In addition, we notice that the
average number of colors required for
100 nodes, density=10 and with the
support of Rule R2 is 31 for Sol1 and
40.8 for Sol2. This means that the most
impacting factor on the number of colors
is the number of conflicting nodes per
node and not the support of Rule R2.
This is because the heuristic that we use
(the priority of any node is equal to the
number of its descendants) tends to
alleviate the impact of rule R2 by
favoring color reuse. Meanwhile, when a
node has more conflicting neighbors,
color reuse will evidently decrease and
hence the number of colors will increase.
3) The average number of nodes that
fail to have a backup:
small, for density=8, it ranges from 3.4

to 10.6 for 25 and 81 nodes respectively.
As expected, the number of nodes
without backup decreases when density
increases. Indeed, increasing the density
allows nodes to have more neighbors
and then the probability to have a
backup is higher. This result witnesses
the fact that Sol1 is suitable for high
density networks.
4) Comparison between the three
solutions:
In this section, we fix the density of
nodes to 10. For Sol1, we focus on the
variant integrating the color order
between the node and its parent backup.
As expected, SERENA+many backups
induces the highest number of colors
(see Figure 7(a)). Consequently, Sol2
ensures the reliability of wireless
communications at the detriment of both
data gathering delays and energy
consumption. Also, Sol2 requires the
highest number of rounds (a round is an
iteration where a node receives a
message from each one-hop neighbors,
processes it and transmits its own
message to its neighbors). This is
explained by the fact that to color itself,
any node collects information that is at
most 3 hops away from it. The
propagation of this information adds a
time overhead for the solution.
Fig. 6. Average number of nodes without parent

backup.
In this section, we focus on Sol1, where

any node has 0 or 1 backup. Figure 6
illustrates the average number of nodes
that fail to have a parent backup for
different configurations. This number is
77
(a) Number of colors
Concerning Sol1, it generates almost the

same number of colors and rounds as
SERENA. This is because in fact, this
solution does not add a high number of
conflicting nodes (nodes depicted in
Figure 4(b)). For instance, the average
number of additional conflicting nodes
in SERENA+one backup compared to
SERENA is 1 for 49 nodes with density
8. Moreover, these nodes are a subset of
the 3-hop neighbors, that is why Sol1
produces a smaller number of colors
than Sol2.
Concerning the average number of
messages sent, we notice a very
negligible difference between SERENA
and Sol1. This results from the
negligible difference between the two
solutions in terms of rounds.
5. CONCLUSION
(b) Number of rounds
The motivation of this paper is to

enhance the reliability of wireless
communications in wireless sensor
networks. Our focus is the coloring
algorithm that should tolerate link
failures in data gathering trees to avoid
the tree recoloring in case of topologies
changes. The solutions we proposed
which are an optimization of SERENA
are based on node backup usage. They
respond to the aforementioned objective
at the cost of a slight increase in the
number of colors for the first solution
and a higher number of colors for the
second solution. Also, we show that the
delays may increase if we want to
tolerate a maximum number of failures.
The future directions of this work are to
integrate these solutions with a routing
protocol and evaluate this integration.
(c) Number of sent messages

Fig. 7. Comparison of Sol1 and Sol2
78
6. REFERENCES
[1] Minet, P.; Mahfoudh, S. SERENA:
SchEduling RoutEr Nodes Activity in
wireless ad hoc and sensor networks. In
Proceedings of IWCMC 2008, IEEE
International Wireless Communications
and Mobile Computing Conference,
Crete Island, Greece, August 2008.
[2] Xu N., A Survey of Sensor Network
Applications, IEEE Communications
Magazine, vol. 40, 2002.
[3]
Garcia-Hernandez
C.F.;
Ibarguengoytia-Gonzalez P.H.; GarciaHernandez J.; Perez-Diaz J. A., Wireless
Sensor Networks and Applications: a
Survey,
International
Journal
of
Computer Science and Network
Security, IJCSNS07, vol. 3, pp. 264273, 2007.
[4] Heinzelman W.R.; Kulik, J.;
Balakrishnan H., Adaptive protocols for
information dissemination in wireless
sensor networks, 5th ACM/IEEE
conference on Mobile computing and
networking,
MobiCom99,
Seattle,
Washington, United States, August
1999.
[5] Heinzelman W.; Chandrakasan A.;
Balakrishnan H.; Energy efficient
communication protocol for wireless
sensor
networks,
International
Conference System Sciences, Hawaii,
January 2000.
[6] Senouci S.M.; Pujolle G., Energy
efficient routing in wireless ad hoc
networks,
IEEE
International
Conference on Communications ICC04,
Paris, France, June 2004.
[7] Levis P.; Clausen T.; Hui J.;
Gnawalli O.; Ko J., The Trickle
Algorithm, RFC 6206, MARS 2011,
http://tools.ietf.org/html/rfc6206
[8] Carle J.; Simplot-Ryl D., EnergyEfficient Area Monitoring for Sensor
Networks, Computer, vol. 37, no. 2, pp.

40-46, February 2004.
[9] Winnie L.L.; Amitava D.; Rachel
C.O., FlexiTP: A FlexibleScheduleBased TDMA Protocol for FaultTolerant and Energy-Efficient Wireless
Sensor Networks, IEEE Transactions on
Parallel and Distributed Systems, vol.
19, pp. 851-864, June 2008.
[10] Gobriel S.; Mosse D.; Cleric R.,
TDMA-ASAP: sensor network TDMA
scheduling with adaptive slot stealing
and parallelism, ICDCS09, Montreal,
Canada, June 2009.
[11] Cardei M.; Du D., Improving
wireless sensor network lifetime through
power aware organization, ACM Journal
of Wireless Networks, May 2005.
[12] Cardei M.; Thai M.; Li Y.; Wu W.,
Energy-efficient target coverage in
wireless
sensor
networks,
IEEE
INFOCOM 2005, Miami, Florida,
March 2005.
[13] Xu Y.; Heidemann J.; Estrin D.,
Geography-informed Energy
Conservation for Ad-hoc Routing, ACM
MOBICOM 2001, Rome, Italy, July
2001.
[14] Ye W.; Heidmann J.; Estrin D., An
Energy-Efficient MAC Protocol for
Wireless Sensor Networks, IEEE
INFOCOM, New York, USA, June
2002.
[15] Dam T. V.; Langendoen K., An
adaptive energy-efficient MAC protocol
for wireless sensor networks, ACM
SenSys03, November 2003.
[16] Lu G.; Krishnamachari B.;
Raghavendra C., An Ada ptive EnergyEfficient and Low-Latency MAC for
Data Gathering in Wireless Sensor
Networks, Parallel and Distributed
Processing Symposium, April 2004.
[17] Lu G.; Krishnamachari B.;
Raghavendra C., O-MAC: An Organized
Energy-Aware MAC Protocol for
79
Wireless Sensor Networks, IEEE ICC,

Glasgow, UK, June 2007.
[18] Rajendran V.; Obraczka K.; GarciaLuna-Aceves J.J., Energyefficient,
collision-free medium access control for
wireless sensor networks, Sensys03,
Los Angeles, California November
2003.
[19] Rajendran V.; Garcia-Luna-Aceves
J.J.; Obraczka K., Energyefficient,
application-aware medium access for
sensor networks, IEEE MASS 2005,
Washington, November 2005.
[20] Mahfoudh S.; Minet P.; Amdouni I.,
Energy Efficient Routing and Node
Activity Scheduling in the OCARI
Wireless Sensor Network, Journal
Future Internet, vol. 2, pp. 308-340,
September 2010.
[21] Brelaz, D. New methods to color
the vertices of a graph, Commun. ACM,
pp. 251256, 1979.
[22] Hansen, J.; Kubale, M.; Kuszner,
L.; Nadolski, A. Distributed largest-first
algorithm for graph coloring. In
Proceedings of EURO-PAR 2004, Pisa,
Italy, August 2004.
[23] Wang Y.H.; Chao C.F., Dynamic
backup routes routing protocol for
mobile ad hoc networks, Journal
Information Sciences, vol. 176, pp. 161185, January 2006.
[24] Lee S.J.; Gerla M., AODV-BR:
Backup Routing in Ad hoc Networks,
Wireless
Communications
and
Networking Conference, WCNC02,
Chicago, USA, September 2000.
[25] Ganesan D.; Govindan R.; Shenker
S.; Estrin D. Highly-Resilient, EnergyEfficient Multipath Routing in Wireless
Sensor Networks, Journal Mobile
Computing
and
Communications
Review, vol. 5, pp. 11-25, October 2001.
[26]http://pages.cs.wisc.edu/~suman/cou
rses/838/papers/zigbee.pdf
80
Presentation and analysis of a new technology for low-power wireless

sensor network
Sabri Khssibi*, Hanen Idoudi**, Adrien Van Den Bossche*, Thierry Val*
and Leila Azzouz Saidane**
*University of Toulouse, UT2, IRIT, BP60073 F-31703 Blagnac, France
**National School of Computer Sciences, University of Manouba, Tunisia
{val, bossche, sabri.khssibi}@irit.fr
{Hanen.idoudi, Leila.saidane}@ensi.rnu.tn
ABSTRACT
The tremendous and rapid development
in sensors technology allowed their
application in various fields requiring
monitoring, such as, transportations, rare
species surveillance, agriculture, military
activities, medical field, etc. Due to their
intrinsic constraints and limitations,
several dedicated MAC protocols have
been designed for wireless sensor
networks and whose main objectives are
bandwidth optimization while keeping
very low energy consumption. One of
the newest proposals is ANT which is
emerging as a widely used MAC
protocol for wellness and sports sensor
devices.
In this paper, we give a comprehensive
overview of ANT/ANT+ and a
comparison between ANT and two
established standards: Bluetooth Low
Energy (BLE) and 802.15.4/4a. We
detail also some practical experiments
we conducted to evaluate some
characteristics of ANT.
KEYWORDS
ANT, 802.15.4, BLE, WSN, MAC protocol.
1 INTRODUCTION
Wireless Sensor Networks (WSNs) have
been proposed for many applications
requiring the collection and aggregation

of environmental data detected in a
certain area of perception. We can
mention some typical applications: the
monitoring of vehicle movements in
hostile area [1], the observation of the
lives of rare species [2], the monitoring
of the structure of infrastructure [3] [4],
medical assistance [5], etc. Several
researches are directed to design new
MAC protocols adapted to the
characteristics of wireless sensor
networks. To this end, several MAC
protocols [6] [7] [8] [9] have been
proposed in the literature to optimize
access to the wireless medium. In recent
years, we have witnessed a major
expansion of the ANT (Advanced and
adaptive Network Technology) [10]
protocol and its extension ANT+. These
protocols are considered by the
industrial community as a well suited
architecture
to
the
monitoring
applications in sports and medical
domains, where it is necessary to use
wireless sensor networks.
ANT is a communication protocol with
ultra low power consumption. It
maximizes battery lifetime and provides
the use of many channels. It has a very
good theoretical baud rate which attends
1Mbps, making this protocol more
suitable for Wireless Personal Area
81
Networks with low bandwidth (LRWPAN).
ANT+ adds network management layers
which enables interoperability, defines
data structure, provides definition of
node profile and manages network keys.
In this paper, we propose to present a
theoretical overview of ANT and ANT+
and some practical tests conducted with
real material.
This paper is organized as follows: In
the beginning, we present the basic
principles of ANT protocol and ANT+.
Then, we present
a
comparison
between ANT, BLE [24] [25] [26] [27]
[28] and 802.15.4/4a. Next, we present a
practical study of ANT. In the end, a
conclusion comes to take stock of our
study and discusses the prospects for
research on which we move.
2 PRESENTATION OF ANT
ANT is a protocol designed for wireless
personal area networks (WPAN) with
low power consumption, low processing
power and low baud rate. The small size
of the ANT protocol stacks allows the
integration into components such as:
PDAs, watches, phones, Smartphone,
etc. It operates in the 2.4GHz ISM
(Industrial, Scientific and Medical radio
bands) band and allows for a flow baud
rate of 1Mbps with transmission
duration less or equal than to 150
microseconds / frame for 8bytes of data.
ANT provides a management of
Physical, Data link, Network and
Transport OSI layers (Fig 1). It's a very
compact stack requiring minim material
resources.
Figure 1. Correspondence between the ANT

model and the OSI model
2.1 ANT Node

ANT defines two types of nodes. The
first type is the master node. It is the
initiator
of
communication,
the
responsible of controlling the channels
and the transmitter of data (the primary
transmitter). Its window of reception is
used to provide the co-existence of many
masters in the same network with the
same frequency and to receive data from
slaves or other masters (Figure 2).
The slave node, the second type of node,
is the primary receiver of a frame from
the Master and responsible of
synchronization with the master (Figure
2 (a)).
In ANT protocol, a node can act as a
master and a slave at the same time with
different nodes. In Figure 2 (b), we have
three nodes where the node 1 is a master
of node 2 (slave in this case) and the
node 2 is the master of the node 3
(slave).
Figure 2. Different roles of a node
82
2.2 ANT channels
2.2.1 Parameters
The ANT protocol uses 2.4 GHz
bandwidth and divides it into 125
channels of width equal to 1 MHz and a
baud rate of up to 1 Mbps.
In order to create a connection between
two nodes, they must have the same
channel configuration. Then the slave
needs to know several parameters we
cite here after [21].
Initially, the default network number is

equal to 0 with a public network key and
no rule for the use of the network.
b- Frequency RF (Radio Frequency)
It represents the frequency where the
network operates. For example, if
Frequency RF is equal to 66, then the
network is operating at 2466 MHz.
In ANT protocol a channel must operate
on a single frequency among the
existing. The RF frequency is coded on 8
bits and varies between 0 and 124.
a- Network
ANT can set up several private and
public networks. In order to establish a
communication between two nodes, they
must be in the same network, which
enables a network service that can be
shared with multiple source nodes. On
the other hand, we can establish a private
network to ensure confidentiality and
restrict access. ANT node can participate
in several networks since the channels
can be assigned independently to
multiple networks. This parameter
contains two information: Network
number and Network key.
Network number is coded on 8 bits and
varies between 0 and a maximum value
fixed during the implementation of
ANT. The default value of the network
number is 0.
The Network Key is coded on 8 bits and
configured by the application. Each
network must have a network key. The
key and the network number allow the
deployment of networks with a varying
level of security and access control.
The Frequency RF value represents the

offset from the 2.4 GHz by 1 MHz
increment.
c- Channel ID
This is the most important information
for the channel. The higher layers are
responsible for the specification of this
parameter. The value of Channel ID is
coded on 4 bytes and contains the
following information:
Transmission type: 8 bits, it presents
the type of transmission
Device Type: 8 bits, this field is
used to differentiate between
different types of nodes in the
network, which allows participating
nodes to know the different classes
of connected nodes, and hence they
can
decode
the
information
received. If the device type is equal
to 0, the receiving node can connect
to multiple types of nodes.
Device Number: 16 bits, supposed
to be unique in the network.
83
Table 1. Transmission types and characteristics
Broadcast
Acknowledged
Burst
Forward direction
-Send the data in each time period.
-The data are not acknowledged.
-If the higher layer does not send
new data, ANT retransmits the last
data received by the HC layer.
-The frames are sent only once.
-The master may decide to send a
frame with the request of
acknowledgement during the step of
establishing a two-way connection.
-The HC layer can choose to send
all data frames with a request of
acknowledgement or it mixes the
broadcast transmission and data
transmission with acknowledgment.
-Communication by request of
acknowledgment of data uses more
bandwidth and consumes more
energy.
-This type of transmission is
important for critical data.
-There
is
no
automatic
retransmission of data packets
without acknowledgment.
-If the ANT protocol doesnt have a
new data in the interval time, the
last frame is retransmitted.
-This mechanism is used by the
master to transmit a large amount of
data to the slave.
-The burst transmission begins in
the transmission period and it is
expected to send a set of frames that
must be acknowledged by the slave.
-All lost frames are automatically
retransmitted.
-The transmission time is not
limited
- Data rate is up to 20kbps.
-The channel that uses this type of
transmission has priority over all
other channels
Reverse direction
-Send at the end of the period
of transmission of the master.
-Frames sent by this type of
communication are never
acknowledged.
-The slave may send a frame

to be acknowledged by the
master at the end of the
master transmission period
during
the
stage
of
establishing a two way
connection.
-There is no automatic
retransmission
of
data
packets
without
acknowledgment.
Data can be sent to the

master
by
the
burst
transmission in the same
manner described in the
master.
84
d- Channel type
This parameter defines the type of
transmission that will occur in the
channel. It is coded on 8 bits and can
define eight types of communication.
e- Channel period
This is the basic frequency used by the
master for transmission of data. By
default, this parameter represents the
interval time between two successive
transmissions of data transmission with
broadcast transmission (Figure 3). The
frequency of data frames in a channel
varies between 0.5Hz and 200Hz and the
default frequency is equal to 4Hz. The
channel period is coded on 16 bits.
Figure 3. Channel transmission
f- Data Type and format

The data type field determines the type
of communication taking place between
the participants of this channel and it is
fixed
by
the
higher
layers.
Communication consists generally of
two levels [21] (Table 1):
Manage the direction (Figure 4)
o Forward direction (from the
master to the slave)
o Reverse direction (from the
slave to the master)
Specification
of
type
of
communication
o Broadcast
o Acknowledge
o Burst
ANT does not provide an algorithm to

ensure the QoS in the network, but it
defines only the acknowledgement
packet. This packet must be transmitted
by the slave node when it receives a
frame
that
requires
an
acknowledgement. This ACK is used
only to inform the master that the frame
was transmitted successfully. ANT does
not retransmit the unacknowledged
frame.
2.2.2 Different methods of channels
management
ANT defines three different ways to use
the channel: Independent channels,
Shared channels and Channels scan.
In the independent channels, the channel
has one master and one slave. The node
can have a different role in another
channel. The node can support a limited
number of independent channels.
The shared and scan channels are used
when an ANT node must receive data
from multiple sources.
Figure 4. Channel communication direction
In the channel scan mode, the slave node

is on receiving state all the time, for this
reason it can receive the data from
multiple masters at any time. The radio
is always in the continuous scan mode,
so the ANT node cannot open other
channels.
85
2.2.3 Establishment of a channel
If two nodes want to communicate, they
must undergo the following process:
First, the master sets the type of
network, the frequency, the message rate
and the channel ID. Secondly, it
establishes the channel by transmitting
the channel ID with the data in a
particular time interval (channel period).
ANT channel will be maintained
indefinitely at the given time interval
and frame baud rate. The higher layers
of the master will eventually provide
new data to the ANT protocol for
continuous transmission. Thirdly, the
slave opens its channel to search for a
specific master or to search a subset of
masters (the type of research depending
on the configuration of slave channel).
When the slave receives the first
message, it selects the type of network,
operating frequency, and frame baud rate
of the master. With this information it
can synchronize with the master and
begin receive data [11][21][17].
In ANT, after the synchronization step,
in the ideal case, we don't have a
transmission failure if the slave node is
in the transmission ranges of the master
nodes. If the slave node is connected to
many master nodes, it switches between
channels to receive the information.
2.3 Device pairing
Pairing action consists of establishing a
relationship between two nodes that
wish to communicate together (master
and slave). This relationship may be
permanent,
semi-permanent
or
transitory.
device. In the case of a permanent

pairing, the slave node must keep the
identity of the master in the permanent
or nonvolatile memory. This ID is used
to open a channel with this ID in all
subsequent communication sessions. In
semi-permanent relationship, the pairing
lasts as long as the channel is maintained
[12][21].
3 PRESENTATION OF ANT+
ANT+ is an enhancement of ANT
protocol by adding the interoperability
function and defining the data structure,
channel parameters and the network
keys in order to enable the
communication with others ANT+
products. A managed network, ANT+
simplifies the collect, transmission and
tracking of health information.
ANT+ allows shared access to the
channel; it manages the definitions of
node profiles and adds it to the managed
networks in order to know the needs of
the participant nodes.
ANT+ is used in any application
including wireless sensor monitoring in
sport, wellness or home health such as:
temperature sensor, heart rate monitor,
speed and distance monitors, bicycle
power, blood pressure, etc.
The Figure 5 presents the layers defined
by ANT+. The use of ANT+ and ANT
together offers an Ultra low power
wireless technology solution and enables
the interoperability between nodes
(Figure 5).
It consists of a slave device acquiring

the unique channel ID of the master
86
transmission of big size of data and it

allows us to allocate time slots and use
TDMA to transmit the data (generally
used for QoS application).
5. PERFORMANCE STUDY
Figure 5. Layer model of ANT/ANT+[20]
4. COMPARISON BETWEEN ANT

