Journal of Computers Contents - Academy Publisher

Journal of Computers
ISSN 1796-203X
Volume 5, Number 1, January 2010
Contents
Special Issue: Selected Papers of the IEEE International Conference on Computer and
Information Technology (ICCIT 2008)
Guest Editors: Syed Mahfuzul Aziz, Vijayan K. Asari, M. Alamgir Hossain, Mohammad A.
Karim, and Mariofanna Milanova
Guest Editorial
Syed Mahfuzul Aziz, Vijayan K. Asari, M. Alamgir Hossain, Mohammad A. Karim, and Mariofanna
Milanova
SPECIAL ISSUE PAPERS
Cutting a Cornered Convex Polygon Out of a Circle
Syed Ishtiaque Ahmed, Md. Ariful Islam, and Masud Hasan
Decision Tree Based Routine Generation (DRG) Algorithm: A Data Mining Advancement to
Generate Academic Routine and Exam-time Tabling for Open Credit System
Ashiqur Md. Rahman, Sheik Shafaat Giasuddin, and Rashedur M Rahman
Anomaly Network Intrusion Detection Based on Improved Self Adaptive Bayesian Algorithm
Dewan Md. Farid and Mohammad Zahidur Rahman
Performance Evaluation for Question Classification by Tree Kernels using Support Vector Machines
Muhammad Arifur Rahman
Recurrent Neural Network Classifier for Three Layer Conceptual Network and Performance
Evaluation
Md. Khalilur Rhaman and Tsutomu Endo
An Enhanced Short Text Compression Scheme for Smart Devices
Md. Rafiqul Islam and S. A. Ahsan Rajon
Design and Analysis of an Effective Corpus for Evaluation of Bengali Text Compression Schemes
Md. Rafiqul Islam and S. A. Ahsan Rajon
A Corpus-based Evaluation of a Domain-specific Text to Knowledge Mapping Prototype
Rushdi Shams, Adel Elsayed, and Quazi Mah- Zereen Akter
Implementation of Low Density Parity Check Decoders using a New High Level Design
Methodology
Syed Mahfuzul Aziz and Minh Duc Pham
REGULAR PAPERS
A Formal Model for Abstracting the Interaction of Web Services
Li Bao, Weishi Zhang, and Xiong Xie
1
4
12
23
32
40
49
59
69
81
91

Performance Evaluation of Elliptic Curve Projective Coordinates with Parallel GF(p) Field
Operations and Side-Channel Atomicity
Turki F. Al-Somani
VPRS-Based Knowledge Discovery Approach in Incomplete Information System
Shibao Sun, Ruijuan Zheng, Qingtao Wu, Tianrui Li
NaXi Pictographs Input Method and WEFT
Hai Guo and Jing-ying Zhao
A Systematic Decision Criterion for Urban Agglomeration: Methodology for Causality Measurement
at Provincial Scale
Yaobin Liu and Li Wan
Application of Improved Fuzzy Controller for Smart Structure
Jingjun Zhang, Liya Cao, Weize Yuan, Ruizhen Gao, and Jingtao Li
Numerical Simulation of Snow Drifting Disaster on Embankment Project
Shouyun Liang, Xiangxian Ma, and Haifeng Zhang
The Simulation of Extraterrestrial Solar Radiation Based on SOTER in Zhangpu Sample Plot and
Fujian Province
Zhi-qiang Chen and Jian-fei Chen
Multiplicate Particle Swarm Optimization Algorithm
Shang Gao, Zaiyue Zhang, and Cungen Cao
Research and Design of Intelligent Electric Power Quality Detection System Based on VI
Yu Chen
99
110
117
125
131
139
144
150
158

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 1
Special Issue on Selected Papers of the IEEE International Conference on Computer and
Information Technology (ICCIT 2008)
Guest Editorial
The unprecedented advances in hardware and software technologies, computer communications, networking
technologies and protocols, the Internet, parallel, distributed and mobile computing are allowing us to enhance the way
we go about our everyday business. Consequently, demand imposed by these new and innovative applications of
computers and information technologies continues to challenge researchers to seek innovative solutions.
This Special Issue presents selected papers from the IEEE International Conference on Computer and Information
Technology (ICCIT 2008) held on December 25-27, 2008 at Khulna University of Engineering and Technology in
Bangladesh. Before introducing the synopsis of the papers, a brief introduction to the history of ICCIT is in order.
ICCIT 2008 was the eleventh annual conference in the series, the first one was held in Dhaka, Bangladesh, in 1998.
Since then the conference has grown to one of the largest conferences in the South Asian region, focusing on computer
technologies, IT and relevant areas, with participation of academics and researchers from many countries. A double
blind review process is followed whereby each paper submitted to the conference is reviewed by at least two
independent reviewers of high international standing. The acceptance rate of papers in recent years has been less than
30%, indicative of the quality of work the papers need to demonstrate to be accepted for presentation at the conference.
The proceedings of ICCIT 2008 were included in IEEExplore.
In 2008, a total of 538 full papers were submitted to the conference of which 158 were accepted for the conference
after reviews conducted by an international program committee comprising 77 members from 12 countries with
assistance from 83 reviewers. Form these 158 only 21 highly ranked papers were invited for this Special Issue. The
authors were invited to enhance their papers significantly and submit the same for review. Of those only nine papers
survived the review process and have been selected for inclusion in this Special Issue. The authors of these papers
represent academic and/or research institutions from Australia, Bangladesh, Japan, and United Kingdom. These nine
papers cover four domains of computing namely, application-driven algorithms, classification, text compression, and
electrical/digital systems.
The first paper by S.I. Ahmed, M.A. Islam, and M. Hasan presents algorithms for efficiently cutting a cornered
convex polygon out of a circle. The problem of cutting small polygonal objects efficiently out of larger planar objects
arises in many applications, such as metal sheet cutting, furniture manufacturing, ceramic industries, ornaments, and
leather industries. One of these algorithms has been shown to have better running time and the other better
approximation ratio compared to the known algorithms. The next paper by A.M. Rahman, S.S. Giasuddin, and R.M.
Rahman presents heuristic based strategies that generate efficient academic routines and exam timetables for an
educational institution that follows open credit system. The algorithm developed, based on decision tree and sequential
search techniques, shows promising simulation results to satisfy both student and teacher preferences. To provide
improved detection of network intrusion, a self adaptive Bayesian algorithm is presented by D.M. Farid and M.Z.
Rahman in the last application-driven paper. The technique proposed in this paper for alert classification is aimed at
reducing false positives in intrusion detection.
The next two papers deal with classification challenges. In the first of these two papers, M.A. Rahman presents an
approach to automatic question classification through machine learning approaches. It provides empirical evaluation of
Support Vector Machine based question classification across three variations of tree kernels as well as three major
parameters. The second classification related paper authored by M.K. Rahman and T. Endo proposes an Elman RNNbased
classifier for disease classification for a doctor patient-dialog system. A three layer memory structure is adopted
to address the challenge of contextual analysis of dialog. It simulates the human brain by discourse information.
Two of the papers next are geared to address text compression issues. The first paper by M.R. Islam, and S.A.A.
Rajon presents a low-complexity lossless compression scheme suitable for smart devices such as cellular phones and
personal digital assistants. These devices typically have small memory and relatively low processing speed. Therefore
these applications are expected to benefit from the proposed compression scheme, which offers lower computational
complexity and reduced memory requirements. Next paper by these same authors proposes a platform for evaluation of
Bengali text compression schemes. It includes a scheme for construction of Bengali text compression corpus. The paper
also presents a mathematical analysis on the data compression performance with structural aspects of corpora. The
proposed corpus is expected to be useful for evaluating compression efficiency of small and middle sized Bengali text
files
The last two papers of this Special Issue either address and/or draw inspiration from electrical/digital systems. The
paper by R. Shams, A. Elsayed, and Q.M. Akter evaluates a domain-specific Text to Knowledge Mapping prototype by
using a corpus developed for the DC electrical circuit knowledge domain. The evaluation of the prototype considers
two of its major components, namely lexical components and knowledge model. The domain-specific corpus is
expected to be useful for developing parsing and lexical component analysis tools and also contribute to domainspecific
text summarization. The final Special Issue paper by S.M. Aziz and M.D. Pham proposes a new high level
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.1-3

2 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
methodology for the design and implementation of error correction decoders for digital communication. It uses
Simulink based design flow and automatic generation of HDL codes using a set of emerging tools. The proposed
methodology significantly reduces design effort and time while providing decoder performances that are comparable to
tedious hand coded HDL-based designs.
Finally, the Guest Editors would like to express their sincere gratitude to the twenty-three reviewers of the Special
Issue from six countries (M.M. Ali, A.A.S. Awwal, K. P. Dahal, M. Erdmann, N. Funabiki, S. Haran, H-Y. Hsu, S.K.
Garg, F. Islam, P. Jiang, J. Kamruzzaman, D. Lai, A.S. Madhukumar, D. Neagu, H. Ngo, S. Pandey, Y.H. Peng, R.
Sarker, M.H. Shaheed, T. Taha, A.P. Vinod, D. Zhang, and M. Zhang) who have given immensely to this process. They
have responded to the Guest Editors in the shortest possible time and dedicated their valuable time to ensure that the
Special Issue contains high-quality papers with significant novelty and contributions.
Guest Editors:
Syed Mahfuzul Aziz
School of Electrical & Information Engineering, University of South Australia, Mawson Lakes, SA 5095, Australia
Vijayan K. Asari
Department of Electrical & Computer Engineering, Old Dominion University. 231 Kaufman Hall. Norfolk, VA
23529, USA
M. Alamgir Hossain
Department of Computing, University of Bradford, Bradford BD7 1DP, UK
Mohammad A. Karim
Office of Research, Old Dominion University, 4111 Monarch Way #203, Norfolk, VA 23508, USA
Mariofanna Milanova
Department of Computer Science, University of Arkansas at Little Rock, Dickinson Hall 515, 2801 S. University
Avenue, Little Rock, AR 72204, USA
Syed Mahfuzul Aziz received Bachelor and Masters Degrees, both in electrical & electronic engineering, from
Bangladesh University of Engineering & Technology (BUET) in 1984 and 1986 respectively. He received a
Ph.D. degree in electronic engineering from the University of Kent (UK) in 1993 and a Graduate Certificate in
higher education from Queensland University of Technology, Australia in 2002. He was a Professor in BUET
until 1999, and led the development of the teaching and research programs in integrated circuit (IC) design in
Bangladesh. He joined the University of South Australia in 1999, where he is currently serving as the inaugural
academic director of the first year engineering program. He was a visiting scholar at the University of Texas at
Austin in 1996 and a visiting professor at the National Institute of Applied Science Toulouse, France in 2006.
During 2001-2003 Dr. Aziz led avionics hardware modelling projects funded by the Australian Defence
Science and Technology Organisation (DSTO). Since 2005 he has been leading an endoscopic capsule project
and a near infrared spectroscopic instrumentation project in collaboration with Women’s and Children’s Hospital and the South
Australian Spinal Cord Injury Research Centre respectively. Recently he has received funding from the Pork CRC, Australia for a
project on Precision Livestock Farming. He has authored over eighty five research papers. His research interests include CMOS IC
design, modelling/synthesis of high performance digital systems, biomedical engineering and engineering education. Dr. Aziz is a
senior member of IEEE. He has received numerous professional awards including international and Australian national teaching
awards. He has served as member of the program committees of many international conferences and was the organising secretary of
the inaugural International Conference on Computer and Information Technology (ICCIT) in 1998. He reviews papers for the IEEE
Transactions on Computer and Electronics Letters, UK. Recently he has been appointed a reviewer of the National Priorities
Research Program, a flagship funding scheme of the Qatar National Research Fund.
© 2010 ACADEMY PUBLISHER
Vijayan K. Asari is a Professor in Electrical and Computer Engineering at Old Dominion University,
Virginia, USA, and Director of the Computational Intelligence and Machine Vision Laboratory (Vision Lab) at
ODU. He received the Bachelor's degree in electronics and communication engineering from the University of
Kerala (College of Engineering, Trivandrum), India, in 1978, the M. Tech and Ph. D degrees in electrical
engineering from the Indian Institute of Technology, Madras, in 1984 and 1994 respectively. He had been
working as an Assistant Professor in Electronics and Communications at the University of Kerala (TKM
College of Engineering), India. In 1996, he joined the National University of Singapore as a Research Fellow
and led the research team for the development of a vision-guided microrobotic endoscopy system. He joined
the School of Computer Engineering, Nanyang Technological University, Singapore in 1998 and led the
computer vision and image processing related research activities in the Center for High Performance
Embedded Systems at NTU. Dr. Asari joined Old Dominion University in fall 2000. He has so far published

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 3
more than 250 research articles including 54 peer reviewed journal papers. His current research interests include signal processing,
image processing, computer vision, pattern recognition, neural networks, and high performance and low power digital architectures
for application specific integrated circuits. Dr. Asari is a Senior Member of the IEEE, Member of the IEEE Computational
Intelligence Society (CIS), IEEE Computer Society, IEEE Circuits and Systems Society, Association for Computing Machinery
(ACM), Society of Photo-Optical Instrumentation Engineers (SPIE), and the American Society for Engineering Education (ASEE).
He was awarded two United States patents in 2008 with his former graduate students.
M. Alamgir Hossain received the Dphil degree from the University of Sheffield, UK. Currently, he is serving
as senior lecturer in the Department of Computing at the University of Bradford. He is an active member of
artificial intelligent (AI) research group. Prior to this he has held academic position at Sheffield University (as
visiting research fellow), Sheffield Hallam University (as senior Lecturer) and University of Dhaka (as
Chairman & Associate Professor of the Computer Science & Engineering Department). He has extensive
research experience in high performance real-time computing, intelligent system, optimisation, system biology
and adaptive control. He is currently leading an EU funded project, eLINK (about 5.5 million EURO) which
has ten partners from Asia and Europe. He is also acting as the UK co-ordinator of a British Council funded
research network project for higher education link programme. Dr. Hossain is currently supervising 11 PhD
students mostly to the area of intelligent systems, optimisation and systems biology. In the past, he had
involvement of many funded research projects and joint research with companies, including Balfour Beaty
Rail, Goodrich Engine Design, Aramco (Saudi Arabia), NEC (Japan) etc. Dr. Hossain acted as programme chair, organising chair
and IPC member of many international conferences. He is currently serving as an editor and member of the editorial board of three
journals. He has reviewed many journal papers, including IEEE transaction on SMC, Networking, Aerospace and Electronic
Systems, IET journals, Elsevier Science etc. Dr. Hossain has published over 120 refereed research articles and 12 books. He received
the "IEE- F C Williams" award for a research article in 1996. He is a member of the IEEE and Secretary of the CLAWAR
Association.
Mohammad Ataul Karim is Vice President for Research of Old Dominion University in Norfolk, Virginia.
Previously, he served as dean of engineering at the City College of New York of the City University of New
York. His research areas include information processing, pattern recognition, computing, displays, and electrooptical
systems. Dr. Karim is author of 15 books, 6 book chapters, and over 350 articles. He is North American
Editor of Optics & Laser Technology and an Associate Editor of the IEEE Transactions on Education. He has
served as guest editor for fifteen journal special issues. Professor Karim is an elected fellow of the Optical
Society of America, Society of Photo-Instrumentation Engineers, the Institute of Physics, the Institution of
Engineering & Technology, and Bangladesh Academy of Sciences. He received his BS in physics in 1976 from
the University of Dacca, Bangladesh, and MS degrees in both physics and electrical engineering, and a Ph.D.
in electrical engineering from the University of Alabama respectively in 1978, 1979, and 1981.
Mariofanna Milanova is Associate Professor of Computer Science in the Department of Computer Science at
the University of Arkansas at Little Rock, USA. She received her M. Sc. degree in Expert Systems and AI in
1991 and her Ph.D. degree in Computer Science in 1995 from the Technical University, Sofia, Bulgaria. Dr.
Milanova did her post-doctoral research in visual perception at the University of Paderborn, Germany. She has
extensive academic experience at various academic and research organizations including the Navy SPAWARS
System Center in San Diego, USA, the University of Louisville, USA, Air Force, Dayton , USA, the National
Polytechnic Institute Research Center in Mexico City, Mexico, the Technical University of Sofia in Bulgaria,
the University of Sao Paulo in Brazil, the University of Porto in Portugal, the Polytechnic University of
Catalunya in Spain, and at the University of Paderborn in Germany. She had grants from the German Research Foundation, the
Brazilian FAPESP State of Sao Paulo Research Foundation, the US National Science Foundation, the European Community, NATO,
and from the US Department of Homeland Security. Dr Milanova is a Senior Member of the IEEE and Computer Society, member of
IAPR, member of the IEEE Women in Engineering, member of the Society of Neuroscience and a member of the Cognitive
Neuroscience Society. Milanova serves as a book editor of two books and associate editor of several international journals. Her
main research interests are in the areas of artificial intelligence, biomedical signal processing and computational neuroscience,
computer vision and communications, machine learning, and privacy and security based on biometric research. She has published
and co-authored more than 60 publications, over 33 journal papers, 11 book chapters, numerous conference papers and 2 patents.
© 2010 ACADEMY PUBLISHER

4 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Cutting a Cornered Convex Polygon
Out of a Circle
Syed Ishtiaque Ahmed, Md. Ariful Islam and Masud Hasan
Department of Computer Science and Engineering
Bangladesh University of Engineering and Technology
Dhaka 1000, Bangladesh
Email: ishtiaque@csebuet.org, arifulislam@csebuet.org, masudhasan@cse.buet.ac.bd
Abstract— The problem of cutting a convex polygon P out
of a piece of planar material Q with minimum total cutting
length is a well studied problem in computational geometry.
Researchers studied several variations of the problem, such
as P and Q are convex or non-convex polygons and the
cuts are line cuts or ray cuts. In this paper we consider
yet another variation of the problem where Q is a circle
and P is a convex polygon such that P is bounded by a
half circle of Q and all the cuts are line cuts. We give two
algorithms for solving this problem. Our first algorithm is
an O(log n)-approximation algorithm with O(n) running
time, where n is the number of edges of P . The second
algorithm is a constant factor approximation algorithm with
approximation ratio 6.48 and running time O(n 3 ).
Index Terms— algorithms, approximation algorithms, computational
geometry, line cut, ray cut, polygon cutting,
rotating calipers
I. INTRODUCTION
The problem of cutting small polygonal objects “efficiently”
out of larger planar objects arises in many
industries, such as metal sheet cutting, paper cutting,
furniture manufacturing, ceramic industries, fabrication,
ornaments, and leather industries. This type of problems
are in general known as stock cutting problems [1].
Different types of cuts and different criteria for efficiency
of cutting are considered in cutting processes, mostly
depending up on the types of the materials. The two most
common types of cuts are the line cuts (i.e., guillotine
cuts) and ray cuts. A line cut is a line that cuts the
given object into several pieces and does not run through
the target object. A ray cut is a ray that runs from the
infinity to a certain point of the given object, possibly
to a boundary point of the target object. A line cut is
feasible for cutting out convex objects since it cuts an
object into many pieces below and above the cut. On the
other hand, ray cuts can be performed by many types
of saws such as scroll saw, band saw, laser saw and wire
saw [2]. Ray cuts can cut out non-convex objects too. But
at the same time they need to make turns in the cutting
process and so, needs some “clearance” for a turn, which
can make it impossible to cut an arbitrary non-convex
polygon. In particular, for applying ray cuts to a nonconvex
polygon it is necessary for the polygon to have no
“twisted pockets”, i.e., part of the polygon boundary that
does not see the infinity. As a whole, a cutting process
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.4-11
that uses only line cuts is much simpler than that uses
only ray cuts.
In a cutting process the main criteria for “efficiency”
of cutting is to minimize the total cutting length, which
is also known as the cutting cost. While cutting a convex
polygon it may be true that the cutting cost for ray cuts are
less than that for line cuts. But due to the above simplicity
line cuts are more popular for cutting convex objects and
are well studied as well, at least theoretically [1], [3]–[9].
Moreover, it can be shown that [9] it is not always possible
to replace line cuts by ray cuts to get better cutting cost.
In this paper we consider the problem of cutting a
convex polygon P out of a circle Q by using line cuts
where P is “much smaller” than Q, namely P is cornered
convex with respect to Q. A cornered convex polygon P
inside a circle Q is a convex polygon which is positioned
completely on one side of a diameter of Q. See Fig. 1(a).
The (cutting) cost of a line cut is the length of the
intersection of the line with Q. After a cut is made, Q is
updated to the piece containing P . A cutting sequence is a
sequence of cuts such that after the last cut in the sequence
we have P = Q. We give algorithms for cutting P out
of Q by line cuts with total cutting cost of the cutting
sequence as small as possible. See Fig. 1(b). In many
applications, such as metal sheet cutting, it is natural to
have the given object as a large circular sheet and the
target object as a sufficiently smaller convex polygon.
A. Known results
If Q is another convex polygon with m edges, this
problem with line cuts has been approached in various
ways by many researchers in computational geometry
community [1], [3]–[9]. Overmars and Welzl first introduced
this problem in 1985 [3]. If the cuts are allowed
only along the edges of P , they proposed an O(n 3 + m)time
algorithm for this problem with optimal cutting
length, where n is the number of edges of P . The problem
is more difficult if the cuts are more general, i.e., they are
not restricted to touch only the edges of P . In that case
Bhadury and Chandrasekaran showed that the problem
has optimal solutions that lie in the algebraic extension of
the input data field [1] and due to this algebraic nature of

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 5
Q
c
P
Q
(a) (b)
Figure 1. (a) A cornered convex polygon P inside a circle Q. (b) Two different cutting sequences (bold lines) to cut P out of Q; Cutting cost of
the sequence in the left figure is more than that in the right figure.
this problem, an approximation scheme1 is the best that
one can achieve [1]. They also gave an approximation
scheme with pseudo-polynomial running time [1].
After the indication of Bhadury and Chandrasekaran [1]
to the hardness of the problem, several researchers have
given polynomial time approximation algorithms. Dumitrescu
proposed an O(log n)-approximation algorithm
with O(mn+n log n) running time [5], [6]. Then Daescu
and Luo [7] gave the first constant factor approximation
algorithm with ratio 2.5 + ||Q||/||P ||, where ||Q||
and ||P || are the perimeter of Q and P respectively.
Their algorithm has a running time of O(n3 + (n +
m) log (n + m)). The best known constant factor approximation
algorithm is due to Tan [4] with an approximation
ratio of 7.9 and running time of O(n3 + m). In the
same paper [4], the author also proposed an O(log n)approximation
algorithm with improved running time
of O(n + m). As the best known result so far, very
recently, Bereg, Daescu and Jiang [8] gave a polynomial
time approximation scheme (PTAS) for this problem with
running time O(m + n6
ɛ12 ).
For ray cuts, Demaine, Demaine and Kaplan [2] gave a
linear time algorithm to decide whether a given polygon
P is ray-cuttable or not. For optimally cutting P out of
Q by ray cuts, if Q is convex and if P is non-convex
but ray-cuttable, then Daescu and Luo [7] gave an almost
linear time O(log 2 n)-approximation algorithm. If P is
convex, then they gave a linear time 18-approximation
algorithm. Tan [4] improved the approximation ratio for
both cases as O(log n) and 6, respectively, but with much
higher running time of O(n3 + m).
B. Our results
All the previous results consider Q as a polygon (either
convex or non-convex). However, to our knowledge, no
algorithm is known when Q is a circle. In this paper,
we consider the problem where Q is a circle and P is a
cornered convex polygon inside Q. We give two approximation
algorithms for this problem. Our first algorithm
1 A ρ-approximation algorithm (similarly, an approximation scheme)
has a cutting length that is ρ times (similarly, (1+ɛ) times, for any value
ɛ > 0) the optimal cutting length. Please refer to [10] for preliminaries
on approximation algorithms.
© 2010 ACADEMY PUBLISHER
P
Q
has an approximation ratio of O(log n) and runs in O(n)
time. Our second algorithm has an approximation ratio of
6.48 and runs in O(n 3 ) time.
C. Comparison of the results
While for an approximation algorithm a constant factor
approximation ratio is preferable to input-dependent
approximation ratio, the linear running time of our first
algorithm is much better than the cubic running time
of our second algorithm. The ratio 6.48 of our second
algorithm is better than that of 7.9 of Tan’s algorithm [4]
(although the later deals with Q as a convex polygon).
When both P and Q are convex polygons, almost
all the existing algorithms on line cuts use two major
steps: cutting a minimum area rectangle or a minimum
area triangle from Q that bounds P and then cutting P
out of that bounding box. Our algorithms also follow
similar approach. However, we observe that the existing
algorithms can not be applied directly to solve our
problem. Moreover, the running time of those algorithms
are too high compared to our algorithms. In particular,
Tan’s [4] constant factor approximation algorithm takes
O(n3 + m) time, the O(log n)-approximation algorithm
of [5], [6] takes O(mn + n log n) time and the PTAS
of Bereg et.al. [8] takes O(m + n6
ɛ12 ) time, where m can
be arbitrarily large. In contrast, the running time of our
algorithms are free of m and one of them is linear. Also
observe that in the existing algorithms, increasing the
value of m for “approximating” Q to a circle makes them
inefficient. See TABLE I for a summary of comparisons
among the existing algorithms and our algorithms.
D. Outline
The rest of the paper is organized as follows. We give
some definitions and preliminaries in Section II. Then we
present our two algorithms in Section III and Section IV
respectively. Finally, Section V concludes the paper with
some future works.
P
II. PRELIMINARIES
A line cut is a vertex cut through a vertex v of P if
it is tangent to P at v. Similarly, a line cut is an edge