AND OTHER PROTOCOLS
Table 2 presents a comparison between
802.15.4/4a, Bluetooth Low Energy
(BLE) and ANT protocol. We note that
802.15.4a have a transmission rate
greater than that of ANT, BLE and
802.15.4. The range of transmission of
802.15.4/4a is greater than that of ANT
and BLE.
The transmission range and rate of ANT
and BLE are equal.
In 802.15.4/4a a node can be an FFD or
RFD, the RFD nodes will be connected
to an FFD node. In the other hand, an
FFD node can be a coordinator or a
router for other nodes and can be
connected to all types of nodes in the
network. In ANT we have two different
nodes: a simple node (sensor) and a
central node. Simple nodes collect the
information and send it to the central
node where it will be conserved.
BLE is not compatible with classical
version of Bluetooth. But, a node can
implement a dual mode to be able to
communicate with BLE or classic
Bluetooth nodes.
In this section, we will present some

radio measurements of ANT protocol.
Our goal is to understand how ANT
transmits messages and uses the
bandwidth and to verify if ANT offers
the same theoretical baud rate with burst
transmission rate. For this reason, we
used two ANT USB keys and two
spectrum analysers (WiSpy [22] and a
classical HP tool) to observe the power
on the physical wireless medium and the
frequency of the signal transmitted by
ANT USB. The WiSpy has a intersting
functionnality for our performance
study: it represents temporal medium
energy density. This functionnality
enables us to view medium activity,
especialy in burst mode.
ANT USB is a computer dongle used to
receive, store, analyze and display data
transmitted by another ANT USB or
other devices like ANT sensor and
watch.
There are two ANT USB generations,
USB1 and USB2. The first generation,
USB 1, is replaced by the second
generation and it is no longer used. The
USB 1 and USB 2 have the same
hardware. The difference between the
two generations is in the software
application only. Where, the new
software offers better functionality and
performance. In our tests, we used the
second generation USB2.
Finally, ANT can be used for

transmission of data with small size.
802.15.4/4a can be used for the
87
Table 2. Comparison between ANT and 802.15.4/4a
Frequency band
Channel
numbers
Max Packet
Play load
(bytes)
802.15.4
3
frequency:
868, 915
and
2400MHz
27
802.15.4a
ANT
BLE
3 frequency:
Sub-Ghz, low
band
and high band
2400MHz
2400MHz
- 16 with
UWB
- 14 with chirp
signal
8 channels
40
20
30
10m and other

ranges
depending on
application
100
Range (meters)
100
Min
configuration
Transceiver
Node Type
A node can be: FFD (Full

Function Device)
or RFD (Reduce Function
Device)
Battery life
6 months
Synchronization
The transmission of beacon

periodically
Mode of
transmission
We have 2 modes: beacon

enabled (where we use
CSMA/CA for transmission
without QoS and TDMA for
the data that need a QoS
and non beacon enabled (we
use CSMA/CA only)
Transmit only or
transceiver
We have two types
of node: central
node (for collecting
the information
from simple nodes )
and simple node
(transmit the
detected
information to the
central
node)
Up to 3 years
The transmission of
the parameters of
synchronization in
all transmitted
packets
Best effort
Transceiver
Single mode
(BLE) or dual
mode
(Bluetooth &
BLE)
1 year
By the master
(AFH)
advertising
event &
connection
event
88
The USB2 stick is characterized by
[13]:
It uses the ISM band and uses the
frequency from
2403 to
2480MHz,
It
supports
the
Broadcast,
acknowledged, or burst data
transmissions,
Its minimum message rate per
channel is 0.5 Hz,
Its burst transfer rate is limited by
20kbps,
It uses up to 8 channels,
It has the scanning mode,
It has up to 3 public, managed
and/or private network keys.
These USB sticks will work in the
2466MHz frequency, the frequency of
transmission is equal to 4 Hz.
We have configured the ANT USB to
use the 2466MHz frequency to transmit
data with a transmission frequency equal
to 4 Hz, these values are the default
values of the ANT product.
We have transmitted 1000Kbytes of data
with burst transmission. Our goal is to
test and verify the correct functionality
of ANT protocol.
The Figures 6 and 7 will help us to find
the center frequency and the spectrum
large that are used by the node. The
Figure 6 represents the signal
transmitted by the ANT USB and
detected by the oscilloscope. This
spectrum is periodic and shows the
effect of this transmission in the
networks. The device uses the full
bandwidth from 2464 to 2468MHz
frequencies to transmit the data and we
have a peak in 2466MHz, so ANT USB
uses a bandwidth with 1 MHz of width
and a center
2466MHz.
frequency
equal
to
Figure 6. Spectrum analysis by the HP tool

during the transmission of 1000Kbytes
In Figure 7, we note the existence of

points with different colors. These colors
represent the power level on every
spectrum portion generated by the
transmitter during the message's
broadcast. As you note, the frames are
sent on 2466MHz (center frequency)
presented with red color and when the
points are moving away from the center,
the transmission power decreases. The
variation of transmission power appears
in Figure 7 by the change of the colors.
The transmission of data has taken some
time. This delay made us deduce the
baud rate and compare it with the
theoretical value.
ANT protocol took 7min and 29seconds
to transmit a payload of 1000Kbytes,
which means it transmitted 20568bits
per second. This result is the same
presented in the theoretical study and
USB sticks characteristic.
89
Now, we will test the possibility of

transmission over 8 channels with the
same frequency.
In Figure 9, we can see the effect of
using 8 channels in the same time. The
transmission is occurred between 2464
and 2468MHz and the center frequency
is 2466MHz.
Figure 7. Spectrum analysis by the WiSpy
during the transmission of 1000Kbytes
We have used a single channel and since

ANT allows the use of multiple
channels, we decided to test this
property and look at the effect of the use
of four channels which have the
following frequencies:
Channel 0: 2420MHz
Channel 1: 2440MHz
Channel 2: 2460MHz
Channel 3: 2480MHz
We note in Figure 8 that the
transmission is not performed on all
channels, we only have signal in 2
channels (channel 0 and channel 3).
Then we deduce that ANT can't transmit
in many different channels at the same
time. This is because the channel period
is not very large and the time of
transmission of one frame is equal to
150microseconds.
Hence, we can deduce that ANT allows

the transmission over multiple channels
in one condition: all channels will be
configured to transmit at the same
frequency as shown in Figure 9. We,
also, note that the results shown by
Figure 9 are close to that presented by
Figure 7 but here we don't use burst
transmission,
we
use
broadcast
transmission instead. In addition, in the
Figure 7 we note that the transmission is
occurs in different frequencies in the
interval [2464MHz, 2468MHz]. This
variation of frequencies permits to avoid
the collision. The points represent the
transmission and the reception periods.
Figure 9. Result of the transmission in 8

channels in the same frequency
Figure 8. Result of the multi-channels tests
In this section, we have presented some

practical test of ANT protocol realized
with the use of two ANT USBs. When
90
we use burst transmission ANT uses all
the bandwidth and stops other
transmission.
we can properly confirm the theoretical

study of ANT and ANT+.
7 REFERENCES
In the second test, we have many

transmissions in the same bandwidth,
despite we haven't interference. This is
explained by the fact that ANT protocol
allows a single channel to be divided
into hundred of timeslots to ensure the
co-existence of many transmission in the
same frequency and the use of an
adaptive
isochronous
network
technology which provides the ability
for each node to have a free time slot
within the used bandwidth. Also, ANT
protocol is used for the transmission of
low data quantities, which limits its use
to some monitoring applications only.
6. CONCLUSION
In this paper, we presented a new
emerging protocol for wireless sensor
networks which is at early stage of
development and improvement. Yet, it is
already widely used in sports and health
monitoring field.
At first, we made a general presentation
of the ANT protocol. Then, we presented
the enhancement and the new
functionality added by ANT+. Next we
compared ANT with BLE and
802.15.4/4a, and we deduced that ANT
can be used only for application where
we don't have a large amount of data
collected and do not require QoS.
Finally, we presented some practical
tests we conducted with real material,
where we compared the theoretical study
with real analysis of ANT protocol.
In our future work, we will perform
more advanced testing with this protocol
with an ANT Development Kit[23], so
1.
T. He et al., Energy-Efficient Surveillance

Systems Using Wireless Sensor Networks,
ACM MobiSys, Boston, USA, June 2004.
2. A. Mainwaring, J. Polastre, R. Szewczyk, D.
Culler, and J. Ander-son,Wireless Sensor
Networks for Habitat Monitoring, First
ACM Work-shop on Wireless Sensor
Networks and Applications, Atlanta, GA,
USA, September 2002.
3. N. Xu, S. Rangwala, K. K. Chintalapudi, D.
Ganesan, A. Broad, R. Govindan, and
D.Estrin, A wireless sensor network for
structural monitoring. The 2nd international
conference on Embedded networked sensor
systems, November 2004.
4. M.J. Whelan, M.V. Gangone and K.D.
Janoyan, Integrated Smart Wireless Sensors
for Bridge Structural Health and Homeland
Security Monitoring, The 3rd International
Conference on Structural Health Monitoring
of Intelligent Infrastructure, Vancouver,
Canada, November 2007.
5. Wireless Medical Sensor Networks,
http://www.cs.virginia.edu/wsn/medical/,
consulted in December 29, 2010
6. IEEE 802.15.4 Strandard on Wireless LAN
Medium Access Control (MAC) an Physical
Layer (PHY) specifications for low Rate
Wireless Personal Area Networks,IEEE
2006
7. J. Polastre, J. Hill and D. Culler. Versatile
Low Power Media Access for Wireless
Sensor Networks. ACM SenSys, 2004.
8. W. Ye, J. Heidemann, and D. Estrin. An
Energy-Efficient MAC
Protocol
for
WirelessSensor Networks. IEEE Infocom,
pp. 67-76, New York, NY, July 2002.
9. W. Ye, J. Heidemann, D. Estrin, Medium
Access Control With Coor-dinated Adaptive
Sleeping for Wireless Sensor Networks,
Networking, IEEE/ACM Transactions, vol.
12, no. 3, pp. 493-506, June 2004.
10. http://www.thisisant.com/, consulted in
January 31, 2012
11. ANT Channel Search and Background
Scanning Channel , Application notes, taken
from http://www.thisisant.com/
12. ANT Device Pairing, Application notes,
taken from http://www.thisisant.com/
91
13. ANT usb2 stick data sheet, taken from
http://www.thisisant.com/
14. ANT+ Devices profiles, ANT presentation,
15. The opportunities, ANT presentation, taken
16. ANT+ Alliance - an overview, ANT
presentation,
taken
from
http://www.thisisant.com/
17. Proximity search, Application notes, taken
18. Implementing A Receiver for Transmit-Only
ANT Devices, Application notes, taken
19. Interfacing with ANT General Purpose
Chipsets and Modules, Applica-tion notes,
20. ULP - advantage ANT, ANT presentation ,
21. ANT Message Protocol and Usage,
Application
notes,
taken
from
http://www.thisisant.com
22. WiSpy
:
Spectrum
Analysis,
http://www.metageek.net/products/wi-spy/
23. kit de developpement : ANTAP2DK1
(Nordic p/n nRF24AP2-DK1), Dy-nastream
Innovations
Inc
,http://www.thisisant.com/pages/products/an
t-dev-kit
24. Comparing
Low
Power
Wireless
Technologies, http://www.digikey.com ,
consulted in march 16, 2013
25. Specification:
Adopted
Documents
https://www.bluetooth.org/Technical/Specifi
cations/adopted.htm, consulted in march 16,
213
26. Bluetooth
low
energy,
http://www.litepoint.com/whitepaper/Blueto
oth%20Low%20Energy_WhitePaper.pdf
27. Specification of the Bluetooth System, Covered
Core Package, Version: 4.0; The Bluetooth
Special Interest Group: Kirkland, WA, USA,
2010.
28. Bluetooth low energy wireless technology
backgrounder,
http://www.nordicsemi.com/jpn/nordic/content_d
ownload/3045/40678/version/1/file/Bluetooth_le
_technology_backgrounder_4.pdf
92
A Comparison of Wireless Sensor Networks Co-Simulation Platforms

for Smart Grid Applications
Samira Chouikhi *, Ins El Korbi *, Yacine Ghamri-Doudane and Leila Azouz Saidane*
*
National School of Computer Science University of Manouba, 2010 Tunisia
ENSIIE & Universit Paris-Est (LIGM Lab), France

*
{ines.korbi, leila.saidane}@ensi.rnu.tn, chk.sam@gmail.com
ghamri@ensiie.fr
ABSTRACT
In the past few years, the smart grids have
attracted more and more attention. The fact
that these systems are based on the
pervasive use of new Information and
Communications Technologies (ICTs) opens
new perspectives in an aim to improve the
electrical grid performances. One of the
most commonly deployed technologies in
such a real environment is the wireless
sensor networks (WSNs) technology. WSNs
are thus used to monitor and control the
electrical grid components. In this paper we
will give an overview on WSNs applications
and challenges for smart grids. One
promising research topic, when speaking
about WSNs in smart grids, is the cosimulation which allows studying and
evaluating any new technique in a near real
grid behavior. Hence, in this paper, we will
also present and discuss the architectures,
the advantages and the drawbacks of the
most important smart grids co-simulation
platforms used to combine the network
behavior with the power grid systems
models.
KEYWORDS
Co-simulation, network simulation, electric
power simulation, smart grid, wireless
sensor networks.
1 INTRODUCTION
The huge development of consumer
electronics (PDAs, cell phones, laptops,
etc.) and the increase of the world

population make the techniques used in
the electricity generation, transmission
and distribution of the current power
grids conceived more than a century
ago- inappropriate to meet the new
power consumption needs. In fact, the
energy consumption is expected to
double by 2020 [1]. Other factors such
as
the
wear
of
the
used
components/equipments
prompt
depletion of the fossil energy used in the
electric energy generation process also
affect the current power grids efficiency.
Due above mentioned causes; many
countries have experienced, in the last
years
serious
blackouts
causing
considerable financial losses. Therefore,
the power renovation grids renovation
process becomes a real need to meet the
efficiency and reliability requirements.
The renovation process targets: i) the
electric equipments construction and
modernization, ii) the renewable energy
integration (wind, photovoltaic, etc.), iii)
the available grid resource management
to avoid the energy wasting, and iv) the
efficient grid components control and
monitoring.
A major power grid objective is to
ensure the permanent electricity
availability regardless all the conditions
(weather, component failure, hacking
attacks, etc.). This objective can be
reached by foreseeing the power
consumption, monitoring the grid
93
components, the grid self-healing
improving and adjusting automatically
the electricity distribution. All these
functionalities become feasible thanks to
the advent of Information and
Communications Technologies (ICTs).
ICTs
guarantee
the
efficient
communication
and
coordination
between all power grid participants
(suppliers, consumers, equipments, etc.).
The power grids equipped with ICTs are
called smart grids.
One
of
the
most
appropriate
technologies for the smart grids control
and monitoring is the wireless sensor
networks (WSNs) technology [2].
Hence, WSNs are deployed all over the
electricity grid parts (generation plants,
substations, power lines, consumers
side, etc.). Then, in this paper, we
propose to present the different
challenges to be raised by WSNs to meet
the smart grids requirements. We also
undertake a study on how a WSN based
smart grid solution is validated through
what we call the co-simulation process
(i.e. combining the network behavior
with the power grid systems models in
order to assess the performance in both
network communication side and
efficient power management side).
The rest of this paper is organized as
follows. In the section II, we focus on
the different WSNs applications in the
smart grids [3] [4] and we present the
new challenges facing the WSNs to
satisfy the smart grids requirements. In
section III, we present the need of using
co
simulation
to evaluate the
performance of WSNs communication
behavior in the context of the smart grids
by considering the characteristics of the
electrical components. Different cosimulation platforms will be presented
with a discussion on the advantages and
the drawbacks of each platform. Finally,

we conclude the paper in section IV.
2 WSNS APPLICATIONS AND
CHALLENGES IN SMART GRIDS
Considering their autonomy, their ease
of deployment, as well as the sensor
node sensing and communication
capacities, new opportunities had raised
in the past decade in the Wireless Sensor
Networks (WSNs) domain. These
opportunities are not linked to specific
applications, and WSNs can be deployed
in an outfit of applications (military
applications,
industry,
e-health,
environment, security, etc.).
On the other hand, in the smart grids
field,
the
power
generation,
transmission,
distribution
and
consumption processes require more and
more control and monitoring to improve
the (power) resource management and to
optimize the electricity consumption by
adapting the energy distribution to the
consumers needs. Such a monitoring
and control can be done only by
deploying a sensing device in each grid
component (transformers, generators,
etc.) and transmitting the gathered
information to either other components
(regulator, actuator, etc.) or to a control
center. This latter will be in charge of
taking adequate decisions. WSNs are the
natural choice to fulfill the sensing and
communication tasks. Then, new types
of sensors can be conceived to satisfy
the smart grid needs such as:
Basic measurements sensors:
voltages, currents, temperature;
Smart voltage sensors;
Smart sensors for fault or failure
detection;
Smart sensors for transformers
monitoring;
94
Temperature sensors for high
voltages lines;
etc.
2.1 Applications of WSNs in Smart
Grids
The electrical power grid can be
partitioned into three main conceptual
segments: the energy generation, the
power transmission & electricity
distribution, and the consumption side.
1) WSNs for energy generation segment:
In the conventional power grids, energy
generators are monitored by a limited
number of high-cost wired sensors
deployed in some critical locations. One
of the main objectives of smart grids is
the usage of intensive renewable energy
sources (wind, photovoltaic cells, etc.)
and their ease of integration in the
generation process. These generation
sources are generally situated in remote
and hard to reach areas within harsh
environment
which
requires
a
continuous monitoring of such zones via
low-cost sensors. Then WSNs represent
the ideal technology to monitor such
energy generators [5] [6] [7] [8].
2) WSNs for transmission and
distribution segment: The transmission
system
(towers,
overhead
and
underground power lines) and/or
distribution
system
(substations,
transformers and wiring to the
consumers) failure may cause blackouts
and may even present a danger to the
public
security.
Hence,
these
components must be monitored in near
real time to troubleshoot any problem. In
addition to the continuous monitoring,
this segment must be protected against
external attacks due to the ease of access
to any component. Once again, WSNs
provide promising solutions for the
electrical power grids monitoring and

securing [9] [10] [11] [12].
3) WSNs for the consumer side: For the
conventional electricity grids, the
substations have been the last mile of the
grid. Indeed, the consumer premises are
not included in the power grid. In smart
grids, a two-way flow of electricity and
information between the supplier and the
consumer become possible with the
usage of smart meters and an Advancing
Metering Infrastructure (AMI). It
becomes then possible to monitor, to
control the energy consumption of each
consumer,
and
even
to
react
automatically by using advanced tools
for home automation based on WSNs
capacities. Then, each appliance in the
modern house can be equipped by some
sensors to determine the energy
consumption cost and communicate it to
a local control center [3] [13].
2.1 WSNs Challenges in Smart Grids
The WSNs deployment in the smart
grids introduces new challenges which
sensor networks must overcome. These
challenges are related to communication,
reliability & robustness, and real time
constraints.
1) Communication challenges: The use
of usual communication technologies for
the WSNs such as ZigBee on IEEE
802.15.4 does not allow taking into
consideration the hostile environment in
which the smart grid components
operate. These components are a target
to noise and important interferences to
which the ZigBee technology is not
adapted since Zigbee shares the same
frequency band with the wireless LANs
using IEEE 802.11 and the Bluetooth
technology.
This may degrade considerably the
network communication performance.
95
2) Reliability and robustness of WSNs:
WSNs failures arising in power grids
may create serious problems. In fact, the
environmental conditions (interferences,
humidity, dust and vibration) may cause
the sensor nodes defection and affect the
network topology. The problems in
WSNs can be caused by the sensor
nodes constraints (limited battery and
CPU, and reduced memory capacity).
3) Real time constraints: Considering the
real time requirements in smart grids
which must react very quickly to a
failure situation, the communication
network must transmit the data in
a near real time to correct the anomaly
arising in the network and restore the
normal functioning of the electricity
grid.
After giving an overview of WSNs
applications in smart grids and some
challenges related to the sensor networks
deployment, the following section will
be dedicated to the description of smart
grids co-simulation platforms.
3 SMART GRIDS CO-SIMULATION
The electric power grid is among the
most critical infrastructures for a nation.
In
fact,
this
infrastructure
is
interdependent
with
other
vital
infrastructures and disruptions in the
power
grid
can
have
severe
consequences for other critical systems
such as the natural gas and the water
supply systems. To avoid any disruption
in that grid, any new mechanism must be
well studied and experimented before
deploying it in the real infrastructure
which is the role of simulation.
3.1 Necessity of Co-Simulation for

Smart Grids
In smart grids, the more accurately a
simulation platform emulates the
behavior and performance of the smart
grid architecture, the better we could
improve this grid. The problem in a
smart grid is to simulate two completely
different systems interacting with each
other. On the one hand, the electric
power system has a continuous time
dynamic behavior. On the other hand,
the communication network including
many types of networks (wireless sensor,
wired, LAN, and WLAN networks) has
a discrete event behavior. Therefore,
simulators conceived for conventional
power grids or only for networking are
unable to emulate the real behavior of
smart grids. This incapability rise the
need for new platforms that can reflect
the real smart grids behavior and provide
tools for the researches to simulate any
new technique or mechanism specific to
these grids. The term co-simulation is
used to define a simulation of two
different systems, which needs the
cooperation of more than one simulator
(the power simulator and the network
simulator in the case of smart grids) to
fulfill the simulation task.
Co-simulation supports many different
methods for connecting simulators. One
method is that each simulator runs
separately from the other and when an
interaction is needed it can send or
receive information by using third tier
(for example pipes). In this method, the
synchronization is a key factor to
guarantee the simulation success. The
second method is to implement new
interface in each simulator ensuring the
interaction with the other simulator
without need of a third tier. Another
method is to integrate one simulators
96
functionalities in the other as a module.
This can avoid the problem of
synchronization and gives a perfect
emulation of the real world.
3.2 Smart
Platforms
Grid
Co-Simulation
In the literature, many co-simulation

platforms were proposed to allow smart
grid efficiency simulation. For the
networking side the most used simulator
is the well known ns-2 (Network
Simulator-2) [14]. For instance, ns-2,
allows the simulation of a large panel of
communication
technologies
in
particular the IEEE 802.11 or IEEE
802.15.4 medium access layers that can
be used to ensure communication in
WSNs.
For the power systems side, many
simulators are experienced in the cosimulation process (adevs, OpenDSS,
PSLF, and Modelica).
1) The ns-2/adevs co-simulator: The
well known network simulator ns-2
provides a rich set of techniques and
protocols for WSNs. So selecting it as
the network simulator in the most cosimulation platforms is explained by its
functionalities covering different types
of sensor networks. In [14], the authors
presented a new co-simulation technique
to integrate power systems models with
the network-based wide area control
schemes. In fact, the difference between
a power system and a network system is
that the physical laws dictating the
power system behavior are defined by
differential-algebraic
equations
(continuous models) with few discrete
events, while the network behavior is
described by chains of significant events
(discrete models). The integration
technique is based on formal semantics
provided by DEVS (Discrete EVent
System Specification) [15] which offers a

large set of mathematical basis for
building and simulating hybrid systems.
adevs (A Discrete EVent System
simulator) [16] is used to implement
discrete events and continuous processes
that do not need communication using
the C++ language. Other processes using
communication are modeled with ns-2.
The events and processes modeled by
adevs are then integrated into the ns-2
simulation model using the adevs
simulation control API. The adevs
simulation tool is encapsulated as an ns2 TclObject and used directly by the ns2 simulation. The ns-2 module invokes
the adevs module if one of two events
occurs: an internal adevs event or a
message receipt at a process modeled
within adevs. When the first type of
events occurs, the ns-2 simulator queries
adevs for any message needed to be sent,
schedules ns-2 events to send messages
returned by the query, instructs adevs to
update its internal state and queries
adevs for its next event time to schedule
a corresponding adevs event in ns-2. The
second type of events causes ns-2 to
inject information into the adevs module.
It injects events into the adevs simulator,
instructs this simulator to update its
internal state according to the injected
events and asks it for its next event time.
ADEVS structure is used to simulate the
continuous processes. A good DEVS
formalization of the power behavior is
essential to reflect the real power
dynamics. The simulation of WSNs
deployment can be evaluated in this
platform thank to the different ns-2
techniques
and
modules
(MAC
protocols, Routing protocols. . .). The
small scale WSNs applications such as
home automation can be well simulated
in this platform.
97
2) The ns-2/OpenDSS co-simulator: In
2010, Godfrey et al. proposed a cosimulation platform [17] using the ns-2
and the OpenDSS (Open Distribution
System Simulator). They studied the
impact of cloud transient (solar
ramping) phenomenon, when clouds
pass in front of photovoltaic (PV)
panels, on the voltages variation. The
communication network is needed to
discharge distributed storage batteries to
compensate the voltage reduction.
OpenDSS is a simulation tool for electric
utility distribution systems. It provides
energy analysis tools for power delivery
and is designed to simulate discrete
events.
In the co-simulation, when the storage
controller detects the solar ramping, it
will attempt to dispatch the storage units
to maintain a smooth voltage. OpenDSS
provides the time of the PV ramp event,
the topology of the storage batteries and
the power load profile for the storage
units. ns-2 simulates the arrival of
controller messages to each storage unit,
which
is
assumed
to
respond
immediately. The arrival times are sorted
and fed back to the OpenDSS engine as a
script. The system voltage can be
computed at any time according to the
storage units reaction determined by
message arrival times.
The PV panels form a WSN that
communicates using the IEEE 802.11
standard. The model used in the ns-2 is
configured to simulate the IEEE 802.11
operating in 915 MHz Industrial,
Scientific and Medical (ISM) band.
Figure 1 gives an overview of the data
flow between the OpenDSS and the ns-2
during the co-simulation. The messages
sending time and the storage units
coordinates are transmitted to the ns-2
module, which computes the messages
arrival times. These times are merged
with the load profiles and then injected

in the OpenDSS engine. The process is
repeated until the storage batteries are
fully discharged.
Figure 1. OpenDSS/ns-2 co-simulation.
The study undertaken in [17] shows

how, using the ns-2/OpenDSS cosimulation, we can evaluate the solar
ramping phenomenon in smart grids
using photovoltaic panels and monitored
by a wireless sensor network.
3) The ns-2/Positive Sequence Load
Flow (PSLF) co-simulator: In [18], Lin
et al. integrate the PSLF (Positive
Sequence Load Flow) and the ns-2 to
implement a co-simulation framework
improving the practical investigation of
smart grid and evaluating wide-area
measurement and control schemes.
The authors explained their choices of
simulators by the fact that the two
simulators are proven in their domains
and have good supports for used defined
extensions. PSLF provides steady state
and dynamic power system simulations.
It has an extensive library of
electromechanical dynamic models and
can simulate a system with up to 60000
buses. PSLF offers also a graphical user
interface to facilitate the interactions
with the users.
The co-simulation is provided by
implementing new interfaces in both
PSLF and ns-2 sides. The ns-2 network
event scheduler is adopted as the global
98
scheduler. In the ns-2 side, a new class is
implemented as a PSLF interface that
initializes the power system dynamic
simulation rounds in ns-2. This class
collects power system data from PSLF
after each dynamic round and saves it in
ns-2 for future use by other network
components. In PSLF, a new stand-alone
dynamic model implemented and
attached to the simulated system. It has
two functionalities; the first is reporting
back the power system state data to the
ns-2 after a demand. The second
functionality is suspending PSLFs
simulation after each dynamic round and
waiting for ns-2 commands to continue.
The second function is crucial for
scheduling co-simulation by a global
scheduler.
Besides the new interface, other
components are implemented in ns-2 to
facilitate the power system applications
management.
Slave
agents
are
implemented in application layer that
stand agents interacting with power
system
and
communication
infrastructure (intelligent electronic
devises or digital relays). A master agent
is also implemented in the same layer to
coordinate the slave agents tasks. This
agent can stand for a central control
module or an operation center. Classic
transport protocols are modified to
support power data transmission. After
describing their co-simulation platform,
the authors studied an agent-based
supervisory backup relay protection
scheme on this platform. This case study
allows the co-simulation framework and
the protection scheme validation.
4)
The
ns-2/Modelica
based-on
simulator: In [19], the co-simulation is
used to evaluate a new communication
network for smart grids named
PowerNet. This network includes
interoperable heterogeneous networks
and aims to provide adequate degrees of

QoS, reliability, and security for the
networked control. It attempts to exploit
the previous deployed networks to
reduce the cost of deploying new
networks (LANs, WLANs, WiMax,
WSNs, etc.).
The co-simulation evaluation benefits
from the communication functionalities
offered by ns-2, the extensive set of
power devices available in Modelica
libraries, and from the graphical
interfaces of both simulators. The
intercommunication between the two
simulators is achieved by UNIX pipes.
The synchronization between the two
simulators is a key feature for the
interconnection to success. The intersimulator connection proceeds as
follows: ns-2 and Modelica start at the
same time (ti). While Modelica is
pausing, ns-2 runs until the first event
needing communication occurs at time
(ti+1). ns-2 pauses running and instructs
Modelica to run until (ti+1). At this time
is the event needs information to be
communicated from Modelica to ns-2,
this latter instructs Modelica to write the
information to a named pipe and then ns2 reads it from this pipe. If the
information is from ns-2, it informs
Modelica to read this information from
the named pipe. The process repeats
until the end of the simulation. At any
given time, we can notice that one
simulator is running, while the other is
pausing. In this co-simulation, ns-2 is
responsible
of
determining
the
communication times. Therefore, this
co-simulation cannot support sending
data between the two simulators in
response of events occurring inside
Modelica. Furthermore, the mechanism
of running and pausing reduces the
exploitation of parallelism available in
co-simulation.
99
The simulator choice is explained by the
fact that ns-2 provides tools to model
different networks (wired, wireless, local
and wide-area networks) and it can
evolve easily. For the power side,
Modelica is an object-oriented modeling
language for large-scale complex
physical systems.
5) Other co-simulation libraries: In
addition to the co-simulators presented
above, many other tools are available for
building new co-simulation platforms,
such as the Toolkit for HYbrid Modeling
of Electric power systems (THYME) [20]
which is a C++ library for building
simulator that integrates power system
modules
with
existing
network
simulators (OMNET++, ns-2). THYME
applications include: the study of control
system based on demand/response, the
impact of specific communication
technologies and intelligent sensors on
situational awareness, control or both,
and the design and analysis of wide area
diagnosis technologies that depend
critically on communication networks.
This tool includes an extensive set of
modules representing the power system
elements and behaviors.
THYME is not a simulator but a set of
libraries integrated in the adevs
simulator. As it is presented in Figure 2,
THYME provides the power model and
the power flow data, while adevs gives
the simulation environment (algorithms
and interfaces). The unlimited set of
network simulators that can be
integrated with THYME offer the
possibility
of
evaluating
WSNs
performance deployed in different
environments and with different network
sizes.
Figure 2. Integration of THYME in adevs

simulator.
Table
1
summarizes
the
advantages/drawbacks of each cosimulation platform described above.
Before choosing which co-simulation
platform is adapted to a given smart grid
application using WSNs as ICT, we first
have to determine the requirements of
such an application based on the
advantages and drawbacks of each cosimulation platform.
For instance, if we choose ns-2/adevs
co-simulation,
we
avoid
the
synchronization problem given that
adevs is integrated as a module in ns-2,
but the complexity of the power
behavior
formalization
decreases
drastically the performance of such a
platform. Moreover, in the case of large
scale sensor nodes deployment, the ns-2
simulator is not appropriate.
In OpenDSS/ns-2 co-simulation, the
scripts writing is facilitated by the
separate simulators running. But the
intercommunication technique is not
defined. In the ns-2/PSFL platform, the
use of new interfaces in both simulators
facilitates the communication of
simulators in spite of the large number
of modules to implement.
The ns-2/Modelica platform can
integrate WSNs with existing networks
technologies
-such
as
WIMAX,
WHANs, WPANs, etc.- which reduces
the deployment cost. Finally, cosimulation with THYME offers a large
100
number
of
WSNs
deployment
opportunities using different network
simulators.
Table 1. A comparative table of co-simulation platforms.

Platforms
Advantages
Drawbacks
Adevs is integrated as a module in ns-2

which avoids the synchronization need.
Each
simulation
is
implemented
separately which facilitate the scripts
writing
Complexity of power behavior

formalization
The
inter-communication
technique is not defined and must
be manually
The use of inter-communication interfaces

in both simulators
Many new modules must be

implemented in ns-2
ns-2/ Modelica
Proven simulation for the most existing

communication networks ( WSNs,
WIMAX, WiFi, WHAN, WPAN)
Modileca cannot determine when

to send data to ns-2;
Aperiodic control and alarm
signals that are generated in
response to events triggered
exclusively inside Modelica are
not accounted for
THYME/ OMNET++
or NS-2
An extensive set of network simulators

can be used with THYME.
THYME will permit the accurate
simulation of network-centric systems
with continuous & discrete-event
components.
THYME is a library and not a

stand-alone simulator, it must be
implemented with adevs
NS2/adevs
NS/ OpenDSS
NS2/PSLF
4 CONCLUSION
Due to the importance of the power grids
and the impact of their failure on the
Human life, new researches attempt to
modernize and smarten these grids so as
to avoid all the problems experienced
with the conventional electric power
systems. The so called smart grid is a
combination of the existing power grid
with
new
Information
and
Communications Technology
(ICT).
In this paper we presented the
application of a technology extensively
used to monitor and control the smart
grids behavior: the wireless sensor
networks
technology.
We
then
enumerated the different challenges that
should be met by the WSNs to satisfy
the smart grids requirements.
In the second part of the paper, we

focused on the smart grids co-simulation
platforms. For instance, in the particular
context of the smart grids, two
simulations engines have to cooperate
together to take into consideration the
communication
and
the
power
components
characteristics
when
evaluating the performance of smart grid
applications. Then, we described the
way of function of the most well known
smart grids co-simulation platforms and
dressed the advantages/drawbacks of
each platform especially for the
particular case of WSNs deployment.
5 REFERENCES
1.
2.
Garrity, T. F.: Getting Smart. IEEE Power

and Energy Magazine, April-March, pp. 3845 (2008).
Akyildiz,I. F., Su, W., Sankarasubramaniam,
Y., Cayirci, E.: A Survey on Sensor
101
Networks.
IEEE
Communications
Magazine, pp.102114 ( 2002).
3. Erol-Kantarci, M. H., Mouftah, T.:
Pervasive Energy Management for the
Smart Grid: Towards a Low carbon
Economy.
Pervasive
Communications
Handbook, Eds. S. I. A. Shah, M. Ilyas, H.
T. Mouftah, CRC Press (2011).
4. Isaac, S. J., Hancke, G. P.: A Survey of
Wireless Sensor Network Applications from
a Power Utilitys Distribution Perspective.
AFRICON (2011).
5. Lu, B., Habetler, T.G., Harley, R.G.,
Gutierrez, J.A., Durocher, D.B.: Energy
evaluation goes wireless. IEEE Industry
Applications Magazine, vol.13, no.2, pp.1723 (2007).
6. Shen, L., Wang, H., Duan, X., Li, X.:
Application of Wireless Sensor Networks in
the Prediction of Wind Power Generation.
4th International Conference on Wireless
Communications, Networking and Mobile
Computing, vol., no., pp.1-4 (2008).
7. Ipakchi, A., Albuyeh, F.: Grid of the future.
IEEE Power and Energy Magazine, vol.7,
no.2, pp.52-62 (2009).
8. Al-Anbagi, I., Mouftah, H. T., ErolKantarci, M.: Design of a Delay-Sensitive
WSN for Wind Generation Monitoring in
The Smart Grid. IEEE Canadian Conference
on Electrical and Computer Engineering
(CCECE), Niagara Falls, ON, Canada
(2011).
9. Leon, R. A., Vittal, V., Manimaran, G.:
Application of Sensor Network for Secure
Electric Energy Infrastructure. IEEE
Transactions on Power Delivery, vol.22,
no.2, pp.1021-1028 (2007).
10. Lim, Y., Kim, H.M., Kang, S. A.: Design of
Wireless Sensor Networks for a Power
Quality Monitoring System. Sensors 2010;
10(11):9712-9725 (2010).
11. Ullo, S., Vaccaro, A., Velotto, G.: The role
of pervasive and cooperative Sensor
Networks in Smart Grids communication.
15th IEEE Mediterranean Electro-technical
Conference
(MELECON),
pp.443-447
(2010).
12. Yang, Y., Lambert, F., Divan, D.: A Survey
on Technologies for Implementing Sensor
Networks for Power Delivery Systems.
IEEE Power Engineering Society General
Meeting, vol., no., pp.1-8 (2007).
14.
15.
16.
17.
18.
19.
20.
infrastructure for the Smart Grid. IEEE

Electrical Power & Energy Conference
(EPEC), pp.1-7 (2009).
Nutaro, J., Kuruganti, P. T., Miller, L.,
Mullen, S., Shankar, M.: Integrated HybridSimulation of Electric Power and
Communications Systems. In: Proceedings
of the 2007 IEEE Power Engineering
Society General Meeting, pp. 1-8 (2007).
Zeigler, B. P., Praehofer, H., Kim, T. G.:
Theory of Modeling and Simulation. 2nd
Edition. Academic Press (2000).
adevs: A Discrete Event system Simulator,
online at http://www.ornl.gov/~1qn/adevs/.
Godfrey, T., Mullen, S., Dugan, R. C.,
Rodine, C., Griffith, D. W., Golmie, N.:
Modeling Smart Grid Applications with CoSimulation. The 1st IEEE International
Conference on Smart Grid Communications
(SmartGridComm 2010), pp. 291-296,
Gaithersburg, MD (2010).
Lin, H., Sambamoorthy, S., Shukla, S.,
Thorp, J., Mili, L.: Power System and
Communication Network Co-Simulation for
Smart Grid Applications. Innovative Smart
Grid Technologies (ISGT), IEEE PES
(2011).
Liberatore, V., Al-Hammouri, A.: Smart
Grid Communication and Co-Simulation.
IEEE Energytech conference (2011).
THYME: Toolkit
for HYbrid Modeling
of Electric power systems, online at
http://www.ornl.gov/~1qn/thyme/docs/index
.html.
13. Sood, V. K., Fischer, D., Eklund, J. M.,

Brown, T.: Developing a communication
102
CMSA: a Centralized Maintenance Strategy with Anticipation for