6 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Cut Type Q P Approx. Ratio Running Time Reference
- Non-convex Ray-cuttable? O(n) [2]
Convex Convex 18 O(n) [7]
Ray cuts Convex Non-convex O(log2 n) O(n) [7]
Convex Convex 6 O(n3 + m) [4]
Convex Non-convex O(log n) O(n3 + m) [4]
Convex Convex O(log n) O(mn + n log n) [5], [6]
Convex Convex 2.5 + ||Q||/||P || O(n
Line cuts
3 + (n + m) log (n + m)) [7]
Convex Convex 7.9 O(n3 Convex Convex (1 + ɛ)
+ m)
O(m +
[4]
n6
ɛ12 Circle
Circle
Cornered convex
Cornered convex
O(log n)
6.48
)
O(n)
O(n
[8]
This paper
� ) This paper
cut through an edge e of P if it contains e. At any time
the edges of P through which an edge cut has passed are
called cut edges of P and other edges of P are called
uncut edges of P . To cut P out of Q all n edges of P
must become cut edges and for that we require exactly
n edge cuts. However, applying only edge cuts may not
give an optimal solution and we need vertex cuts as well.
In the rest of this section we give some elementary
geometry that plays important role in our paper. Let c be
the center of Q. An edge e of P is visible from c if for
every point p of e the line segment cp does not intersect
P in any other point. So, if e is collinear with c, then
we consider e as invisible. Similarly, a vertex v of P is
visible from c if the line segment cv does not intersect P
in any other point.
In this paper we do not consider the diameter of Q as
a chord, i.e., a chord is always smaller than a diameter.
Let ll ′ be a chord of Q. ll ′ divides Q into two circular
segments, one is bigger than a half circle and the other
one is smaller than a half circle. Let tt ′ be another chord
intersecting ll ′ at x such that tx is in the smaller circular
segment of ll ′ . See Fig. 2(a). The following two lemmas
are obvious and their illustration can be found in Fig. 2.
Lemma 1: xt is no bigger than ll ′ .
Lemma 2: Let △abc be an obtuse triangle with the
obtuse angle � bac. Consider any line segment connecting
two points b ′ and c ′ on ab and ac, respectively, possibly
one of b ′ and c ′ coinciding with b or c respectively. Then
the angle � bb ′ c ′ and � cc ′ b ′ are obtuse.
l
t
x
l ′
(a)
t ′
b
b ′
b ′
b ′
a
(b)
TABLE I.
COMPARISON OF THE RESULTS.
c ′
c ′
Figure 2. Illustration of Lemma 1 and Lemma 2.
© 2010 ACADEMY PUBLISHER
c
III. ALGORITHM 1
Our first algorithm has four phases : (1) D-separation,
(2) triangle separation, (3) obtuse phase and (4) curving
phase. In D-separation phase, we cut out a small portion
of Q (less than the half of Q) which contains P (and
looks like a “D”). Then in triangle separation phase we
reduce the size of Q even more by two additional cuts and
bound P by almost a triangle. In obtuse phase we assure
that all the portions of Q that are not inside P are inside
some obtuse triangles. Finally, in curving phase we cut
P out of Q by cutting those obtuse triangles in rounds.
Let C ∗ be the optimal cutting length to cut P from Q.
Clearly C ∗ is at least the length of the perimeter of P .
A. D-separation
A D of the circle Q is a circular segment of Q which
is smaller than an half circle of Q. By a D-separation
of P from Q we mean a line cut of Q that creates a
D containing P . In general, for a circle Q and a convex
polygon P there may not exist any D-separation of P .
But in our case since P is cornered, there always exists a
D-separation of P . We first find a D-separation that has
minimum cutting cost.
Lemma 3: A minimum-cost D-separation, C1, can be
found in O(n) time.
Proof: Clearly C1 must touch P . So, C1 must be a
vertex cut or an edge cut of P . Observe that any vertex
cut or edge cut that is a D-separation must be through a
visible vertex or a visible edge. Let e be a visible edge
of P . Let ll ′ be a line cut through e. Let cp be the line
segment perpendicular to ll ′ at p. If p is a point of e,
then we call e a critical edge of P and ll ′ a critical edge
cut of P . Since P is convex, it can have at most one
critical edge. Similarly, let v be a vertex of P and let
tt ′ be a vertex cut through v. Let cp be the line segment
perpendicular to tt ′ at p. If tt ′ is such that p = v, then
we call v a critical vertex of P and tt ′ a critical vertex
cut of P . Again, P can have at most one critical vertex.
Moreover, P has a critical edge if and only if it does
not have a critical vertex. See Fig. 3. Now, if P has
the critical edge e (and no critical vertex), then C1 is
the corresponding critical edge cut ll ′ . C1 is minimum,
because any other vertex cut or edge cut of P is either

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 7
closer to c (and thus bigger) or does not separate c from
P . On the other hand, if P has the critical vertex v, then
C1 is the corresponding critical vertex cut tt ′ of P . Again,
C1 is minimum by the same reason.
l
C1
e
c
Q
P
l ′
t
C1
Figure 3. Critical edge and critical vertex.
For running time, all visible vertices and visible edges
of P can be found in linear time. Then finding whether
an edge of P is critical takes constant time. Over all
edges finding the critical edge, if it exists, takes O(n)
time. Similarly, for each visible vertex v we can check in
constant time whether v is critical or not by comparing
the angles of cv with two adjacent edges of v, which takes
constant time. Over all visible vertices, it takes linear
time.
Lemma 4: Cost of C1 is at most C ∗ .
Proof: Consider any optimal cutting sequence C with
cutting cost C ∗ . C must separate P from c. However, it
may do that by using a single cut (Case 1) or by using
more than one cut (Case 2). In Case 1, if it uses a single
cut then it is in fact doing a D-separation. By Lemma 3,
it can not do better than C1, and therefore, cost of C1 is
at most C ∗ .
In Case 2, there are several sub-cases. We will prove
that in any case the cutting cost of separating P from c is
even higher than that for a single cut. Let C be the first
cut in C that separates c from P . C can not be the very
first cut of C, otherwise, it is doing a D-separation and we
are in Case 1. It implies that C is not a (complete) chord
of Q. For the rest of the proof please refer to Fig. 4. Let
a and a ′ be the two end points of C. Let bb ′ be the chord
of Q that contains aa ′ where b is closer to a than to a ′ . At
least one of a and a ′ is not incident on the boundary of
Q. We first assume that a is not incident to the boundary
of Q but a ′ is (Case 1a). Let Cx = xx ′ be the cut that
was applied immediately before C and intersects C at a.
Since Cx does not separate c from P , the bigger circular
segment created by Cx contains P , c and a. Now if x ′ is
an end point of Cx, then by Lemma 1 ab is smaller than
xx ′ . It implies that having Cx in addition to C1 increases
the cost of separating P from c (see Fig. 4(a)). Similarly,
if x ′ is not an end point of Cx, and thus another cut is
involved, then by Lemma 1 the cost will be even more
(see Fig. 4(b)).
Now consider the case when both a and a ′ are not
incident on the boundary of Q (Case 1b). So at least two
cuts are required to separate ab and a ′ b ′ from bb ′ . Let
those two cuts be Cx and Cy respectively. Let xx ′ and
© 2010 ACADEMY PUBLISHER
c
Q
v
P
t ′
yy ′ be the two chords of Q along which Cx and Cy are
applied. Now if Cx and Cy do not intersect, then for each
of them by applying the argument of Case 1a we can say
that the cutting cost is no better than that for a single
cut. But handling the case when Cx and Cy intersect is
not obvious (see Fig. 4(c, d)). Let z ′ be their intersection
point. Again, there may be several sub-cases: x and y may
or may not be end points of Cx and Cy. Assume that both
x and y are end points of Cx and Cy. Remember that none
of Cx and Cy separates P from c. So, the region bounded
by xz ′ y must contain P and c inside of it. In that case
the total length of xz ′ and yz ′ is at least the diameter of
Q, which is bigger than bb ′ (see Fig. 4(c)). For the other
cases, where at least one of x and y is not an end point
of Cx and Cy, respectively, by Case 1a the cost is even
more (for example, see Fig. 4(d)).
B. Triangle separation
In this phase we apply two more cuts C2 and C3 and
“bound” P inside a “triangle”. From there we achieve
three triangles inside which we bound the remaining uncut
edges of P . (In the D-separation phase, at most one edge
of P becomes cut).
Let C1 = aa ′ be the cut applied during the Dseparation.
We apply two cuts C2 = at and C3 = a ′ t ′
such that both of them are also tangents to P . If C2 and
C3 intersect (inside Q or on the boundary of Q), then
let z be the intersection point (see Fig. 5(a)), otherwise
let z be the point outside Q where the extensions of C2
and C3 intersect (see Fig. 5(b)). We get three resulting
triangles Ta, Ta ′, Tz having a, a ′ and z, respectively, as
a peak. We only describe how to get Ta; description for
Ta ′ and Tz are analogous. If C1 is an edge cut, then let
rr ′ be the corresponding edge such that a is closer to r
than to r ′ (see Fig. 5(a)). If C1 is a vertex cut, then let
r be the corresponding vertex of P . Let s be the similar
vertex due to C2 (see Fig. 5(b)). Then Ta = △ars. The
polygonal chain of P bounded by Ta is the edges from
r to s that reside inside Ta.
Lemma 5: Total cost of C2 and C3 to achieve Ta, Ta ′
and Tz is at most 2C ∗ . Moreover, C2 and C3 can be found
in linear time.
Proof: Whether z is within Q or outside Q, the
length of at and a ′ t ′ can not be more than twice the
length of aa ′ . By Lemma 3, aa ′ is no more than C ∗ .
Therefore, the total cost of C2 and C3 is at most 2C ∗ .
To find at linearly, we can simply scan the boundary
of P starting from the vertex or edge where aa ′ touches
P and check in constant time whether a tangent of P
is possible through that vertex or edge. Similarly we can
find a ′ t ′ within the same time.
C. Obtuse phase
Consider the triangle Ta = △ars obtained in the
previous phase. We call the vertex a the peak of Ta and
the edge rs the base of Ta. Observe that the angle of Ta
at a is acute. Similarly, the angle of Ta ′ at its peak a′ is

8 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Ta
C1
r ′
r
Q
x ′
a ′
a
b
C3
a
s
C2
z
Cx
c
x
(a)
P
C
a′ = b ′ Q
x ′
b
a
Cx
c
x
(b)
C
a ′ = b ′
y ′
x ′
Q
b
z ′
Cx
x
b ′
x ′
Q
z
y
′
y ′
a ′ Cy
Cy
c
C Cx
c
C
P a P a P
(c) (d)
Figure 4. Separating P from c by using more than one cut. Bold lines represent the cutting cost.
t
t ′
Ta
a ′
C1 C3
C2
(a) (b)
Figure 5. Triangle separation.
also acute. However, the angle of Tz at its peak z may
be acute or obtuse.
For each of Ta, Ta ′ and Tz (if Tz is acute) we apply
a cut and obtain one or two triangles such that the
angle at their peaks are obtuse and they jointly bound
the polygonal chain of the corresponding triangle. We
describe the construction of the triangle(s) obtained from
Ta only; description for the triangles obtained from Ta ′
and Tz are analogous.
Please refer to Fig. 6. Let Pa be the polygonal chain
bounded by Ta. Length of ar and as may not be equal.
W.l.o.g. assume that as is not larger than ar. Let s ′ be the
point on ar such that length of as ′ equals to the length
of as. Connect ss ′ if s ′ is different from r. We will find
a line segment inside Ta such that it is tangent to Pa and
is parallel to ss ′ . We have two cases: (i) ss ′ itself is a
tangent to Pa, and (ii) ss ′ is not a tangent to Pa.
Case (i): If ss ′ itself is tangent to Pa (at s), we apply
our cut along ss ′ . This cut may be a vertex cut through
s or an edge cut through the edge of Pa that is incident
to s. If it is a vertex cut then we get the resulting obtuse
triangle △rss ′ and the polygonal chain bounded by this
triangle is same as Pa (see Fig. 6(a)). On the other hand,
if the cut is an edge cut, let u be the other vertex of the
cut edge. Then our resulting obtuse triangle is △rus ′ and
the polygonal chain bounded by this triangle is the edges
© 2010 ACADEMY PUBLISHER
r
a
s
t ′
t
z
b
from r to u (see Fig. 6(b)). Since as and as ′ are of same
length, in either case the triangle △rss ′ or △rus ′ has
obtuse angle at s ′ .
Case(ii): For this case, let u, u ′ be two points on as
and ar, respectively, such that uu ′ is tangent of Pa and
is parallel to ss ′ . We apply the cut along uu ′ . Again,
this cut may be a vertex cut or an edge cut. If it is a
vertex cut, let g be the vertex of the cut. Then we get
two obtuse triangles △ru ′ g and △sug and the polygonal
chains bounded by them are the sets of edges from r to g
and from g to s respectively (see Fig. 6(c)). If it is an edge
cut, then let gg ′ be the edge of the edge cut with u being
closer to g than to g ′ . Then we get two obtuse triangles
△ru ′ g ′ and △sug and the polygonal chains bounded by
them are the sets of edges from r to g ′ and from g to s
respectively (see Fig. 6(d)). Again, since au and au ′ are
of same length, in either case the pair of triangles have
obtuse angles at u and u ′ respectively.
Lemma 6: Total cost of obtaining obtuse triangles from
Ta, Ta ′, and Tz is at most C ∗ . Moreover, they can be
found in O(log n) time.
Proof: Consider the construction of the obtuse
triangle(s) from Ta. Length of the cut ss ′ or uu ′ is at
most the length of rs, which is bounded by the length
of Pa. Over all three triangles Ta, Ta ′ and Tz, the total
cutting length is bounded by the length of the perimeter
of P . Since C ∗ is at least the length of the perimeter of
P , the first part of the lemma holds.
For running time, in Ta, we can find the tangent of Pa
in O(log |Pa|) time by using a binary search, where |Pa|
is the number of edges in Pa. Therefore, over all three
triangles Ta, Ta ′ and Tz, we need a total of O(log n) time.
D. Curving phase
After the obtuse phase the edges of P that are not
yet cut are partitioned and bounded into polygonal chains
with at most six obtuse triangles. In this phase we apply
the cuts in rounds until all edges of P are cut. Our cutting
procedure is same for all obtuse triangles and we describe
for only one.
Let Tu = △gus with peak u and base gs be an obtuse
triangle. (See Fig. 7). Let the polygonal chain bounded by
Tu be Pu. Let the edges of Pu be e1, . . . , ek with k ≥ 2.
x
a ′
b ′
y

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 9
r
s ′
Pa
(a)
a
s
r
a
′
s u
Ta Ta Ta
We shall apply cuts in rounds and all cuts will be edge
cuts. In the first round we apply an edge cut C ′ along
the edge e k/2. Then we connect g and s with the two
end points of e k/2 to get two disjoint triangles. Since Tu
is obtuse, by Lemma 2 these two new triangles are also
obtuse. In the next round we work on each of these two
triangles recursively and continue until all edges of Pu
are cut.
g
u
ek/2
Tu
Figure 7. Curving phase
Lemma 7: In curving phase, to cut all edges of Pu
there can be at most O(log k) rounds of cuts. Moreover,
the total cost of these cuts is |Pu| log k, where |Pu| is the
number of edges of Pu, and the running time is O(|Pu|).
Proof: At each round the number of triangles get
doubled and the number of edges that become cut also
get doubled. So after log k rounds all k edges are cut.
Let the length of Pu be Lu. At each round the total
length of the bases plus the length of the edges that are
being cut is no more than Lu. So, the total cutting cost
is Lu log k over all rounds.
For running time, finding the edge e k/2 takes constant
time. Moreover, once an edge becomes cut, it will not be
considered for an edge cut again. So, there can be at most
|Pu| edge cuts, which gives a running time of O(|Pu|).
Corollary 1: Total cutting cost of curving phase is
C ∗ log n and the running time is O(n).
Proof: Over all six obtuse triangles � Lu is no more
than the perimeter of P , which is bounded by C ∗ , and
� |Pu| = n. Therefore, for all six triangles, the total cost
is at most � Lu log k = C ∗ log n and running time is
© 2010 ACADEMY PUBLISHER
s
r
s ′
u ′
g
a
(b) (c) (d)
Figure 6. Obtaining obtuse triangle(s) from Ta.
s
u
s
r
O( � |Pu|) = O(n).
Combining the results of all four phases, we get the
following theorem.
Theorem 1: Given a circle Q and a cornered convex
polygon P within Q, P can be cut out of Q by using line
cuts in O(n) time with a cutting cost of O(log n) times
the optimal cutting cost.
s ′
u ′
g ′
IV. ALGORITHM 2
In this section we present our second algorithm which
cuts P out of Q with a constant approximation ratio of
6.48 and running time of O(n 3 ). This algorithm has three
phases: (1) D-separation, (2) cutting out a minimum area
rectangle that bounds P and (3) cutting P out of that
rectangle by only edge cuts.
The D-separation phase is same as that of our first
algorithm and we assume that this phase has been applied.
A. Cutting a minimum area bounding rectangle
We will use the technique of rotating calipers, which
is a well known concept in computational geometry and
was first introduced by Toussaint [11]. We use the method
described by Toussaint [11].
A pair rotating calipers remain in parallel. It rotate
along the boundary of an object with two calipers being
tangents to the object. If the object is a convex polygon
P , then one fixed caliper, which we call the base caliper,
is tangent along an edge e of P and the other caliper is
tangent to a vertex or an edge of P . In the next step of the
rotation, the base caliper moves to the next edge adjacent
to e and continue. The rotation is complete when the base
caliper has encountered all edges of P .
For our case we use two pairs of rotating calipers,
where one pair is orthogonal to the other. We fixed only
one caliper, among the four, as the base caliper. As we
rotate along the boundary of P , we always place the
base caliper along an edge of P and adjust other three
calipers as the tangents of P . The four calipers give us a
bounding rectangle of P . After the rotation is complete,
we identify the minimum area rectangle among the n
bounding rectangles. For that rectangle we apply one cut
a
Ta
g
u
s

10 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
along each of its edges that are not collinear with the
chord of the D.
The above technique can be done in O(n) time [11].
Once the base calipers is placed along an edge, the other
three calipers are also rotated and adjusted to make them
tangent to P . Notice that each caliper “traverses” an edge
or a vertex exactly once.
P
=⇒
Figure 8. Rotating two pairs of orthogonal calipers. The broken lines
are the total cutting cost of this phase.
Lemma 8: The cost of cutting a minimum area rectangle
out of the D achieved from the D-separation phase is
no more than 2.57C ∗ .
Proof: There can be at most four pieces, other than
the one inside the bounding rectangle, resulting from four
cuts applied to the D. The length of each cut is no more
than the portion of the perimeter of D that is separated by
that cut. So, as a whole the total cutting cost is no more
than the perimeter of D. Also see Fig. 8
Now the perimeter of D is CD + Rθ, where CD is
the length of the chord of D and θ is the angle made by
the arc of D at the center c. Since CD is the cost of Dseparation,
which by Lemma 4 is bounded by C ∗ , θ is at
most π, and C ∗ can not be more than 2R, the maximum
perimeter of D is C ∗ + (C ∗ /2)π = 2.57C ∗ .
B. Cutting P out of a minimum area rectangle by only
edge cuts
For this phase we simply apply the constant factor
approximatoin algorithm of Tan [4]. If P is bounded by a
minimum area rectangle, then Tan’s algorithm cuts P out
of the rectangle by using only edge cuts in O(n 3 ) time
and with approximation ratio (1.5 + √ 2) [4].
We summerize the result of our second algorithm in
the following theorem.
Theorem 2: Given a circle Q and a cornered convex
polygon P of n edges within Q, P can be cut out of Q
by using line cuts in O(n 3 ) time with a cutting cost of
6.48 times the optimal cutting cost.
Proof: We have the cutting cost of C ∗ for D-
Separation, 2.57C ∗ for cutting the minimum area rectangle,
and (1.5 + √ 2)C ∗ for cutting P out of the rectangle,
which give a total cost of 6.48C ∗ .
© 2010 ACADEMY PUBLISHER
P
V. CONCLUSION
In this paper we have given two algorithms for cutting
a convex polygon P out of a circle Q by using line cuts
where P resides in one side of a diameter of Q. Our first
algorithm is an O(log n)-approximation algorithm with
O(n) running time, where n is the number of edges of P .
Our second algorithm is a 6.48-approximation algorithm
with running time O(n 3 ).
While there exist several algorithms when Q is another
polygon, we are the first to address the problem where Q
is a circle. Our first algorithm has a better running time
and the second one has better approximation ratio than
the best known previous algorithms that deal with Q as
a convex polygon.
There remain several open problems and directions for
future research:
1) The general case of this problem where P is not
necessarily in one side of a diameter of Q is still
to be solved.
2) We also think it would be interesting to see approximation
schemes for this problem.
3) Finally, in many industry applications it is common
to have both P and Q as 3D objects, for which we
do not know any algorithm. In future it is important
to design similar algorithms for the 3D case.
REFERENCES
[1] J. Bhadury and R. Chandrasekaran, “Stock cutting to
minimize cutting length,” European Journal of Operations
Research, vol. 88, pp. 69–87, 1996.
[2] E. D. Demaine, M. L. Demaine, and C. S. Kaplan, “Polygons
cuttable by a circular saw,” Computational Geometry:
Theory and Applications, vol. 20, pp. 69–84, 2001.
[3] M. H. Overmars and E. Welzl, “The complexity of cutting
paper,” in Proc. 1st Annual ACM Symposium on Computational
Geometry (SoCG’85), 1985, pp. 316–321.
[4] X. Tan, “Approximation algorithms for cutting out polygons
with lines and rays,” in Proc. 11th International Conference
on Computing and Combinatorics (COCOON’05),
ser. LNCS, vol. 3595. Springer, 2005, pp. 534–543.
[5] A. Dumitrescu, “The cost of cutting out convex n-gons,”
Discrete Applied Mathematics, vol. 143, pp. 353–358,
2004.
[6] ——, “An approximation algorithm for cutting out convex
polygons,” Computational Geometry: Theory and Applications,
vol. 29, pp. 223–231, 2004.
[7] O. Daescu and J. Luo, “Cutting out polygons with lines and
rays,” International Journal of Computational Geometry
and Applications, vol. 16, pp. 227–248, 2006.
[8] S. Bereg, O. Daescu, and M. Jiang, “A PTAS for cutting
out polygons with lines,” in Proceedings of the 12th Annual
International Conference on Computing and Combinatorics
(COCOON’06), ser. LNCS, vol. 4112. Springer,
2006, pp. 176–185.
[9] R. Chandrasekaran, O. Daescu, and J. Luo, “Cutting out
polygons,” in Proc. 17th Canadian Conference on Computational
Geometry (CCCG’05), 2005, pp. 183–186.
[10] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein,
Introduction to Algorithms. Cambridge, Massachusetts:
MIT Press, 2001.
[11] G. Toussaint, “Solving geometric problems with the rotating
calipers,” in Proc. 2nd IEEE Mediterranean Electrotechnical
Conference (MELECON’83), Athens, Greece,
1983.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 11
Syed Ishtiaque Ahmed was born in 1986 in Pabna,
Bangladesh. He completed his B.Sc. Engineering in Computer
Science and Engineering from Bangladesh University
of Engineering and Technology (BUET) in March,
2009. Currently he is an M.Sc. student in the same
department. He is also working as a Lecturer in United
International University, Dhaka, Bangladesh. During his
undergraduate studies, he got University Merit Scholarship,
Deans List Award, Imdad Siatara Khan Foundation
Scholarship etc. for outstanding academic performances.
His research interests have focused on the problems of
Computational Geometry.
Md. Ariful Islam was born in Bogra, Bangladesh in
1987. He received his B.Sc. Engineering in Computer
Science and Engineering from Bangladesh University of
Engineering and Technology (BUET) in March, 2009.
While studying as an undergrad student, he got University
Merit Scholarship, Dean’s List Award, Educational Board
Scholarship. He was the moderator of BuetNet (Association
of BUET internal network). He is currently working
as a Quantitative Software Developer at Stochastic Logic
Limited, Bangladesh. At the same time, he is doing his
M.Sc. in computer science and engineering in BUET. His
research interests include Computational Geometry, Data
Mining, and Artificial Intelligence.
© 2010 ACADEMY PUBLISHER
Masud Hasan was born in Kurigram, Bangladesh in
1973. He received his B.Sc. and M.Sc. Engineering in
Computer Science and Engineering from Bangladesh University
of Engineering and Technology (BUET) in 1998
and 2001, respectively. He received his PhD in Computer
Science from the University of Waterloo, Canada in 2005.
In 1998, he joined the Department of Computer Science
and Engineering of BUET as a Lecturer, and since 2001
he is an Assistant Professor in the same department.
His undergraduate and graduate teaching includes Algorithms,
Graph Theory, Mathematical Analysis of Computer
Science, Computational Geometry, and Algorithms
in Bioinformatics. His main research interests focus on
Algorithms, Computational Geometry, and Algorithms in
Bioinformatics. He is the author of more than 25 peerreviewed
international journal and conference papers. He
has served as a reviewer for many international journal
and conferences.