Wireless Sensor Networks
Skander Azzaz and Leila Azouz Saidane
ENSI, Tunis, Tunisia
{skander.azzaz, leila.saidane}@ensi.rnu.tn
ABSTRACT
1 INTRODUCTION
Providing a continuous service is the main

requirement for many application types
deployed in Wireless Sensor Networks
(WSNs). In this paper, we study two
maintenance strategies using a small number
of mobile maintainer robots to restore the
WSN coverage and connectivity upon a
sensor failure: the Centralized Maintenance
Strategy (CMS) and the Centralized
Maintenance Strategy with Anticipation
(CMSA). The CMS and CMSA are based on
the Centralized Manager Algorithm [1] used
to detect, report sensor failures and
coordinate the movement of robots.
In CMSA, the predictive version of CMS, a
selected robot is chosen as a manager to
anticipate the sensor failures and schedule
the available maintainer robots to repair
them before they happen. To predict the
lifetime of a sensor node, we propose an
energy state model that represents the
behavior of a sensor node based on Markov
Chain. We use this model, validated by
simulations, to predict the energy
consumption by a sensor node and
consequently the lifetime of a wireless
sensor node. The simulation results show
that the CMSA ensures a null dysfunction
network time and a message overhead lower
than the classical centralized manager
strategy.
Generally, the Quality of Service (QoS)

in a wireless sensor network (WSN) is
measured by the sensing area coverage
degree and the connectivity between the
deployed sensors. To protect these WSN
QoS parameters, many approaches have
been proposed. Exploiting redundancy
of nodes [2] is one technique to fill holes
appeared in sensor network upon a
sensor node failure. This approach
requests mobile sensors to maintain the
coverage and connectivity in WSNs.
Knowing that a sensor node able to
move is expensive, Mei and all in [1]
propose using a small number of mobile
robots in a static sensor network to
replace failed sensors. They have
introduced
three
algorithms:
a
centralized manager algorithm, a fixed
distributed manager algorithm, and a
dynamic distributed manager algorithm
to detect, report sensor failure and
coordinate the movement of robots.
To detect failure, a relation guardianguardee is established between all
sensors. Each sensor node (guardee)
selects its guardian (a one-hop
neighbor). If a failure occurs, a guardian
of the failed node sends a report to its
manager robot. In the centralized
manager algorithm, all failure reports in
the network are addressed to a central
manager robot. Once the failure report is
received, the manager robot sends the
order to handle the failure to the
KEYWORDS
Energy Model; Fault-Tolerance; Markov
Chain; Mobile Robot; Wireless Sensor
Networks.
103
appropriate robot (a maintainer robot). In
the two distributed algorithms, the
management responsibility is distributed
over the robots, and each robot operates
as both a manager and a maintainer. In
fact, in the fixed distributed algorithm,
the area covered by the sensor nodes is
divided into equal-size subareas. Each
robot is assigned to a particular subarea
as both a manager and a maintainer
robot. Using this algorithm, a sensor
node selects its manager robot based on
its position in the network. If a manager
robot receives a failure report, it moves
to the failure location and replaces the
failed node. In the dynamic distributed
algorithm, the manager robot is selected
as the closest robot, and the failure is
handled by the manager robot that
received the report.
In their study, the authors of [1] have
shown that the centralized manager
algorithm has a low message overhead
and the smallest average robot traveling
distance per failure among the three
algorithms.
In this paper, we present a first strategy,
called Centralized Maintenance Strategy
(CMS), to repair failures in the WSNs
based on the Centralized Manager
Algorithm [1]. We detail our
improvements on the original algorithm
in Section 3.
However, failure handling according to
CMS still presents some drawbacks in
WSNs. In fact, the connectivity and
coverage of the network can be greatly
affected if the failure involves a critical
node during the repair time of nodes. In
this paper, we focus on the improvement
of the CMS. Our goal is to provide a
fault-tolerant WSN for real-time
applications. Our idea is to anticipate the
sensor failures. To reach this objective,
we need an energy model to predict the
energy consumption of sensor node. In
the literature, many analytical energy

models have been proposed. For
example [3] introduce mathematical
models to model different MAC (Media
Access Control) protocols such as IEEE
802.11 [4], IEEE 802.15.4 (ZigBee) [5]
or SMAC (Sensor Media Access
Control) [6].
Many problems were present in the
deployment of these models. Firstly,
these models introduce unrealistic
assumptions like Poisson traffic [7] and
saturation traffic assumption. Secondly,
these models depend on the MAC
protocol used by sensor nodes. And
finally, for complexity reasons, these
models cannot be implemented on the
sensor nodes with limited computing
capability, memory and energy supplies.
In this paper, we propose a simple
analytical energy model dedicated to be
implemented in a sensor network. In the
next step, we use the analytical model to
improve the CMS to provide a faulttolerant sensor network for real-time
applications. The new version of CMS is
called Centralized Maintenance Strategy
with Anticipation (CMSA).
The remainder of the paper is organized
in the following way. In section 2, we
present our analytical model to represent
the energy consumption of a sensor
node. In section 3, we present our
repairing strategies CMS and CMSA
dedicated to real-time applications
supported by WSNs. We validate by
simulation our analytical energy model
in section 4 and evaluate the
performances of our strategies. In
section 5, we conclude the paper by
giving directions for future work.
2 ANALYTICAL MODEL
The energy dissipated by a sensor node
can be divided in two parts: the energy
104
spent for radio communications by the
sensor radio interface, denoted in this
, and the sensing energy
paper by
(the energy it costs to take
measurements) denoted by
.
We focus first on the evaluation of
.
We claim that the status of the MAC
layer of the radio interface of a sensor
node can be modeled by a Markov
Chain. The states of this Markov Chain
depend on the MAC layer used by the
sensor node. For example, if we use the
IEEE 802.11 MAC layer we have three
modes of operation (states): Transmit,
Receive and Idle. Each state corresponds
to a different power consumption level
[8].
: node is transmitting
- State 1
.
a frame with transmission power
- State 2
: node is receiving a
frame with reception power
.
- State 3
: even when no messages
are being transmitted over the medium,
the node stays idle and listening the
.
medium with idle power
If the sensor node uses IEEE 802.15.4 or
SMAC as a MAC protocol, we have a
fourth state:
- State 4
: The radio is turned off,
and the node is not capable to detect
signals. We suppose that the node uses a
in this state which is largely
power
smaller than in any other state.
In conclusion, we model any MAC layer
of a sensor node as a discrete time
Markov Chain with four states:
,
,
and
.
the random variable related
We note
to the state of the MAC layer at time .
We choose a time step (unit) such that
the duration of any action (e.g;
transmission/reception of a frame) is a
multiple of this time step and we
suppose that all state transitions occur at
the beginning of the time step.
The notation
means that the
MAC layer is in state at time step .
Let
be the probability that a node in
state will enter in state at the next
transition.
(1)
Then, we note by
the transition
probability matrix (the element of row
column represents the probability ).
be the
Let
probability vector with
represents the probability that the MAC
layer is in state
at time step
(
).
Then, we have:
(2)
And:
(3)
Knowing that at time step 0 the node is

), we have:
generally in the state 3 (
(4)
the energy consumed by
Let
radio communication at the
transition given that the node is in state
at time step :
(5)
Where:
is obtained from equation 3 and
(the element of column of the vector
) is the energy consumed to transit
from state to state .
105
(6)
Where:
is the length in seconds of a
time step.
We note
the energy consumed
for radio communication by a node at
transition:
the
(7)
that represents the total

Then,
energy consumed by the radio interface
until the time step is:
(8)
In addition to radio communication
energy, the sensor consumes an amount
of energy for event sensing. We note by
the average consumption
energy in sensing mode (to take
measurements).
(9)
Where:
: is the average rate of events detected
by sensor node, : is the sensing mean
time of an event and
: is the
energy consumed by node in sensing
mode (the energy it costs to take one
measurement).
the sensing energy
We note
consumed by the sensor until the timestep :
(10)
Finally, the total energy consumed by a

is
sensor node until the time step
:
designated by
(11)
3 COORDINATION STRATEGIES
To coordinate the movement of
maintainer robots upon occurred failures
in the WSNs, [1] has proposed a
centralized manager algorithm to detect,
report and handle failure. We start with
this algorithm that we modify to obtain
the Centralized Manger Strategy in the
following subsection. Then, we propose
CMSA to achieve fault-tolerance and
provide a continuous service for
applications supported by WSNs.
3.1 The Centralized Maintenance
Strategy
In our study, we assume a connected
network and an area covered by static
wireless sensors. We suppose also that
all nodes know their location by means
of a localization technique like the GPS
(Global Positioning System) [9].
At the beginning, robots are uniformly
distributed over the considered area. One
robot is selected to act as a central
manager. Failures are detected by the
guardian nodes and reported to the
central manager robot. The manager then
forwards the report to a selected
maintainer robot to handle the failure.
Like the centralized manager algorithm,
we distinguish three stages in this
strategy:
In the first stage (the initialization stage),
we have an exchange of three types of
messages: (i) Sensor node broadcast
messages: this message containing the
identity of sensor node and its location is
periodically broadcast to the one-hop
neighbors sensor nodes. Upon receiving
such messages, any sensor node of the
network can determine the list of its
106
guardees (its one-hop neighbors). Hence,
the
guardian-guardee
relationship
between sensor nodes is established, (ii)
Manager broadcast messages: The
manager robot broadcasts periodically its
identity over the network (to all sensor
nodes and maintainer robots) and (iii)
Maintainer reply messages: When a
maintainer robot receives a manager
broadcast message, it responds with a
message containing its identity and
location. Hence the manager can
determine the location of all maintainer
robots in the network.
To establish a guardian-guardee relation
between nodes in WSN, the centralized
manager algorithm [1] supposes that a
guardee node selects the nearest node as
a guardian. This choice presents some
drawbacks. Indeed, if a failure occurs on
the guardee node that is the relay point
for the guardian to the manager robot,
the report failure cannot be sent to the
manager robot. In our strategy, we
propose that any node is guarded by all
its one- hop neighbors. In this case we
guarantee that at least one report failure
is received by the manager.
After initialization, any sensor node
periodically sends its broadcast message
(message type 1) to its one-hop neighbor
nodes. If a guardian has not received any
messages from a guardee for a certain
amount of time, the guardian deduces
that the guardee has failed and sends a
report failure containing the identity and
the location of failed node to the
manager robot (the failure detection and
reporting stage).
To report failures, Mei and all in [1]
have used a geographic routing protocol
(knowing that we assume network
connectivity and area coverage). But
upon multiple simultaneous failures, this
assumption is not realistic. So, using the
nearest node to the manager as a relay
point by a guardian node to send the

report failure may not be a good choice
to communicate with the manager. For
this reason, we have used in CMS an
adhoc routing protocol such as AODV
(Ad hoc On Demand Distance Vector)
[10] to ensure that the report failure is
received by the manager (if a route
exists).
In the last stage (Failure handling) and
upon the reception of a failure report, the
manager selects the maintainer robot
whose current location is the closest to
the failure, and sends an order for this
robot to handle the failure. Once the
maintainer has finished the task, it sends
to its manager robot a message
containing its new location.
In [1], to handle failure, the manager
selects the nearest maintainer robot to
the failure regardless of its state (busy or
available) and the requests are served by
the maintainer robot as first-come-firstserved. Using this technique, we risk an
overload on robots which are close to the
sink node, given that the network
activity of nodes (and consequently the
failure by energy depletion) increases in
neighborhood of the sink node. In our
strategy, we suppose that the repair
requests queue is managed by the
manager robot. Then, to handle failure,
the manager robot selects the closest
available maintainer robots to repair a
given failure to ensure load sharing over
robots.
3.2 The Centralized Maintenance
Strategy with Anticipation
The main goal of the CMSA is to predict
the energy failure. To achieve this goal
we must compute the energy consumed
by the sensor in the future. We use then
our analytical energy model presented in
section 2. Each sensor node must first
107
measure
the
different
transition
probabilities (
). Once
the
transition
probabilities
are
determined, they are sent with the value
of the residual energy
and the node
location to the manager robot. With the
reported measurements, the manager can
compute the lifetime of sensors using the
analytical model. Indeed, it is equivalent
corresponding to:
to finding the time
(12)
Where:
is calculated according
to the equation 11.
Therefore, with our proposed analytical
model, the manager robot can first
estimate the N expected sensor failures
in the network designated by
). To determine the maintainer
(
, the
robot of an estimated failure
robot manager identifies the list of
robots able to handle the anticipated
failure at its required replacement time
with a null off-service time and
selects among them the nearest robot as
the
maintainer robot. To ensure a null
off-service time for the failure
, the
must be
selected maintainer robot
scheduled to repair the expected failure
at:
(13)
With:
-
the distance between the
and .
current position of
represents the speed of
Similarly to the CMS, we have three

stages: the initialization stage, failure
detection and reporting and failure
handling. However, in the initialization
stage, we have an additional type of
messages:
Measures
transition
probabilities messages that contain the
transition probabilities matrix sent by

sensor nodes to the manager robot. In the
failure handling stage, the manager robot
can send an order to repair the failure
before it occurs. This order is based on
the analytical model computed for each
node sensor.
4 EXPERIMENTS
In this section, we validate our analytical
model to estimate the energy
consumption of a sensor node with
different MAC layer protocols. We also
compare the performances of the two
coordination strategies presented in this
paper.
4.1 Experimental Setup
We have implemented our analytical
model and the two coordination
strategies in the NS-2 Simulator [11].
We have implemented also an ondemand mobility model in which robots
move on demand after receiving a failure
report. The failure detection, failure
report, failure repair request and the
analytical model are implemented at the
application level. We have used AODV
as a routing protocol.
We have selected the following
simulation parameters: the sensor area is
x
with a total of
sensor
nodes; the robots speed is
, based
on the specification of Pioneer 3DX
robots [12]; the number of robots varies
from to
robots; the simulation time
seconds and the sensor node
is
generates a traffic with a constant bit
rate and a constant packet size of 128
bytes.
IEEE 802.15.4 is used as a sensor MAC
Layer in the WSN. The sensor power
consumption in each MAC layer state
(
,
,
and
)
108
used in simulation is equal to (0.1404
Watt, 0.1404 Watt, 0.0018 Watt,
0.000018 Watt). References values are
taken from a ZigBee node implementing
IEEE 802.15.4 medium access.
4.1 The Network Dysfunction Time
In a first scenario, we suppose that
sensor failures are caused only by energy
depletion. Experiments are run with
different bit rates per sensor node: 1, 2
and 3 kb/s. With CMS and CMSA,
simulation has shown that we must have
at least 2 robots to guarantee the network
connectivity and coverage of the
considered area. We designate by
network dysfunction time, the sum of
duration for which the network
connectivity and the coverage are not
guaranteed. In other words, its the
period during which at least one node is
failed in the WSN. Figure 1 represents
the variation of the percentage of the
network dysfunction time when CMS
and CMSA are used for different number
of maintainer robots with different node
bit rates (1, 2 and 3 Kb/s).
the network connectivity and coverage

of the considered area with a null
network dysfunction time for the
considered node bit rates: 1,2 and 3
Kb/s.
Minimizing the network dysfunction
time is the primary requirements for
real-time applications. To achieve this
goal by using CMS, the solution, as the
figure 1 shows, is to increase the number
of maintainer robots in the WSN. This
solution cannot be retained not only
because it is expensive but also because
the improvement given by this solution
(on the network dysfunction time)
remains always limited by the failure
detection time and the failure handling
time spent by robot to move to the
occurred failure.
On the other hand, simulation results
show that CMSA can provide a null
network dysfunction time with the
minimal number of robots used by CMS
if only energy depletion failures are
considered.
2. CMS vs CMSA: The network dysfunction

time versus the number of maintainer robots
with unpredictable failures.
1. CMS vs CMSA: The network dysfunction

time rate versus the number of maintainer
robots with energy depletion failures only.
Using CMSA, figure 1 shows that only 2

robots are fully sufficient to guarantee
But generally, failures occurred in

WSNs are caused by many reasons other
than the energy depletion such as
hardware failures. To reflect this kind of
failures in the simulation, we propose a
109
second scenario where we add random
failures generated according to a Poisson
distribution with an average of 2 failures
per hour. Therefore, two types of failures
can occur: failures caused by energy
depletion which are anticipated and
repaired before they happen and
unpredictable failures which are detected
by guardian nodes and reported to the
manager.
We represent in figure 2 the network
dysfunction time given by CMS and
CMSA for different number of
maintainer robots and a fixed node bit
rate equal to 3Kb/s. This figure shows
that CMSA reduces considerably the
network dysfunction time using the
minimal number of robots. In fact, the
network dysfunction time provided by
CMSA is due only to the repair time of
unpredictable failures generated during
the simulation, which explains the low
percentage of the network dysfunction
time compared to CMS.
the initial sensor number plus the

number of replaced nodes.
3. Message overhead versus node bit rate.
However, the Centralized Manager

Strategy gives a high rate of signaling
messages that increases considerably
with the failure number. In fact, using
the original strategy, all guardians of a
failed node must send a report failure to
the manager robot. This explains the
important number of signaling messages.
4.1 Message Overhead

6 CONCLUSION
The message overhead is measured as

the number of transmissions incurred by
failure detection, failure reporting, and
coordination messages. Since the two
strategies are similar in number of
coordination messages, we focus on the
overhead introduced by failure detection
and failure reporting.
For CMS, when a failure happens, all the
guardians (the one-hop neighbors) of
faulty node report the failure to the
manager robot. However, the use of
CMSA, implies one message for each
sensor node sent to the manager
containing the probability transition
matrix at the initialization stage.
As the figure 3 shows, the number of
messages induced by CMSA is equal to
In this paper, we have presented a

proactive maintenance strategy for faulttolerant wireless sensor networks. We
have presented an analytical energy
model to anticipate sensor failures, in
which each node of the sensor network
is modeled by a Markov Chain. The state
of the model depends on the used MAC
layer.
Based on the proposed energy model, we
have investigated two versions of failure
repair strategies: CMS and CMSA. Two
types of robots are distinguished: a
manager robot and several maintainer
robots. In the CMS version, the manager
receives a report failure from a guardian
node upon a failure. Thereafter, it selects
the closest maintainer robot to handle the
occurred failure. For a real-time sensor
110
network, the reaction time upon a failure
cannot be tolerated as simulation results
have shown. To remedy this problem,
we have proposed an anticipated version
of Centralized Manager Strategy:
CMSA. In this technique, the manager
receives the transition probabilities
matrix from each node sensor in the
network. The transition probabilities are
used by the analytical model to estimate
the lifetime of sensor nodes. Before the
estimated sensor lifetime expires, the
manager sends an order to a maintainer
robot to replace the corresponding
sensor. Simulation results have shown
that using this technique, we obtain a
null dysfunction time of the wireless
sensor network with a minimal number
of robots.
In a future work, we propose to improve
the two distributed strategies given in [1]
to coordinate the movement of robots
based on fixed distributed manager
algorithm and dynamic distributed
manger algorithm. Thanks to our
analytical model predicting the lifetime
of a sensor node, these strategies will be
made able to support real-time
applications.
5.
6.
7.
8.
9.
B. Bougard, F. Catthoor, C. Daly, A.

Chandrakasan, and W. Dehaene, Energy
efficiency of the IEEE 802.15.4 Standard in
dense wireless microsensor network:
Modeling and improvement perspectives ,
Design,
Automation
and
Test
in
Europe,2005.
W. Ye, J. Heidemann and D. Estrin , An
Energy-Efficient MAC
Protocol for
Wireless Sensor Networks, IEEE Infocom
2002
S. Tang, An analytical traffic flow model for
cluster-based wireless sensor networks,
International Conference on Pervasive
Computing, 2006.
S. Mahfoudh, Energy effciency in wireless
ad hoc and sensor networks: routing, node
activity scheduling and cross-layering, thesis
INRIA 2010.
A. Nemra and N. Aouf, Robust INS/GPS
Sensor Fusion for UAV Localization Using
SDRE Nonlinear Filtering, Sensors Journal ,
2010.
10. C. Perkins, E. Belding-Royer and S. Das, Ad hoc

On demand Distance Vector (AODV) Routing,
IETF RFC 3561, 2003.
11. The Network Simulator ns-2,
http://www.isi.edu/nsnam/ns/
12. Y. Mei, Y.-H. Lu, Y. C. Hu, and C. G. Lee, A
Case Study of Mobile Robots Energy
Consumption and Conservation Techniques, In
International Conference on Advanced Robotics,
pages 492497, 2005.
6 REFERENCES
1.
2.
3.
4.
Y. Mei, C. Xian, Y. Charlie , Repairing

Sensor Network Using Mobile Robots,
Workshop on Wireless Ad hoc and Sensor
Networks, 2006.
Y. Gao, K.Wu, and F. Li, Analysis on the
redundancy of wireless sensor networks,
ACM international conference on Wireless
sensor
networks
and
applications,
September 2003.
A.Kumar, E. Altman, D. Miorandi, and M.
Goyal, New Insights from a Fixed Point
Analysis of Single Cell IEEE 802.11
WLANs, In Proceedings of IEEE
INFOCOM, 2005.
IEEE 802.11 standard for Wireless Local
Area Networks,
http://standards.ieee.org/getieee802/802.11.html.
111
Roman Law and Contemporary Legislation

Wolfgang Kahlig, Eleonora Kahlig
CONTAKT-EDV AG
Academy of software-development and structuring of legal relationships
kahlig@attglobal.net wolf.kahlig@edu.fw-wien.ac.at
ABSTRACT
Simple rules and laws are necessary for
the understanding of the meaning and of
the content. This is a fundamental claim of
all citizens. The instruments of legislation
are deeply connected in many countries to
the principles of the Roman law and to the
ideas of the former centuries. Modern
effective aspects, such as visualization,
structuring and object-orientation are not
focused. One method for handling
complex projects is UML, the Unified
Modeling Language. This tool has served
for some years as an instrument in
understanding the complex structures of
the law, but could also be applied to the
complete establishing, admini-stration and
development of a new and modern
building of law.
KEYWORDS
Rules for legislation, visualization,
structuring, object-orientation, UML
1 INTRODUCTION [ 1]
The law primarily has a linguistic
representation
especially
within
language-constructs. This is changing,
and the representation is becoming
more multi-layered, both in terms of
the form (text corpus as a legal
information system, citizen portals,
formalization,
structure,
images,
graphics, input forms, etc.) as well as
the legal regime (international law, EU
law, national law, etc.). A linguistic
transformation is taking place in the
concretization of general norms by
practice and courts. The law is

becoming multi-layered and multisensory: To the language with its
different terms are added pictures,
flow charts, videos, graphics etc. but
also semantic, automatically processed
versions (legal ontologies, logical
structures, etc.). This diversity makes
it possible to reach all population
groups and to support semantic and
automatic processing of content by
using
the
means
of
electronic instruments.
2 MULTI-LAYERED
REPRESENTATION
Traditional Legislative Process
In the former centuries according to
the rules of the Roman law - the law
primarily has a linguistic representtation especially with languageconstructs. Examples of this linguistic
representation can be found in the
whole contemporary law in the most
counties.
Table 1
E.g. the Constitution of the United
States, the supreme law of the U.S.A.
[ 1 ] The leading topic of IRIS2012, Erich

Schweighofer, august 23th, 2011
112
Table 2
E. g. the German Tenancy Law:
Table 3
E.g. the French Code Napoleon:
In these examples you dont have to

read or to understand the text, they
only will demonstrate the principle of
using just words and sentences.
In the majority of cases there arent
any pictures of the structural context,
no images, no graphics or any forms.
According to the notation in former
centuries we often get very long and
nested
sentences,
which
are
incomprehensible for jurists too. The
average of our citizen has to spend a
lot of time reading the legal text over
and over to get an idea, what the writer
is trying to say or express.
3 POSSIBLE METHODS OF
ANALYZING
3.1.1 Methods Of Other Sciences
Object Orientation is an
attempt to develop structures of
organi-zation as well as
software by classification of
information
using
their
characteristics and the possible
operations. An object-oriented
design is part of object oriented
methodology and forces the
analyst to think more in terms
of objects and less in
procedures. An object contains
encapsulated
data
and
procedures which are grouped
together to represent an entity.
Therefore
the
interface
between the objects is exactly
defined.
Methods such as UML, the

Unified Modeling Language
are models of standardized
visual specification language
for object modeling and a
general-purpose
mode-ling
language that includes a
graphical notation. This tool
allows you to create an abstract,
Table 4
E.g. the Italian Tenancy Law:
113
but standar-dized model of
complex systems.
3.1.2
Fundamental Aspects.
Information, what should be done or

what should be omitted in certain
situations must be described in such a
clear and unmistakable manner, that
both,
persons
with
different
educational background and also
computers, can understand the rules
and know how to carry them through.
We can recognize governmental efforts
in creating rules for rules and
establishing
a
more
variable
communication
and
information
system. A lot of administrative
functions were built. But all these
efforts need urgent completion in the
realm of legislation and execution.
The form, the logical processing and
the semantic structuring of many parts
of the existing set of rules, regulations
and laws are often based on a
traditional shaping. The knowledge of
the modern sciences is not or only in a
minimal
degree
taken
into
considerations. Overall views and
structures are not implemented enough.
Instead of clear surveys and
instructions the citizen often gets
excessively long and nested sentences.
The following example should
illustrate the problem (Austrian Law,
46c MRG, just the first sentence). You
dont need to understand this text (in
language or meaning), please consider
only the syntax and the length:
46c. Wenn (IF) die Voraussetzungen des 16 Abs. 1 nicht
vorliegen, sind dennoch (THEN)
Verein-barungen ber die Hhe des
Hauptmietzinses fr eine Wohnung
ohne die Beschrnkungen des 16
Abs. 2 bis 4 und 6 bis zu dem fr die
Wohnung nach Gre, Art,
Beschaffenheit, Lage, Ausstattungs-
und Erhaltungszustand angemessenen Betrag auch weiterhin zulssig,

wenn (IF) der Mietgegenstand eine
Wohnung der Ausstattungskategorie
A oder (OR) B in ordnungsgemem Zustand ist, deren
Standard vom Vermieter nach dem
31.
Dezember
1967
durch
Zusammenlegung von Wohnungen
der Ausstattungs-kategorie B, C
oder D (OR / OR / OR), durch eine
andere bautechnische Aus- oder
Umgestaltung greren Ausmaes
einer Wohnung oder mehrerer
Wohnungen
der
Ausstattungskategorien B, C oder D oder sonst
(OR / OR / OR) unter Aufwendungen
erheblicher
Mittel
angehoben wurde, oder (OR) wenn
(IF) der Mietgegenstand eine
Wohnung
der
Ausstattungskategorie C in ord-nungsgemem
Zustand ist, deren Standard vom
Vermieter nach dem 31. Dezember
1967 durch Zusammenlegung von
Woh-nungen der Ausstattungskategorie D, durch eine andere
bautechnische
Ausoder
Umgestaltung greren Ausmaes
einer Wohnung oder (OR) mehrerer
Wohnungen
der
Ausstattungskategorie D oder (OR) sonst unter
Aufwendung erheblicher Mittel
angehoben wurde, sofern (IF) der
Vermieter
die
Arbeiten
zur
Standardanhebung vor dem 1.
Oktober 1993 tatschlich begonnen
hat.
Let us look only at this first sentence
of the paragraph. MS WORD counts
in the original text 182 words and 1134
signs (without space characters). This
sentence alone is composed of 20 lines
(in the original text) and shows
multiple insertions, interlocks and
cross-references. However, it is almost
impossible to realize the structure, any
connections and the content of such a
representation.
114
3.1.3
Objective Facts
The main goal for a citizen with an

average ability must ultimately be to
be able to quickly and surely recognize,
what to do and what to leave. A
judicial system should offer a citizen a
simple route on how to answer the
main questions, without having longer
analyses or studies.
This goal can naturally be pursued by
the simplification of the laws.
However
this
may
cause
a
disadvantage, such as a simple law
may be an unfair law. [2]
3.2 Suggested Solutions
Meta-logical structures have to be
formed - logical structures and
connections, a warehouse of all similar
and/or connected cases and very
comfortable searching algorithm,
which gives not only answers to the
key-words but also to the complex
connections and comparable facts.
Establishing of searching - catalogues
with an application of standardized
words (words in the infinitive, by
ignoring
of
conjugation
and
declension).
References to fundamental law,
ordinances, guidelines, general rules
are important.
Requirements for the inquiry by direct
and superior structures are relational
databases, networked computer in
network-structures, both national and
worldwide a networking of universities
as well as research centers.
It is essential that logical connections
could be available, surveys and above
connections can be formed. Each
regulation-process, but also each rule
must be inserted into a logical
structure. The basic idea was that the
publication of the law must be
2
[ ] speech at Kepler-University, Linz,

December 2005, Einfachere oder gerechtere
Gesetze, Kahlig Wolfgang
extended with electronic means. The

further mechanism should integrate the
new rule into a logical concept.
3.3 Representation Of Structures
Already with a structured order, the
items and the parts of sentences can
easily be reached. The connections are
more easily recognized and logical
structures can be obvious so that the
content can quickly be surveyed. For
instance, the
1 MRG could be
presented in such a structure (without
change of mind):
1. (1) MIETE /
GELTUNGSBEREICH
Dieses Bundesgesetz gilt fr die Miete
von
- Wohnungen,
- einzelnen Wohnungsteilen oder
- Geschftsrumlichkeiten aller Art
(wie im besonderen von
Geschftsrumen,
Magazinen,
Werksttten,
Arbeitsrumen,
Amts-und Kanzleirumen)
Now there is no problem to recognize
that the items
Wohnungen
einzelnenWohnungsteilen
Geschftsrumlichkeiten
do have the same importance or value
in the first level, but the items
Geschftsrumen,
Magazinen,
Werksttten,
Arbeitsrumen,
Amts-und Kanzleirumen)
show the same importance or value in
the second level.
115
3.4 Strategies
would be identical with:
The further automation, streamlining
and simplification will form the basis.
Fundamentally, all rules should be
formulated in the manner of a daily or
with a "check list" and should be
simply readable. The principle of the
avoidance of lengths and nesting
should be considered consistently, as
most of the Legistik demand. New
techniques are necessary, like charts
and tales, displayed clearly on Excelsimilar pictures. The visualization [3]
by surveying pictures and the means of
structure-diagrams, flow charts and
graphic diagrams must be raised to the
principle.
Also, the important logic - notation
with conscious and visible marking of
the logical AND, the logical OR
and the building of parenthesis is of
decisive meaning.
It must be
considered, that the combination
between sentences and/or parts of
sentences is made with different
conjunctions. So an AND (=UND)
connects more intensively than an
OR (=ODER). A reversal of this
rule is possible, as shown in the
following example:
If the event A happens and
the event B happens or
event C happens, then event
XY would happen.
[3]
Heindl / Kahlig / Stingl Wohn- und Steuerrecht

anschaulich, Manz, 2004, Wien
Heindl / Kahlig: Mietrecht automatisch, Manz, 2004,
Wien CD-Ausgabe
Heindl / Kahlig: WEG2002 automatisch, Manz, 2004,
Wien CD-Ausgabe
Kahlig: Mietrecht einfach, Eppenberg, Wien, 1997
Kahlig:
Handbuch
des
Immobilienwesens,
CONTAKT, 2004, Loseblattausgabe
Lachmayer: Mglichkeiten einer Verwendung
normentheoretischer Analysen fr die Gesetzgebung,
Studien zu einer Theorie der Gesetzgebung, SpringerVerlag Berlin-Heidelberg-New York, 1976
Lachmayer: Regeln zur grafischen Darstellung des
Rechts, sterreichisches Anwaltsblatt, 1976
Lachmayer:
Grafische
Darstellungen
im
Rechtsunterricht, Zeitschrift fr Verkehrsrecht, 1976
If (the event A happens and

the event B happens)
or
(the event C happens), then
event XY would happen.
A completely different sense is yielded
after changing the parenthesis:
If (the event A) and (the event B
happens or the event C happens),
then event XY would happen.
In this case the event C would now
not more determining (not more
sufficient)!
3.5 Modelling And Structuring
In literature it is often pointed out, that
the conjunctions like AND / OR /
AS
WELL
AS
are
not
unambiguously used. It is problematic
because the AND - connection is
sometimes used logically (within the
meaning of a logical content) and
sometimes linguistically, for instance
sometimes as careless (in connection
with logical aspects) but usual and/or
completion of a chain of OR
connections.
Example of 1 MRG:
1. (1) Dieses Bundesgesetz
gilt fr die Miete von
Wohnungen,
einzelnen
Wohnungsteilen
oder
aller Art () samt den etwa
mitgemieteten () Hausoder Grundflchen () UND
fr die genossenschaftlichen
Nutzungsvertrge
ber
derartige Objekte ()
116
The
last
connection-word
(UND/AND) represents only a
linguistic usage and not a logical
AND.
The logical content should mediate us
really a logical enumeration, therefore
the above-mentioned AND is just a
connection for the enumeration. The
law is intended to inform you, - in
spite of this not-logical AND
(und) that this law is made for
AND
fr die genossenschaftlichen
Nutzungsvertrge
But if you are following the exact
logistical rules, you get: [4]
1. (1) Dieses Bundesgesetz gilt
nur, wenn es sich handelt um
(this law is only valid, if)
die Miete von Wohnungen
OR (VEL)
einzelnen Wohnungsteilen
OR (VEL)
Geschftsrumlichkeiten aller
Art
OR (VEL)
fr die genossenschaft-lichen
Nutzungsvertrge

OR (also)
einzelnen Wohnungsteilen
OR (also)
aller Art ()
OR (also) !!!
fr die genossenschaft-lichen
Nutzungsver-trge
ber
derartige Objekte
OR has 2 meanings, namely a
EXCLUDING
OR
(formerly,
Roman Law: AUT) and the
LISTING OR (formerly, Roman
Law: VEL)
Often the logical connections are
mixed up with the linguistic
connection. Therefore you can hear the
idea, that you could write the sentence
thoroughly with AND.
This assertion appears in the first
moment OK, if you read:
1. (1) Dieses Bundesgesetz
gilt fr die
Miete vonWohnungen
AND
einzelnen
Wohnungsteilen
AND
aller Art
Ultimately with the necessary /

sufficient rules you get:
1. (1) Es ist NOTWENDIG,
dass es sich um (it is necessary that..)
OR (VEL)
die
Miete
von
einzelnen
Wohnungsteilen
OR (VEL)
OR (VEL)
fr
die
genossenschaftliche
Nutzungsvertrge
handelt, damit das Bundesgesetz gilt.
(Just in these cases you get legal
force)
And still more clearly with the
conditional if / then notation:
1. (1) WENN es sich handelt um
(IF you have a .)
[
] Rechtsmodellierung im E-Government,
Fallbeispiele zur Legistik (2005), Diss, Kepler-Universitt
Linz
117
OR (VEL)
die Miete von einzelnen
Wohnungsteilen
OR (VEL)
..
DANN
(THEN)
ist
das
Bundesgesetz anzuwenden.
(Just in these cases you get legal
force)
Thereby it would be clear, that there
is only the logical listing OR (=
VEL) and not the excluding OR (=
AUT) and not a logical connection
with AND. (The logical AND would
necessitate,
that
you
SIMULTANEOUSLY must have all
the above conditions, as
Never use the connections

AND/ OR, but always
ET (logical AND)
AUT (logical excluding OR)
VEL (logical listing OR)
Nevertheless it is possible, that this
stringent rule takes some time to get
used to. Therefore a minimal demand
could be:
AND usage just as the
logical AND
OR usage just as the
logical excluding OR
comma usage just as the
logical listing OR
3.6 Conditional Model

AND SIMULTANEOUSLY
die Miete von einzelnen
Wohnungsteilen
AND SIMULTANEOUSLY
die Miete von
Geschftsrumlichkeiten..
AND SIMULTANEOUSLY
genossenschaftliche
Nutzungsvertrge
But this would not be possible,
because the terms are absolutely
excluding terms.
The mentioned example above, at
which in the construction the and
appears seems in the first moment
plausible, but this and is actually a
and also, which would be to be
equated with the logical OR
(=lat.VEL).
Now we are able to postulate:
To avoid brackets it is necessary

that the authors of law make
intellectual acrobatics, which are not
just beneficial to the legibility.
Another example for the severe logical
arrangement would be the imperative
introduction of the IF / THEN / ELSE
clause.
3.7 Models Selection
There are a lot of methods to show the
logical structure or to visualize the
content of
3.7.1 Institutions
(reference: www.univie.ac.at)
Every year the International Legal
Informatics Symposium (IRIS) took
place at the Law Faculty of the
University of Salzburg, in this year
from 21-23 February 2013. Already
being in its 14th year, IRIS has been
established as the largest and most
important academic conference on
computers and law in Austria and
Central Europe.
The leading topic of IRIS2013 was:
118
Abstraction and Application
In law, abstraction should give a
focused view on the essential, i.e. the
concrete
elements
and
their
characteristics receive a certain
conceptual meaning.
Generalization and their formal
representation are the basis for the
multi-layered representation of the law
(text, commentary, legal ontologies,
logical structures, etc.).
The application is an application
software program to solve user
problems. The application is the
practice of theory and abstraction.
The software can be very complex
reflecting its importance in the
knowledge society. The app is an
application for smart phones and
tablet computers
(e.g. RIS: app to search in the
consolidated version of the federal
law).
While the abstraction has a very long
tradition, the application and the app
are a recent development. Complex
processes should be supported by
means of a user-friendly smart phone
application, which in turn requires an
abstraction (and simplification). In law,
the challenge of the app is presenting
complex search queries and documents
in a simple and user-friendly form.
One of the goals of the IRIS2013 was
to highlight the interaction between
abstraction and application in law and
to find appropriate practical solutions.
The IRIS conference is well known for
its interdisciplinary approach and the
involvement
of
administration,
business and civil society.
structured into single steps and could

be restructured. Specialized sym-bols
derivated from UML are able to
structure with the following structureelements:
Figure 1: Begin of the facts of the case.
Figure 2: YES/NO - decision
Figure 3: Action
Figure 4: End of the facts of this case.
3.7.2 Practical Techniques

The dismantling by little steps leads
the project to its intended goal. Every
Computer works only with the
YES/NO decision, therefore it is
obvious, that the whole law can be
119
Jurisprudence is underdeveloped,
which limits itself to dogmatic)
4.2.2 Initiatives
Figure 5: Workflow- element
Often a pretext is pleaded to block

innovations.
A classical, historic
example is the rejection of printing
books in the former centuries. Each
new idea naturally has advantages and
disadvantages, which must be weighed
out reciprocally. From the current
view the printing would not have been
stopped with such extreme resistance.
Outdated fears, which look to block
the progress, must be disassembled.
4.2.3 Concrete And Neutral Controlelements
It would be essential that a control
authority would be arranged for a very
broad basis.
Figure 6: Logical connection
5 CONCLUSION
These 6 elements could manage all

interrelations.
The Legistik - rules must be completed

thus after the above-mentioned
postulates. [5]-[26] Creation of a
thesaurus on the basis of objectoriented models will be helpful. A
powerful modeling language, as UML
would help to create the whole system.
After the installation and training of
the described methodologies the
running care and a relative simple
expansion would possible. Thereby
the rules could become simpler and
more understandable and will help to
avoid some of the expenses. New
models, as shown above will have a
central position.
4 OUTLOOK AND CONCRET

STEPS
4.1 National And International
Defined, Logical Elements Of Law.
The education of teams and the
development of superior, international
elements will be inevitable.
4.2 Principals
4.2.1 Acceptance
The embarking on of unusual ways
must be allowed and it must become
the rule. The pertinent and neutral
judgment of the foreign opinion and/or
conviction is essential. (Quotation
Ludwig
Adamovich:
This
120
Additional References
1. Schweighofer, E.: The leading topic of
IRIS2012, , august 23th, Salzburg, (2011)
2. Kahlig, W.: Speech at Kepler-University,
Linz, Einfachere oder gerechtere Gesetze
(2005)
3. Lachmayer, F.: Mglichkeiten einer
Verwendung normentheoretischer
Analysen fr die Gesetzgebung, Studien zu
einer Theorie der Gesetzgebung, SpringerVerlag Berlin-Heidelberg-New York
(1976)
4. Kahlig, W.: Rechtsmodellierung im EGovernment, Fallbeispiele zur Legistik,
Diss, Kepler-Universitt Linz (2005)
5. Adamovich, L.: Probleme einer modernen
Gesetzestechnik, in: Winkler, G. /
Schilcher, B.(Hsg.), Gesetzgebung, Wien
(1981)
6. Heindl, P., Kahlig, W.: Mietrecht
anschaulich, Manz, Wien, (2001)
anschaulich, Manz, Wien (2002)
8. Heindl, P., Kahlig, W.: Wohnrecht
anschaulich, Manz, Wien (2003)
9. Heindl, P., Kahlig, W., Stingl W.: Wohnund Steuerrecht anschaulich, Manz, Wien
(2004)
10. Heindl, P., Kahlig, W., Stingl W.: Wohnund Steuerrecht anschaulich Gesamtausgabe, Manz, Wien (2004)
automatisch, Manz, Wien CD-Ausgabe
(2004)
12. Heindl, P., Kahlig, W.: WEG2002
automatisch, Manz, Wien CD-Ausgabe
(2004)
13. Kahlig, W.: Mietrecht einfach, Eppenberg,
Wien (1997)
14.Kahlig, W.: Handbuch des
Immobilienwesens, CONTAKT,
Loseblattausgabe (2004)
15. Kahlig, W.: Steuer-Handbuch des
Immobilienwesens, CONTAKT,
Loseblattausgabe (1996)
16. Stingl,W., Kahlig, W.: Beibltter im
Steuerrecht anschaulich, Manz, (2004)
17. Stingl, W., Kahlig, W.: Beibltter im
Steuerrecht automatisch, Contakt (2005)
18. Lachmayer, F., Reisinger, P.: Legistische
Analyse der Struktur von Gesetzen, Manz,
Wien (1976)
19. Lachmayer, F.: Grafische Darstellungen als
Hilfsmittel des Gesetzgebers,
Gesetzgebungs-theorie, Juristische Logik,
Zivil- und Prozessrecht, Springer-Verlag
Berlin-Heidelberg-New York (1978)
20. Lachmayer, F.: Visualisierung des Rechts,