12 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Decision Tree Based Routine Generation (DRG)
Algorithm: A Data Mining Advancement to
Generate Academic Routine and Exam-time
Tabling for Open Credit System
Ashiqur Md. Rahman
North South University, Computer Science and Engineering Department, Dhaka, Bangladesh
Email: ashiq_rahman@yahoo.com
Sheik Shafaat Giasuddin and Rashedur M Rahman
North South University, Computer Science and Engineering Department, Dhaka, Bangladesh
Email: rshafaat@gmail.com, rashedurrahman@yahoo.com
Abstract—In this paper we propose and analyze techniques
for academic routine and exam time table generation for
open credit system. The contributions of this paper are
multi-folds. Firstly, a technique namely Decision tree based
Routine Generation (DRG) algorithm is proposed to
generate an academic routine. Secondly, based on the DRG
concept, Exam-time Tabling algorithm (ETA) is developed
to implement conflict free exam-time schedule. In open
credit course registration system any student may choose
any course in any semester after completion of pre-requisite
course(s). This makes the research more challenging and
complex to accomplish. Academic routine and exam timetable
generation are in general NP-Hard problems, i.e., no
algorithm has been developed to solve it in reasonable
(polynomial) amount of time. Different methods based on
heuristics are developed to generate good time-table. In this
research we developed heuristic based strategies that
generate an efficient academic routine and exam time-table
for a university that follow open credit system. OLAP
representation helps to classify the courses along with the
proposed algorithm to eliminate some constraints. Daybased
pattern, minimum manhattan distance between
courses of same teacher; minimum conflicted course
distribution has been stage-managed to classify the courses.
Our ETA algorithm is based on decision tree and sequential
search techniques.
Index Terms— OLAP, Crosstable, Conflict List, Favorite
Slot, Faculty Choice, Course Color, Day-time slot pattern.
I. INTRODUCTION
This paper depict Decision Tree based Routine
Generation (DRG) algorithm to generate a university
class routine within a tolerable range of some constraints
and conflict free Exam-Time Tabling algorithm (ETA). A
decision-tree based classification algorithm has been
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.12-22
introduced to solve this NP-Hard Problem [22]. CPL
(constraint logic programming) is a respected technology
for solving hard problems which include many (nonlinear)
constraints [1]. Constraints propagation technique
has been applied to overcome the preferential
requirements for slots of teachers, courses from preadvising
by students and class room allocation. Versatile
choices for courses may lead to a deadlock situation.
Golz used priorities heuristic ordering [2] where
Abdennadher introduced an optimized cost-based rule
mining [3,4] to solve these type of problems. On the other
hand, knowledge based in a hyper heuristic course
scheduling using case based reasoning is used to
maximize the rule covering area [5]. Further expansion is
possible to accomplish the exam-time tabling using
OLAP technique [16]. Exam-time tabling is another
highly constrained combinatorial optimization problem.
The major objective is to confirm 100% conflict free
exam schedule with a fixed interval of days. Limited
room capacity and room availability problems must be
overcome to place exams on each time slot. The
computational time is reduced by using heuristic based
search in comparison with the permutation of courses for
exam-time tabling. Identification of a novel heuristic is
the most challenging task. Using OLAP, proposed
conflict free Exam-Time Tabling algorithm (ETA)
produces substantial results to accommodate all students
with zero conflict tolerance.
DRG presents key features of generating class routine
with minimum computational time. Heterogeneous
distribution of courses is classified with maximum
satisfaction of all constraints. Section II describes about
previous related works and preliminaries in details.
Section III illustrates the problem dimensions with the
data filtering technique used in the paper to summarize
further manipulation; pursued by the proposed algorithm,
DRG, and the classification procedure to find the feasible
solution in Section IV followed by Exam-Time Tabling
algorithm (ETA) in section V. Extensive computational

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 13
results are conducted to study the performance analysis
for both algorithms in section VI. Finally the paper
concludes the work in section VII.
II. PRELIMINARIES &RELATED WORK
A university class routine generation problem – as
considered in this work – consists of assigning each
course in a set of slots (classes) in a limited class rooms
within the teachers’ favorite time slots. This highly
constrained problem is optimized by simulated annealing
and genetic algorithm [6,7]. Seven different major and
minor objectives are discovered and deadlock situation is
overcome by randomly exploring the composite
neighborhood [8]. The most closely related attempt with
this work appears to be the constraint programming
approach used by Boizumault [9] and the simulated
annealing approaches explored by Dowsland and
Thompson [10,11]. The principal innovation in DRG is
the sequential use of these two methods. DRG may select
some poor slots, with respect to teachers or students,
under a tolerable conflict range. A similar sequential
approach has been taken in DRG on other problems:
White and Zhang [12] used constraint programming to
second a starting point for tabu search in solving course
timetabling problems. For a high school timetabling
Yoshikawa tested several combinations of two stage
algorithms, including a greedy algorithm followed by
simulated annealing and a constraint programming phase
followed by a randomized greedy hill climbing algorithm
(which is deemed to be the best combination of those
used). In a similar vein, Burke, Newell and Weare [13]
used their work on sequential construction heuristics [14]
to generate initial solutions for their memetic algorithm
[15].
Wide variety of courses increases the emergence to
provide adequate exam-time tables for the educational
institutions. The development of an examination
timetable requires the institution to schedule a number of
examinations in a given set of time slots, so as to satisfy a
given set of constraints. A common constraint for any
educational institution is that none of the students can
have more than one exam scheduled at the same time.
Many other constraints were presented by Marlot in [17].
Sequential construction heuristics have been applied to
the publicly available data in a variety of forms by
selecting exams from a randomly chosen subset of all
exams by Burke [14] where Carter [19, 20] allow limited
backtracking de-allocation of exams. On the other hand
Caramia [18] includes an optimization step after each
exam allocation. Sequential construction heuristics order
the exams in some way and attempt to allocate each exam
to an ordered session by satisfying all the constraints.
Using a memetic algorithm for exam timetabling Burke,
Newall and Weare [15] proposed a hybrid algorithm
consist of a simulated annealing phase to improve the
quality of solution, and a hill climbing phase for further
improvement. To avoid local maxima problem these
solutions require random jitter [21] whereas the proposed
algorithm has no impact on randomization.
© 2010 ACADEMY PUBLISHER
III. PROBLEM DESCRIPTION &DATA FILTERING
The routine maps a set of courses chosen by students
and teachers to a specific room and time-slot. A major
objective, in developing an automated system, is to
minimize the hassle of separating conflicted courses from
choices by students. In this paper the major identified
problems are (a) Number of lectures per week for each
course are fixed, (b) Room overlapping is prohibited, (c)
Fitting the routine with teacher’s favorite timeslots, and
(d) trying to assert different timeslots to same level of
courses. On the other hand, the minor objectives are (e)
day-timeslots pattern for the course, (f) room capacity,
(g) avoiding gaps between classes of same teacher, if
possible, (h) single class for student per day and (i)
ensuring compactness of interclass time difference.
The required scattered data contains total courses
(course choices from pre-advising by students) C = {c1,
c2, c3, …, cn} where the dependencies between courses
are also maintained. Here course dependency can be
defined as �Ci, Ci � Cj where Ci, Cj � C. For this paper
the students’, S = {S1, S2, S3, …, Sz}, course choices can
be derived as Sj = {ci} where � i, ci �C and 1 � i � n and
|Sj| = max_course_choice for the student as shown in Fig.
1. Teachers’ favorite timeslots are grouped according to
day-timeslots pattern. Here group A and B is formed for
the teacher tk, where T = {t1, t2, t3, …, tm}, A(tk) =
{favorite time_slots of tk | sequential time slots for
Saturday, Monday and Wednesday} and B(tk) = {favorite
time slots of tk | sequential time slots for Sunday, Tuesday
and Thursday}, whereas time_slots = {1, 2, …, 30}
contains 5 sequential slots per day starting from Saturday.
Here Friday is considered as an off-day. Priority of the
teacher has been introduced by P(tk) = {1, 2, …, 10}
where a higher value represents higher priority. An
exceptional priority is also introduced as 11 reflecting
part-time faculty, whose projected time cannot be
changed. Fig. 2. and Fig. 3. shows the teachers’ wish-list
and the course distribution among the teachers’. Target of
this work is to find the values of the “class slot routine”
field of the Fig. 3. The resultant routine vector, V =
{{ci}q} � i, ci �C and 1 � i � n and 1 � q � 30, consists of
the course classification as per day required for each
course and class room availability.
In this paper, this huge dimensionality of dataset is
reduced by initiating an OLAP (On-Line Analytical
Processing) representation [16]. Here a Crosstable (Cr) of
(n × n) courses are initialized. Cr (n × n) = {conflicti, j}
where 1 � i,j � n and ‘n’ is the total number of courses
requested by the student (or is ‘n’ number of courses
offered by the department). Here the conflict of Cri,j is a
positive integer that reflects the common students
between ci and cj. The diagonal values of Cri,i show the
total number of students requesting for the course ci.
�Cri,j, [1 � j � n] is the total conflict for the course ci [�
i, i � j]. Maximum “chaos” (conflict) courses can be
easily sorted out from this two dimensional conflict
distribution (Crosstable).
To minimize the potential for time conflict, an
admissible heuristic (h) is imposed to regroup the courses
according to their dissimilarity. Graph Coloring

14 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
algorithm is used to cluster where the conflict in
Crosstable confirms the weights of the edges in the
graph. Here the admissible heuristic is applied as the
maximum number of colors needed to find the minimum
number of groups of courses. Here faculty redundancy is
also considered as weight, that is, same faculty of
different courses cannot be in the same group.
Student
ID
Course ID Semester ID
S1 C1 1
S1 C3 1
S2 C1 1
S2 C3 1
S3 C2 1
S3 C4 1
S3 C5 1
S4 C4 1
Figure 1. Courses Pre-Advised by the Students.
Teacher ID
Favorite
Slots
Priority
t1 7,8,9,13,17,18,19,22,23,24 7
t2 2,12,22,29 11
t3 5,10,15,20,25,30 9
t4 2,5,8,9,12,15,18,22,23,24,25 10
Figure 2. Teachers’ Slots Preferences along with Priority.
Teacher
ID
Course ID
Class
Slot
Routine
Sem.
ID
Class
per
Week
t1 c1 7,17 1 2
t1 c4 8,18,24 1 3
t2 c2 2,12 1 2
t3 c3 5,15 1 2
t4 c5 5,15,25 1 3
Figure 3. Courses and Classes Distribution for the Teachers.
The output of Graph Coloring Algorithm assigns a color
to all courses individually in Fig. 5(a); same colored
courses are treated as a group. Fig. 4. shows the resultant
Crosstable. Each group may consist more than the
threshold limit members with tolerable conflict range.
Here the number of the rooms is considered as the
threshold value. In this work the tolerable conflict range
is set to 0.
Course
ID
c1 c2 c3 c4 c5
Total
Conflict
c1 40 5 7 0 0 12
c2 5 50 0 2 1 8
c3 7 0 35 5 0 12
c4 0 2 5 40 1 8
c5 0 1 0 1 45 2
t1 t2 t3 t1 t4
Figure 4. Crosstable for n × n Courses.
This easy formation of coloring may lead to a measure
of the undesirability of having classes overlapped in the
routine. It will be effective to try to fit the most “chaos”
courses of high priority teachers in the routine first.
Random selection may be used to select teacher having
© 2010 ACADEMY PUBLISHER
same priority. The data used in this work to test the
algorithm is real.
Constraint programming model and data filtering
techniques for routine generation motivate the increasing
interest to develop an exam-time tabling. In routine
generation algorithm the colored courses refer the non-
conflicting sets of courses. Faculty redundancy is not
considered as constraint any more but the room capacity.
For ETA the graph consisting edge weight between two
courses Dy,z = Cy + Cz – Cri,j where Cy and Cz represents
the number of students of Ci and Cj courses respectively.
(a) (b)
Figure 5(a). Courses Graph Color for DRG. (b). Courses Graph
Color for ETA.
�i,j �Dy,z � total_room_capacity and |Dy,z| �
room_capacity shown in Fig. 5(b). Important factor of
grouping courses is that the number of members in each
group must not exceeds the total number of room
availability. The minimum requirement of days for 100%
concurrent conflict free exam is greater than or equal to
the number of groups that is the numbers of color
requires coloring the graph. If each day consists of more
than one slot of examination and the provided day is less
than the numbers of color then the Crosstable is able to
ensure the numbers of consecutive examinations for an
individual or groups of students on the day. Each exam
slot holds a group of courses only, with zero conflict. But
the consecutive slot may embrace some conflicts among
the groups due to differ in color. By using dynamic data
structure the number of consecutive examinations
between different groups of courses can be easily sorted
out described in section V.
IV. THE DECISION TREE BASED ROUTINE GENERATION
(DRG) ALGORITHM
The aim of the proposed DRG algorithm is to classify
all the courses with a degree of satisfaction. Enormous
permutation of courses may lead to a time consuming
process. So a standardized branch & bound condition
may be applied to reduce the problem surface area. The
DRG sequentially follows 4 sets of cascading decision
trees. Depending upon the emergence and success rate,
the result of one tree is propagated to another tree as
shown in Fig. 6. These transitions may lead to a solution
but also may degrade the satisfaction threshold. A control
portion helps to justify the problem solution needed to be
more explored or not.
Each transition from one decision tree to another
shrinks the overall problem surface area by eliminating
the classified courses. Classical decision tree takes certain
decision depending upon some gain factor. Beside this,
the proposed trees concentrate on the reduced problem

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 15
dimension which helps to classify the unclassified
courses with tolerable time complexity.
The key factor for each of the four decision trees is (a)
PDRG: Teacher Priority, (b) CDRG: Highest conflicted
course, (c) TDRG: Tolerable conflict and (d) NTRG:
Neighbor slots by ignoring teacher’s wish list. A
university routine is created and remains unchanged for a
particular semester. If the placed courses do not match
the favorite slots of the teacher, the evolved
dissatisfaction is much higher for a higher prioritized
teacher. So, we used PDRG as our first decision tree.
From an observation it is clear that the placing conflicted
course in a dense routine vector is difficult as it can
introduce student conflict in the routine. So, it will wise
to consider the higher conflicted courses first as they are
the principle component which reflects the major
problem surface area. By doing this the problem surface
area is reduced easily. For this reason CDRG is the
second selection. The decision tress TDRG and PDRG
are quite similar. But TDRG introduces considerable
student conflict in to the routine vector.
Figure 6. Program flow of DRG
Therefore, TDRG plays the third role in the whole
algorithm. After all the three decision trees, the
unclassified and partially classified courses shows, there
is no place (slot) according to the teacher’s favorite slots.
Exploring over the contour of the teachers favorite slots
is necessary in order to achieve the course’s class per
week constraints. Only student conflict is considered in
TDRG where the day-time pattern is maintained strictly.
On the other hand, NTDRG will dissatisfy the teachers &
may not follow the day-time pattern. Hence, NTDRG is
the final decision tree.
A. Priority regulated DRG (PDRG)
The first decision tree accumulates the high prioritized
teacher tk and most conflicted course ci of tk to the routine
© 2010 ACADEMY PUBLISHER
vector with maximum fitness value. The max_fit function
tries to discover the day-time pattern for the course. If the
consequential day-time slots are already occupied by
other courses, it checks the corresponding course color. If
the color of the courses is same it classifies the course ci
with a degree. Here the fitness value of a course is
referred as degree. The highest returned degree of a daytime
slot, as maximum fitness value, is selected as the
course class for ci. This operation provides a patternbased
course distribution, A(tk) or B(tk), in the routine
although some courses may be placed partially as per
their class_per_week and max_class_per_slot constraints.
In this manner, the low priority teachers may suffer by
not getting the classes in a sequential manner. But in
practice 26% of the total courses can be placed with zero
conflict and with a high level of satisfaction. The level of
satisfaction is a quantitative measure of placement for a
course with respect to the teachers’ favorite slots. The
definitive PDRG tree is shown in Fig. 7. The partially
placed and not yet placed courses then elected as
cascaded input to second level of exploration. The pseudo
code presented in algorithm 1 describes actions to be
taken by Priority regulated DRG algorithm (PDRG). The
computational time of this operation requires O (n × m)
where the maximum number of courses per teacher is ‘n’
and the number of faculties is ‘m’.
PDRG( )
{
// select the high prioritized teacher;
// select most conflicted course of the teacher;
// select the corresponding faculty’s low frequent favorite
// time slot
1. Find max-patterned day time slots;
IF the time slot is empty PLACE the course;
2. ELSE find the color of the course that already placed
on the slot;
2.1. IF the color is same AND on the range of the
room AND the course is not already placed on
that day before, PLACE the course;
2.2. ELSE select the next max-pattern;
3. every time after PLACEing the course, remove the
slot from the faculty favorite slot list;
4. repeat the step 1,2 until the course slot remain
unchanged;
}
Algorithm 1. Priority regulated DRG (PDRG)
B. Chaos eradication DRG (CDRG)
In this decision tree, the less demanded time slots (with
respect to number of rooms) are labeled as “cold”
whereas high demanded ones are labeled as “hot”.
Among the remaining most conflicted courses (may not
or partially placed) with low frequent time slots of the
corresponding teacher are chosen for the second decision
tree. If the considerable slot is empty then the course is
placed in that slot, otherwise the colors have to be
matched. If the color of the courses placed in the slot
matches with the concerning course, the latter course is

16 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
classified, and if not other options are taken into
consideration for the teacher. The considerable courses in
CDRG are the overlooked courses from PDRG, where
PDRG confirms the day-time pattern is not possible for
the courses due to color and the scattered choice of slots
by the teachers. So day-time slots pattern are ignored in
this operation. After using this second decision tree few
courses may remain unchanged. Nevertheless this time
the number of remaining courses is much less than the
previous one. Around 32% of the remaining courses are
placed successfully by CDRG. The combined effort of
decision trees still provides high confidence. Fig. 8.
demonstrates the Chaos eradication DRG. The pseudo
Figure 7. Decision Tree of PDRG
code presented in algorithm 2 describes actions to be
taken by Chaos eradication DRG (CDRG). Now this tree
holds time complexity of O((n – d) × (m – l)) where the
maximum number of courses per teacher is ‘n’, the
number of faculties is ‘m’ and ‘d’ is the already classified
courses of each teacher by the PDRG and ‘l’ is the
number teacher whose all courses were placed in the
routine by PDRG.
CDRG( )
{
// select the most conflicted courses not yet placed or
// partially placed;
// select the remaining low frequent favorite slot of the
// faculty;
1. IF the slot is empty PLACE the course;
2. ELSE find the color of the course that already
placed on the slot;
2.1. IF the color is same AND on the range of the
room AND the course is not already placed on
that day before, PLACE the course;
2.2. ELSE select the next low frequent favorite slot
of the faculty;
3. Repeat the step 1,2 until the course remain
unchanged state;
© 2010 ACADEMY PUBLISHER
}
Algorithm 2. Chaos eradication DRG (CDRG)
C. Tolerable DRG (TDRG)
The first two decision trees are aimed at automated
generations of a better assignment. The second approach
seeks to find an assignment of vector which may be more
difficult to locate in the search space using the already
assigned vector. The third decision tree allows the
remaining courses according to the priority of the
teachers to find a place into the routine within a tolerable
conflict range of the subsequent teachers’ favorite slots.
Here the course color is overlooked. This classification
now introduces errors into the system by considering the
tolerable students conflicts only. Important issue is that,
this manipulation may iterate several times to include as
many courses possible to place in to the routine. 22% of
the unclassified and partially classified courses are
labeled with a tolerable error.
Figure 8. Decision Tree of CDRG
The flowchart presentation of this decision tree is
shown in Fig. 9. The pseudo code presented in algorithm
3 describes actions to be taken by Tolerable DRG
(TDRG). The average time complexity of TDRG is
approximately O (t × (n – (d + o)) × (m – (l + p)) where
the maximum number of courses per teacher is ‘n’, the
number of faculties is ‘m’ and ‘d’ is the already classified
courses of each teacher by the PDRG and ‘l’ is the
number teacher whose all courses were placed in the
routine by PDRG. ‘o’ is the number of courses placed by
CDRG and ‘p’ is the number of teachers whose courses
are completely placed by CDRG. Here ‘t’ is the number
of TDRG iterates.
TDRG( )
{
// select the most conflicted courses not yet placed or

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 17
// partially placed of a high prioritized teacher;
// select the remaining low frequent favorite slot of the
// faculty;
1. IF the slot is empty PLACE the course;
2. ELSE IF on the range of the room AND conflict
among the courses are in tolerable range AND the
course is not already placed on that day before,
PLACE the course;
3. ELSE select the next low frequent favorite slot of the
faculty;
4. Repeat the step 2,3 until the course remain
unchanged state;
}
Algorithm 3. Tolerable DRG (TDRG)
D. Neighboring Tour DRG (NTDRG)
The fourth possibility search allows the courses to look
over the contour of the teachers’ favorite slots to find the
least conflicted slots. Although the students’ scattered
choices is overlooked in TDRG but the placed courses do
not displease teachers’ preference. The final decision tree
is modeled in such a manner so that the rest of unplaced
courses are going to be graded into the routine vector
within a minimum distance of the teachers’ choice.
Manhattan Distance (MD) is calculated for this placement
of courses. MD = total slot gap of tk on each day i.e. the
total unused slots of a teacher on a particular day.
Manhattan Distance is a vital performance measuring tool
to find the slot gaps per day for an individual teacher. The
optimization can be done by keeping Cumulative
Manhattan Distance (cMD) for the teachers as low as
possible where cMD = �all used day by t MD(t) It is assured
that the courses were not yet placed on that day earlier.
From the remaining courses, a course is selected
according to the priority of the teacher.
Figure 9. Decision Tree of TDRG
Complement of the intersection between the classification
value of that course and the corresponding teachers’ used
slots are considered as new host slots. The neighboring
slots of the new hosts are the most likely candidates.
© 2010 ACADEMY PUBLISHER
Among the candidates the most “cold” (less desired slots)
slots are considered as candidates.
Considerable issues in this placement are tolerable
conflict range, allocable number of rooms and day-time
misjudgment. Fig. 10. demonstrates the overall scenario
of Neighboring Tour DRG (NTDRG). The approximate
time complexity is O(2 × (n – (d + o + u)) × (m – (l + p +
v)) � O((n – (d + o + u)) × (m – (l + p + v)) where the
maximum number of courses per teacher is ‘n’, the
number of faculties is ‘m’ and ‘d’ is the already classified
courses of each teacher by the PDRG and ‘l’ is the
number of teacher whose all courses were placed in the
routine by PDRG. ‘o’ is the number of courses placed by
CDRG and ‘p’ is the number of teachers whose courses
are completely placed by CDRG. ‘u’ is the number of
courses placed by TDRG and ‘v’ is the number of
teachers whose courses are completely placed by TDRG.
The pseudo code presented in algorithm 4 describes
actions to be taken by Neighboring Tour DRG (NTDRG).
NTDRG( )
{
// select the courses not yet placed or partially placed;
// select the corresponding faculty routine;
1. Find the candidate slot (where candidate slot is the
neighboring slots of the used slots of the faculty);
2. IF on the range of the room AND conflict among the
courses are in tolerable range AND the course is not
already placed on that day before, PLACE the
course;
3. ELSE select the next candidate slot of the faculty;
4. Repeat the step 2,3 until the course remain
unchanged state;
}
Algorithm 4. Neighboring Tour DRG (NTDRG).
Figure 10. Decision Tree of NTDRG
These four sequential decision trees feed forward to an
acceptable solution. It is ensured that the course is not yet

18 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
place on that very day whenever the “PLACE” decision
is taken. The Overall Complexity is shown in equation
(1).
= PDRG + CDRG + TDRG + NTDRG
= O(mn) + O( (n-d)(m-l) ) + O( t(n-(d+o))(m-
(l+p)) ) + O( (n-(d+o+u))(m-(l+p+v)) )
= mn + mn – nl – md + dl + t(mn – nl – np – md
+ dl + dp – mo + ol + op) + mn – nl – np – nv
– md + dl + dp + dv – mo + ol + op + ov – mu
+ ul + up + uv
= (t + 3)mn - (d + td + to + d + o + u)m – (l + tl
+ l + tp + v)n + (dl + tdl + tdp + tol + top + dl
+ dp + dv + ol + op + ov + ul + up + uv)
= A.mn – B.m – C.n + K --------------------- (1)
where,
A = t + 3;
B = (t + 2)d + (t + l)o + u;
C = (t + 2)l + tp + v;
and K = dl + tdl + tdp + tol + top + dl + dp + dv + ol
+ op + ov + ul + up + uv
The cumulative time complexity of DRG is O(A.mn -
B.m - C.n), The cumulative time complexity of DRG
mainly depend upon the number of iterations in TDRG
algorithm and number of courses placed by each and
every decision tree. The algorithm degrades due to the
fact that the students have a freedom to choose any
course (assuming that the prerequisite course is
completed).
Firstly PDRG faces problem if same prioritized
teachers focus into same favorite time slots. By using
CDRG this situation may prevail over considering a level
of discontinuity into the day-time pattern for the courses.
In second run, if the low “chaos” course holds high
prioritized teacher then the classification may dissatisfy
the teacher. For the third rotation the conflict may arise
for the students for not considering the color. For
NTDRG, if the host slot is elected as the first (V{{ci}q} �
i, ci�C and 1 � i � n and 1 � q � 30 where q = 6, 11, 16,
21, 26 is the first slots of the day) and last (V{{ci}q} � i,
ci�C and 1 � i � n and 1 � q � 30 where q = 5, 10, 15, 20,
25 is the last slots of the day) slot of the day, the previous
and the next consecutive slots are from different days. So,
this day jump increases huge distance for the teacher
which may lead to an unfeasible classification. After
NTDRG a few courses may remain partially classified or
unclassified due to three major factors (1) Number of
rooms not adequate, (2) Teacher’s preferred time slot is
not applicable and (3) Student conflict may cross
tolerable conflict range.
The resultant unclassified or partially classified
courses, after all decision trees exploration, represent the
problem can not be solved without the extensive
relaxation of the constraints mentioned earlier. Unclassification
or partially classification may tag to a
course not only for the conflict but also for room
availability. So, such event may occur where there is no
conflict among the courses but the low prioritize course
(due to teacher’s priority in PDRG or low conflicted
score in CDRG or little bit higher considerable conflict
score in TDRG ) is not able to be placed in to the routine
© 2010 ACADEMY PUBLISHER
vector for room limitation. Same situation may arise for
other two constraints. So, to eliminate the partially
classified or unclassified courses the above mentioned
factors have to be compromised that is by increasing the
room capacity or expanding teacher’s favorite time or
ignoring students’ conflicts.
V. EXAM-TIME TABLING ALGORITHM (ETA)
Typical constraint programming method is applied to
the above exam-time tabling problem. By considering the
equal or less number of members only from the power set
of the groups of courses, the ETA calculates the total
number of conflicts and the total number of students on
each group. The proposed algorithm also tries to place the
course groups on a specific exam slot. The total numbers
of valid power set are the combination of the groups
along with the given slots for the exam. Valid range in
calculating the number of members is 1 to equal to given
slot value ‘s’ i.e. Gn Cs + Gn Cs-1 +… + Gn C1 where ‘Gn’ is
the total number of course groups. By sorting the groups
in descending order according to their total student to
conflict ratio, the ETA tries to judge the most appropriate
groups to place in to the exam-time table first. Here if the
conflict signifies zero then the total student to conflict
ratio remains equal to the total number of the students.
Among the sorted power set list the most supported
single member is selected to be placed on each day of the
exam-time table by using greedy algorithm i.e. by
choosing the most profitable groups first. Support vector
is calculated from the ratio of the total student and
conflict. If the provided day for exam scheduling is less
and the groups of courses and number of slots is more
than one on each day, then the placed group will try to
calculate the most appropriate unplaced groups and select
the group to form the pair. This situation may produce
concurrent conflicts but conflicts among the courses on
each slot remain zero. These newly paired groups are
eliminated from the list so that other placed groups can
select their feasible member groups. Considering the
scenario described in Fig. 4, the resultant colorized
course groups are G1 = {c1, c4} 80 and G2 = {c2, c3} 85 and
the power set of the groups are {G1}, {G2}, {G1, G2}.
Here the highest considerable set are those, whose
number of members are less or equal to the provided slots
for exam per day. Total student to conflict ratio of the
selective power sets are 80, 85 and 165/19, where the
ETA mainly concentrates on. If the provided days for the
exam are equal to the number of groups then the most
appropriate exam schedule is Day 1 = G1 and Day 2 =
G2. There is an 11.5% possibility of having concurrent
exam, if the provided days for exam are less then the
number of colorized course groups. The pseudo code
presented in algorithm 5 describes actions to be taken by
Exam-Time Tabling Algorithm (ETA).
ETA( )
{
1. Color the courses and from the group;

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 19
2. Find the power set of the groups according to the
provided exam slots per day;
3. Sort the power set in descending order according to
their total student to conflict ratio treated as “Gain”;
4. PLACE the highest “Gain”ed course on each day of
exam;
5. Find the most appropriate pair for the group;
6. Optimize the exam routine by imposing the MST
(Minimum Spanning Tree – by using Prims
algorithm; Here the graph of the groups contains
edges with conflicts) to spread away the concurrent
exams for the student.
}
Algorithm 5. Exam-Time Tabling Algorithm (ETA).
For better exam-time table ETA also calculates the
conflicts among the each day-placed group and considers
these groups as a vertex. Further optimization can be
done by using PRIMS algorithm. In this case PRIMS
algorithm will increase the Manhattan distance (MD) of
the closest conflicting groups where the non conflicting
groups also hold an edge with zero weight. This
optimization stage is used to maximize the day gap of
exams for the students. So by using PRIMS algorithm,
ETA founds the minimum spanning tree of group of
courses where the nearest group exams hold less common
students. Manhattan Distance (MD) represents the value
of total student to conflict ratio between the groups of the
days. The cumulative time complexity of ETA is � O(s Gn
+ Gn 2 ) depends upon the provided slots per day (s)
strictly.
A. Algorithm Analysis
VI. EXPERIMENTAL RESULTS
The Decision tree based Routine Generation algorithm
(DRG) is a set of 4 sequential feed-forwarded trees.
Important observation has been made by considering the
fitness value where it represents the best matching score
among these selected day-time pattern slots for a selected
teacher. An example is shown in Fig. 1, 2, 3, 4 and 5(a)
where the first decision tree (PDRG) selects t2 because of
high priority and t2 (11) holds the courses {c2}. The daytime
pattern shows {{2,12,22},{29}} = {2,12}, {12,22}
or {2,22} as the best match for the course c2 from t2
favorite time slots. Here after imposing the day-time
pattern with respect to number of class per week for the
course, the calculated ordered sets of fitness value are
{2,12}, {12,22}, {2,22} and {29} where the first three
sets hold highest and equal fitness value corresponding
with class per week. As no courses have been placed yet
vector V is empty i.e.; V({ci}2) = { } & V({ci}12) = { } &
V({ci}22) = { } & within room limit hence, the fitness
value({2,12}) = fitness value({12,22}) = fitness
value({2,22}), where the fitness value is an integer
number representing the maximum possibility to take the
classes on the provided time slots. So, PDRG places c2 in
slot {2,12}, i.e. V({c2}2) and V({c2}12) where {2,12} is a
random selection.
© 2010 ACADEMY PUBLISHER
Again PDRG selects t4 with the priority 10 as its next
candidate which holds the courses {c5}. Here the daytime
pattern is, {{2,12,22},{5,15,25}..} =
{2,12,22},{5,15,25} extracted from the t4’s favorite slots.
As V({ci}2) = {c2} & V({ci}12) = {c2} V({ci}22) = { } &
V({ci}5) = { } & V({ci}15) = { } V({ci}25) = { } and the
fitness value({2,12,22}) < fitness value({5,15,25}).
Therefore c5 is placed in slot {5,15,25}, i.e. V({c5}5) and
V({c5}15) and V({c5}25) where V({ci}2) and V({ci}12) is
not empty. In similar approach the for the teacher t3 the
course c3 is placed V({c3}10) and V({c3}20) where
V({c3}5), V({c3}15) and V({c3}25) is ignored due to
different course color. And for t1 the courses {c1, c4} hold
12 and 8 as total conflicts respectively. Selected course c1
can be placed {7,17}, {8,18}, {9,19}, {13,23}, {22},
{24} accordingly to the t1 favorite slots. Course c1 is
placed in V({c1}7) and V({c1}17). And from the
remaining time slots none the generated pattern provide
sufficient classes for the course c4 where the required
number of classes is 3. So the course c4 is placed on
V({c4}8) and V({c4}18) and c4 is tagged as partially placed
course needed to be explored more. Random selection
among courses for exploration is acceptable if more than
one course holds same conflict score.
This class based reasoning left 1 partially placed
course, need to be walked around more. Next decision
tree CDRG is now in operation with fewer courses as
compared with the beginning and slot 24 will be allocated
for the course c4. The important issue is, although the
classes are placed in zigzag fashion but all the selected
slots for the course c4 are from the favorite slots of the
faculty respectively. So, the denoted term teacher
satisfaction is 100% for the case and overlapping classes
for the student is zero (i.e., student satisfaction). The
results of above example are shown in Fig. 11. In practice
the situation may be more complex with many courses.
By using the same data filtering technique for Exam-
Time tabling the proposed algorithm ETA generates the
power set of courses according to their course color
extracted from the cross_table(Cr) where the teacher
redundancy is ignored Fig. 5(b). So, the resultant set is,
{c1}, {c2}, {c3}, {c4}, {c5}, {c1, c4} and {c2, c3}. If the
provided exam days are 3 and exam slots per day are 2
the for 100% conflict free exam per slots are day 1 : {c1,
c4}, day 2 : {c2, c3} and day 3 : {c5}.
PDRG CDRG TDRG NTDRG
Final
Finding
Unclassified 0 0 - - 0
Partially
classified
Day-time
1 0 - - 0
patterned
classified
4 1 - - 5
Student
conflict
0 0 - - 0
Unsatisfied
time
0 1 - - 1
Figure 11. Simulation result of DRG
Further more if the provided days for exam are 2 and
slots are 2 then the exam-time tabling is like, day 1 : slot
1 {c1, c4}, slot 2 {c5} and day 2 : slot 1 {c2}, slot 2 {c3},

20 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
where day 1 consists zero conflict of exam on each slots
but 1 consecutive exam of an student. Fig. 12 shows the
overall outcome of the ETA algorithm for this special
scenario.
B. Experimental Results
Test results for DRG algorithm are carried out on a PC
with Pentium IV/1.6 GHz processor and 256 MB of
memory. Table I. shows the computational results for
semester 1 and semester 2 with 66 and 61 courses
correspondingly. Where for semester 1, around 24
teachers, with minimum 10 classes per week and at least
3 courses and 120 students with 430 combinations of
choices of courses are considered as input. For semester
2, around 23 teachers, with minimum 10 classes per week
and at least 3 courses and 106 students with 414
combinations of choices of courses are considered as
input.
Day
Slot 1
Cr.
(total
std)
Slot 2
Cr.
(total
std)
Slot 3
Cr.
(total
std)
Concurrent
Exam
Conflict
Overall
Gain
Day = 3
Slot = 2
1
2
3
c1 (40)
c2 (50)
c5 (45)
c4 (40)
c3 (35)
-
-
-
-
0
0
0
80
85
45
Day = 2
Slot = 2
1
2
c2 (50)
c3 (35)
c1 (40)
c5 (45)
c4 (40)
-
-
1
0
130
80
Day = 1
Slot = 2
1 - - - -
Not
Possible
Day = 1
Slot = 3
1
c1 (40)
c4 (40)
c5 (45)
c2 (50)
c3 (35)
21 10
Figure 12. Simulation result of ETA
In Table I. and Fig. 11 unclassified courses refers to the
number of courses that were not classified by any
decision trees, partially classified courses gives the
number of courses partially classified (the number classes
already placed into the routine is less then the required
classes). Day-time pattern shows the numbers of courses
that followed the day-time pattern. student conflict
estimates the percentage of unsatisfied requirements for
courses by students. Unsatisfied time refers the number of
time slots automatically generated beyond teachers’
favorite choice. Final Finding refers the final output for
every subsection after using the all cascade trees. The
main objective of this classification is to achieve the state
where the value the unclassified course is equal to zero.
Final value of unsatisfied time and student conflict shows
the teacher and student satisfaction respectively.
Randomization or prediction on classification is not used
in DRG. So the output of this � O(A.mn - B.m - C.n)
based deterministic algorithm strictly depends upon the
input.
Test results for ETA are generated depending upon 61
courses with average 25 students per course within 9
exam days and 2 slots per day for semester 2. Table II. as
well as Fig. 12 describes the outcome of the Exam-time
Tabling Algorithm. Here the “slots” represents the total
numbers of consecutive exam on a single day. Time
© 2010 ACADEMY PUBLISHER
duration is 3 hours for exam-time tabling whereas same
slot holds 1 hour as class duration in DRG. Total student
to conflict ratio on a particular exam day is referred as
overall gain. Around 3.8% of the total students hold
concurrent exam schedule whereas scheduled courses on
each slot are 100% conflict free. The overall gain also
confirms the profit for taking the course groups together
as a candidate on a single day. High satisfaction of the
students attests a high-quality exam-time tabling.
VII. CONCLUSION
Timetabling problem usually varies significantly from
institution to institution in terms of specific requirements
and constraints [22]. Many current successful university
timetabling systems are often applied only in the
institutions where they were designed. The metaheuristic,
heuristic and hybrid methods are used to
solving timetabling problems so as case base reasoning.
The main idea is to try and design an algorithm that will
choose the right decision tree to carry out a certain task in
a certain situation.
This paper outlines the algorithm Decision Tree based
Routine Generation (DRG) using OLAP representation,
to construct a university class routine and conflict free
Exam-Time Tabling algorithm (ETA) to produce conflict
free exam schedule with a fixed interval of days. It
should be noted that the DRG algorithm brings the
complexity to a considerable level and this solution
classifies 96% - 97% of the courses as well 93% - 95%
satisfaction for teacher. For this data set, students’
satisfaction is 100% but in general 90% - 93%
satisfaction may be achievable by using DRG.
Preferential requirements (teacher satisfaction) on time
variables are met around 93%. Again the ETA algorithm
provides satisfactory exam-time table with 100%
satisfaction. The results also illustrate that the proposed
algorithms achieve significant performance gains over
different data set.
The proposed algorithms are designed in such a manner
so that they are easy to code and imply significant
importance to construct generalized automated timetabling
software. Author(s) of the paper realize the
difference in constraints level in different institutes;
however in future generalized automated time-tabling
software will be examined. This paper does not consider
any classical benchmark problems. It is important to
analyze the performance with other established
algorithms. Incorporating those heterogeneous constraints
with the proposed data structure will also be examined in
future.
ACKNOWLEDGEMENT
The proposed algorithm is employed to construct a
class and exam routine for a reputed university in
Bangladesh. Author(s) of the paper express gratitude to
the university authority for providing indispensable
information.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 21
Semester-1
Semester-2
Semester-2
REFERENCES
[1] Christelle Guért. Narendra Jussien, Patrice
Boizumault and Christain Prins, “Building university
timetable using constraint logic programming,”
Springer-Verlag LNCS 1153, pp. 130-145, 1996.
[2] Hans-Joachim Goltz, Georg Küchler and Dirk
Matzkae, “Constraint-based timetabling for
universities,” INAP’98, pp. 75-80, 1998.
[3] Slim Abdennadher and Michael Marte, “University
course timetabling using constraint handling rules,”
Journal of Applied Artificial Intelligence, vol. 14, no.
4, pp. 311-326, 2000.
[4] Thom Früwirth, “Constraints handling rules,”
Constraint Programming: Basic and Trandes, LNCS
910, Springer, 1995.
[5] E.K. Burke, B.L. MacCarthy, S. Petrovic and R. Qu,
“Knowledge Discovery in a Hyper-Heuristic for
Course Timetabling Using Case-Based Reasoning,”
PATAT 2002, 4 th International Conference, pp. 90-
103, August 2002.
[6] D. Abramson, “Constructing school timetables using
simulated annealing: Sequential and parallel
algorithms,” Management Science, vol. 37, pp. 98-
113, 1991.
[7] A. Schaert, “Tabu search techniques for large highschool
timetabling problems,” 13 th National
Conference on Artificial Intelligence AAAI’96, pp.
363-368, 1996.
© 2010 ACADEMY PUBLISHER
TABLE I.
COMPUTATIONAL RESULTS OF DRG
PDRG CDRG TDRG NTDRG Final Finding
Unclassified courses 5 3 1 0 0
Partially classified courses 45 32 2 2 2
Day-time patterned classified 16 31 63 64 64
Student conflict 0 0 0 0 0
Unsatisfied time 0 0 3 7 9
Unclassified courses 2 2 1 0 0
Partially classified courses 41 26 2 2 2
Day-time patterned classified 18 33 58 59 59
Student conflict 0 0 0 0 0
Unsatisfied time 0 0 5 8 11
TABLE II.
COMPUTATIONAL RESULTS OF ETA
# of
Courses on
Slot 1
# of
Courses on
Slot 2
Concurrent
Exam Conflict
Total
Student on
Slot 1
Total
Student on
Slot 2
Overall Gain
Day 1 3 4 3 72 63 45
Day 2 3 3 2 77 63 70
Day 3 2 2 5 32 53 17
Day 4 4 4 5 76 87 32.6
Day 5 5 5 11 91 64 14
Day 6 3 2 3 46 18 21.3
Day 7 4 4 7 42 74 16.5
Day 8 2 5 2 28 71 49.5
Day 9 2 3 2 16 68 42
[8] Luca Di Gaspero and Andra Schaerf, “Multi-
Neighbour Local Search for Course Timetabling,”
PATAT 2002, 4 th International Conference, pp. 128-
132, August 2002.
[9] P. Boizumault, Y. Delon, and L. Peridy, “Constraint
logic programming for examination timetabling,”
Journal of Logic programming, vol. 26, pp. 217-233,
1996.
[10] K. Dowslan, “Using simulated annealing for efficient
allocation of students to practical classes,” Applied
Simulated Annealing – Lecture Notes in Economics
and Mathematical System, Springer-Verlag, vol. 396,
pp. 125-150, 1993.
[11] K. Dowslan, “Off-the peg or made-to-measure?
Timebaling and scheduling with sa and ts,”
PATAT’97, Springer – Verlag, pp. 37-52, 1998.
[12] G.M. White and J. Zhang, “Generating complete
university timetables by combining tabu search with
constraint logic,” PATAT’97, Springer – Verlag, pp.
187-198, 1998.
[13] E.K. Burke, J. Newall, and R.F. Weare,
“Initialization strategies and diversity in evolutionary
timetabling,” Evolutionary Computation Journal, vol.
6.1, pp. 81-103, 1998.
[14] E.K. Burke, J. Newall, and R.F. Weare, “A simple
heuristically guided search for the timetable
problem,” Proceeding of the International ICSC

22 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Symposium on Engineering of Intelligent System,
pp. 574-579, 1998.
[15] E.K. Burke, J. Newall, and R.F. Weare, “A memetic
algorithm for university exam timetabling,” Lecture
Notes in Computer Science, Springer-Verlag, vol.
1153, pp. 241-250, 1996.
[16] Tan, Steinbach and Kumar, Introduction to Data
Mining, pp. 130-139, 2004.
[17] Liam T.G. Merlot, Natashia Boland, Barry D.
Hughes, and Peter J. Stuckey, “A Hybrid Algorithm
for the Examination Timetabling Problem”,
Proceedings of the 4 th International Conference on
the Practice and Theory of Automated Timetabling -
PATAT’2002, Springer – Verlag, pp. 348-371.
[18] M. Caramia, P. Dell'Olmo, and G.F. Italiano, “New
algorithms for examination timetabling”,
Proceedings: Algorithm Engineering 4 th International
Workshop, WAE 2000, Germany, September 2000.
“Lecture Notes in Computer Science 1982”,
Springer-Verlag, Berlin Heidelberg New York, pp.
230-241, 2001.
[19] M. Carter, G. Laporte, and J. Chinneck, “A general
examination scheduling system”, Interfaces, pp.
109-120, 1994.
[20] M. Carter, G. Laporte, and S.T. Lee, “Examination
timetabling: algorithmic strategies & applications”,
Journal of the Operational Research Society, pp. 373
- 383, 1996.
[21] Bart Selman, Henry A. Kautz, and Bram Cohen,
“Noise strategies for improving Local search”, 12 th
National Conference on Artificial Intelligence
(AAAI-94), pp. 337– 343, 1994.
[22] T. B. Cooper, J. H. Kingston, “The Complexity of
Timetable Construction Problems”, Practice and
Theory of Automated Timetabling, Springer Verlag,
pp. 283-295, 1996.
Ashiqur Md. Rahman received his
B.Sc. Degree in Computer Science and
Engineering from American International
University Bangladesh, Dhaka in
January, 2004. He is currently perusing
his M.Sc. degree from North South
University, Dhaka since January 2006.
He has authored in 4 national and international journal and
conference papers in the area of Data Mining, VHDL,
Cryptography and PVc module design. His current research
interest is in Grid Computing especially in large Grid
Environment.
© 2010 ACADEMY PUBLISHER
Shafaat S. Giasuddin received his B.Sc.
Degree in Computer Science from
Ahsanullah University of Science and
technology, Dhaka in November, 2005.
He is currently perusing his M.Sc. degree
from North South University, Dhaka
since January 2006. He has authored in 3
national and international journal and conference papers in the
area of Data Mining, VHDL and Cryptography. His current
research interest is in Data Mining especially in corporate sector
like banking and telecommunication.
Rashedur M. Rahman received his
Ph.D. Degree in Computer Science from
University of Calgary, Canada in
November, 2007. He has received his
M.Sc. degree from University of
Manitoba, Canada in 2002 and Bachelor
degree from Bangladesh University of
Engineering and Technology (BUET) in 2000 respectively. He
is currently working as an Assistant Professor in North South
University, Dhaka, Bangladesh. He has authored more than 25
international journal and conference papers in the area of
parallel, distributed, grid computing and knowledge and data
engineering. His current research interest is in data mining
especially on financial, educational and medical surveillance
data, data replication on Grid, and application of fuzzy logic for
grid resource and replica selection.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 23
Anomaly Network Intrusion Detection Based on
Improved Self Adaptive Bayesian Algorithm
Dewan Md. Farid
Dept. of CSE, Jahangirnagar University, Dhaka-1342, Bangladesh
Email: dmfarid@uiu.ac.bd
Mohammad Zahidur Rahman
Dept. of CSE, Jahangirnagar University, Dhaka-1342, Bangladesh
Email: rmzahid@juniv.edu
Abstract—Recently, research on intrusion detection in
computer systems has received much attention to the
computational intelligence society. Many intelligence
learning algorithms applied to the huge volume of complex
and dynamic dataset for the construction of efficient
intrusion detection systems (IDSs). Despite of many
advances that have been achieved in existing IDSs, there are
still some difficulties, such as correct classification of large
intrusion detection dataset, unbalanced detection accuracy
in the high speed network traffic, and reduce false positives.
This paper presents a new approach to the alert
classification to reduce false positives in intrusion detection
using improved self adaptive Bayesian algorithm (ISABA).
The proposed approach applied to the security domain of
anomaly based network intrusion detection, which correctly
classifies different types of attacks of KDD99 benchmark
dataset with high classification rates in short response time
and reduce false positives using limited computational
resources.
Index Terms—anomaly detection, network intrusion
detection, alert classification, Bayesian algorithm, detection
rate, false positives
I. INTRODUCTION
Network intrusion detection is the problem of detecting
unauthorized use of computer systems over a network,
such as the Internet. IDSs were introduced by James P.
Anderson in 1980, an example of an audit trails would be
a log of user access [1]. IDSs have become an integral
part of today’s information security infrastructures. In
order to detect intrusion activities, many machine
learning (ML) algorithms, such as Neural Network [2],
Support Vector Machine [3], Genetic Algorithm [4],
Fuzzy Logic [5], and Data Mining [6], etc have been
widely used to the huge volume of complex and dynamic
dataset to detect known and unknown intrusions. It is
very important for IDSs to generate rules to distinguish
normal behaviors from abnormal behavior by observing
dataset, which is the record of activities generated by the
operating system that are logged to a file in
chronologically sorted order. IDSs using ML algorithms
aim to solve the problems of analyzing huge volumes of
dataset and realizing performance optimization of
detection rules. IDSs detect attacks on computer systems
and signal an alert to the Computer Emergency Response
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.23-31
Team (CERT). The network intrusion detection area is
the arrival of new, previously unseen attacks, because
hackers are very inventive and they use newer ways to
disrupt the normal operation of servers and users.
Anomaly detection detects new attacks which has not
presented in the dataset by observing system activities
and classifying it as either normal or anomalous. An
important research challenge today is to develop adaptive
IDSs to improve classification rates, and reduce false
positives.
The Bayesian algorithm (BA) provides a probabilistic
approach for classification [7], [8], which provides an
optimal way to predict the class of an unknown example.
It is widely used in many fields of data mining, image
processing, bio-informatics, and information retrieval etc.
It calculates conditional probabilities from a given dataset
and used these conditional probabilities to find out the
probabilities of belongingness to different classes. The
unseen example is then categorized to that class, which
assumes the maximum value. In this paper, based on a
comprehensive analysis for the current research
challenges we propose a new algorithm to address the
problem of classification rates and false positives using
improved self adaptive Bayesian Algorithm. This
algorithm correctly classifies different types of attacks of
KDD99 dataset with high detection accuracy in short
response time, and also maximizes the detection rate
(DR) and minimizes the false positives (FP). In our
experiments proposed algorithm reduced the number of
false positives by up to 90% with acceptable
misclassification rates.
The remainder of this paper is organized as follows.
Section II provides a review of IDSs and section III
provides IDSs using machine learning algorithms.
Section IV presents our proposed new algorithm.
Experimental results are presented in section V. Finally,
section VI makes some concluding remarks along with
suggestions for further improvement.
II. REVIEW OF INTRUSION DETECTION SYSTEMS
A. History of IDSs
In 1980, the concept of IDSs began with Anderson’s
seminal paper, introduced the notion that audit trails

24 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
contained vital information that could be valuable in
tracking misuse and understanding user behavior. His
work was the start of host-based IDSs (HBIDSs). In
1986, Dr. Dorothy Denning published a model which
revealed the necessary information for commercial IDSs
development [9]. In 1988, multics intrusion detection and
alerting system (MIDAS), an expert system using P-
BEST and LISP was developed [10]. Haystack was also
developed in this year using statistics to reduce audit
trails [11]. In 1989, wisdom & sense (W&S) was a
statistics-based anomaly detector developed, that created
rules based on statistical analysis, and then used those
rules for anomaly detection [12]. In 1990, Heberlein first
introduced the idea of network IDSs, development of
Network Security Monitor (NSM), and hybrid IDSs [13]
and Lunt proposed SRI named intrusion detection expert
system (IDES), a dual approach of a rule-based expert
system and a statistical anomaly detection, which ran on
Sun Workstations and could consider both user and
network level data [14]. Also in the early 1990s the
commercial development of IDSs were started and the
Time-based inductive machine (TIM) did anomaly
detection using inductive learning of sequential user
patterns in Common LISP on a VAX 3500 computer
[15]. In 1991, distributed IDSs (DIDS), an expert system
created by the researchers of University of California
[16], and the network anomaly detection and intrusion
reporter (NADIR) a statistic based anomaly detector and
also an expert system developed by Los Alamos National
Laboratory’s Integrated Computing Network (ICN) [17].
In 1993, Lunt proposed the Next-generation Intrusion
Detection Expert System by developing SRI followed
IDES using artificial neural network [18]. The Lawrence
Berkeley National Laboratory introduced rule language
called Bro for packet analysis from libpcap dataset in
1998 [19]. The audit data analysis and mining IDSs used
tcpdump to build profiles of rules for classifications in
2001 [20].
B. Types of IDSs
Intrusion is a set of actions that attempt to compromise
the confidentiality, integrity or availability of computer
resources. IDSs collect information from a variety of
systems and analyze the information for signs of
intrusions. In general, IDS categorize into five types,
which are described below:
1) Network IDSs (NIDSs) responsible for detecting
attacks related to the network [21], [22]. NIDSs
investigate incoming and outgoing network traffic by
connecting with network devices to find suspicious
patterns. If a NIDS has no additional information about
the protected host, the malicious attacker can easily avoid
detection by taking advantage of different handling by
overlapping IP/TCP fragments by IDS and a target host
[23].
2) Host-based IDSs (HBIDSs) usually are located
in servers to examine the internal interfaces [24]. HBIDSs
can either use standard auditing tools [25], or specially
instrumented operating system [26], or application
platforms [27]. It detects intrusions by analyzing system
© 2010 ACADEMY PUBLISHER
calls, application logs, file-system modifications, and
other host activities related to the machine.
3) Protocol-based IDSs (PIDSs) monitor the
dynamic behavior and state of the protocol used the web
server. PIDSs sit at the front end of a web server,
monitoring and analyzing the HTTP protocol stream. It
understands the HTTP protocol to protect web server by
filtering IP address or port number.
4) Application protocol-based IDSs (APIDSs)
monitor and analysis on a specific application protocol or
protocols between a process, and group of servers that is
used by the computer system. APIDSs can be sitting
between a web server and the database management
system that monitoring the SQL protocol specific to the
business logic. Generally, APIDSs look for the correct
use of the protocol.
5) Hybrid IDSs (HIDSs) combines two or more
intrusion detection approaches. HIDS provide alert
notification from both network and host-based intrusion
detection devices.
C. Detection Models of IDSs
Detection rate (DR) is defined as the number of
intrusion instances detected by the system divided by the
total number of intrusion instances present in the dataset,
and false positives (FP) is an alarm, which rose for
something that is not really an attack. There are two types
of detection models for IDSs, which are described below:
1) Misuse or signature-based IDSs are also known
as pattern-based IDSs. It performs simple pattern
matching to match a pattern corresponding to a known
attack type in the database. DR of these IDSs is relatively
low, because attacker will try to modify the basic attack
signature in such a way that will not match the known
signatures of that attack and it cannot detect a new attack
for which a signature is not yet installed in the database.
2) Anomaly-based IDSs tries to identify new
attacks by analyzing strange behavior from normal
behaviors. It has a relatively high detection rate for new
types of intrusion. The disadvantage is that in many cases
there is no signal “normal profile” and anomaly-based
systems tend to produce many false positives.
D. Functions of IDSs
IDSs are automated systems detecting and alarming of
any situation where an intrusion has taken or is about to
take place. According to the Common Intrusion Detection
Framework (CIDF), generally IDSs consists of four
components, like: sensors, analyzers, database, and
response units. Most modern IDSs use multiple intrusion
sensors, which obtain alerts from the large computational
environment to maximize their trustworthiness.
Analyzers use the input of the sensors, analyze the
information gathered by these sensors, and return a
synthesis or summary of the input. Today, machine
learning algorithms have become an indispensable tool in
the analyzers of IDSs. Database stores all alerts and
support the analysis process. The response units carry out
prescriptions controlled by the analyzers. The functions
of IDSs are follows as [28]:
� Monitoring user’s activity.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 25
� Monitoring systems activity.
� Auditing system configuration.
� Assessing the data files.
� Recognizing known attack.
� Identifying abnormal activity.
� Managing audit data.
� Highlighting normal activity.
� Correcting system configuration errors.
� Stores information about intruders.
III. IDS USING MACHINE LEARNING
A. Bayes Rule
The Bayes rule provides a way to calculate the
probability of a hypothesis based on its prior probability
[7],[8]. The best hypothesis is the most probable
hypothesis, given the observed data D plus any initial
knowledge about the prior probabilities of the various
hypotheses h (h is a hypothesis space containing possible
target function). Bayes rule is defined in equation “(1)”.
Bayes Rule:
P�D|
h�
P h|
D)
P�D�
P(
h)
( � (1)
Here P(h|D) is called the posterior probability, while
P(h) is the prior probability associated with hypothesis h.
P(D) is the probability of the occurrence of data D and
P(D|h) is the conditional probability. In many learning
scenarios, the learner considers some set of candidate
hypothesis, H and is interested in finding the most
probable hypothesis h € H given the data D. Any such
maximally probable hypothesis is called maximum
posterior (MAP) hypothesis. The MAP hypothesis use
Bayes rule to calculate the posterior probability of each
candidate hypothesis. More exactly, hMAP is a MAP
hypothesis provided:
hMAP � argmaxh �H
P�h|
D�
=
P�h|
D�P(
h)
argmaxh �H
P�D�
= argmax P�D|
h�
P(
h)
(2)
h� H
Finally, dropped the term P(D) because it is a constant
dependent of h. P(D|h) is also called likelihood of the
data D given h any hypothesis that maximizes P(D|h) is
called a Maximum Likelihood (ML) hypothesis, hML:
hML � maxh
H P�D|
h�
arg � (3)
B. Naïve Bayesian Classifier
Naïve Bayesian (NB) classifier is a simple
probabilistic classifier based on probability models that
incorporate strong independence assumptions which often
have no bearing in reality. The probability model can be
derived using Bayes rule. Depending on the precise
nature of the probability model, NB classifier can be
trained very efficiently in a supervised learning. In many
practical applications, parameter estimation for naïve
Bayesian models uses the method of maximum
© 2010 ACADEMY PUBLISHER
likelihood. In spite of naïve design and apparently oversimplified
assumptions, naïve Bayesian classifiers often
work much better in many complex real-world situations.
The NB classifier is given as input a set of training
examples each of which is described by attributes A1
through Ak and an associated class, C. The objective is to
classify an unseen example whose class value is unknown
but values for attributes A1 through Ak are known and
they are a1, a2,.…, ak respectively. The optimal prediction
of the unseen example is the class value c such that
P(C=ci|A1=a1,…Ak=ak) is maximum. By Bayes rule this
probability equals to
argmax
ci
�C
P
�A�a1,... Ak
�ak
| C�c
i�
P�A�a,....
A �a
�
�C�c� 1
P i
(4)
1 1 k k
Where P(C=ci) is the prior probability of class ci,
P(A1=a1,…Ak=ak) is the probability of occurrence of the
description of a particular example, and
P(A1=a1,…Ak=ak|C=ci) is the class conditional
probability of the description of a particular example ci of
class C. The prior probability of a class can be estimated
from training data. The probability of occurrence of the
description of particular examples is irrelevant for
decision making since it is the same for each class value
c. Learning is therefore reduced to the problem of
estimating the class conditional probability of all possible
description of examples from training data. The class
conditional probability can be written in expanded from
as follows:
P(A1=a1,…Ak=ak|C=ci)
= P(A1=a1| A2=a2 ^…Ak=ak ^ C=ci)
* P(A2=a2| A3=a3 ^…Ak=ak ^ C=ci)
* P(A3=a3| A4=a4 ^…Ak=ak ^ C=ci)
* P(A4=a4 ^…Ak=ak ^ C=ci) (5)
In NB, it is assumed that outcome of attribute Ai is
independent of the outcome of all other attributes Aj,
given c. Thus class conditional probabilities become:
P(A1=a1,…Ak=ak|C=ci) =� P(
Ai
� ai
| C � ci
) If the
i � 1
above value is inserted in equation “(4)” it becomes:
� arg maxci
� C P(C=c)� P(
Ai
� ai
| C � ci
) (6)
i � 1
In Naïve Bayesian learning, the probability values of
equation “(6)” are estimated from the given training data.
These estimated values are then used to classify unknown
examples.
C. Decision Tree Algorithms
The ID3 technique builds decision tree using
information theory [29]. The basic strategy used by ID3
is to choose splitting attributes from a data set with the
highest information gain. The amount of information
associated with an attribute value is related to the
probability of occurrence. The concept used to quantify
k
k