Akten des 2.Semiotischen Kolloquiums,
Regensburg (1978)
21. Wimmer, M.,,Traunmller, R., Lenk, R.:
Electronic Business invading the Public
Sector, Proceedings of the 34th Hawaii,
International Conference on System
Sciences, Hawaii (2001)
22. Kahlig, W.: Rechtsmodellierung im eGovernment, vdm Mller (2008)
23. Kahlig, W.: Klagenfurter
Legistikgesprche, ABGB 2011
strukturiert, Klagenfurt (2009)
24. Heindl, P., Kahlig, W., sterreicher, T.,
Sommer, A.:WGG strukturiert, Manz
(2011)
Sommer, A.: WGG II strukturiert, Manz
(2012)
Sommer, A.: WGG - Navigator, Manz
(2012)
Additional Speeches/ Lectures

27. Kahlig, W.: Sequentiell strukturiertes
Mietrecht, Juridicum Wien, (1999)
28. Kahlig, W., Heindl P.: Linear
strukturiertes Wohnrecht, Universitt
Salzburg, Salzburger
Rechtsinformationsgesprche (2001)
29. Kahlig, W., Heindl, P.: Mietrecht einfach,
Kitzbhel, Kongress des sterr. Verbandes
der Immobilientreuhnder (2001)
30. Kahlig, W.: Artikulations-methoden und
Strukturierungs-verfahren im 3.
Jahrtausend, Universitt Wien (2001)
31. Kahlig, W.: Visualisierung und logische
Strukturen am Beispiel des Wohnrechts,
Universitt Salzburg, IRIS-Kongress
(2002)
32. Kahlig, W.: Der juristische Thesaurus,
(2003)
33. Heindl, P., Kahlig, W.: Das Wohnungseigentumsgesetz plakativ,
Vorarlberg, Schruns, VI Winterseminar
(2003)
34. Kahlig W.: Strukturelle Visualisierung
eines juristischen Fallprfungsschemas,
Universitt Innsbruck (2003)
35. Kahlig, W.: Logik und Rechtssetzung,
Universitt Wien, Juridicum (2003)
36. Kahlig, W.: Logik und Rechtssetzung,
Bundeskanzleramt, Wien (2003)
37. Kahlig, W.: Legistische Richtlinien 1990,
Problemaufriss, mgliche Ziele, BKA,
Freskensaal, im Rahmen der Vortragsreihe:
121
Reform der Standards fr die Gestaltung
von Rechtsvorschriften (2003)
38. Kahlig, W.: Content Visualisierung,
Johannes Kepler Universitt Linz (2003)
39. Kahlig, W.: Netzplananalysen von
Gesetzesentwrfen, Krntner
Landesregierung, Klagenfurter LegistikGesprche (2003)
40. Kahlig, W.: Visualisierung des
Wohnrechts im Spannungsfeld zwischen
dem Rm. Recht und den Kantschen
Thesen, Universitt Salzburg, IRISKongress (2004)
41. Kahlig, W.: NetzplanVisualisierung,
Universitt Innsbruck, Arbeitswissenschaftliches Institut (2004)
42. Heidl P., Kahlig, W.: Legistik im 3.
Jahrtausend, Prtschach, Bundestag der
Immobilientreuhnder (2004)
43. Kahlig W.: Rechtsmodellierung am
Beispiel des Wohnrechts, Johannes
Kepler -Universitt Linz (2004)
44. Kahlig, W.: Begriffsvisualisierung am
Beispiel des Wohnrechts, Universitt
Salzburg, IRIS-Kongress (2005)
45. Kahlig, W., Stingl, W.: Das Wohnrecht im
3. Jahrtausend, Winterseminarwoche des
VI (sterr. Verband der Immobilientreuhnder), Schladming 2005
46. Kahlig, W.: Das Wohn- und ImmobilienSteuerrecht im 3. Jahrtausend, FHW,
Lehrgang fr Immobilientreuhnder
(2005)
47. Heindl, P., Kahlig, W., Stingl, W.: Das
Wohn- und Immobilien-Steuerrecht
anschaulich Der neue, vereinfachte
Zugang zu den Wohnrechtsgesetzen,
Gewerbehaus der Handelskammer Wien
(2005)
48. Kahlig, W., Stingl, W.: Vorteile im
Immobilien-Steuerrecht durch
CONTHAUS-Software, IBM-Symposium
Messegelnde Wien (2005)
49. Kahlig, W.: Ab wann werden nur mehr
Computer Gesetze machen?, Universitt
Innsbruck (2005)
50. Kahlig, W.: Automatisierte Verfahren im
Immobilien-Wohnrecht, FHW,
Fachhochschule Wien, (2005)
51.Kahlig, W., Stingl, W.: Das ImmobilienSteuerrecht,
Bundestag der Immobilientreuhnder in Loipersdorf (2005)
52.Kahlig, W.: Wohn- und ImmobilienSteuerrecht fr Praktiker Der neue
Landesinnung Tirol, Innsbruck,
Fachgruppe (2005)
53. Kahlig, W.: Wohn- und ImmobilienSteuerrecht fr Praktiker Der neue
Landesinnung Krnten, Klagenfurt,

Fachgruppe (2005)
54. Kahlig, W.: Gesetzesstrukturen und
Legistikregeln, Landhaus Krnten, 3.
Krntner Legistikgesprche (2005)
55. Kahlig, W.: Unsere Gesetze im Spiegel
der Legistik, Universitt Linz, Festvortrag
(2005)
56. Kahlig, W.: Das Wohn- und ImmobilienSteuerrecht anschaulich Der neue,
vereinfachte Zugang zu den
Wohnrechtsgesetzen,
Landesinnung Vorarlberg Fachgruppe
(2005)
57. Kahlig W.: Alternative
Visualisierungskonzepte am Beispiel des
Wohnrechts, Universitt Innsbruck (2006)
58. Heindl, P., Kahlig W.: Abrechnungsfragen
im Wohnrecht, Winterseminarwoche des
VI, Tirol, tztal (2006)
59.Kahlig W.: Gesetzgebung einfach, Vortrag
vor Landeshauptmann O, Dr. Phringer,
Landesregierung O (2006)
60. Kahlig W.: Vereinfachung von Gesetzen in
der Europischen Union, Europaparlament
Brssel (2006)
61. Kahlig W.: Prinzipien formaler
Konstruktion von juristischen Texten,
62. Kahlig, W.: Automatisierte Verfahren im
Immobilien-Wohnrecht, Fachhochschule
Wien (2006)
63. Kahlig, W., Malloth T.:
Frderungs-Automatismen bei Immobilien,
Vortragsreihe der ERSTEBANK, Wien
(2006)
64. Kahlig, W.: Strukturierungsverfahren im
Wohnrecht, IBM-Forum, Wien (2007)
65. Kahlig W.:
Strukturimmanente Denkanstze in der
Legistik, Universitt Salzburg, IRISKongress (2007)
66. Kahlig, W.:
Strukturierungsverfahren im Wohnrecht,
,IBM-Forum, Igls/Innsbruck (2007)
67. Heindl P., Kahlig, W.:
Abrechnungsfragen im Wohnrecht,
WISEWO des VI, Schruns, Vorarlberg
(2007)
68. Kahlig, W.: bersichtlichere Gestaltung
des Wohnrechts, Richtertagung Tulbinger
Kogel bei Wien (2007)
69. Kahlig, W.: Juristische e-Visualisierung im
Wohnrecht, Universitt Innsbruck (2007)
70. Kahlig, W.: Vereinfachung von Gesetzen
in der Europischen Union, Vortrag vor
EU-Abg. Karas, Brssel, Belgien (2007)
71. Kahlig W.: Vereinfachung von Gesetzen,
Vorbereitungskurs fr die
122
Konzessionsprfung fr Immobilientreuhnder, WIFI, Wien (2007/08)
Vorbereitungskurs fr die Konzessionsprfung fr Immobilientreuhnder, WIFI,
Wien (2007)
73. Kahlig, W.: Das strukturierte Wohn- und
Immobilien-Steuerrecht, Bundestag der
Immobilientreuhnder in Wien (2007)
74. Kahlig, W.: Strukturiertes Wohn- und
Immobilien-Steuerrecht, FH-Wien (2007)
75. Kahlig, W., Stingl W.: Automatisierungsverfahren im Immobilien
Steuerrecht, Vortragsreihe der
ERSTEBANK, Wien (2007)
76. Kahlig W.: Juristische Visualisierung mit
UML, Universitt Innsbruck (2007)
77. Kahlig, W.: Automatisierungsverfahren fr
Immobiliengesetze, Vortragsreihe der
Wiener Landesinnung Fachgruppe fr
Immobilientreuhnder (2008)
78. Kahlig, W.: Unified Modeling Language
fr juristische Anwendungen, Universitt
Salzburg, IRIS- Symposium (2008)
79. Kahlig, W.:
http://lawgical.jura.unisb.de
/index.php?/entry/343IRIS2008-Wolfgang-KahligUnified-Modeling-LanguageUML-fuer-juristischeAnwendungen.html
80. Kahlig, W.:
http://www.jurablogs.com/de/b
ersicht-iris-2008
81. Kahlig, W.: Restrukturierung des
Rechtsgebudes mit formalen Mitteln,
82. Kahlig, W.: Kahlig C.: Entwicklung des
Wohnrechts, Festveranstaltung Hotel
Marriott, Wien (2008)
Konzessionsprfung fr
Immobilientreuhnder, WIFI, Wien (2008)
84. Kahlig, W.: Ontologie und Logiken in der
Versicherungswirtschaft, Universitt
Wien, Juridicum (2008)
85. Kahlig, W.: Legal Cognitive Decision
Processing Towards a Dynamic Legal
Taxonomy, London East University,
London, GB (2008)
86. Kahlig, W.: Juristische Visualisierung
beyond UML, Universitt Innsbruck,
Arbeitswissenschaftliches Institut (2009)
Wohnrechts Gesamtbild und Statistiken,
88. Kahlig, W.: Strukturimmanente
Denkanstze in der Legistik, Universitt
89. Heindl P., Kahlig, W., Tschtscher:

Abrechnungstechniken und die NORM
A4000, Winterseminar-woche des VI,
Schladming (2009)
90. Kahlig, C., Kahlig, W.: Rechtsinformatik
Strukturierung / Visualisierung, IBMSymposion, Wien (2009)
91. Heindl, P., Kahlig, W.:
Sockelsanierung und 18-Verfahren,
Vortragsreihe ERSTEBANK, (2009)
92. Kahlig, W.: Einfhrung in die
Rechtslogik, Uni Innsbruck,
93. Kahlig, W.: Rechtliche Neustrukturierung
durch formale Mittel, eGov-Kongress,
Linz (2009)
94. Kahlig, W.: Vereinfachung von Gesetzen,
Immobilientreuhnder, WIFI Wien (2009)
Wohnrechts, Uni Innsbruck,
96. Kahlig, W.: Erweiterte Notation zur
Erhhung der globalen Rechtssicherheit
und als Beispiel fr den proaktiven Staat,
(2010)
97.Kahlig, W.: Rechtsvisualisierung,
98.Kahlig, W.: Erweiterte Notation der
linearen Strukturierung, Universitt
Innsbruck (2010)
Arbeitswissenschaftliches Institut
99.Kahlig, W.: Vereinfachung von Gesetzen,
100.Kahlig, W.: Rechtsvisualisierung am
Beispiel des DTV-Atlasses zum Recht,
Universitt Innsbruck,
101.Kahlig, W.: Privatrecht 2011
Strukturierung und Visualisierung,
(2011)
102.Kahlig, W.: Petri-Netze, Universitt
Innsbruck (2011)
103.Kahlig, W.: Vereinfachung von Gesetzen,
104.Kahlig, W.: Rechtsinformatik aus der
Sicht der Technik, Universitt Innsbruck
(2011)
105.Kahlig, W.: Digital Structuring Of
Complex Legal Interrelations, Universitt
123
106.Kahlig, W.: IT-untersttzte Rechtsvisualisierung am Beispiel des WEG,
107.Kahlig, W.: Recht und Logik, Techn.
Universitt TU Mnchen, Germany (2012)
108.Kahlig,, W.: Vereinfachung von
Gesetzen, Vorbereitungskurs fr die
Konzessionsprfung fr Immobilientreuhnder, WIFI Wien (2012)
109.Kahlig, W.: Strukturierung des
Rechtswissens, Universitt Innsbruck
(2012)
110.Kahlig, E., Kahlig W.:
Legistical Efforts By E-Government,
Dubai, I.A.U. University (2013)
111.Kahlig, W.: Strukturierung des
Rechtswissens -ex ante und ex post,
(2013)
124
Wireless Networks: Developments, Threats and Countermeasures

Mardiana Mohamad Noor and Wan Haslina Hassan
Communication System and Network (iKohza) Research Group,
Malaysia Japan International Institute of Technology (MJIIT),
Universiti Teknologi Malaysia.
mardianamnoor@yahoo.com, wanhaslina@ic.utm.my
ABSTRACT
This paper discusses current threats in wireless

networks and some academia research reviews
regarding the matters. Significant and
persistent threats discussed are sniffing, Man
In the Middle Attack (MITM), Rogue Access
Points (RAP), Denial Of Services (DoS) and
social engineering attacks. Some current
developments of wireless communication
technology
such
as
short
range
communication, cloud computing, bring your
own device policy (BYOD), devices tethering
and convergences of WiFi and cellular
network technology are also presented. Some
practical
suggestion
and
advanced
countermeasures are also reviewed in this
paper. The findings from reviewing these
research papers proved that the complexity of
the attacks had increased by time and the
attacks in WiFi network are passive and more
dangerous to the end users.
KEYWORDS
Current threats in WiFi, risks mitigation in

WiFi, network scanning, password cracking,
MITM, jamming attack, Rogue Access Points,
devices tethering, BYOD, countermeasures of
wireless threats.
1 INTRODUCTION
Wireless networks are susceptible
and exposed to attack because of its
borderless nature. It is easy to penetrate
any wired network via wireless network as
Access Point (AP) is bridging between
wireless and wired network. Packet
sniffing can be done passively because of
the hub-based configuration of the APs.
This activity initiates more dangerous
attacks such as Man In The Middle
(MITM) attack. MITM attacks such as
session hijacking and MAC spoofing are
some of the critical threats for wireless
networks. Moreover, hacking tools are

largely available in the market and online.
These tool which are usually meant to be
used by penetration testers and for
educational purposes are being misused
and abused by underground or even novice
hackers. Therefore, the sophistication and
frequency of attacks have increased; by
just using ready to use tools. On the other
hand, the flexibility and ubiquity of mobile
devices such as smartphones, tablets,
phablets and laptops are the main reason of
the popularity of hotspots which are
exposed of the rogue access points.
Another feature of newly developed
smartphones which is tethering introduces
more security issues to the end users.
From the data gathered by
Malaysian
Communication
and
Multimedia Commission (MCMC) [1]
until the second quarter of 2012, 1.2
million of registered hotspots were
recorded. For the purpose of comparison,
in 2011, only 0.4 million hotspots were
registered. These statistics show that the
number of hotspots subscription in 2012 is
rather large. In other word, in Malaysia
particularly, internet users are moving
towards wireless connectivity. This
scenario will definitely raising the bar of
security measures that should be taken
especially in curbing intrusion into
wireless networks.
Cases of misuse, incidents and
threats of internet have been reported in
Malaysia since 1997 to Malaysian
Computer Emergency Response Team
(MyCERT) [2]. Numbers of cases reported
keep increasing tremendously year by year
which confirm the upward trend of internet
125
threats. The statistic from the report shows
that the pattern of attacks changes from
time to time which basically follows the
development of the internet. From 1997
until 2003 most of the incidents reported
were due to spam and virus cases, but from
2004 until 2011 there is tremendous
incline in the attempt to intrude and
intrusion of the network attack. In 2003
only 60 cases due to intrusion were
reported, but in 2004, 368 cases were
reported. In the same report, the
inclination is very obvious from 2007 until
2011 when attacks due to intrusion
increase sharply by 861%.
In this paper we present current
and persistent threats to wireless networks
and accompanied by some active
researches from academia regarding the
advancement and some counter measures
to the threats.
The second section of the paper
discusses some current threats in wireless
networks and researches of developments
and countermeasures to the attacks. The
third part discusses security risks to the
wireless network due to the advancement
of cloud computing, and the fourth section
presented some new hacking tools
available on the market. The fifth section
is regarding the latest developments of
wireless networks, section six presents a
case study in Malaysia, findings and
analysis in section seven, some proposed
countermeasures and mitigation in section
eight, criteria for secure wireless network
in section nine and conclusion in the last
section.
2
THREATS
NETWORKS
OF
WIRELESS
2.1
Vulnerability
Scanning
and
Password Cracking
Vulnerability scanning is a process
when hackers use certain tools such as
Kismet and InSSider to scan the network.
The objectives of this activity are to:
a) find the vulnerabilities and security
level of the network
b) determine signal strength
c) determine the accessibility of the target
network
d) map the target network

After scanning the network, the
attacker might proceed to get into the
network. Despite of the known weaknesses
of WEP, it is still in use because of the
several reasons which involves some
issues of installation, interoperability,
convenience and flexibility.
In the following researches are the
evidences of vast availability of network
scanning tools and some of them are open
sources. In [3] and [4] war driving
activities were carried out by using
different tools and platform. In [3] it has
been found that Cain and Able
outperformed Netstumbler and Kismet in
terms of functions because it possess ARP
poison, VoIP logger, password crackers
and built in WiFi scanner in the expense of
volume of access points detected. In [4]
wardriving activity using Windows and
Mac operating systems were conducted
concurrently and comparison and analysis
of the best scanning tools using both
platforms were presented. This research
concluded that inSSIDer is the best tool
(shows vendor of the access point and
indicate signal strength graphically) using
Windows operating system and KisMAC
is the best tool for Mac (detects WiFi
silently, channel tuning capability and
detects wireless clients connected to the
access point).
In [4], a war driving was conducted
in several neighbourhoods in Dubai, UAE
in order to investigate the current WiFi
security issues. A laptop running MAC OS
X and a WiFi scanning tool and a car were
used in the war driving. From the war
driving in four different neighbourhoods a
total number of 1,228 WiFi networks were
found. Four categories of WiFi networks
have been found which are:
a) Open Network WiFi networks that did
not implement any protection - 35%
b)
WiFi implementing
Wired
Equivalence Privacy (WEP) - 26%
c) WiFi implementing WiFi Protected
Access (WPA) - 30%
d) WiFi implementing WiFi Protected
Access 2 - 9%
126
From the war driving experience more
than 50% of the residents have no security
or implementing weak security protocol.
Researches to attempt cracking
WEP and WPA/2 were done in [5-8].
Series of attempts by Fluhrer, Mantin and
Shamir (also known as FMS), and later in
2004 a person under pseudoname KoreK
made a second attempt and succeeded.
Tews, Weinmenn and Pyshkin (known as
PTW) launched new generation attack in
2007 followed by a Chopchop attack.
Based on these successful attempts,
cracking tools were developed. In WEP
mode even though the length of the
passphrase is increased or complicated,
only 30 minutes were taken to break the
code. Nevertheless, users which are using
WEP because of the convenience of
setting and interoperability, are advised to
set proper passphrase which will take
longer time to break and will create noise
in the network.
Even though WPA/WPA2 is said
to be robust, it is still protected by a
passphrase which can be cracked by using
Dictionary Attack. In [8], a new
proposed space-time trade off solution is
used where the Pair Master Keys (PMK)
are pre calculated for each passphrase in
the library and store them into another
library called Hash Library. This research
also
suggested
employing
cloud
computing to generate possible passphrase
and to take GPU parallel computing into
consideration to effectively calculate PMK
and proofread the Hash Library.
This section concludes that with
the matured amount of wardriving and
password
cracking
activities
had
developed the advanced and powerful
tools for network scanning and password
cracking.
2.2 Man In the Middle Attack (MITM)

and Packet Sniffing
MITM attack is to position the
attacker between two hosts in order to
hijack connection and injecting traffic. In
wireless networks, MITM can occur as
jamming by consistently transmitting
signals to the existing wireless access
points while providing clear signal from

another fake access points. Another MITM
attack is by using a spoofed de-association
or de-authentication frames to hijack the
connection between legitimate AP and the
users [9]. Wireless networks are more
susceptible to the kind of attacks because it
causes less disturbances if the attacker
poses as one of the client hosts in order to
access the network and launch attack to a
single host.
Packet
sniffing
is
another
significant threat to wireless networks by
using packet sniffer such as Wireshark,
Network Miner or Cain and Able.
Wireshark for example is a user friendly
and free software for sniffing activity. In
addition to that, its functions can be
enhanced and integrated with other
software as well. During this attack
attacker usually sniff the content of
packets and access unencrypted usernames
and passwords.
In [10], the author has listed out the
security risks from this activity such as
eavesdropping, breaching the credentials,
session hijacking by stealing the victims
session websites cookie and revealing
ones internet activities. By using tools
such as Wireshark, Ethercap or
NetworkMiner, sniffing activities can be
done by anybody by little practice. The
author of [10] also stated that most of the
Internet runs in the plaintext, making it
readable by packet sniffers, but if the
conversation is run through encrypted
connection such as site using SSL
encryption, data is less vulnerable. It is
also revealed that session hijacking is also
possible in sniffing activity by stealing
victims session cookie for a particular
website, especially when the websites do
not encrypt their traffic to the end user.
Packet sniffing can be a handful
task to perform in large networks because
of incapability of the tools to sniff large
amount of packets. A sophisticated form of
packet sniffing is presented in [11], where
a passive monitoring system for complex
wireless network is designed. This
research is to design a robust framework to
monitor real time network passively on a
large scale WiFi network. In [11], volume
127
of data gathered from sniffing activities
was reduced, so the system was capable to
listen to the traffic in the larger radius.
2.3 Rogue Access Points (RAP)