26 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
information is called entropy, which is used to measure
the amount of randomness from a data set. When all data
in a set belong to a single class, there is no uncertainty
then the entropy is zero. The objective of decision tree
classification is to iteratively partition the given data set
into subsets where all elements in each final subset
belong to the same class. The entropy calculation is
shown in equation “(7)”. The value for entropy is
between 0 and 1 and reaches a maximum when the
probabilities are all the same. Given probabilities p1,
p2,..,ps where �i=1 pi=1,
Entropy: H(p1,p2,…ps) = ��
s
i 1
(pi log(1/pi)) (7)
Given a data set, D, H(D) finds the amount of subset of
data set. When that subset is split into s new subsets S =
{D1, D2,…,Ds}, we can again look at the entropy of those
subsets, A subset of data set is completely ordered if all
examples in it are the same class. ID3 chooses the
splitting attribute with the highest gain. The ID3
algorithm calculates the gain by the equation “(8)”.
Gain (D,S) = H(D)- ��
s
i 1
p(Di)H(Di) (8)
The C4.5 algorithm improves ID3 through GainRatio
[30]. For splitting purpose, C4.5 uses the largest
GainRatio that ensures a larger than average information
gain.
GainRatio(D,S) =
Gain.(
D,
S)
�|
D1
| | Ds
| �
H��
,..., ��
� | D|
| D|
�
The C5.0 algorithm improves the performance of
building trees using boosting, which is an approach to
combining different classifiers. But boosting does not
always help when the training data contains a lot of noise.
When C5.0 performs a classification, each classifier is
assigned a vote, voting is performed, and the example of
data set is assigned to the class with the most number of
votes. CART (classification and regression trees) is a
process of generating a binary tree for decision making
[31]. CART handles missing data and contains a pruning
strategy. The SPRINT (Scalable Parallelizable Induction
of Decision Trees) algorithm uses an impurity function
called gini index to find the best split [32]. “Equation
(10), defines the gini for a data set, D”.
(9)
gini (D) = 1-� pj 2 (10)
Where, pj is the frequency of class Cj in D. The
goodness of a split of D into subsets D1 and D2 is defined
by
ginisplit(D) = n1/n(gini(D1))+ n2/n(gini(D2)) (11)
The split with the best gini value is chosen. A number
of research projects for optimal feature selection and
classification have been done, which adopt hybrid
strategy involving evolutionary algorithm and inductive
decision tree learning [33], [34], [35], [36].
© 2010 ACADEMY PUBLISHER
D. K-Nearest Neighbors
The K nearest neighbors (KNN) is a classification
algorithm based on the use of distance measures [37]. It
finds k examples in training data that are closest to the
test example and assigns the most frequent label among
these examples to the new example. When a classification
is to be made for a new example, its distance to each
attribute in the training data must be determined. Only the
K closest examples in the training data are considered
further. The new example is then placed in the class that
contains the most examples from this set of K closest
examples. KNN can be considered decision making
technique as equivalent to Bayesian classifier in which
the number of neighbors of each example is used as an
estimate of the relative posterior probabilities of class
membership in the neighborhood of a sample to be
classified.
IV. IMPROVED SELF ADAPTIVE BAYESIAN ALGORITHM
A. Adaptive Bayesian Algorithm
Adaptive Bayesian algorithm creates a function from
KDD99 benchmark intrusion detection training data [38],
which first estimate the class conditional probabilities for
each attribute value based on their frequencies over the
weights with match of same class in the training data. In a
given training data, D = {A1, A2,…,An} of attributes,
where each attribute Ai = {Ai1, Ai2,…,Aik} contains
attribute values and a set of classes C = {C1, C2,…,Cn},
where each class Cj = {Cj1, Cj2,…,Cjk} has some values.
Each example in the training data contains weight, W =
{W1, W2…, Wn}. Initially, all the weights for each
example of training data have equal unit value that set to
Wi = 1.0. Then calculate the sum of weights for each
class from the training data by counting how often each
class occurs in the training data and the sum of weights
for each attribute value with respect to same class in the
training data. Next calculate the class conditional
probabilities for each attribute value using equation
“(12)” from the training data.
P(Aij|Cji) =
�
�
W Aij
(12)
Here P(Aij|Cji) is the class conditional probability, WA
is the sum of weights for each attribute value, and WC is
the sum of weights for each class. After calculating the
class conditional probabilities for each attribute value
from the training data, the algorithm classify the each test
example using equation “(13)”.
W
Cji
TargetClass = � Ci P(Aij|Cji) (13)
If any test example is misclassified, the algorithm
updates the weights of training data. The algorithm
compare each of test examples with every training
example and compute the similarity between them, and
then weights of training data are increased by a fixed

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 27
small value multiplied by the corresponding similarity
measure.
Wi = Wi + (S*0.01); (14)
Here S is the similarity between test and training
examples. If the test example is correctly classified, then
the weights of training data will remain unchanged. After
weights adjustment, the class conditional probabilities for
attribute values are recalculated from the modified
weights in training data. If the new set of probabilities
correctly classifies all the test examples, the algorithm
terminates. Otherwise, the iteration continues until all the
test examples are correctly classified or the target
accuracy is achieved. At this stage the algorithm
terminates, the class conditional probabilities are
preserved for future classification of seen or unseen
examples.
B. Improved Self Adaptive Bayesian Algorithm
Improved self adaptive Bayesian algorithm (ISABA) is
the modification of adaptive Bayesian algorithm. Given a
training data ISABA initializes the weights of each
example Wi set to 1.0 and estimates the prior probability
P(Cj) for each class by summing the weights how often
each class occurs in the training data. For each attribute in
training data, Ai the number of occurrences of each
attribute value Aij can be counted by summing the
weights to determine the probability P(Aij). Similarly, the
probability P(Aij | Cj) can be estimate by summing the
weights how often each attribute value occurs in the class
in the training data. An example in the training data may
have many different attributes A = {A1, A2,…,An}, and
each attribute have many values Ai = {Ai1, Ai2,…,Aik}. The
conditional probability P(Aij | Cj) estimate for all values
of attributes. Then the algorithm uses these conditional
probabilities to classify all the training examples. When
classifying a training example, the conditional and prior
probabilities generated from the training data are used to
make the prediction. This is done by multiplying the
probabilities of the different attribute values from the
example. Suppose the training example ei has p
independent attribute values {Ai1, Ai2,…,Aip}, the
algorithm has P(Aik | Cj), for each class Cj and attribute
Aik, and then estimate probability P(ei | Cj) by using
equation “(15)”.
P(ei | Cj) = P(Cj) �k=1�p P(Aij | Cj) (15)
To calculate the probability P(ei), the algorithm
estimate the likelihood that ei is in each class. The
probability that ei is in a class is the product of the
conditional probabilities for each attribute values. The
posterior probability P(Cj | ei) is then found for each
class. The class with the highest probability is the one
chosen for the example in training data. Now the
algorithm updates the weights for each example in the
training data with the highest value of posterior
probability P(Cj | ei) for that example and also changes
the class value associate with highest posterior
probability. If any example in the training data is
© 2010 ACADEMY PUBLISHER
misclassified then the algorithm again calculates the prior
probability P(Cj) and conditional probability P(Aij | Cj)
from the training examples using the updated weights,
and again classify the training examples and updates the
weights of training examples. This iteration will continue
until all the training examples are correctly classified or
the target accuracy is achieved.
After classifying the training examples, the algorithm
classify the test examples using the conditional
probabilities P(Aij | Cj). If any test example is
misclassified, the algorithm again updates the weights of
training examples. The algorithm compares each of test
example with every training examples and compute the
similarity between them (out of four attributes two
attribute values are same then similarity becomes 0.5 or
one attribute value is same then similarity is 0.25 and so
on), and then weights of training examples are increased
by a fixed small value multiplied by the corresponding
similarity measure. If the test examples are correctly
classified, then the weights of training examples will
remain unchanged. After weights adjustment, the
conditional probabilities P(Aij | Cj) for attribute values are
recalculated from the modified weights of training
examples. If the new set of probabilities correctly
classifies all the test examples, the algorithm stores the
conditional probabilities and builds a decision tree by
information gain using last updated weights. Otherwise,
the iteration continues until all the test examples are
correctly classified or the target accuracy is achieved. At
this stage the algorithm correctly classifies all the test
examples, the conditional probabilities P(Aij | Cj) are
preserved for future classification of seen or unseen
intrusions, and builds a decision tree by information gain
using the last updated weights of training examples,
which is also used for classification of seen or unseen
intrusions. The main procedure of ISABA algorithm is
described as follows.
Algorithm ISABA
Input: training data D, testing data
Output: intrusion detection model
Procedure:
1. Initialize all the weights in D, Wi=1.0.
2. Calculate the prior probabilities P(Cj) for each
class Cj from D.
�W
i
P(Cj) =
Ci
n
�
i�1
3. Calculate the probabilities P(Aij) for each
attribute value Aij from D.
W
P(Aij) =
�� Wi
Aij
Cj
4. Calculate the conditional probabilities P(Aij | Cj)
for each attribute values from D.
i

28 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
P(Aij | Cj) = P ( Aij)
W
5. Classify each example in D.
�
Ci
P(ei | Cj) = P(Cj) � P(Aij | Cj)
6. Initialize all the weights in D with Maximum
Likelihood (ML) of posterior probability P(Cj|ei)
and change the class value associated with
highest posterior probability.
Wi= PML(Cj|ei)
Cj = Ci� PML(Cj|ei)
7. If any example in D is misclassified, then go to
step 2, else go to step 8.
8. Classify all test examples with the conditional
probabilities P(Aij | Cj).
9. If any test example is misclassified then update
the weights of D using similarity S.
Wi = Wi + (S+0.01)
10. If weight update, then recalculate the
probabilities P(Aij) and P(Aij | Cj) using updated
weights of D and go to step 8.
11. If all test examples are correctly classified then
the algorithm stores conditional probabilities for
future classification of seen or unseen intrusions.
12. The algorithm builds decision tree by
information gain using final updated weights of
training examples.
V. EXPERIMENTAL RESULTS
A. Experimental Data
The KDD99 dataset is a common benchmark for
evaluation of intrusion detection techniques [39]. In the
1998 DARPA intrusion detection evaluation program, a
simulated environment was set up to acquire raw TCP/IP
dump data for a local-area network (LAN) by the MIT
Lincoln Lab to compare the performance of various
intrusion detection methods. The DARPA98 was
operated like a real environment, but being blasted with
multiple intrusion attacks and received much attention in
the research community of adaptive intrusion detection.
In 1999, based on the DARPA98 data the Third
International Knowledge Discovery and Data Mining
Tools Competition established the KDD99 benchmark
dataset for intrusion detection based on data mining. In
the KDD99 data set, each data example represents
attribute values of a class in the network data flow, and
each class is labeled either as normal or as an attack with
exactly one specific attack type. The KDD99 data records
are all labeled with one of the following five types:
1) Normal connections are generated by simulated
daily user behavior such as downloading files, visiting
web pages, etc.
2) Denial of Service (DoS) attack causes the
computing power or memory of a victim machine too
© 2010 ACADEMY PUBLISHER
i
busy or too full to handle legitimate requests. DoS attacks
are classified based on the services that an attacker
renders unavailable to legitimate users. Example of DoS
attacks are Apache2, Land, Mail bomb, Back, etc.
3) User to Root (U2R) is a class of attacks that an
intruder/hacker begins with the access of a normal user
account and then becomes a super-user by exploiting
various vulnerabilities of the system. Most common
exploits of U2R attacks are regular buffer overflows,
Loadmodule, Fdformat, and Ffbconfig.
4) Remote to User (R2L) is a class of attacks that a
remote user gains access of a local account by sending
packets to a machine over a network communication,
which include Sendmail, and Xlock.
5) Probing (Probe) is an attack scans a network to
gather information or find known vulnerabilities. An
intruder with a map of machines and services that are
available on a network can use the information to look for
exploits.
B. Eeperimental Analysis
Correct classification of known and unknown
intrusions is one of the central problems for networkbased
intrusion detection. Many supervised and
unsupervised learning algorithms already applied to
classifying intrusions but their performance were not very
satisfactory due to the challenging problem of detection
novel attacks with low false alarms. In order to evaluate
the performance of improved self adaptive Bayesian
algorithm (ISABA) for intrusion detection, we performed
5-class classification using KDD99 benchmark dataset.
The training and testing data are taken randomly from the
KDD99 dataset with different ratios of positive versus
negative instance. The training data are used to train the
algorithms, and the test data are used to evaluate the
performance of the algorithms. Table I shows the number
of examples in 10% training data and 10% testing data of
KDD99 dataset. There are some new attack examples in
testing data, which is no present in the training dataset.
TABLE I.
NUMBER OF EXAMPLES IN TRAINING AND TESTING DATA
Attack Types
Training
Examples
Testing
Examples
Normal 97277 60592
Probing 4107 4166
Denial of Service 391458 237594
User to Root 52 70
Remote to User 1126 8606
Total Examples 494020 311028
In the experiments, first we performed the
classification on KDD99 testing data using naïve
Bayesian classifier (NB), adaptive Bayesian algorithm
(ABA), and improved self adaptive Bayesian algorithm
(ISABA) and compare their results shown in Table II. It
is clear that the proposed new algorithm is approximately
2 times faster in training and testing than conventional
algorithms.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 29
TABLE II.
TRAINING AND TESTING TIME COMPARISON
Naïve Bayesian
Classifier
Adaptive Bayesian
Algorithm
Improved Self Adaptive
Bayesian Algorithm
Training Time (s) Testing Time (s)
42.9 18.4
24.6 9.6
28.4 7.5
The performance comparison between NB and ISABA
on KDD99 (using 41 attributes) testing data is listed in
Table III, which shows that the ISABA performs
balanced and high classification rates on 5 attack classes
of KDD99 data and minimize the false positives.
TABLE III.
PERFORMANCE COMPARISON BETWEEN NB AND ISABA
Normal Probe DoS U2R R2L
Naïve Bayesian
Classifier (DR %)
99.25 99.13 99.69 64 99.11
Naïve Bayesian
Classifier (FP %)
Improved Self Adaptive
0.08 0.45 0.04 0.14 8.02
Bayesian Algorithm
(DR %)
Improved Self Adaptive
99.62 99.22 99.49 99.17 99.15
Bayesian Algorithm
(FP %)
0.05 0.36 0.03 0.10 6.91
We tested our proposed algorithm using the reduced
dataset of 12 attributes and 17 attributes in KDD99
dataset, which increase the classification rate for intrusion
classes that are summarized in Table IV.
TABLE IV.
PERFORMANCE OF ISABA USING REDUCED DATASETS
12 Attributes 17 Attributes
Normal 99.97 99.96
Probe 99.91 99.95
DoS 99.99 99.98
U2R 99,36 99.46
R2L 99.53 99.69
We also compare the detection performance among
Support Vector Machines (SVM), Neural Network (NN),
Genetic Algorithm (GA), Naïve Bayesian Classifier
(NB), and improved self adaptive Bayesian algorithm
(ISABA) on KDD99 dataset [40],[41],[42]. In total, 40
attributes of KDD99 dataset have been used. Each
connection can be categorized into five main classes (one
normal class and four main intrusion classes: probe, DoS,
U2R, and R2L). The experimental setting uses 494020
data samples for training and 311028 data samples for
testing. The comparative results are summarized in Table
V.
TABLE V.
COMPARISON OF SEVERAL ALGORITHMS
SVM NN GA NB ISABA
Normal 99.4 99.6 99.3 99.55 99.82
Probe 89.2 92.7 98.46 99.43 99.72
DoS 94.7 97.5 99.57 99.69 99.49
U2R 71.4 48 99.22 64 99.47
R2L 87.2 98 98.54 99.11 99.35
© 2010 ACADEMY PUBLISHER
Detection Rate
1
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6
False Positives
Figure 1. ROC curves for alert classification systems.
The ROC (relative operating characteristic) curve
shows the relationship between detection rate and false
positives on 10% KDD99 testing data in figure 1.
VI. CONCLUSION AND FUTURE WORKS
The main advantage of this proposed algorithm is to
generate a minimal rule set for network intrusion
detection, which can detect network intrusions based on
previous activities. Proposed algorithm analyzes the
large volume of network data and considers the complex
properties of attack behaviors to improve the performance
of detection speed and detection accuracy. This paper
presents a new intrusion detection algorithm based on
intelligent machine learning algorithms. In this paper we
have concentrated on the development of the performance
of naïve Bayesian classifier, which adjusts the weights of
training examples until either all the test examples are
correctly classified or the target accuracy on test
examples is achieved. The experimental results marked
that this algorithm minimized false positives, as well as
maximize balance detection (classification rates) on the 5
classes of KDD99 data. The future research issues will be
applying domain knowledge of security to improve the
detection accuracy, and visualizing the procedure of
intrusion detection in real world problem domains.
ACKNOWLEDGMENT
Support for this research received from Ministry of
Science and Information & Communication Technology,
Government of Bangladesh. We would like to thank Prof.
Dr. Chowdhury Mofizur Rahman, United International
University, Bangladesh for fruitful discussions and
valuable help in the implementation of this algorithm.
REFERENCES
ISABA
NB
GA
NN
SVM
[1] James P. Anderson, “Computer security threat monitoring
and surveillance,” Technical report, James P. Anderson
Co., Fort Washington, Pennsylvania, April 1980.
[2] Cannady J., “Artificial neural networks for misuse
detection,” Proceedings of the ’98 National Information
System Security Conference (NISSC’98), Arlington:
Virginia Press, 1998, pp. 443-456.

30 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
[3] Shon T, Seo J, and Moon J, “SVM approach with a genetic
algorithm for network intrusion detection,” Proceedings of
the 20 th International Symposium on Computer and
Information Sciences (ISCIS 05), Berlin: Springer Verlag,
2005, pp. 224-233.
[4] Yu Y, and Huang Hao, “An ensemble approach to
intrusion detection based on improved multi-objective
genetic algorithm,” Journal of Software, Vol.18, No.6,
June 2007, pp.1369-1378.
[5] J. Luo, and S. M. Bridges, “Mining fuzzy association rules
and fuzzy frequency episodes for intrusion detection,”
International Journal of Intelligent Systems, 2000, pp. 687-
703.
[6] W.K. Lee, and S. J. Stolfo, “A data mining framework for
building intrusion detection model,” Proceedings of the
IEEE Symposium on Security and Privacy, Oakland, CA:
IEEE Computer Society Press, 1999, pp. 120-132.
[7] Kononenko I, “Comparison of inductive and naïve
Bayesian learning approaches to automatic knowledge
acquistion,” in Wieling, B. (Ed), Current trend in
knowledge acquistion, Amsterdam, IOS press, 1990.
[8] Langely, P., Iba, W., Thomas, and K., “An analysis of
Bayesian classifier,” in Proceedings of the 10 th national
Conference on Artificial Intelligence (San Matro, CA:
AAAI press) 1992, pp. 223-228.
[9] Denning, and Dorothy E., “An Intrusion Detection Model,”
Proceedings of the Seventh IEEE Symposium on Security
and Privacy, May 1986, pp. 119-131.
[10] Sebring, Michael M., Whitehurst, and R. Alan., “Expert
Systems in Intrusion Detection: A Case Study,” The 11 th
National Computer Security Conference, October, 1988.
[11] Smaha, and Stephen E., “Haystack: An Intrusion Detection
System,” The 4 th Aerospace Computer Security
Applications Conference, Orlando, FL, December, 1988.
[12] Vaccaro, H.S., and Liepins, G.E., “Detection of Anomalus
Computer Session Activity,” The 1989 IEEE Symposium
on Security and Privacy, May, 1989.
[13] Heberlein, and Let al., “A Network Secutiry Mnitor,”
Proceedings of the IEEE Computer Society Symposium,
Research in Security and Privacy, May 1990, pp. 296-303.
[14] Lunt, and Teresa F., “IDES: An Intelligent System for
Detecting Intruders,” Proceedings of the Symposium on
Computer Security; Threats, and Countermeasures; Rome,
Italy, November 22-23, 1090, pp.110-121.
[15] Teng, Henry S., Chen, Kaihu, Lu, and Stephen C-Y,
“Adaptive Real-time Anomaly Detection using Inductively
Generated Sequential Patterns,” 1990 IEEE Symposium on
Recearch in Security and Privacy, Oakland, CA, pp. 296-
304.
[16] Snapp, Steven R, and Ho., “DIDS (Distributed Intrusion
Detection System) – Motivation, Architecture, and An
Early Prototype,” The 14 th National Computer Security
Conference, October, 1991, pp.167-176.
[17] Jackson, Kathleen, DuBois, David H., Stallings, and Cathy
A., “A Phased Approach to Network Intrusion Detection,”
14 th National Computing Security Conference, 1991.
[18] Lunt, and Teresa F., “Detecting Intrusion in Computer
Systems,” 1993 Conference on Auditing and Computer
Technology, SRI International.
[19] Paxson, and Vern, “Bro: A System for Detecting Network
Intruders in Real-Time,” Proceedings of The 7 th USENIX
Security Symposium, San Antonio, TX, 1998.
[20] Barbara, Daniel, Couto, Julia, Jajodia, Sushil, Popyack,
Leonard, Wu, and Ningning, “ADAM: Detecting Intrusion
by Data Mining,” Proceedings of the IEEE Workshop on
Information Assurance and Security, West Point, New
York June 5-6, 2001.
© 2010 ACADEMY PUBLISHER
[21] Jackson, T., Levine, J., Grizzard, J., Owen, and H., “An
investigation of a compromised host on a honeynet being
used to increase the security of a large enterprise network,”
Proceedings of the 2004 IEEE Workshop on Information
Assurance and Security, IEEE, 2004.
[22] Krasser, S., Grizzard, J., Owen, H., and Levine. J., “The
use of honeynets to increase computer network security
and user awareness,” Journal of Security Education,
Volume 1, pp. 23-37, 2005.
[23] Thomas H. Ptacek, and Timothy N. Newsham, “Insertion,
evasion and denial of service: Eluding network intrusion
detection,” Technical report, Secure Networks Inc., 1998.
[24] D.Y. Yeung, and Y.X. Ding, “Host-based intrusion
detection using dynamic and static behavioral model”,
Pattern Recognition, 36, pp. 229-243, 2003.
[25] SunSoft. SunSHIELD Basic Security Model. SunSoft.
1995.
[26] Diego Zamboni. “Using Internal Sensors for Computer
Intrusion Detection,” PhD thesis, Purdue University, 2001.
[27] Tadeusz Pietraszek, and Chris Vanden Berghe. “Defending
against injection attacks through context-sensitive string
evaluation”, In Recent Advances in Intrusion Detection
(RAID2005), volume 3858 of Lecture Notes in Computer
Science, pp. 124-145, Seattle, WA, 2005. Springer-Verlag.
[28] Sebastiaan Tesink, “Improving Intrusion Detection System
throung Machine Learning,” Technical Report Series no.
07-02, ILK Research Group, Tilburg University, March
2007.
[29] J. R. Quinlan, “Induction of Decision Tree,” Machine
Learning Vol. 1, pp. 81-106, 1986.
[30] J. R. Quinlan, “C4.5: Programs for Machine Learning,”
Morgan Kaufmann Publishers, San Mateo, CA, 1993.
[31] L. Breiman, J. H. Friedman, R. A. Olshen and C.J. Stone,
“Classification and Regression Trees,” Statistics
probability series, Wadsworth, Belmont, 1984.
[32] John Shafer, Rakesh Agarwal, and Manish Mehta,
“SPRINT: A Scalable Parallel Classifier for Data
Maining,” in Proceedings of the VLDB Conference,
Bombay, India, September 1996.
[33] D. Turney, “Cost-Sensitive Classification: Empirical
Evaluation of a Hybrid Genetic Decision Tree Induction
Algorithm,” Journal of Artificial Intelligence Research, pp.
369-409, 1995.
[34] J. Bala, K. De Jong, J. Haung, H. Vafaie and H. Wechsler,
“Hybrid Learning using Genetic Algorithms and Decision
Trees for Pattern Classification,” in Proceedings of 14 th
International Conference on Artificial Intelligence, 1995.
[35] C. Guerra-Salcedo, S. Chen, D. Whitley, and Stephen
Smith, “Fast and Accurate Feature Selection using Hybrid
Genetic Strategies,” in Proceedings of the Genetic and
Evolutionary Computation Conference, 1999.
[36] S. R. Safavian and D. Landgrebe, “A Survey of Decision
Tree Classifier Methodology, ” IEEE Transactions on
Systems, Man and Cybernetics 21(3), pp. 660-674, 1991.
[37] Duda, R., P.E. Hart, and D.G. Stork, “Pattern
classification,” Second edn. John Wiley & Sons, 2001.
[38] Dewan Md. Farid, and Mohammad Zahidur Rahman,
“Learning Intrusion Detection Based on Adaptive Bayesian
Algorithm,” Proceedings of 11 th International Conference
on Computer and Information Technology (ICCIT 2008),
Khulna, Bangladesh, 25-27 December, 2008, pp.652-656.
[39] The KDD Archive. KDD99 cup dataset. 1999.
http://archive.ics.uci.edu/ml/datasets/KDD+Cup+1999+Da
ta
[40] Mukkamala S, Sung AH, and Abraham A, “Intrusion
dection using an ensemble of intelligent paradigms,”

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 31
Proceedings of Journal of Network and Computer
Applications, 2005, 2(8): pp. 167-182.
[41] Chebrolu S, Abraham A, and Thomas JP, “Feature
deduction and ensemble design of intrusion detection
systems.” Computer & Security, 2004, 24(4), pp. 295-307.
[42] Lee WK, Stolfo SJ, and Mok KW, “A data mining
framework for building intrusion detection models,”
Proceedings of the ’99 IEEE Symp. On Security and
Privacy, Oakland: IEEE Computer Society. 1999, pp. 120-
132.
Dewan Md. Farid was born in 1979. He received the B.Sc.
Engineering in Computer Science and Engineering from Asian
University of Bangladesh in 2003 and Master of Science in
Computer Science and Engineering from United International
University, Bangladesh in 2004. He is continuing Ph.D. at
Department of Computer Science and Engineering,
Jahangirnagar University, Bangladesh. His major field of study
is artificial intelligence, machine leaning, and data mining.
He is a faculty member at Department of Computer Science
and Engineering, United International University, Bangladesh.
He has published two conference papers, which include
Learning Intrusion Detection Based on Adaptive Bayesian
Algorithm, etc.
Mr. Farid is a member of Bangladesh Computer Society and
Research Scholar Association, Jahangirnagar University. In
2008, he received the Fellowship of National Science and
Information & Communication Technology (NSICT) from
Ministry of Science and Information & Communication
Technology, Government of Bangladesh.
© 2010 ACADEMY PUBLISHER
Mohammad Zahidur Rahma is currently a Professor at
Department of Computer Science and Engineering,
Jahangirnager University, Banglasesh. He obtained his B.Sc.
Engineering in Electrical and Electronics from Bangladesh
University of Engineering and Technology in 1986 and his
M.Sc. Engineering in Computer Science and Engineering from
the same institute in 1989. He obtained his Ph.D. degree in
Computer Science and Information Technology from University
of Malaya in 2001. He is a co-author of a book on E-commerce
published from Malaysia. His current research includes the
development of a secure distributed computing environment and
e-commerce.

32 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Performance Evaluation for
Question Classification by Tree Kernels using
Support Vector Machines
Abstract — Question answering systems use information
retrieval (IR) and information extraction (IE) methods to
retrieve documents containing a valid answer. Question
classification plays an important role in the question answer
frame to reduce the gap between question and answer. This
paper presents our research work on automatic question
classification through machine learning approaches. We
have experimented with machine learning algorithms
Support Vector Machines (SVM) using kernel methods. An
effective way to integrate syntactic structures for question
classification in machine learning algorithms is the use of
tree kernel (TK) functions. Here we use SubTree kernel,
SubSet Tree kernel with Bag of words and Partial Tree
kernels. Trade-off between training error and margin, Costfactor
and the decay factor has significant impact when we
use SVM for the above mentioned kernel types. The
experiments determined the individual impact for Trade-off
between training error and margin, Cost-factor and the
decay factor and later the combined effect for Trade-off
between training error and margin, Cost-factor. For each
kernel types depending on these result we also figure out
some hyper planes which can maximize the performance.
Based on some standard data set outcomes of our
experiment for question classification is promising.
Index Terms — Precision, Recall, kernel, SubSet Tree,
SubTree, Partial Tree, SVM, Question Classification,
Question Answering.
I. INTRODUCTION
The World Wide Web continues to grow at an
amazing speed. So, there are also a quickly growing
number of text and hypertext documents. Due to the huge
size, high dynamics, and large diversity of the web, it has
become a very challenging task to find the truly relevant
content for some user or purpose. The open-domain
question answering system (QA) has been attached great
attention for its capacity of providing compact and
precise results for users.
The study of question classification (QC), as a new
field, corresponds with the research of QA. At present the
studies on QC are mainly based on the text classification.
Though QC is similar to text classification in some
aspects, they are clearly distinct in that: Question is
usually shorter, and contains less lexicon-based
information than text, which brings great trouble to QC.
Therefore to obtain higher classifying accuracy, QC has
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.32-39
Muhammad Arifur Rahman
Department of Physics, Jahangirnagar University
Dhaka-1342, Bangladesh
arifmail@gmail.com
to make further analysis of sentences, namely QC has to
extend interrogative sentence with syntactic and semantic
knowledge, replacing or extending the vocabulary of the
question with the semantic meaning of every word.
In QC, many systems apply machine-learning
approaches [1-3]. The classification is made according to
the lexical, syntactic features and parts of speech.
Machine learning approach is of great adaptability, and
90.0% of classifying accuracy is obtained with SVM
method and tree kernel as features. However, there is still
the problem that the classifying result is affected by the
accuracy of syntactic analyzer, which need manually to
determine the weights of different classifying features.
Some other systems adopting manual-rule [4].
This paper presents our approaches at question
classification to improve the f1-measure [6]. of question
categorization. Our experiment tried to determine the
optimum value for trade-off between training error and
margin, cost-factor by which training errors on positive
examples outweighs errors on negative examples. It is
used to adjust the rate between precision and recall on the
development set [5-6]. and the decay factor which has the
role to penalize larger tree structures by giving them less
weight in the overall summation [7]. To represent
questions, the paper uses the standard Li and Roth [8]
question classification dataset.
After this overview, question classifying method, a
brief description about the used algorithm and the kernel
methods are introduced, and then impact of different
parameters and parameters combination methods has
been investigated. The comparisons are testified in
experiments based on precision, recall and f1-measure.
The last part of the paper is about the conclusion of the
present research and about the introduction of the further
work could be done on this issue.
II. QUESTION CLASSIFICATION
Question classification means putting the questions
into several semantic categories. Approaches to question
classification can be divided in two broad classes, namely,
rule-based and machine learning methods. Most recent
studies have been based on machine learning approaches.
Li and Roth proposed 6 coarse classes and 50 fine classes
for TREC factoid question answering. The UIUC QC

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 33
dataset, which they developed, contains 5,500 training
questions and 500 test questions, and it is now the
standard dataset for question classification [8-9]. We used
this dataset for our training and testing purpose.
In Table I each coarse grained category contains a
non-overlapping set of fine grained categories. Most
question answering systems use a coarse grained category
definition. Usually the number of question categories is
less than 20. However, it is obvious that a fine grained
category definition is more beneficial in locating and
verifying the plausible answers.
TABLE I: THE COARSE AND FINE GRAINED QUESTION
CATEGORIES.
Coarse Fine
ABBR Abbreviation, expansion
DESC description, definition, manner, reason
ENTY body, color, event, food, creation, currency,
animal, disease/medical, instrument, language,
letter, other, plant, product, religion, sport,
substance, symbol, technique, term, vehicle,
word
HUM description, group, individual, title
LOC city, country, mountain, other, state
NUM code, count, date, distance, money, order,
other, percent, period, speed, temperature, size,
weight
III. SUPPORT VECTOR MACHINE (SVM)
Support vector machine are based on the structural
risk minimization principle from statistical learning
theory [10]. The idea of structural risk minimization is to
find a hypothesis for which we can guarantee the lowest
true error. The bounds are connected on the true error
with the margin of separating hyper planes. In their basic
form support vector machines find the hyper plane that
separates the training data with maximum margin. Since
we will be dealing with very unbalanced numbers of
positive and negative examples in the following, we use
cost factors C+ and C- to be able to adjust the cost of false
positives vs. false negatives as [11]. Finding this hyper
plane can be translated into the following optimization
problem:
1 2
Minimize: � � � i � � � j
2
i:
yi
�1
i:
y j ��1
C
C w �
�
�
� �
Subjected to: �k : yk
�w. xk
� b � 1�
�k
�
Here xi is the feature vector of example i. yi equals +1 or -
1, if example i is in class ‘+’ or ‘-’ respectively.
Using support vector machine we tried to determine
the optimal value of trade-off between training error and
margin (c), cost-factor (j) by which training errors on
positive examples outweighs errors on negative examples.
It is used to adjust the rate between precision and recall
on the development set [11]. Here lambda (�) the decay
factor which has the role to penalize larger tree structures
by giving them less weight in the overall summation.
© 2010 ACADEMY PUBLISHER
IV. KERNEL METHODS
One of the most difficult tasks on applying machine
learning for question classification is the feature design.
Feature should represent data in a way that allows
learning algorithm to separate positive from negative
examples. In SVMs, features are used to build the vector
representation of data examples and the scalar product
between example pairs quantifiers how much they are
similar. Instead of encoding data in the features vectors,
kernel functions can be designed that provide such
similarity between example pairs without using an
explicit feature representation [6].
The kernels we considered in this paper represent
trees in terms of their substructure. Such fragments define
the feature space which, in turn, is mapped into a vector
space. The kernel function measures the similarity
between trees by counting the number of common
fragments. These functions have to recognize if a
common tree subpart belongs to the feature space that we
intended to generate. Here we considered three important
characterization: SubTrees (STs), SubSet Trees (SSTs)
and a new tree class Partial Trees (PTs) described as in
[6].
In case of syntactic parse trees each node with its
children is associated with a grammar production rule,
where the symbol at left hand side corresponds to the
parent and the symbol at right hand side are associated
with the children. The terminal symbols of the grammar
are always associated with the leaves of the tree. For
example, Fig 1. illustrate the syntactic parse of the
sentence-
“Adib brought a cat to the school”.
N
The root
a leaf
a subtree
S
VP
Adib V NP
brought D N
Figure. 1. A syntactic parse tree.
a cat
S
S
N VP
Adib V NP
brought D N
a cat
VP
VP
V NP
N VP
brought D N
a cat
V NP PP
PP
PP
IN N
N
to school
Figure. 2. A syntactic parse tree with its SubTrees (STs).
A SubTree (ST) can be defined as any node of a tree
along with all its descendants. For example, in the Fig. 1.
circle the SubTree rooted in the NP node. A SubSet Tree
(SST) is a more general structure which necessarily not
NP
D N
a cat
IN N
school
V
brought
D
a
N
cat
N
Adib

34 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
includes all the descendants. The only restriction is that
an SST must be generated by applying the same
grammatical rule set which generated the original tree, as
pointed out in [12]. Thus, the difference with the SubTree
is that the SST’s leaves can be associated with nonterminal
symbols. For example, [S [N VP]] is an SST of
the tree in the Fig. 1. and it has the two non terminal
symbols N and VP as leaves.
Figure. 3. A tree with some of its SubSet Trees (SSTs).
VP
V NP
VP
V NP
brought D N
brought D N
a cat
a cat
VP
V NP
VP
V NP
D N
a cat
D N
a cat
Figure. 4. A tree with some of its Partial Trees (PTs).
If we relax the constraints over the SST’s, we obtain a
more general form of substructure that we defined as
Partial Trees (PTs). These can be generated by the
application of partial production rules of the original
grammar. For example, [S [N VP]], [S [N]] and [S
[VP]] are valid PTs of the tree in the Fig. 1. Here Fig. 2.
shows the parse tree of the sentence “Adib brought
a cat” together with its 6 STs. The number of SSTs is
always higher. For example, Fig. 3. shows 10 SSTs (out
of all 17) of the SubTree of the Fig. 2. rooted in VP. Fig.
4. shows that the number of PTs derived from the same
tree is still higher (i.e. 30 PTs). These different
substructure numbers provide an intuitive quantification
of the different information level among the tree based
representations [6].
The main idea of the tree kernel is to compute the
number of the common substructures between two trees
T1 and T2 without explicitly considering the whole
fragment space. From [6]. given a tree fragment space {f1,
f2,….} = F, here we denote the indicator function Ii(n)
which is equal 1 if the target f1 is rooted at the node n and
0 otherwise. It followed that-
K�T1 , T2
��� ���n1,
n2
�…
(1)
n1�N
T n2
�N
1 T2
where N T and N
1
T are the sets of the T1’s and T2’s
2
F
nodes, respectively and ��n1,
n2
���Ii�n1�Ii�n2�. i�1
This letter is equal to the number of common
fragments rooted at the n1 and n2 nodes. We can compute
� as follows-
© 2010 ACADEMY PUBLISHER
VP
NP
D N
a cat
VP
V NP
D N
NP
D N D
a
VP
NP
D N D N D N
D N a cat a
NP D N
a
VP
NP
NP
VP VP
NP NP
D
NP
V
a cat brought
D N D
NP
cat
VP
NP
N
NP NP
N
1. if the production at n1 and n2 are different
the �� n ��0 n 1,
2 ;
2. if the production at n1 and n2 are the same, and
n1 and n2 have only leaf children (pre-terminal
symbols) then �� n ��1 n 1,
2 ;
3. if the production at n1 and n2 are the same, and
n1 and n2 are not pre-terminals then
nc�n1�
j j
� , �� � � ��
, ��
c
n
� 1 2 � n1 n2
j �1
c
n � … (2)
where � ��0,
1�,
nc �n1� is the number of the children
of n1 and j
cn is the j-th child of the node n. As the
production is same nc(n1)= nc(n2).
When � =0, � �n1,n2� is equal 1 only if
j j
� ��
, �� 1
n n c c j , i.e. all the productions associated
1 2
with the children are identical. By recursively applying
the property, it follows the SubTrees in n1 and n2 are
identical. Thus Eq. 1 evaluate the SubSet Tree
kernel .When � =1, � 1 2 � ,n n � evaluates the number of
SSTs common to n1 and n2 as proved in [12].
To include the leaves as fragments it is enough to add,
to the recursive rule set for the � evaluation, the
condition is- if n1 and n2 are leaves and their associated
symbols are equal then ��n 1,
n2
��1. By [6]. such
extended kernel can be referred as SST+bow (bag-ofwords).
Here the decay factor (�) is defined as
��n x, nz
���and nc�nx�
j j
��nxnz����
� ��c��
n c
1 n2
j�1
,
, � �
.
To evaluate all possible substructures common to two
trees, we can (1) select a child from the both trees, (2)
extract the portion of the syntactic rule that contains such
subset, (3) apply Eq. 2 to the extracted partial productions
and sum the contributions of all the children subsets.
Such subsets corresponds to all possible common (non
continuous) node subsequence and computed efficiently
by means of sequence kernel [13]. Let
�
�
J1 � �J11, .., J1r
� and J2 � �J21, .., J2r
� be the
index sequence associated with the order child sequence
of n1 and n2 respectively. Then the number of PTs is
evaluated by the following � function:
�
l�J1�
�
��
���
� � ��
� J1
n1
, n2
1
c ,
� � � � � n1
J , J , l J �l J i�1
where �J� 1
2
� � � �
1
2
�
i J 2i
cn2
l 1
� indicates the length of the target child
� �
and J2 i are the i th children in the
sequence, whereas J1i two sequences.
�
�
�

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 35
V. RESULTS AND ANALYSIS
We evaluate the performance of Question
Classification using SVM with SubTree kernel, SubSet
Tree kernel with Bag of words (SST+BOW) and Partial
Tree kernel. For SubSet Tree kernel with bag of words
first we investigate the optimum value of trade-off
between training error and margin (c) keeping cost-factor
(j) and the decay factor (�) constant. Then we investigate
the impact of cost factor (j) for constant trade-off between
training error and margin (c) and decay factor (�).
F1-Measure
100
90
80
70
60
50
40
30
For Cost Factor (j)=0.5
0 1 2 3 4
Training error and margin Trade-off (c)
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 5 (a). Training error and Margin Vs F1 Measure
for j = 0.5 and SST+BOW
F1-Measure
100
90
80
70
60
50
40
30
For Cost Factor (j)=1.5
0 1 2 3 4
Training error and margin Trade-off (c)
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 5 (b). Training error and Margin Vs F1 Measure
for j = 1.5 and SST+BOW
F1-Measure
100
90
80
70
60
50
40
30
For Cost Factor (j)=3.5
0 1 2 3 4
Training error and margin Trade-off (c)
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 5 (c). Training error and Margin Vs F1 Measure
for j = 3.5 and SST+BOW
© 2010 ACADEMY PUBLISHER
Then for constant factor (�) we determine the hyper plane
for trade-off between training error and margin (c) and
cost-factor (j) which have the maximum f1-measure.
Finally we determined the impact of decay factor (�) for a
constant trade-off between training error and margin (c)
and cost-factor (j). The similar process was repeated for
SubTree kernel and partial tree kernel.
Fig. 5 (a-c). shows the trade-off between training error
and margin (c) versus performance (measured by f1measure)
curve for different types of QC data set using
different Cost Factor (j) and fixed decay factor (�=0.01).
F1-Measure
100
90
80
70
60
50
40
30
For Trade-off between Training Error & Margin C=0.5
0 2 4 6 8
Cost Factor (j)
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 6. (a) Cost factor Vs F1 measure for C=0.5 and
SST+BOW.
F1-Measure
100
90
80
70
60
50
40
30
For Trade-off between Training Error & Margin C=1
0 2 4 6 8
Cost Factor (j)
Figure. 6 (b). Cost factor Vs F1 measure for C=1 and
SST+BOW.
F1-Measure
100
90
80
70
60
50
40
30
For Trade-off between Training Error & Margin C=1.5
0 2 4 6 8
Cost Factor (j)
ABBR
DESC
ENTY
HUM
LOC
NUM
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 6 (c). Cost factor Vs F1 measure for C=1.5 and
SST+BOW.

36 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Here it is notable for same set of parameters different
data types have different performance level. For different
cost factor highest f1-measure can be achieved by setting
the trade-off between training error and margin (c)
around 2, Afterwards its performance becomes almost
constant.
Fig. 6. (a-c) shows the cost factor (j) versus
performance (measured by f1-measure) curve for
different types of question classification data set using
different of Trade of between training error and margin
(c) and fixed decay factor (�=0.01). For a constant tradeoff
between training error and margin different types of
questions have highest f1-measure.
For different value of c highest f1-measure can be
achieved by setting the cost factor (j) less than 2. Even
for some cases if the value of cost factor is increased then
the f1-measue is decreased a certain amount and after a
stage it becomes almost constant.
Fig. 7. shows the decay factor (�) versus f1-measure
curve for a constant values of trade-off between training
error and margin (c=1.5) and cost factor (j=1.5). To get a
very high f1 measure decay factor should have very small.
Starting after 0, up to 0.5 its performance is almost steady
but if we increase the decay factor more the performance
decreased drastically.
F1-Measure
F1-Measure
100
90
85
80
75
90
80
70
60
50
40
30
70
4
For fixed Cost Factor (j=1.5) and Trade-off (c=1.5)
0 0.5 1 1.5 2
Decay Factor (lambda)
Figure. 7. Impact of Decay Factor and SST+BOW.
3
2
Trade-off (c)
1
ABBR
0
0
1
2
3
4
5
Cost FActor (j)
ABBR
DESC
ENTY
HUM
LOC
NUM
Figure. 8(a). Hyper plane for ABBR data type and SST+BOW.
© 2010 ACADEMY PUBLISHER
6
Fig. 8(a-f). show the hyper plane by which maximum
f1-measure can be achieved. For a constant decay factor
(�) a set of values of trade-off between training error and
margin (c) and cost factor (j) can provide the highest f1measure.
Among the six type of data set ‘LOC’ type of
question have to classified by some fixed combination of
trade-off between training error and margin (c) and cost
factor (j) because there is no hyper plane but some peaks.
F1-Measure
94
92
90
88
86
84
82
4
3
2
Trade-off (c)
1
0
HUM
0
2
4
Cost FActor (j)
Figure. 8(b). Hyper plane for HUM data type and SST+BOW.
F1-Measure
100
90
80
70
60
50
40
4
3
2
Trade-off (c)
1
DESC
0
0
2
4
Cost FActor (j)
Figure. 8(c). Hyper plane for DESC data type and SST+BOW.
F1-Measure
90
80
70
60
0
1
2
Trade-off (c)
3
4
6
5
LOC
4
3
2
Cost FActor (j)
Figure. 8(d). Hyper plane for LOC data type and SST+BOW.
1
0
6
6

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 37
F1-Measure
80
70
60
50
40
30
20
4
3
2
Trade-off (c)
1
0
ENTY
0
2
4
Cost FActor (j)
Figure. 8(e). Hyper plane for ENTY data type and SST+BOW.
F1-Measure
95
90
85
80
75
70
65
60
4
3
2
Trade-off (c)
1
0
NUM
0
2
4
Cost FActor (j)
Figure. 8(f). Hyper plane for NUM data type and SST+BOW.
Figure. 9. Training error and Margin Vs F1 Measure
for j = 0.5 and PT
Fig. 9. shows the training error and margin versus F1-
measure curve for a constant cost factor (j = 0.5) using
partial tree. Like SST+BOW the performance is very low
when the training error and margin is very low. After a
certain level (here it is 2) of training error and margin the
performance tends to become almost constant.
© 2010 ACADEMY PUBLISHER
6
6
Figure. 10. Cost factor Vs F1 measure for C=0.5 and PT.
Figure. 11. Impact of Decay Factor for PT.
Figure. 12. Hyper plane for NUM data type and PT.
Fig. 10. shows the Cost factor versus F1 measure
curve for a constant training error and margin (C=0.5).
While Fig. 11. shows the impact of decay Factor. It is
clear that at the lower range it doesn’t have much impact
on performance like SST+BOW. Fig. 12 and Fig.13.
respectively shows the hyper plane for NUM and ENTY
data type.