The purpose of RAP is to hijack the
connection of legitimate users is in order
to sniff the activities or to steal
confidential credentials of the users and
later launch further attacks or simply to
penetrate the Ethernet. With the
availability and competitive price of access
points (AP) in the market, anyone can set a
fake access point especially in free WiFi
hotspots. Moreover, nowadays most of the
laptops can function as a soft AP.
According to EC Council [12], there are at
least four available APs nowadays, which
are:
a) Compact and pocket sized RAP
device plugged into an Ethernet port
of corporate network
b) Software-based RAP running on a
corporate Windows machines
c) RAPs connected to corporate network
over a WiFi link
d) USB-based RAP access point device
plugged into a corporate machine
RAPs are usually placed behind a firewall
to avoid network scanner.
Counter measures to RAP is an
active area of research which concentrate
at two end points which are client side and
administrator
side
solutions.
The
advantage of having RAP solution in the
network administrator side is users are
warned about the safety of the connection
automatically every time they are using
one particular wireless network.
In [13] a full automated concept
which is to detect and eliminate RAP at
the network administrator side by using
mobile agents, namely master and slave
agents was introduced. A master agent is
generated on the server and then generated
slave agents according to the numbers of
APs. If any new AP exists slave agents
will be cloned. When a client find a new
AP, information of the packet will be
created and the clone slave agent will
bring the information to the slave agent
and the information will be sent to the
master agent to be verified. Master agent

will match the information with the
repository and if it is not matched then the
AP will be eliminated. This method is
claimed to be easy to implement, reliable
and cost effective. This method is seen as a
robust method to detect and block RAPs.
In [14], another RAP counter
measure at network administration side
solution was proposed which is a
centralized passive Indirect Rogue Access
Points Detection System (RAPiD). RAPiD
is to discover and verify RAP by collecting
data at routers or gateways and send the
packets to the network wireless host
engine which will be used to track each
unique local network host and determine
whether a host is using the wireless
network or not. However, the results of
RAPiD dependent on wireless host
discoveries and authorization verification.
In [15] developed an intrusion
detection scheme based on social network
and biomimetic approach. This first part of
this research is to propose sneeze
algorithm which is analogous to human
immune system. In this algorithm it is
assume that all APs are virtually connected
to each other and aware of the presence of
each AP. Each AP periodically acts as a
node and monitors the presence of APs by
tracking the Basic Service Set Identifier
(BSSID). If any RAP is placed within the
subnet, it will detect its presence and
availability of the beacon signal and its
BSSID. Then it will check the AP with the
list of APs it has in its list. If the AP is not
in the list, it changes its access key and
alerts the network administrator to
physically remove the RAP from the
network. This discovery has reduced the
work of network administrator from
constantly monitoring the network. Once
the RAP was found by the sneeze
algorithm, network administrator will
physically remove the RAP.
Consistent monitoring and the list
of known legitimate APs are the criteria
needed in order to implement network
administrator side solution. Another
obvious weakness of this solution is
because of the scale of the network is
usually immense, accuracy in 128
the
executed algorithm might be jeopardized
and users might get false alarm of RAP or
wrong information about the illegitimate
APs.
In [16], an end user solution has
been proposed in order to detect RAPs
especially in public hotspots. In this
proposed work, a flexible and practical
solution is suggested, especially for mobile
users or travellers to protect their
credentials where the security monitoring
is not reliable. The technique proposed
was based on the knowledge that in the
existence of RAP, the client has to
communicate with a remote server through
an evil twin AP and a normal AP. In this
case, compared to the normal scenario, the
twin evil case has one more wireless hop.
To distinguish these two cases (one and
two wireless hops), Inter-packet Arrival
Time (IAT) statistic, which is a time
interval between two consecutive data
packets sent from the same devices has
been adopted. A prototype system called
Evil Twin Sniffer (ET Sniffer) which has
been evaluated in real wireless network
including 802.11b/g was introduced by
using two novel algorithms which are
Trained Mean Matching (TMM) and Hop
Differentiating Techniques (HDT). The
analysis from the experimental set up
yields ET Sniffer can detect illegitimate
APs quickly with high accuracy and
resilient
to changes in wireless
environment.
In [17], Somayeh et al proposed a
novel method for RAP detection on the
client-side which is able to detect both
MITM attack and evil twin attack. This is
a client side detection method where
clients mobile will be warned on the
safety of the network by indicating
different colours of lights namely green
(safe), yellow (RAP is present) and red
(MITM attack). In this research, a method
to decide whether the network is exposed
to RAP, tricked by RAP or is it a safe
network is designed by comparing SSIDs,
MAC and IP address of the public APs.
Green light will be indicated if the network
is safe to connect to, yellow indicates the
existence of RAP and red light is the sign
to warn the users of the MITM hazard in

the network.
These are solutions which are
suitable for mobile users whom always use
their devices in places prone to RAP such
as airports and hotels. The disadvantages
of client side solution are users have to
implement and be familiar of the
introduced algorithm and simple WiFi
enabled devices might not support the
proposed algorithm.
2.4
Denial of Service
One of the weaknesses of wireless
network is the restricted usage of
bandwidth. This situation is giving great
advantage to hackers to launch Denial of
Service attack by replaying packets in
order to generate noise or by sending deauthenticate packets to the legitimate users
in the subnet. In fact, according to Belly
Rachdianto [18], this is the technique used
by a large group of hackers, Anonymous
to bring down governments web servers in
the recent attacks.
Practically, jamming attack is a
popular attack in military battlefield
usually by exploiting the weaknesses of
Wireless Sensor Network (WSN) which
have limited power resources. Jamming
attack can be launched either from outside
or inside the network by using jamming
devices to block the wireless medium and
prevent
wireless
devices
from
communicating with each other in order to
degrade the performance or to paralyze the
network. Some researches presented
herewith are advancements in jamming
attack which could be launched from
internal and external of the network.
Liu et al. in [19] have proposed
mobile jamming attack to Wireless Sensor
Networks (WSNs). In this attack, the
mobile jammers are placed randomly in
the network, listen to the network traffics
and detect the cluster head based on
received signal strength (RSS). According
to [19], the attack to the cluster head node
is an effective method because if the
communication between this node and sink
node is interrupted the base station may
not get any information. The results show
that mobile jammers has longer lifetime
129
corrupted and less energy is needed for
mobile jammers to bring down a WSN.
In selective jamming/ dropping as
proposed in [20], the targeted channel
which was defined by separate frequency
band, time slot or Pseudo Noise (PN) code
will be under attack or the attacker will
select data of high importance by
observing specific packet identifiers. Due
to the nature of selective jamming, this
type of attack is not easy to detect as
aggregate behavioural metrics are unable
detect attacks of selective nature, when
only a small fraction of packets were under
attacks.
Random
Packet
Distribution
(RPD) is a type of external jamming attack
when the specified attack model is wellprotected with various security measures
[21]. RPD jammer is put at the area of APs
in a Wireless Mesh Network (WMN). At
this point, the jammer will destroy the
packets by transmitting a short period of
noise signal when packets transmission is
detected. According to [21], any kind of
encryption and authentication may not
mitigate the risk of such attack because the
attacker does not need to participate in the
network.
In this investigation, the analysis of
damages caused by this attack was
conducted using Qualnet Network
Simulator. Performance metric such as
goodput, delay, jitter and Mean Opinion
Score (MOS) were analysed. The
simulation also showed that real time
applications such as (Voice over Internet
Protocol) VoIP will be adversely affected.
MOS score shows that at 5% attack rate
can affect the voice quality. Based on the
analysis of TCPs throughput in [21],
packet lost rate at 20% can bring down a
network, which is equivalent to 40% RPD
attack rate.
In another research, Distributed
Jammer Network (DJN), utilizes the
advantages of nano-technology, where the
jammers are not visible in the naked eyes
is proposed by Hong et al [22]. These low
powered jammers are claimed to have less
self-interference. DJN has few distinguish
advantages which are:
a) because DJN is composed of a large

number of devices of ample redundancy, it
is claimed to be robust
b) DJN emits low power, which is
advantageous because of health
c) It is hard to detect due to its very small
size
To measure the impact of DJN, a scenario
is simulated in QualNet and the result
revealed that DJN can cause phase
transition even though the power is held
constant.
From these researches it has been
proven that jamming attacks are
potentially energy efficient, silent and hard
to trace and robust means of attacks.
2.5 Social Engineering

Social engineering is the art of
utilizing human behaviour to breach
security without the victim realizing that
they have been manipulated.
Social
Engineering can be categorized into two
categories which are technology based
deception and human based deception
[23].
Study in [23] revealed that here are
several behaviours that are vulnerable to
this attacks which are trust, carelessness,
curiosity
and
ignorance.
Social
engineering attacks mostly will affect big
companies and organization where the
workers
and
people
inside
the
organizations are the victims of the threats.
The effects of this attack can be in
monetary form, where a company can
suffer billions of loss due to data leakage
and distrust from the existing customers
and compromised reputation.
The only way to mitigate security
risks from social engineering is to educate
the entities in any corporation in order to
create awareness to the social engineering
threats.
2.6
Threats in Bluetooth Technology

Saroj et al [24], has discussed threats
in piconet and medium range wireless
networks such as Bluetooth, RFID and
Wireless Sensor Network (WSN).
130
Bluetooth technology which enabling
file transfer and other data communication
between devices such as printer, PDA and
smartphones are vulnerable to threats
which are classified as in ABOTT (A
Bluetooth Threat Taxonomy) which is
listed below.
a) Surveillance Bluprinting, bluefish, sdp
tool
b) Range Extension BlueSnipping,
bluetone
c) Obfuscation- Spooftooth
d) Fuzzer BluePass, BlueSmack
e) Sniffing BlueSniff, Wireshark
f) Denial of Service Signal Jamming,
Battery exhaustion
g) Malware BlueBag, Caribe
h) Unauthorized direct data access
Bloover, BluBug
i) Man in the middle BlueSpoof
2.7 Threats in RFID
RFIDs are used as smart tags
which are widely deployed and it also
exposed to many kind of threats; which is
similar to attacks in WiFi. Some of the
common attacks are sniffing, tracking,
spoofing, tracking, replaying and Denial of
Service attack. The most critical attack in
RFID is replay attacks which enable the
attacker to intercept and retransmit RFID
queries.
3
CLOUD COMPUTING: THE

SECURITY RISKS TO WIRELESS
NETWORKS
Services offered by cloud are data
storage,
software,
platform
and
infrastructure [25]. According to Tim
Pierson, a professional PenTester and
consultant in Hacker Halted Conference in
Kuala Lumpur, internet users have been
using cloud especially as storage
mechanism for years for example Gmail,
Amazon and YAHOO [26].
One of the latest developments of
cloud computing is making ones WiFi
accessible from anywhere. This is a
practical and flexible solution as it comes
with many applications developed by third
party. In June 2012, Cisco has launched its
smart Wi Fi product which is Cisco
Connect Cloud [27]. On the other hand,
this development raises other privacy and

security issues since there is no guarantee
that the cloud service, provider will not
share users WiFi configuration and other
confidential credentials.
Cloud computing also might solve
some problems of the retailers such as
distributed natures, scalability, security
monitoring and reduces the cost of WiFi
deployment [28]. Since cloud is shared
between tenants, this might lead to data
intrusion and motivate hacking Wi Fi
through cloud. One example of cloud
abuse has been reported in 2011, where
some powerful software leased from
Amazon has been used to hack WiFi WPA
through brute force attack [29]. Not to
mention some security issues in the cloud
itself which are very appealing to the
hackers to launch attack. If hackers can
attack the cloud, the network associated
with it is jeopardized.
4
HACKING TOOLS
The attacks in wireless network are
aided by tools and some of the tools have
been used in attacks explained in section
two. Furthermore, the devices proliferation
for wireless hacking tools online which
require minimal technical skills and easy
to execute such as Raspberry Pwn, Nokia
N900 PwnPhone, and Wi Fi Hot Spot
Honeypot trigger more hacktivism. In the
recent Hacker Halted Conference on 19
November 2012 in Kuala Lumpur, Wayne
M Burke, the CEO of Sequrit [30], has
presented some of the mobile hacking
tools as said above and found that those
tools are being sold from as low as USD
99.99. These tools are merely for the
security professionals and for the
educational purposes, but also have great
potential to be misused, abused and
exploited
by hackers.
The
vast
development of these mobile devices
which have permanent connectivity to the
internet can be used to spy, drop malicious
code (malware), monitoring person by
GPS, redirected the email and read
messages.
4.1
Raspberry PwnPi
131
This tool is a fully-featured
security penetration and auditing platform
[31]. It has comprehensive hacking tools
which will lure the hacker to attack
wireless
medium
which
include
information gathering, network mapping,
penetration
and
vulnerability
identification. It also releases free
community edition, which is definitely an
excitement to all the hackers.
Nokia N900 PwnPhone
Nokia N900 is mobile-phonebased-pentesting platform which includes
Aircrack-NG, Metasploit, Kismet and also
featuring oneclick evil AP, WEP Cracker
and a packet capture. This comprehensive
device is available in the market at the
price of USD 995.00.
Corporation in which will make the

affected smartphones prone to DoS attack
[33]. Broadcom BCM4325 and BCM4329
wireless chipsets have been reported to
contain out-of-bound read error condition
that might be exploited to produce a denial
of service condition. Upon the writing of
this survey paper among the smartphones
which have been planted with these
chipsets are Apple iPhones, Motorola,
Samsung, HTC, LG and Sony Ericsson.
4.2
4.3
WiFi Hot-Spot Honeypot

The WiFi Pineapple Mark IV is
one of the favourite multi-functional
penetration testing platform since 2008
[32]. Especially used to launch the man-inthe-middle attack, which is also enabling it
to observe the connected client, perform
site surveys, customize attacks and capture
data in a few clicks. It also functions as a
mobile broadband, Android tether, WiFi
relay and provide internet connectivity to
PC.
Despite of its comprehensive
capability and functions, this device is
available at a very affordable cost, which
is USD 99.99.
4.4
Open Source Intelligent (OSINT)

One of the famous hacker namely
The Jester has revealed his favorite
hacking tools as OSINT. The tools such as
Maltego, Creepy, Spokeo, CaseFile and
FoxOne Scanner (to name a few) can be
downloaded easily and exploited by
malicious hackers. In fact, all of the
downloadable programs even come with
step-by-step tutorials and provide details
information of the said programs.
4.5
Mobile Vulnerabilities
Recently, it also has been reported
that there is a flaw in wireless chipset
design
provided
by
a
wireless
semiconductor manufacturer, Broadcom
LATEST DEVELOPMENTS
According to [34], found that there
is a tremendous increment in terms of sales
in tablet and smartphones market. In 2012,
821 million of these devices were sold
worldwide, but in the first quarter 2013
itself the sell is 1.2 billion. There is a 46%
increase of sales and it is predicted that
these numbers will be increasing from
time to time. The huge number of sales in
these mobile devices is supported by
flexible
operating systems
mainly
Android. Android out-casted other players
such as Apple and Microsoft by
conquering 72% of the worldwide market.
In fact, during the third quarter of 2012
Android device sales exceeded PC
shipment by 40%.
5.1
Bring Your Own Device (BYOD)

As the mobile devices getting more
portable and smarter, it completes what
Bring Your Own Devices (BYOD) has to
offer. According to [34], the main
objective of BYOD is for the employee to
have better control of their work in the
office. Most of the workers opt to install
personal and work related matters in the
same smartphones. This requires the
employer to come up with a detailed and
effective policing to safeguard companys
credentials to be accessed by unauthorized
personnel. It looks like employer can never
ignore this wind of change, as if the
employer resist the change then there will
be many RAP planted in the whole
network.
The fear of this BYOD is the level
of security awareness of the employee. In
the same research as above indicates that
65% of the employees do not verify
132
security settings of online shopping sites,
which was the main target for SQL
Injection and cross scripting threats. This
is a significant threat to the local network
as those devices were synchronized with
the local network.
BYOD is also the uprising trend in
schools and hospitals. The main concern in
applying BYOD in these environments is
the proper and solid segregation between
authorized and unauthorized territory for
the users. Students especially, by nature
are very curious homo-sapiens, so the
tendency for them to test their level of
authorized access. If they reach the core
network there are many possibilities of
malicious activities such as data
modification, forgery, cheating and any
other possible unethical doings. The
network must be very properly designed,
so that all the users are allocated
reasonable bandwidth, unless the users
will get frustrated and bring their own AP
to access the network.
network in United States called Republic

Wireless [36]. This service offers great
advantages to the users whereby they
enjoy unlimited voice, texting and internet
access via WiFi, and whenever the WiFi
service is not available the phone will
automatically migrate to cellular network,
in this case Sprint. The initial package is
rather expensive but in the long run it is
said to benefit the users as the charge is
only as low as USD 19 per month.
As in Malaysia, Telekom Malaysia
(TM), has launched its own hotspots which
is TM WiFi [37]. Users can subscribe the
service from as low as MYR 15 per month
depends on the speed of the connection.
As for the security measure, only three
devices were allowed to be registered
under one login ID.
5.2
Devices Tethering
In [35] states one useful and
flexible feature of tablets and smartphone
is the capability to tether or distribute their
bandwidth of cellular provider to any other
nearby devices such as laptops or any
other mobile devices. The 3G/4G cellular
network provider will charged this service
with certain amount of bill. The main
concern here is whether the end devices
are consistently secured. The risk of illegal
signal interception (MiTM) is high if the
level of security is not properly set. Some
apparent threats are DoS, virus threats and
compromised privacy due the possible
illegal GPS chips activation. The worst is
when central corporate network is
bypassed through an unsecured mobile
connection. Another issue that might arise
is the lacking of encryption levels in jailbroken phones.
5.3
Convergence of WiFi and
Cellular Network
In the wider scale, there is a
massive integration of WiFi and cellular
133
6
WIRELESS NETWORK
SECURITY: A CASE STUDY IN
MALAYSIA
In Malaysia, cyber space is
governed and monitored by three
incorporated bodies which are Malaysia
Communication
and
Multimedia
Commission (MCMC), Cybersecurity and
Malaysia Computer Emergency Response
Team (MyCERT). These bodies are in
affiliation with Ministry of Science,
Technology and Innovation (MOSTI).
MCMC is a body in which regulating rules
and is authorized to take action against any
wrongdoings subjected to cyber security,
while Cybersecurity and MyCERT are the
bodies investigating and receiving incident
reports or breaches.
According to a Security Assurance Analyst
from Cybersecurity, incidents reported due
to wireless and mobile devices threats are
relatively small, which is only four cases
reported in 2012. The reported incidents
are due to scam, password cracking,
bandwidth leeching and compromised
access point. However, it does not indicate
the wireless threats are not significant in
Malaysia but there is a huge possibility
that other incidents were not reported
accurately in which did not mentioning the
means of network used when incidents

happened. One more reason is since some
wireless threats are in passive form such as
sniffing, the end users were not aware that
their privacy was compromised. Web
defacement or phishing websites are the
most popular incidents reported in 2012
might be aggravated by the ease of
accessibility to wireless networks. Hackers
do not need to sit in the office or to have in
proximity with the target, making those
malicious activities more active and
untraceable. Other cases which are
categorized as frauds might also due to the
unawareness of the end users whom are
having inadequate knowledge about the
threats in wireless networks accessing their
bank accounts or doing monetary
transactions in public WiFi, which is very
vulnerable to packet sniffing. Figure 1
shows the total of cyber incidents reported
from 1999 to 2012. As illustrated in the
Figure 1, in 2004 the most incidents
reported and the highest case is due to
spam which is 14, 371 cases but starting
from 2008 until 2011 a steady increment
of incidents were recorded. Within this
interval the most reported cases are due to
fraud and intrusion.
Figure 1: Number of Incidents Reported to