38 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Fig. 13. Hyper plane for ENTY data type and PT.
TABLE II: MAXIMUM F1- MEASURE OBTAINED USING
SUBSET TREE KERNEL WITH BAG OF WORDS (SST+BOW) FOR
�=0.01.
Coarse F1 P(%) R(%) A(%) j c
ABBR 88.89 99.6 88.89 88.89 6 1
DESC 96.77 98.2 95.74 97.83 1 2.5
ENTY 82.87 93.8 86.21 79.79 6 2
HUM 95.24 98.8 98.36 92.31 4 3.5
LOC 87.8 96 86.75 88.89 5.5 0.5
NUM 94.01 97.4 98.08 90.27 5 1
TABLE III: MAXIMUM F1- MEASURE OBTAINED USING
SUBSET TREE KERNEL FOR �=0.01.
Coarse F1 P(%) R(%) A(%) j c
ABBR 88.89 99.6 88.89 88.89 3.5 2.5
DESC 97.12 98.4 96.43 97.83 1 3.5
ENTY 79.78 92.8 84.52 75.53 6.5 3
HUM 93.75 98.4 95.24 92.31 2 2.5
LOC 88.48 96.2 86.9 90.12 5.5 1
NUM 94.01 97.4 98.08 90.27 5 1.5
TABLE IV: MAXIMUM F1- MEASURE OBTAINED USING
PARTIAL TREE KERNEL FOR �=0.01.
Coarse F1 P(%) R(%) A(%) j c
ABBR 87.5 99.6 100 77.78 3 0.5
DESC 96.06 97.8 95.04 97.1 1 3.5
ENTY 85.56 94.6 86.02 85.11 4 3.5
HUM 81.33 94.4 71.76 93.85 5.5 0.5
LOC 88.48 96.2 86.9 90.12 5.5 1.5
NUM 94.44 97.6 99.03 90.27 4.5 3
Table II shows the maximum F1- measure (F1)
obtained using SubSet tree Kernel-Bag of words
(SST+BOW) for �=0.01 and the corresponding precision
(P), recall (R), accuracy (A) for specific cost factor (j)
and trade-off between training error and margin (c). Table
© 2010 ACADEMY PUBLISHER
III and Table IV shows the performance for SubTree
kernel and Partial Tree kernel respectively.
VI. CONCLUSION
Question classification is one importance role in the
Question Answering frame to reduce the gap between
question and answer. It can conduct answer choosing and
selection. Our question classification method based on
the use of linguistic knowledge and machine learning
approaches and it exploit different classification features
and combination method, also. Though among all the
experiments one or two data set did not provide
distinguishable hyper plane in every cases thereafter the
outcome of experiments done using the tool SVM_light
on Li and Roth question classification data sets
demonstrate some optimal set of values which can
maximize the performance. In the future we aim to
investigate the impact of different parameters on
constituent trees using different types of kernel as well as
other different classifiers.
REFERENCES
[1]. Hovy, E., Hermjakob, U., Lin, C.-Y. and Ravichandran,
D. "Using Knowledge to Facilitate Factoid Answer
Pinpointing", Proceedings of the 19th International
Conference on Computational Linguistics (COLING),
Taipei, Taiwan, 2002.
[2]. Ittycheriah, A., Franz, M., Zhu, W.-J. and Ratnaparkhi,
A. "IBM’s Statistical Question Answering System",
Proceedings of the TREC-9 Conference, Gaithersburg,
MD: NIST, 2000, p. 229.
[3]. Zhang, D. and Lee, W. S. 2003. "Question
Classification using Support Vector Machines",
Proceedings of ACMSIGIR Conference on Research
and Development in Information Retrieval (SIGIR),
Toronto, Canada, 2003.
[4]. LI Xin, HUANG Xuan-Jing, WU Li-de, “Question
Classification by Ensemble Learning”, IJCSNS
International Journal of Computer Science and Network
Security, VOL.6 No3, March 2006.
[5]. Alessandro Moschitti, Silvia Quarteroni, Roberto Basili
and Suresh Manandhar, “Exploiting Syntactic and
Shallow Semantic Kernels for Question/Answer
Classification”, Proceedings of the 45th Conference of
the Association for Computational Linguistics (ACL),
Prague, June 2007.
[6]. Roberto Basili, Alessandro Moschitti, “Automatic Text
Categorization From information Retrival to Support
Vector Learning”, ARACNE editrice, November 2005.
[7]. Stephan Bloehdorn and Alessandro Moschitti,
“Exploiting Structure and Semantics for Expressive
Text Kernels”, Proceeding of the Conference on
Information Knowledge and Management, Lisbon,
Portugal, 2007.
[8]. Li, X., Roth, D, “Learning question classifiers”,
Proceedings of the 19th International Conference on
Computational Linguistics, Taipei, Taiwan, 2002.
[9]. M.-Y. Day, C.-H. Lu, C.-S. Ong, S.-H. Wu, and W.-L.
Hsu, "Integrating Genetic Algorithms with Conditional
Random Fields to Enhance Question Informer
Prediction", Proceedings of the IEEE International
Conference on Information Reuse and Integration,
Waikoloa, Hawaii, USA, 2006, pp. 414-419.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 39
[10]. V. Vapnik, “Statistical Learning Theory”, Wiley, New
York, USA 1998.
[11]. K. Morik, P. Brockhausen, and T. Joachims,
“Combining statistical learning with a knowledge-based
approach – A case study in intensive care monitoring”,
International Conference on Machine Learning (ICML),
Bled, Slovenia, 1999.
[12]. M. Collins, N. Duffy, “New Ranking algorithm for
parsing and Tagging: Kernels over Distance structure,
and the Voted Perception”, Association for
Computational Linguistics (ACL), Philadelphia, USA,
2002.
[13]. Huma Lodhi, Craig Saunders, John Shawe-Taylor,
Nello Cristianini, Christopher Watkins, “Text
Classification using string kernels”, in NIPS, pp 563-
569, 2000.
© 2010 ACADEMY PUBLISHER
Muhammad Arifur Rahman has
completed his Masters degree in Human
Language Technology and Interfaces
(HLTI) under the Department of
Information Engineering and Computer
Science at University of Trento, Italy. He
received his B.Sc. (Honors) degree from the
Department of Computer Science and Engineering,
Jahangirnagar University, Bangladesh at 2001. His major areas
of interest include Natural Language Processing, Data
Communication, and Digital Signal Processing.
He has authored a book titled “Fundamentals of
Communication” and more than 10 international journal and
conference papers. At present he is serving as a Lecturer in the
Department of Physics, Jahangirnagar University, Dhaka,
Bangladesh.

40 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Recurrent Neural Network Classifier for Three
Layer Conceptual Network and Performance
Evaluation
Md. Khalilur Rhaman
Department of Computer Science and Engineering, BRAC University, 66 Mohakhali, Dhaka-1212,
Bangladesh.
Email: khalilur@bracu.ac.bd
Tsutomu Endo
Kyushu Institute of Technology, 680-4 Kawazu, Iizuka, Fukuoka, 820-8502, Japan.
Email: endo@pluto.ai.kyutech.ac.jp
Abstract— Natural language has traditionally been handled
using symbolic computation and recursive processes. Classification
of natural language by using neural network is
a hard problem. Past few years several recurrent neural
network (RNN) architectures have emerged which have
been used for several smaller natural language problems.
In this paper, we adopt Elman RNN classifier for disease
classification for a doctor patient-dialog system. We find that
the Elman RNN is able to find a representation for natural
language. Contextual analysis in dialog is also a major
problem. A three layers memory structure was adopted to
address the challenge which we referred to as ”Three Layer
Conceptual Network” (TLCN). This highly efficient network
simulates the human brain by discourse information. An
extended case structure framework is used to represent
the knowledge. We used the same case frame structure to
train and examine the RNN classifier. This system prototype
is based on doctor-patients dialogs. The over all system
performance achieved 84% accuracy. Disease identification
accuracy depends on number of disease and number of
utterances. The performance evaluation is also discussed in
this paper.
Index Terms— Three Layer Conceptual Network, Knowledge
Representation, Recurrent Neural Network.
I. INTRODUCTION
In this paper we present a Neural Network Classifier
to classify the diseases for our system prototype. EL-
MAN [1] [2] POLLACK [3] and LAWRENCE [4] are
considered to apply in this application. In this initiative
we introduced Neural Network in Three Layer Conceptual
Network (TLCN). To implement this novelty a number of
researches was studied [5] [6] [7] [8] [9] [10] [11] [12]
[13] [14] [15] which integrated RNN with NLP. TLCN
architecture was developed by RHAMAN [16] from the
memory model idea of NOMURA [17]. NOMURA’s
model was a framework for a memory management but
RHAMAN developed architecture to address a real time
dialog system. A discourse algorithm was introduced to
represent discourse information which was related to [18],
[19], [20]. It was a model which could simulate the
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.40-48
human brain memory. Human brain consist three types
of memories Long-term, short-term and mid-term which
was studied in [21], [22] and [23]. In our system episodic
memory module simulates short-term memory, discourse
memory module simulates mid-term memory and ground
memory module simulates long-term memory. But in case
of disease identification the system performance was not
satisfactory. So, We use RNN Classifier and get better
result. We used the traditional modules for input and
response generation but developed an extended case frame
model of BRUCE [24] and SHIMAZU & NOMURA [25]
for handling the knowledge database. Case frame network
is very useful for unstructured and ambiguous languages
like Japanese. However, most systems deal with rather
simple sentences, in which the analysis is a relatively easy
task. Our extension of case frame architecture reduced the
complexity using formalism. We use the same structure
for RNN classifier.
II. MOTIVATION
Patient come back to the doctor repeatedly with the
same symptoms and expects the doctor to fix it. In
another case most of patients come to the doctor with
same symptoms. In the decade of 1980 and 1990, a
number of medical expert systems were designed using
medical knowledge to diagnose the diseases. Medical
expert systems have evolved to provide physicians with
both structured questions and structured responses within
medical domains of specialized knowledge or experience.
Most of these systems [26], [27], [15] and [28] were
based on doctor. Only medical experts could enter the
predefined data set to find the solution. Our proposed
system is very different in this point. Our system is
designed to communicate with patient directly. Patients
can communicate our system by their natural language.
Natural language has traditionally been handled using
symbolic computation and recursive process. The most
successful stochastic models have been based on finitestate
descriptions models. However finite-state model can

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 41
not represent hierarchical structures found in natural language.
In past few decades most of the NLP systems [29],
[30], [31] [32], [33], [34], [35]and [36] were developed
based on syntactic phrase structure. These systems are
good for structured language but not good for encoding
linguistic information of unstructured language like
Japanese. So, our motivation was to implement a system
that can communicate directly with patients and also can
handle the process of sequential sentences and complex
sentence.
III. OVERVIEW OF TLCN
Three Layer Conceptual Network model is a framework
to use the knowledge that simulates human brain.
Discourse information is playing the vital role for this
kind of simulation. It consists of three layered memory
modules: 1) Ground Memory Module (GMM), 2) Discourse
Memory Module (DMM) and 3) Episodic Memory
Module (EMM). In our proposed model, the knowledge
and their relationships accommodate into the GMM.
DMM represents the most basic, useful and recently used
discourse knowledge. EMM is a User interface. It contains
linguistic and non-linguistic knowledge for understanding
input dialogs and generates response dialogs. However,
memory modules are not simply an accumulation of the
contents of other memories. It is a structured memory
coherently organized by merging old and new knowledge.
The process relates to what is called knowledge learning.
Each memory module consists of a knowledge database
and its manager. Fig. 1 depicts the simple process flow
of TLCN.
A. Architecture
Figure 1. Processing of TLCN
The TLCN-Dialog Processor (DP) architecture is illustrated
in Fig. 2. This architecture provides an efficient
integration of component modules. Text input utterances
are sent to dialog understanding unit of EMM. To understand
the input dialog, the entire linguistic knowledge
is assimilated to episodic memory from ground memory
in advance. After normalizing the input text, it generates
case frames that are sent to classifier unit. Here, input
© 2010 ACADEMY PUBLISHER
case frames are classified into some predefined classes.
DMM manager receive these classified case frames from
EMM manager and searches in the discourse memory to
match some diseases. DM will be acknowledged whether
any disease is identified or not. If the information does
not match enough in discourse memory, DM sends the
input case frames to GMM manager. GMM manager
does the same task as DMM manager have done and
assimilates the diseases to discourse memory. If the input
case frames do not match with any of the diseases,
DM generates a ”not identified response”. If it could
match with some diseases but not enough to satisfy
the identification condition, it generates case structures
of question response for the dialog generator. Dialog
generator then asks a natural language question to the
patient. If the matching result satisfies the identification
condition, it generates identification response including
advice case frames, carefulness case frames and treatment
case frames. Then, the dialog generator generates natural
language text dialog from the case frames.
Figure 2. TLCN Dialog Processing System
B. Ground Memory Module (GMM)
The ground memory of GMM contains all kinds of
knowledge. This is the main data source of the system. It
stores information related to linguistic knowledge such as
dictionary, grammar, and non-linguistic knowledge which
includes disease, symptom, cause, treatment, effect etc.
Procedural knowledge which relates inferring a fact from
a collection of facts also described here. We considered
[37], [38], [39], [40], [41] to implement ground
memory. It also includes discourse information like most
common diseases, seasonal diseases, previous record of
a patient and recently processed disease information. In
ground memory, knowledge is represented by a case frame
structure form. Fig. 3 is showing the hierarchy of case
frame representation where the nodes represent the case
structures and the edges represent their relationships. In
this figure, X and Y axis represent the case frames and
Z axis denotes level of their relationships. More than one
case frame of a previous level can be connected with one
or more case frames of the next level and vice-verse. In
the following sub-section we will discuss the details of
knowledge representation in ground memory.

42 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Figure 3. Case structure and their relationship
The knowledge is classified in some predefined classes.
For this system prototype the nodes are classified in
diseases. A class can be a property of other class. The
GMM manager is an algorithm that searches the disease
information in ground memory when requested by the
Dialog Manager and also finds the discourse knowledge
requested by discourse memory module. Ground memory
always updates itself after every conversation and never
deletes anything.
1) Case frame Model: The extended case frame model
we used is an understanding model consists of predicates,
semantic case relations (roles), modalities, and
conjunctive relations that were previously proposed by
SHIMAZU [25]. BRUCH [24] model also studied to
design our extended model.
sentence = simple sentence|sentence +
conjunctive relation + sentence
simple sentence = predicate + case relations +
modalities
case relations = object|method|direction|timeand
space|supplement|modification
modalities = tense|aspect|manner|intention|guess
|attitude|negation|ability|necessity| . . .
conjunctive relation = time|cause|reason|result|
contrast|goal|assumption|
1) Case Relations: one characteristic of this model is
the relatively large collection of case relations. The
collection includes: 1) basic cases like Fillmore’s
case system, 2) case relations such as extent, manner,
and degree, which appear as adverbial phrases,
and 3) additional cases represented by the inflection
of a certain verb.
2) Modalities: semantic structures must include information
on modality such as tense, aspect, intention,
manner, attitude and assumption. Real sentences
convey much of this type of information. Unfortunately,
however, most understanding systems and
linguistic theories have not dealt sufficiently with
modalities.
3) Conjunctive Relations: A sentence generally consists
of several sub-sentences, each of which represents
a unit event. In the model, event relations
© 2010 ACADEMY PUBLISHER
are expressed by conjunctive relations such as time,
cause, reason, result, goal, assumption, contrast and
circumstance. These are essential for understanding
a sentence, and must be organized into semantic
structures.
2) Extended Case Frame Representation: An extension
of case structure model is proposed as a linguistic
model for representing the text meaning structure. Thus
the case frames act as a representation scheme for ground
memory. The traditional case structure is a structure for a
unit sentence which consists mainly in relations between
noun and verb. This is not sufficient to represent structures
of real sentences which sometimes have complex noun
phrase and compound sentences. Also our proposed case
structure has to have facilities for representing other
structures involving relations between two nouns by verb
and preposition. It has been designed to integrate those
structures into one linguistic model. Its nature is hierarchical
with respect to the way constituents are connected;
iterative with respect to conjunction, and recursive with
respect to embedding. Using this formalism, the syntactic
and semantic structures of sentences can be represented
uniformly.
In our proposed structure, every noun has six properties:
adjective, delimiter, preposition, auxiliary verb,
adverb and verb. Noun can be related by either verb
or preposition. This case frame representation makes
the matching algorithm simple and efficient because the
same structure can represent both case frames and their
relationships. The basic structure of case frame is depicted
in Fig. 4.
Figure 4. Case frame structure
Fig. 5 shows the representation of case frame. One case
structure node can have more than one connection node
and one connection node can have more than one case
structure node. There is no restriction about consequence
case structure and connection node to represent a complex
or compound knowledge.
Figure 5. shows the relation of case structures
Fig. 6. shows an example of knowledge representation

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 43
in ground memory where the disease is considered as a
head class and others are considered as daughter classes.
This also explains the representation of knowledge in
a real database. Here the example disease is cholera.
Symptom, cause and others are sub-nodes of cholera.
The sub-nodes of symptom and cause represent the case
frames for the natural language texts written in the box
located at the right bottom of the figure.
Figure 6. Knowledge representation
3) Training to the Ground Memory: The TLCN-DP
system is based on doctor-patient dialog conversation.
First the system was trained under the supervision of a
professional doctor. At the time of training, the trainer
carefully considered the discourse information and the
level of nouns. The statistics of training corpus are given
in Table I. The information about these diseases was
collected from medical web links [42], [43] and [44]. The
corpus of RHAMAN [45] is also considered to train the
system.
TABLE I.
TRAINING CORPUS STATISTICS
Number
Diseases 30
Sentences for each disease 95
Normalized Sentences for each disease 220
Vocabulary Size 2500
C. Discourse Memory Module (DMM)
Discourse memory contains discourse information that
provides situational and contextual information for an
utterance environment. It contains nodes with different
frequency. The frequency is calculated depending on the
uses of knowledge. In our proposed system we used three
frequencies: medium, high and very high. We standardize
the frequency three after examining with other frequencies.
We found better response using the frequency three.
The very high frequency is determined by the most used
case frames, recently used classes and seasonal disease
case frames. The high frequency nodes are determined
by the basic information of the patient and his past
© 2010 ACADEMY PUBLISHER
medical record. Medium frequency nodes are determined
by the medium uses of case frames. The knowledge and
their frequencies of discourse memory are different for
different patients.
1) Discourse Information: Some discourse information
is pre-defined and some discourse information depends
on real time information. After getting the basic
information of patient and environment, the system decides
the discourse knowledge considering the following
criteria. The discourse frequency is always predefined by
the trainer of the system.
1) Most common diseases: In every area or age there
are some common diseases. This information is predefined
by the trainer.
2) Seasonal disease Information: There are some common
diseases for each season. This discourse information
is identified by current date and the predefined
seasonal information.
3) Recently treated disease information: The diseases
and case frames are mostly handled by the current
system in recent time. Its discourse frequency is
determined by the frequency of handled disease.
4) Previous record of current patient: If the patient’s
basic information repeats, then case frames of entire
previous records are considered as discourse. All
previous dialog conversations are uploaded with
very high discourse frequency.
If any disease matches with one of above criteria, the
frequency will be medium. If it matches with two criteria,
frequency will be high. If match with three criteria the
discourse frequency will be very high.
2) Dialog Matching: First the DMM manager compares
the patients symptoms and causes with all the
symptoms and causes of discourse memory. If the symptom
does not match it searches in ground memory with
the help of GMM manager. It selects all of them as
a candidate disease and sends all these case frames to
dialog manager. DM calculates the highest probability
between these diseases with a probability function. This
function will be discussed in DM section. Depending on
the probabilities, DM requests the responses from DMM.
3) Response Generator: In our TLCN-DP architecture,
we have incorporated a response module to help patients
to find the goal disease. The dialog manager analyzes
the result and selects the type of system response. DMM
manager provides case frames for responses to the language
generator. The system responses are divided in
three types: 1) response before disease identification, 2)
after disease identification and 3) failed to identify.
1) Response before disease identification: in this type
of response, first it considers the symptom and
the causes of two diseases which have the highest
probability between selected diseases. Find the dissimilarities
of symptoms and causes between them.
Response module then generates case frames for
these. After 10 conversations it select case frames
for pharmaceutical tests (if any test exists in the
selected diseases).

44 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
2) After Disease Identification: once the disease has
been identified, the dialog manager provides case
frames for treatment including medicines, effects,
and preventions for the specific disease to the
language generator. It can also provide more case
frames for details of this disease according to the
patient request.
3) Failed to identify: if the dialog manager does not
obtain any information about any diseases within 3
conversations and does not match more than 75%
of the important symptoms and causes, it generates
a failure notification.
D. Episodic Memory Module (EMM)
EMM receives input from the user and provides output
to the user. The input processor receives the input text
dialog and understands meaning. Episodic memory stores
the meaning of the ongoing segment of utterances. As
understanding proceeds, the essence of episodic memory
assimilated into discourse memory and the essence of
discourse memory assimilated into ground memory. The
same way the output processor receives case frames from
DMM manager and generates the appropriate natural
language text for user. To understand the dialog the
input processor uses some modules like text normalization
unit, converter for case structure and semantic classifier.
Episodic memory contains all the linguistic information
and the recent conversation with the current user. EMM
manager includes the entire previous dialog and responses
with the case structure of the new dialog and classifies
it in the pre-defined classes. It sends these classified
case structures to the discourse memory module. Episodic
memory is a temporary memory. It updates only for the
ongoing process requirement. After every conversation
with a user, it refreshes the episodic memory.
1) Components: EMM consists of language understanding
module and the language generation module. The
understanding module normalizes the input text and converts
them to a language-neutral meaning representation –
a case frame. Finlay it classifies the case frames and send
them to DMM. After getting the response from DMM, the
language generation module produces the response in text
form. In this section we will discuss four components:
1) text normalization, 2) case frame representation, 3)
classifier and 4) Language generator.
1) Text normalization: this is an essential step for
minimizing ”noise” variations among words and
utterances. The text normalization component is essentially
based on using synonyms and other forms
of syntactic normalization. The main normalization
factors include stemming using a synonyms dictionary,
removal of confusions, non-alphanumeric and
non-white space characters.
2) Case frame representation: it generate the case
frames using the same case frame model discussed
in GMM section.
3) Classifier: in TLCN-DP the nodes are classified in 9
properties: 1) Disease, 2) Symptom, 3) Environment
© 2010 ACADEMY PUBLISHER
Cause, 4) Physical Cause, 5) Other Cause, 6) Tests,
7) Treatment & Medicine, 8) Future Effect and 9)
Prevention.
Disease is the head node. All the other nodes are
related disease. Fig. 7 explains symptom and cause
used to find the goal disease. Then pharmaceutical
Figure 7. Relationship of knowledge
tests confirm the disease (if any tests exist). After
that, the TLCN-DP will be able to provide detailed
information to the patient including treatment,
medicine, effect, carefulness and prevention.
If any utterance from patient’s dialog contains any
data related to the body then it will be classified
as symptom class. If any utterance contains any
data about treatment or medicine then it will be
classified as treatment class. If any utterance contains
any data about test then it will be classified as
test class. Otherwise it will be classified as cause
oriented. Classifier sends the classified utterances to
the DMM manager.
4) Language Generator: It organizes the sentences by
using the parts of speech and following grammatical
rules from the response case frames, and generates
the natural language output text to the user.
IV. NEURAL NETWORK
Natural language has traditionally been handled using
symbolic computation and recursive processes. The most
successful stochastic language models have been based
on finite-state descriptions such as n-grams or hidden
Markov models. However, finite-state models cannot represent
hierarchical structures as found in natural language.
In the past few years several RNN architectures have
emerged which have been used for grammatical inference.
RNNs have been used for several smaller natural language
problems. Neural network models have been shown to be
able to account for a variety of phenomena in phonology,
morphology and role assignment. It has been shown
that RNNs have the representational power required for
hierarchical solutions, and that they are Turing equivalent.
The Elman RNNs investigated in this paper to classify
diseases.
A. Recurrent Neural Network
Recurrent Neural Network is a class of neural network
where connections between units form a directed cycle.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 45
This creates an internal state of the network which allows
it to exhibit dynamic temporal behavior. RNN have
feedback connections and address the temporal relationship
of inputs by maintaining internal states that have
memory. RNN are networks with one or more feedback
connection. A feedback connection is used to pass output
of a neuron in a certain layer to the previous layer(s).
The different between NLP and RNN is RNN have feedforward
connection for all neurons (fully connection).
Therefore, the connections allow the network show the
dynamic behavior.
B. Elman Recurrent Neural Network
Elman Recurrent Neural Network: A recurrent network
with feedback from each hidden node to all hidden nodes.
When training the Elman network back propagationthrough-time
is used rather than the truncated version used
by Elman, i.e. in this paper Elman network refers to the
architecture used by Elman but not the training algorithm.
Back-propagation-through-time has been used to train the
recurrent networks.
V. CLASSIFIER
Recurrent Neural Network Classifier is used to identify
the probability of each disease. The inputs of this
classifier are the indexes of classified case structures. As
we mentioned before, each disease is a combination of
some specific classes (symptom, cause etc.), each class is
a combination of some case frames and each case frame
is a combination of some case structures. All of these
elements are indexed in advance.
A. Architecture
In our proposal we used an Extended Elman RNN
for classification. According to the experiment of Pollack
[3] and Lawrence-Giles-Fongs [4], Elman RNN classifier
provides better performance for Natural Language Processing.
Our classifier is using four word inputs, 7 hidden
nodes. We did experiment with varying the number of
hidden nodes and found optimal performance for 7 hidden
nodes considering time and accuracy. The quadratic cost
function, the learning rate schedules are shown below. An
initial learning rate of 0.2 and we use the random weight
initialization strategy. The Recurrent Network we used for
this application is depicted in Fig. 8. Where S indicates
Symptom, EC indicates Environment Cause, PC indicates
Physical Cause, OC indicates Other Cause, T indicates
Tests, H indicates Hidden node, D indicates Disease and
NI indicates Not Identified.
1) Target Output: Target outputs were 0.1 and 0.9
using the logistic activation function. This helps
avoid saturating the sigmoid function.
2) Weight Initialization: Random weights are initialized
with the goal of ensuring that the sigmoid do
not start out in saturation but are not very small
(corresponding to a flat part of the error surface).
In addition, several (20) sets of random weights
© 2010 ACADEMY PUBLISHER
Figure 8. Disease Classifier
are tested and the set which provides the best
performance on the training data is chosen. In our
experiments on the current problem, it was found
that these techniques do not make a significant
difference.
3) Learning rate schedule:We used the learning rate
scheduled by LAWRENCE [4] that was first proposed
by Darken and Moody [46].
4) Activation Function: Symmetric sigmoid functions
often improve convergence over the standard logistic
function. For our particular problem we found
that the difference was minor and that the logistic
function resulted in better performance.
5) Cost Function: The relative entropy cost function
has received particular attention and has a natural
interpretation in terms of learning probabilities.
We investigated using both quadratic and relative
entropy cost functions.
E = 1
2
�
(yk − dk) 2
k
(1)
E = �
[
k
1 1 + yk
(1+yk)log +
2 1 + dk
1 1 − yk
(1−yk)log ]
2 1 − dk
(2)
Where y and d correspond to the actual and desired
output values, k ranges over the outputs. We found
the quadratic cost function to provide better performance.
A possible reason for this is that the use
of the entropy cost function leads to an increased
variance of weight updates and therefore decreased
robustness in parameter updating.
6) Sectioning of the Training Data: We investigated
dividing the training data into subsets. Initially, only
one of these subsets was used for training. After
100% correct classification was obtained or a prespecified
time limit expired, an additional subset
was added to the working set. This continued until
the working set contained the entire training set.
The data was ordered in terms of sentence length
with the shortest sentences first. This enabled the

46 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
networks to focus on the simpler data first. Elman
suggests that the initial training constrains later
training in a useful way.
B. Training and evaluation of RNN
For each disease, we used different RNN. Each of
these networks is trained by positive and same number
of random negative set of index. Positive set of index
denotes its own disease case frames where negative set
of indexes denote all other case frames that are not the
member of this disease. These case frames are already
discussed in ground memory module section. We were
able to train an RNN up to 100% correct classification
on the training data. Generalization on disease identification
resulted in 89% correct classification on average.
This is better than the performance obtained form other
researches. A possible reason for this is that there are
similarities between input dialog and trained dialog that
means the set of index of input dialogs and the set of
index of disease information. The output range of this
classifier is from 0.9 to 0.1.
C. Simulation Details
The network contained three layers including the input
layer. The hidden layer contained 7 nodes. Each hidden
layer node had a recurrent connection to all other hidden
layer nodes. All inputs and outputs were within the range
zero to one. Bias inputs were used. The best of 50 random
weight sets was chosen based on training set performance.
Targets outputs were 0.1 and 0.9 using the logistic output
activation function. The quadratic cost function was used.
The search then converge learning rate schedule used was
Equ.(3)
Where
η =
n
N/2 +
η0
c1 �
max(0,c
max 1,(c1− 1 (n−c2N)) (1−c2 )
η = learningrate
η0 = initiallearningrate = 0.2
N = totaltreaningepochs
n = currenttrainingepoch
Cl = 50, c2 = 0.65
The training set consisted of 220 normalized sentences
for each disease. The number of disease were 30.
D. Dialog Manager
Dialog manager is a controller for this system. It
maintains the data flow between memory modules. Also it
calculates the RNN classifier. After classification it selects
the disease and system response.
Pseudo code for Disease Selection:
1.Select diseases for: Patients symptoms
or causes = Disease(n)s symptoms
© 2010 ACADEMY PUBLISHER
�
(3)
or causes.
if not found then Goto GMM Manager.
Select diseases for: Patients symptoms
or causes=Disease(n)s
symptoms or causes.
if found then upload the disease &
goto DMM.
Else Generate "Not found" response.
2.if number of request dialog < 15
then
if (MAX(f(input dialog)/f(n))