MyCERT from 1999-2012
Figure 2: The Graphs Showing the Trends of

some Significant Threats
134
As indicated in Figure 2, intrusion
and fraud have inclining pattern towards
2011. It can be deduced that malicious
activities that causing monetary loss and
privacy intrusion developed rapidly and
this is might due to the vast development
of Wifi, especially public Wifi. An
example of intrusion is network and
system intrusion and examples of fraud
cases are web defacements and website
phishing. In many organizations, such as
hotels and hospitals public Wifi is offered
to ensure comfort and accessibility to the
users, but the network might not properly
segmented and isolated that will cause
outsiders reaching the internal network
intentionally or not. But in 2012 it is
shown that cases due to these two
categories were declining. This is due to
the
awareness
of
the
network
administrators to increase the encryption
level of their own network.
The investigation by Cybersecurity
when doing penetration testing in some
companies in Malaysia yields some
common threats in wireless network. The
most popular threat found is WEP
cracking. It was found that the network
administrator tend use default setting or
easy-to-guess password to secure their
wireless network. Rogue access point
(RAP) is another common threats found in
the investigation. RAPs were installed to
the local network and it bridges the local
network to the outer network. Some
authorized officers might intentionally
brought AP for his personal use, without
knowing that a hacker can get access
easily from his RAP.
In some companies, they tend to
ban the officers to have access to some
websites that might contain virus, so this is
when the officers will bring their own
broadband dongle and plugged it into the
laptops and desktop. This action might
transfer the virus into the companys
system and infected the whole network.
Another finding is signal leakage which is
due to inappropriate planning of signal
radius. While doing network scanning
using Nessus or Kismet, this network will

appear anywhere outside the authorized
area and this will give an advantage to
password cracking activity.
WiFi service for public is very
popular and this is one type of selling
point in the marketing strategy in which
advantaging the end users. The
investigation found that most of the
organization which offers this service did
not isolate their network appropriately. In
some
companies,
there
is
poor
segmentation in the local network and
wireless network where some people can
intrude local area network by using
network mapping, ping and port scanning
tools. Even though there is proper
segmentation, the outsider can still get into
the internal network by using VLAN
hopping.
It can be concluded that the threat
in Wifi is more significant to the end users
because the awareness of wireless threats
in Malaysia is considered low. In order to
raise the awareness, Cybersecurity formed
a team called Cybersafe to approach the
end users to raise awareness regarding
internet security.
FINDINGS AND ANALYSIS

Hackers intent to target a wireless
network is to exploit the network users
privacy in a MITM attack, to perform
privilege escalation and to use hotspot as a
platform to attack a nearby target and
remain untraceable. The main reasons for
hacking activities are the opportunities to
exploit vulnerabilities and the ease to
execute attacks. The seamless boundary of
wireless network is making the networks
easy to penetrate while the availability of
free public WiFi is making these malicious
activities easy to launch and possibly done
by novice hackers. One interesting fact
learnt from reviewing the research papers
is the sophistication of the attacks is
increasing while the skill needed is very
minimal. This is probably due to the
misuse of the penetration testing tools
135
which are affordable and easy to use and
carry and the vast availability of hacking
software on the internet.
RAPs are usually found in public
hotspots. With the availability and
competitive price of physical access points
and with a little practice and training,
anyone could install a RAP to any public
hot spot. Furthermore, RAP is almost
invisible especially when the attacker uses
his own laptop as a soft access point. This
threat is actually leads to another
dangerous attack which is Man In The
Middle (MITM). In wireless network
MITM is dangerous because it is a passive
and silent attack which listens and sniffs
users conversation and most probably will
intercept and modify the packets being
relayed. MITM is almost impossible to
prevent as long as the attacker find a
hole to access the network, legally or
not. MITM can lead to another disastrous
attack, for example session hijacking.
People are always the weakest link
in security. The threats posed by social
engineering are almost destructive as other
technical threats. The art of deception is
the most useful skill in social engineering,
while insiders ignorance and negligence
always open the door for the opportunist
attackers to cause security breaches.
On the other hand, a lot of
researches on improvisation of the attacks
and counter measures against the attacks
have been carried out. For example, Denial
of Service attacks especially jamming
attacks have gone through many phases of
advancement in terms of capability and
invisibility,
while
countermeasures
techniques to evade RAPs are also on the
rise.
The recent developments of
smartphones such as tethering capabilities,
convergence with 3G/4G and BYOD
policy add more security vulnerabilities.
The service and users must be more
responsible
to
ensure
secured
communication by adhering to strict policy
and technical countermeasures. The power
to choose, buy and use is in consumers
means, as the choice and price of the

devices
are
within
consumers
affordability.
Internet security education is very
crucial, as this is the only way to overcome
the weakest link in security which is
people. Hopefully, with this education the
cases due to misuse of internet will be
greatly reduced.
8
COUNTERMEASURES
AND
MITIGATION
Everybody is exposed to threats in
wireless network as no network is fully
guaranteed as secured. Hence, network
administrators and users must be more
serious in curbing security issues in
wireless
networks
and
apply
countermeasures to lessen the risks of
security issues. Here are some practical
recommendations for countermeasures to
the threats in wireless networks from
service providers side:
8.1
Change the default Service Set
Identifier (SSID).
Having known the value for the
SSID for the router and access point is the
initial step to break through a wireless
network. Usually, when an attacker sees a
SSID with a default value, he will assume
that the network is poorly configured.
Sometimes, the default SSID is also used
as the password to the network.
8.2
Turn On the Encryption

Be it WEP or WPA/2, encryption
must be turned on. It is much better than
leaving the network as the open network.
No network is guaranteed secure, but at
least precautions can be taken so that the
attack is less likely to occur and more
difficult to launch. Stronger algorithm
such as CCMP is recommended for the
encryption or even better to invest using
WPA the Enterprise mode. Network
administrators are also advised to set
stronger password to avoid novice hackers
from cracking it. It also recommended
changing the password regularly.
136
8.3
MiTM attack. Activating firewall is also

another method to secure the network.
8.4
8.8 Secure Your Confidentialities

By disabling file sharing feature
will mitigate the risk of the threats. By
encrypting and setting privacy to important
folders are other precautions when using
public WiFi.
Disable SSID Broadcast

By broadcasting the SSID into the
air, it tells the existence of the network
without any effort. It is like calling hackers
to break into the network. Broadcasting
SSID is less likely necessary in the home
network.
Place the Access Point Securely
In order to control access to our
WiFi network, the access point must be
placed appropriately so that the signal in
the desirable radius.
8.5
Policy Enforcement
By
having
clear
and
understandable policy or agreement of
prohibited actions in wireless network
such as packet sniffing or placing any
device which can act as a RAP might help
to reduce malicious activities in wireless
network. As to handle threats that might
occur because of social engineering, a
written agreement of policy between
employees and employers must be taken
place from the first day of work.
Employees might be given certain unique
password to login into companys WiFi
network and repeatedly reminding them
not to share the password and other
sensitive credentials to others.Precautions
must also be taken when the users are
using wireless network outside of their
homes or offices, such as hot spots to
safeguard confidential credentials from
any attackers. Here are some advices:
8.6
Disable the WiFi adapter

This is important to prevent auto
connection from the malicious access point
in the network. It is also important to
always monitor the access point that will
connect to the PC by configuring the
setting in the PC.
8.7 Secure Your Network
Users are advised to use Virtual
Private Network especially to prevent
8.9 Prevent Auto Connection to Open

WiFi
Some computers will automatically
connect to the open wireless network
without giving you any notification. It is
also advised not to control the wireless
network remotely and always disable the
network when not in use in a long period
of time.
9
CRITERIA FOR A SECURED
NETWORK
The rapid developments of wireless
networks demand consistent improvements
of network management to ensure better
connectivity and accessibility without
putting security on the line. As suggested
in [38] here are some suggestions of what
a secured wireless network should have.
9.1
Flexible Connectivity
A network must be scalable and able to
expand. It also should be capable to shift
users from one congested AP to any
available AP. This automated system is
very important to avoid frustrated users
whom might bring their own unregistered
AP. A good network should be scalable
and define its parameter very well, for
both indoor and outdoor access. This is to
avoid signal leakage to any unintended
territory.
9.2
Efficient Devices Management,
Control and Monitoring
Todays wireless networks are
much more complex which consist of
numerous access points. A secured
network should have all the APs to be
137
registered and an automated system must
be deployed to enabling APs updating
activities.
Role Based Access Control
(RBAC) is a system that will assign a role
to any device based on how they were
authenticated. RBAC also defines devices
accessibility level and what they can
access from a network. In order to avoid
misuse and abuse of the network and other
security breach, a secured wireless
network management should be able to
monitor users activity in real time. This is
important as there will be hundreds or
thousands of devices connected to the
network, so it is important to monitor and
manage the applications and programs that
they are using in your network. Network
administrator also can even remotely wipe
the device if it lost or stolen.
All devices are also need to be
registered and assigned unique IP from the
network. To avoid hassle, network access
control can allow the users to register
themselves to the network. As for the
network administrators, this is when they
can check the latest updates and validate
security level of the devices.
9.3
Filtering and Firewall

As a network administrator we do
not want any user to abuse the network to
visit malicious websites or run prohibited
programmes, so filtering process must take
place. This is also to reduce the possibility
the users spreading virus to the core
network. A comprehensive security system
which integrated firewall, intrusion
prevention system and application
identification and control is what every
secured network is advised to have.
9.4
Proper Network Segmentation
and Segregation
A secured network should have
clear and solid boundaries of core network
and wireless network that can be accessed
by the users. This is to avoid unauthorized
personnel to reach the core network
intentionally or not. It is also important to

eliminate device which tries to cross the
limit out from the network.
10
CONCLUSIONS
As presented in this paper, wireless
networks are susceptible to attacks and
prone to many types of threats. This is due
to its seamless nature and its popularity.
Users in public hotspots are claimed to be
more exposed to such threats because of
the ease to deploy MITM by sniffing
techniques and planting RAPs and most of
the time victims do not aware that their
data or privacy has been compromised due
to passive nature of the attacks.
Vast availability of the hacking
tools and devices opens up wide path for
hacktivism, while the advancement of
cloud computing making Wi Fi more
vulnerable to threats. The security of cloud
computing as well must be managed very
well to avoid leakage and breach of
information.
Wireless network is an easy access
to local system if the network is not
properly segregated. The developments of
wireless network give flexible access
especially under BYOD policy. For this to
work in a secure environment set of rules
must be followed and adhered.
Acknowledgement: This research is in
affiliation with Ministry of Higher
Education, Universiti Teknologi Malaysia,
Malaysian Japan International Institute of
Technology
(MJIIT)
and
i-Kohza
Computer System and Network (CSN).
.
11
REFERENCES
1.
Number of registered hotspots in Malaysia

Q2, Pocket Book of Statistics Q2
(SKMM), 2012. www.skmm.gov.my.
Reported Internet Incidents from 19972012 (MyCERT), www.mycert.org.my.
Reddy, S.V., Rijutha, K., Ramani, K.S.,
Ali, S.M. : Wireless Hacking A WiFi
Hack By Cracking WEP. In: Proc 2010
Second International Conference on
2.
3.
138
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
Education Technology and Computer

(ICETC) 2010.
Said, H., Guimaraes, M., Mutawa, N.A,
Awadhi, I.: Forensics and War Driving on
Unsecured Wireless Network. In: Proc
2011 6th International Conference on
Internet
Technology and
Secured
Transactions, Abu Dhabi (2011).
Begh, G.R, Mir, A.H.: Quantification of
the Effect of Security on Performance in
Wireless LANs. In: Proc. 2009 Third
International Conference on Emerging
Security Information, Systems and
Technologies (2009).
Mavridis, I.P., Androulakis, A.I, Halkias,
A.B., Mylonas, P.: Real-life Paradigm of
Wireless Network Security Attacks. In:
Proc. 2011 Panhellenic Conference on
Informatics (2011).
Beck, M., Tews, E.: Practical Attacks
Against WEP and WPA. In: Proc. 2009
Second ACM Conference on Wireless
network Security (Wisec) 2009..
Yin, D., Cui, K.: A Research into The
Latent Danger of WLAN. In: Proc.2011
The 6th International Conference on
Computer Science and Education ICCSE,
Singapore (2011).
Andrew, A. Vladimirov, Gavrilenko,
K.V, Mikhailovsky A.A.: WI-FOO: The
Secrets of Wireless Hacking: Breaking
Through.
Pp155-173
Pearson
Education (2004).
Hannah, A.:
Packet Sniffing Basics.
http://delivery.acm.org
Benmoshe, B., Berliner, E., Dvir A.,
Gorodisher, A.: A Joint Framework of
Passive Monitoring System for Complex
Wireless Networks. In: Proc. 2011 First
International
IEEE
Workshop
on
Emerging Densely Connected Networks
(EDCN) 2011.
EC Council (2012). Certified Ethical
Hacker: Hacking Wireless Networks,
Module 15. pg 91-93.
Sriram, V.S.S, Sahoo, G., Agawal, K.K.:
Detecting and Eliminating Rouge Access
Points in IEEE-802.11 WLAN A MultiAgent Sourcing Agent Methodology. Birla
Institute of Technology, India (2010).
Qu, G., Nefcy, M.M.: RAPiD - An
Indirect Rogue Access Point Detection
System. Oakland University, Michigan
USA (2010).
Sampangi, R.V, Dey, S., Viswanath,
V.N.: The Sneeze Algorithm: A Social
Network and Biomimetic Approach for
intrusion Detection in Wireless Networks.
University of Mysore, Mysore India
(2010).
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
Song,, Y., Yang C., Gu, G.: Who is

Peeping at Your Password at Starbucks?
To Catch an Evil Twin Access Point. In:
Proc. 2010 IEEE/ IFIP International
Conference on Dependable Systems and
Networks (DSN).
Nikbakhsh, S., Zamani, M., Abdul Manaf,
A., Janbeglou, M.: A Novel Approach for
Rogue Access Point Detection on the
Client Side. In: Proc. 2012 26th
International Conference on Advanced
Information Networking and Application
Workshop.
Rachdianto, B.: Threats in Wireless
Networks: In Tutorial Ethical Certified
Hacker, EC Council (2012) pp 139
Zhiping, L., Hui, L.:Mobile Jamming
Attack in Clustering Wireless Sensor
Network. In: Proc. 2010 International
Conference on Computer Application and
System Modelling (ICCASM) 2010.
Lazos, L., Krunz, M.: Selective Jamming/
Dropping Insider Attack in Wireless Mesh
Network (WMN). University of Arizona
2011.
Zhou, B., Marshal, A., Zhou, W., Yang,
K.A.: Random Packet Destruction DoS
Attack for Wireless Networks. In: Proc.
2008 ICC.
Huang, H., Ahmed, N., Karthik, P.: On a
New Type of Denial of Service Attack in
Wireless Networks - The Distributed
Jammer Network. In: 2011 Proc. IEEE
Transactions
On
Wireless
Communications.
Gulati, R :The Threat of Social
Engineering and Your Defence Against It.
In SANS Institute Info Sec Reading Room
2003.
Panigrahy, S.K, Jena, S.K, Turuk, A.K.:
Security in Bluetooth, RFID and Wireless
Sensor Network. In: Proc. 2011
ICCCS11, India (2011).
Yu, H., Powel, N., Stembridge, D., Yuan,
X.: Cloud Computing and Security
Challenges. In: Proc. 2012 50th Annual
Southeast Regional Conference (2012).
Pierson, T.: Cloud Computing- APAC
Hacker Halted Module: In Hacker Halted
Conference 2012.
Cisco SmartWiFi.
http://newsroom.cisco.com/press
Why retailers embrace cloud WiFi Access
http://blog.airtightnetworks.com/why
Amazon Cloud Could Help Hack WiFi
Networks
http://www.reuters.com/article/2011/01/07
/us-amazon-hackingidUSTRE70641M20110107
139
30.
31.
32.
33.
Wayne M Burke in Hacker Halted

Conference November 19, 2012, Kuala
Lumpur.
Advance Persistent Pentesting
http://pwnieexpress.com/
Hotspot Honeypot Pentesting Platform
http://hakshop.myshopify.com/collections/
gadgets/products/wifi-pineapple
Vulnerabilities in Mobile Devices
http://thehackernews.com
34.
35.
36.
37.
38.
BYOD
http://www.securedgenetworks.com
The Security Implications of Tethering a
Mobile Phone To Your Computer
http://www.webologist.co.uk
Republic
Wireless
http://www.republicwireless.com
TM WiFi www.tm.com.my
11 Features Every Secure Wireless
Network
Should
Have
http://www.securedgenetworks.com
140

IJDIWC - Volume 3, Issue 1

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

IJDIWC - Volume 3, Issue 1

Uploaded by

Copyright:

Available Formats

BICLUSTERING GENE EXPRESSION DATASET USING ENHANCED

IMPROVING AND MEASURING

COLLABORATIVE PROBLEM SOLVING USING PUBLIC SOCIAL

IMPACT OF A FORM OF ONLINE MATERIALS ON THE QUALITY OF Blanka Frydrychova Klimova,

SEPARATION IN DATA MINING BASED ON FRACTAL NATURE OF

ADAPTIVITY OF A COLORING ALGORITHM TO UNRELIABLE

WITH Skander Azzaz, Leila Azouz

ROMAN LAW AND CONTEMPORARY LEGISLATION

AND Mardiana Mohamad Noor, Wan

A New Method to Channel Estimation in OFDM Systems Based on Wavelet

instantaneously used to data transmission. Also,

presented in section (3). The proposed method

simple and low order sub matrix so that they can

Figure 1. The OFDM System in baseband model

At first, in transmitter the binary inputs are

subcarriers will be modulated by the obtained

Where L is an oversampled factor which can

Where Nc denotes the cyclic prefix length.

Figure 2. Cyclic Prefix adding process in OFDM symbol

Finally, the obtained OFDM signal is

In the frequency domain we can obtain the

and hyperLAN2 use pilots in frequency domain

transmitting OFDM symbols contains data and

Where P denotes the pilot repetition rate or

Figure 3. Block type arrangement of OFDM symbols

According to the type of channel fading, slow

2.3 Least Square Method

Where W(kp)=W(kp)/X(kp) is the noise

Where L denotes distance between two

Where h0(k) and h1(k) are impulse response of

Where k and j are translation and scale

Where a is called the scaling factor and

Figure 5. Wavelet decomposition in 2 levels

Where L can be controlled by the pilot spaces

estimation. According to [14], the calculated

Where Di denotes the detail coefficients.

Finally by taking the DFT of the obtained

Figure 6. Block diagram of the proposed method

are not used and also in order to better

In the following we want to compare the

Figure 7. BER Performance comparison with 16QAM

Figure 8. BER performance comparison in case of QPSK

Figure 9. MSE performance comparison with 16QAM

Figure 10. MSE performance comparison with QPSK

Figure 11. BER performance comparison with 16QAM

Figure 12. MSE performance comparison with 16QAM

In the following we want to compare the

on wavelet decomposition is used in order to

The obtained results are shown in the figure

Figure 13. The Estimated Impulse Response of The

B. Muquet, Z. Wang, et al. Cyclic prefixing or zero padding

11. H. Minn, and V.K. Bharagava, An Investigation into Time

BICLUSTERING GENE EXPRESSION DATASET USING ENHANCED

There are two main categories of bi-clustering

DNA is the secret of life. It is the blue print for

The main problem of graph-based bi-clustering

DNA contains the genes that carry the phenotypes

1. Gene-based clustering in this approach genes

2.1 CHENG AND CHURCH

Sub Column Average

Sub Matrix Average

SPHier [7] extract clusters from dataset using