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 47
A. PERFORMANCE EVALUATION
Evaluation of dialog system performance is a complex
task and depends on the purpose of the desired dialog
metric. We have trained our system with mostly common
30 diseases from [42], [43], [44] where 11 are mainly
related with age, 10 are related with season and the other
of 9 diseases are related with food, accident, habitual,
health condition and location. The real doctor-patient
dialog conversation is collected for this evaluation. The
dialogues have been collected from various books on
medical interviews and some frequently asked questionsanswering
medical web links. The entire information and
output was verified by a professional doctor. 85 patients
dialogs were simulated with this system prototype where
10 patients were tested 3 times and 15 patients 2 times
and 25 patients only once. Table II presents the existing
disease identification accuracy with the number of dialog
conversation. According to the result analysis table; we
found that the system achieved a very satisfactory result
for 10 to 20 diseases. For 15 diseases we found 79% accuracy
of disease identification. For less than 10 diseases the
system can identify with very few dialog conversations.
82% accuracy achieved only for 7 utterances. Before
utilizing RNN we implemented another system where we
used a comparatively simple probability function. More
than 10% overall accuracy gained from that system. The
result of previous system is depicted in table III. We did
not verify more than 20 diseases because the performance
was already not satisfactory for 20 diseases.
TABLE II.
DISEASE IDENTIFICATION EVALUATION
Utterances
Diseases 5 7 9 11 13 15
10 57% 82% 83% 83% 84% 84%
15 35% 60% 78% 78% 79% 79%
20 19% 41% 59% 66% 67% 68%
25 14% 37% 51% 53% 55% 56%
30 10% 23% 33% 42% 48% 48%
TABLE III.
PREVIOUS SYSTEM EVALUATION
Utterances
Diseases 7 9 11 13 15
10 56% 59% 62% 64% 66%
15 49% 55% 58% 60% 63%
20 46% 50% 53% 57% 61%
VII. CONCLUSION
The main underlying strategy is to adopt Elman RNN
with Three Layer Conceptual Network. Uses of extended
case frame, dialog matching algorithm, disease identification
algorithm and response generation criteria are also
presented. In addition of these fundamental strategies,
we discussed the characteristic of Elman RNN with our
system. Finally we presented a comparative study with
our previous system.
© 2010 ACADEMY PUBLISHER
Still, there are lots of scopes to improve in neural
network application. We are looking for some scope to
implement TLCN using efficient neural network. Adopting
some module, it might be possible to train this system
by the real time conversation of a doctor and patient. And
in future we plan to use voice dialog rather then text
dialog.
This is only the beginning of Neural Network with
Three Layer Conceptual Network. Our goal is to make
human-computer successful interaction. We strongly believe
this research is a new breed of handshaking between
TLCN and neural network.
ACKNOWLEDGMENT
We thank Hirosato Nomura for numerous discussions
concerning this work, Teigo Nakamura and Manuel Medina
González for their assistance and the reviewers for
their detailed comments.
REFERENCES
[1] J. L. Elman, “Distributed representations, simple recurrent
networks, and grammatical structure,” Machine Learning,
no. 7, pp. 195–226, 1991.
[2] ——, “Finding structure time,” Cognitive Science, vol. 14,
pp. 179–211, 1990.
[3] J. B. Pollack, “Recursive distributed representations,” Artificial
Intelligence, vol. 46, pp. 77–105, 1990.
[4] S. Lawrence, C. L. Giles, and S. Fong, “Natural language
grammatical inference with recurrent neural networks,”
IEEE Transactions on Knowledge and Data Engineering,
vol. 12, pp. 126–140, 2000.
[5] M. Towsey, J. Diederich, I. Schellhammer, S. Chalup, and
C. Brugman, “Natural language learning by recurrent neural
networks: a comparison with probabilistic approaches,”
pp. 3–10, 1998.
[6] A. Graves, N. Beringer, and J. Schmidhuber, “A comparison
between spiking and differentiable recurrent neural
networks on spoken digit recognition.”
[7] D. Chen, G. Z. Sun, H. H. Chen, and Y. C. Lee, “Extracting
and learning an unknown grammar with recurrent neural
networks,” Advance in neural Information Processing Syatem,
1992.
[8] R. J. Williams and J. Peng, “An efficient gradient–based
algorithm for on–line training of recurrent network trajectories,”
Neural Computation, vol. 2, no. 4, pp. 490–501,
1990.
[9] G. Scheler and T. Munchen, “With raised eyebrows or the
eyebrows raised? a neural network approach to grammar
checking for definiteness,” pp. 160–170, 1996.
[10] C. Fabrizio, F. Paolo, L. Vincenzo, and S. Giovanni,
“Towards incremental parsing of natural language using
recursive neural networks,” Towards Incremental Parsing
of Natural Language using Recursive Neural Networks, p.
Paginated, 2002.
[11] Y. Li and J. G. Harris, “A spiking recurrent neural network,”
VLSI, IEEE Computer Society Annual Symposium
on, p. 321, 2004.
[12] S. C. Kwasny, S. Johnson, and B. L. Kalman, “Recurrent
natural language parsing 1.”
[13] R. Miikkulainen, “Natural language processing with subsymbolic
neural networks,” pp. 120–139, 1997.
[14] M. H. Christiansen and N. Chater, “Natural language
recursion and recurrent neural networks,” 1994.

48 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
[15] S. Wermter and V. Weber, “Screen: Learning a flat syntactic
and semantic spoken language analysis using artificial
neural networks,” Journal of Artificial Intelligence
Research, vol. 6, pp. 35–85, 1997.
[16] R. M. Khalilur and E. Tsutomu, “Three Layer Conceptual
Network Dialog Processor,” in The Twelfth IASTED International
Conference on AI and Soft Computing, 2008, pp.
92–97.
[17] H. Nomura, “Modeling and representative framework for
linguistic and non-linguistic knowledge in natural language
understanding,” in Germany-Japan Science Seminar, 1986,
pp. 1–10.
[18] M. A. Walker, “Evaluating discourse processing algorithms,”
in Proceedings of the ACL, 1989.
[19] H. Nomura, “Experimental machine translation systems
lute,” in Second Joint European-Japanese Workshop on
Machine Translation, 1985, pp. 1–2.
[20] D. Marcu and A. Echihabi, “An unsupervised approach
to recognizing discourse relations,” in The 40th Annual
Meeting of the Association for Computational Linguistics,
July 2002, pp. 368–375.
[21] E. Tulving and F. I. M. Craik, “The oxford handbook of
memory.”
[22] T. K. Landauer, “How much do people remember? some
estimates of the quantity of learned information in longterm
memory.” Cognitive Science, vol. 10, pp. 477–493,
1986.
[23] R. L. Buckner, “Beyond hera: Contributions of specific
prefrontal brain areas to long-term memory retrieval.”
Psychonomic Bulletin and Review, vol. 3, pp. 149–158,
1996.
[24] B. Bruce, “Case systems for natural language,” Artificial
Intelligence, vol. 6, no. 4, pp. 327–360, 1975.
[25] A. Shimazu, S. Naito, and H. Nomura, “Japanese language
semantic analyzer based on an extended case frame
model,” in International Joint Conference on Artificial
Intelligence, 1983, pp. 717–720.
[26] S. Tsumoto, “Automated extraction of medical expert
system rules from clinical databases based on rough set
theory.” Information Sciences, p. 112, 1998.
[27] J. H. Frenster, “Expert systems and open systems in
medical artificial intelligence.” Am Assoc Medical Systems
and Informatics, vol. 7, pp. 118–120, 1989.
[28] F. Shahbaz, F. Maqbool, S. Razzaq, K. Irfan, and T. Zia,
“The role of medical expert systems in pakistan.” World
academy of science engineering and technology, vol. 27,
2008.
[29] Y.-S. Lee, D. J. Sinder, and C. J. Weinstein, “Interlinguabased
english-korean two-way speech translation of
doctor-patient dialogues with cclinc,” Machine Translation,
vol. 17, pp. 213–243, 2002.
[30] J. F. Allen, D. K. Byron, M. Dzikovska, G. Ferguson,
L. Galescu, and A. Stent, “Conversational humancomputer
interaction,” in IASTED International Conference
on AI and Soft Computing, vol. 12, Sept. 2001, pp.
628–802.
[31] M. Kipp, J. Alexandersson, R. Engel, and N. Reithinger,
“Dialog processing,” W. Wahlster (ed.) Verbmobil: Foundations
of Speech-to-Speech Translation, pp. 452–465, 2000.
[32] E. Tsutomu and K. Tsuneo, “Cooperative understanding
of utterances and gestures in a dialogue-based problem
solving system.” Computational Intelligence, pp. 152–169,
1999.
[33] K. Shimada, Y. Uchida, S. Sato, S. Minewaki, and T. Endo,
“Speech understanding using confidence measures and
dependency relations,” Proc. of PACLING2005, pp. 278–
283, 2005.
[34] N. Okada and T. Endo, “Story generation based on dynamics
of the mind.” Computational Intelligence, pp. 123–160,
1992.
© 2010 ACADEMY PUBLISHER
[35] K. Shimada, K. Iwashita, and T. Endo, “A case study of
comparison of several methods for corpus-based speech intention
identification,” Proceedings of the 10th Conference
of the Pacific Association for Computational Linguistics
(PACLING2007), pp. 255–262, 2007.
[36] S. Minewaki, K. Shimada, and T. Endo, “Interpretation of
utterances based on relevance theory: Toward the formalization
of implicature with the maximum relevance,” Proc.
of PACLING2005, pp. 214–222, 2005.
[37] A. Graves, N. Beringer, and J. Schmidhuber, “Grammar
learning for spoken language understanding,” Automatic
Speech Recognition and Understanding, pp. 292–295,
2001.
[38] R. E. Schapire and Y. Singer, “Boostexter: A boostingbased
systemfor text categorization,” Machine Learning,
2000.
[39] C. Noam, “Three models for the description of language,”
IRE Transactions on Information Theory, p. 113124, 1956.
[40] ——, “On certain formal properties of grammars,” Information
and Control, p. 1959, 137-167.
[41] C. Noam, Schtzenberger, and M. P., “The algebraic theory
of context free languages,” Computer Programming and
Formal Languages. Amsterdam: North Holland, pp. 118–
161, 1963.
[42] “http://www.netdoctor.co.uk/,” Weblink, Online Medical
Solution.
[43] “http://www.doctoronline.nhs.uk/ver02/index.asp,”
Weblink, Online Doctor.
[44] “http://www.nzdoctor.co.nz/default.aspx,” Weblink, online
Medical Solution.
[45] R. M. Khalilur and E. Tsutomu, “Recurrent neural network
classifier for three layer conceptual network and performance
evaluation,” in 11th International Conference on
Computer and Information Technology, 2008, pp. 747–
752.
[46] D. C. and M. J., “Note on learning rate schedules for
stochastic optimization,” In Neural Information Processing
Systems, vol. 3, pp. 832–838, 1991.
Md. Khalilur Rhaman received his Dr.Eng. degree from
Kyushu Institute of Technology in 2009, Japan and B.Sc.
and M.Sc. degrees from Institute of Science and Technology,
National University, Bangladesh, in 1997 and 1998, respectively.
Currently he is an Assistant Professor of Department of Computer
Science and Engineering, BRAC University, Bangladesh.
In 2002, he joined Uttara University,Bangladesh, where he held
the position of Lecturer of Department of Computer Science
and Engineering. His research interests include Natural Language
Processing, Neural Network, Image Processing and Data
Mining.
Tsutomu Endo received the B.Eng., M.Eng., and Dr.Eng.
degrees from Kyushu University in 1972, 1974, and 1979
respectively.
Currently, he is a Professor of the Department of Artificial
Intelligence, Kyushu Institute of Technology. His research interests
include natural language understanding, computer vision
and multimodal interface. He is a member of the IPSJ, JSAI,
JSSST, RSJ and ANLP.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 49
An Enhanced Short Text Compression Scheme
for Smart Devices
Md. Rafiqul Islam
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh.
Email: dmri1978@yahoo.com
S. A. Ahsan Rajon
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh.
Email: ahsan.rajon@gmail.com
Abstract — Short Text Compression is a great concern
for data engineering and management. The rapid use of
small devices especially, mobile phones and wireless sensors
have turned short text compression into a demand-of-thetime.
In this paper, we propose an approach of compressing
short English text for smart devices. The prime objective of
this proposed technique is to establish a low-complexity
lossless compression scheme suitable for smart devices like
cellular phones and PDAs (Personal Digital Assistants)
having small memory and relatively low processing speed.
The main target is to compress short messages up to an
optimal level, which requires optimal space, consumes less
time and low overhead. Here a new static-statistical context
model has been proposed to obtain the compression. We use
character masking with space integration, syllable based
dictionary matching and static coding in hierarchical steps
to achieve low complexity lossless compression of short
English text for low-powered electronic devices. We also
propose an efficient probabilistic distribution based
content-ranking scheme for training the statistical model.
We analyze the performance of the proposed scheme as well
as the other similar existing schemes with respect to
compression ratio, computational complexity and
compression-decompression time. The analysis shows that,
the required number of operations for the proposed scheme
is less than that of other existing systems. The experimental
results of the implemented model give better compression
for small text files using optimum resources. The obtained
compression ratio indicates a satisfactory performance in
terms of compression parameters including better
compression ratio, lower compression and decompression
time with reduced memory requirements and lower
complexity. The compression time is also lower because of
computational simplicity. In overall analysis, the simplicity
of computational requirement encompasses the compression
effective and efficient.
Index Terms — Short Text Compression, Syllable,
Statistical Model, Text-ranking, Static Coding, Smart
Devices.
I. INTRODUCTION
Twenty-First century is the age of information and
communication technology. Science through its
Contact Author: S. A. Ahsan Rajon
E-Mail: ahsan.rajon@gmail.com
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.49-58
marvelous achievements has converted this world into
information and communication based global village.
The prime aspect of present technology is to ensure a
better way of communication throughout the world in a
more convenient, easy and cost-effective way. With the
aspects of cost, facility and reliability a new trend of
introducing small sized devices with some sorts of
computing and communicating power have established
its place in the arena of research. With the voyage of
introducing smart devices, the challenge of adorning
them with greater and effective use has come into
question. It is now a great concern to embed maximum
applications within these smart devices where it is an
extreme problem to provide a low-complex and lowmemory
consuming version for smart devices of some
prime necessary applications like data compression,
which generally requires large memory and greater
processing speed. Mobile communication that is a great
gift of modern technology introducing the era of digital
communication also suffers from the same limitation.
Though crossing the boundary of voice communication,
short messages communication has established its robust
place in the arena of digital communication, Short
Message Service (SMS) providers (usually
Telecommunication Companies) have a constraint that
each message should be not more than of 160 characters.
This constraint is really a great limitation for frequent
communication using SMS. In order to overcome this
limitation, compression of the short message is a well
policy. That is why; our aim is to make “short” messages
“shorter”, expressing “larger” feelings in “smaller”
expenses.
Here we introduce a scheme of compressing short
English text for smart devices like cellular phones and
wireless sensors having small memory and relatively low
processing speed communicating with lower bandwidth
i.e. channel capacity. We have employed a new statistical
model with a novel approach of integrating text ranking
or component categorization scheme for building the
model. Modified syllable based dictionary matching and
static coding is used to obtain the compression.
Moreover, we have employed a new theoretical concept

50 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
of choosing the multi-grams, which has facilitated us to
obtain mentionable compression ratio using a small
number of knowledgebase entries than other methods,
consuming less resource. Besides of experimental results
we have provided a comprehensive theoretical analysis
of compression ratio of proposed scheme with similar
existing scheme.
II. RELATED LITERATURE
Though a number of researches have been performed
regarding large-scale data compression, in the specific
field of short text compression, the number of available
research work is small. Business issue of mobile phone
service providers may be indicated as a reason behind the
unavailability of research material. The following
sections give a glimpse of the most recent research
developments on short text compression issues for small
devices.
A. Compression of Short Text on Embedded Systems
The recent literature regarding short text compression
titled “Compression of Short Text on Embedded
Systems” by Rein et al. [1] proposes a low-complexity
version of PPM (Prediction by Partial Matching). A hash
table technique with one-at-a-time hash function is
employed in this method to design the data structure for
data and context model. They use statistical models with
hash table lengths of 16384, 32768, 65536 and 131072
elements requiring two bytes for each element, which
result an allocation of 32, 64, 128 and 256 Kbytes of
RAM respectively. If this memory requirement may be
substantially decreased, we may achieve more efficient
compression and hence may make the scheme usable to
even very low-quality cellular phones. Another
concerned approach by Rein et al. is “Low Complexity
Compression of Short Message” [2] with Low Complex
and Power Efficient Text Compressor for Cellular and
Sensor Networks [3] are variations of [1].
B. Compression of Small Text Files Using Syllables
“Compression of Small Text Files Using Syllables”
proposed by Lansky et al. [4, 5] concerns on
compressing small text files using syllables. To
implement their concept they created database of
frequent syllables. Here, condition for adding syllable to
database is that, its frequency is greater than 1:65000. In
this scheme, the primary knowledge-base size is more
than 4000 entries initially. For low memory devices, it is
obviously difficult to afford this amount of storage as
well as to facilitate a well suited mechanism of searching;
which leads our proposed scheme to redefine the
knowledge-base span as short as possible and hence to
reduce the scope of loosely choosing the syllables or ngrams.
Moreover, in formation of the syllables, space is
not considered with any special concern. But, as in any
text document, it is a common assumption that, at least
20% of the total characters may be spaces, it may be a
© 2010 ACADEMY PUBLISHER
good idea to have specific consideration of syllable
involving spaces. In [4, 5], all the training syllable entries
are stored without any categorization. This often results
for coding redundancy, which can be handled by
integrating text ranking or component categorization
scheme with syllable selection.
C. Modified Greedy Sequential Grammar Transform
based Lossless data Compression
The model proposed by M. R. Islam et al. [6] uses the
advantages of greedy sequential grammar transform with
block sorting to compress data. However, this scheme is
highly expensive in terms of memory consumption and
thus not suitable for low memory devices.
D. Two-Level Directory Based Compression
Dictionary based text compression techniques are an
important and mostly adapted data compression schemes.
A dictionary based text preprocessing scheme titled
TWRT (Two-level Word Replacing Transformation) has
been proposed by P. Skibinski [7]. They use several
dictionaries and divide files into various kinds, which
improve the compression performance.
TWRT can use up to two dictionaries, which are
dynamically selected before the actual preprocessing
starts. For some types of data like programming
languages, references etc. first level dictionaries (small
dictionaries) are specified whereas second level
dictionaries (large dictionaries) are specific for natural
languages (e.g., English, Russian, French). While
concerned with any source text, if no appropriate first
level dictionary is found, then it is not used. Selection of
the second level dictionary is analogous. When TWRT
has selected only the one (the first or the second level)
dictionary, it works like WRT (Word Replacing
Transformation) [7]. If TWRT has selected both the
first and the second level dictionaries, then the
second level dictionary is appended after the first
level dictionary. That is, the dictionaries are
automatically merged. If the same word exists in the first
and the second level dictionaries, then the second
appearance of word is ignored to minimize length of
code-words. Only the names of the dictionaries are
written in the output file, so the decoder can use
the same combination of the dictionaries.
TWRT preprocesses the input file step by step with all
dictionaries and finally to choose the smallest output file.
Nevertheless, this idea is very slow. They propose to
read only the first f (e.g., 250) words from each of n
dictionaries (small and large) and create one joined
dictionary which is completely impossible to afford for
low-memory devices. If there are same words in different
dictionaries, then all occurrences of this words are
skipped which is too an extremely infeasible technique
for smart device platform-aware compression. The main
problem for TWRT is selection of the dictionaries before
preprocessing, which hampers the processing time for
concerned devices [3]. Moreover the dictionary length is

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 51
too huge to be used for small memory devices. For small
text compression it is also not feasible and to some extent
efficient to have field-specific dictionary. Above all,
being the code words are determined on the fly, it is great
doubt to cope with the memory and energy constraints as
well as time consumptions of the low-memory devices.
E. Other Schemes
H. Kruse and A. Mukherjee [8, 9] proposed a
dictionary based compression scheme named star
encoding. According to this scheme, words are replaced
with sequence of * symbol accompanied with reference
to an external dictionary. The dictionary is arranged
according to the length of words and is known to both
sender and receiver. Proper sub-dictionary is selected by
the length of the sequence of * symbols. Length Index
Preserving Transformation (LIPT) is a variation of the
star encoding by the same authors. This algorithm
improves the PPM, BWCA and LZ based compression
schemes [9]. Another related literature known StarNT
works with ternary search tree and is faster than the
previous. The first 312 words of the dictionary are the
most frequently used words of the English language. The
remaining part of the dictionary is filled up by words
sorted by their lengths first and then by their frequencies.
This scheme also does not take the use of substring
weighting approach. Moreover, the scheme requires that,
the dictionary should be transmitted first in order to set
up the knowledge-base. It is completely in-feasible to be
used for compressing small texts for low-powered and
small memory devices.
Prediction by partial matching (PPM) is a major
lossless data compression scheme, where each symbol is
coded by taking account of the previous symbols [9]. A
context model is employed that gives statistical
information about the symbol with its context. In order to
signal the decoder on the context, specific symbols are
used by the encoder. The model order in PPM is a vital
parameter of compression performance. However, PPM
is computationally more complex and the overhead too is
greater [1, 9, 12].
In [1] the compression starts for text sequences larger
than 75 Bytes, and in [10] the starting point is 50 Bytes.
If it is possible to make the lower threshold value into
less than ten characters, the compression may really be a
“very small text file” supported one that may place a new
milestone in very small text file compression ensuring
“short text gets shortest”. Our prime aim is to design
such an effective and efficient “very short text
compression” scheme.
III. SHORT TEXT COMPRESSION FOR SMART DEVICES
The prime concern of this paper is to implement a
lossless compression of short English text for smart
devices in a low complexity scheme. The idea behind this
is to make it possible to communicate more efficiently by
© 2010 ACADEMY PUBLISHER
making utilization of minimum required bandwidth of
wireless and sensor networks for small and embedded
devices. More precisely saying, the main concern is to
develop a low-complexity compression technique
suitable for low-power consuming smart devices with
small-memory; especially for wireless sensors and
mobile phones. The proposed scheme is concerned with
two parts. The first one consists of training the statistical
model and the second provides a compressiondecompression
technique. Specifically, in analyzing step
we count the frequency of represent-able ASCII
characters. Then, the proposed scheme proceeds by
identifying the syllables of length two, three, four and
five. We consider as a distinct vowel and
include this while counting the frequency of syllables,
that is, searches for either at the beginning or
ending of syllable. In the step of boosting the statistical
model, the substrings (which were not grabbed in the
phase concerning syllables) with length two to five are
considered and the frequency of each are counted. In the
second step, we employ the provided text-ranking
scheme for each entry and calculate the entry index. For
entries with same index, we simply sort them. In the
phase of choosing entries, emphasis is give on
probability distribution based text ranking, which is
computed with the help of its neighbor characteristics.
As in most cases, it is unusual to have frequent match
of substrings with length more than four characters,
maximum of five (extra one for ) levels has been
considered to train the statistical model. In each level,
multiple entries with same weightage are simply sorted
over. Resultant entries are assigned with non-conflicting
binary stream. When we are to compress any text, the
input text is successively compared with the statistical
model and for any match, the binary stream is returned as
output and for mismatch in any level, the level below it is
forwarded.
The compression and decompression are expected to
be performed in the following manner.
A. Compression Process
In the first step, we plan to employ Modified Multigrams
or syllable matching proposed by Lansky et al. [4,
5]. A static dictionary based compression scheme uses
approximately the same concept as that of character
masking. It reads input data and searches for symbols or
groups of symbols that are previously coded or indexed
in the dictionary. If a match is found, a pointer or index
into the dictionary can be output instead of the code for
the symbol. Compression occurs if the pointers or index
requires lower space (in terms of memory usage) than the
corresponding fragments [1, 12, 15]. Though it is the
basic idea behind multi-grams, we use it in a slightly
modified fashion. The Knowledgebase for the multigrams
is constructed with the help of the statistics
obtained by analyzing the corpuses.

52 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
In the second step of compression process, static
coding has been used. However, here the modification is
made in such a way that, in spite of calculating on-stage
codes predefined codes are employed in order to reduce
the space and time complexity. The codes for dictionary
entries are also predefined. The total message is encoded
by a comparatively small number of bits and hence, we
get a compressed outcome. The prime fact of static
coding is that, in case of dynamic coding we are to
calculate the frequencies in an elegant manner.
Moreover, as the dynamic codes are determined and the
lengths are varied through the frequencies of the symbols
it is a must to submit the codes along with the
compressed data. This is a limitation of dynamic coding
if the total arena of compression is low spanned and the
total number of symbols is not huge. Compression of
short message also suffers in this context while using
other similar semi-dynamic coding. Consequently, it
motives to move towards static coding scheme to obtain
compression of short messages. For this concerned study,
we have analyzed the corpora- Bib, Book1, Book2,
News, Paper1, Paper2, Paper4 and Paper5 from Calgary
Corpus and Canterbury Corpus [16]. We have also
analyzed 124 collections of small text for the same. A
detail overview of the texts is presented in [10, 15].
B. Decompression Process
The decompression process is performed through the
following steps:
Step 1: Grab the bit representation of the message,
Step 2: Identify the character representation,
Step 3: Display the decoded message.
As all the symbols are to be coded in such a fashion that,
by looking ahead several symbols (Typically the
maximum length of the code) we can distinguish each
character (with the attribute of Static Coding). In step 1,
the bit representation of the modified message is
performed. It is simply analyzing the bitmaps. The
second step involves recognition of each separate bitpatterns
and indication of the characters or symbols
indicated by each bit pattern. This recognition is
performed on the basis of the information from fixed
encoding table used at the time of encoding. The final
step involves simply representing i.e. display of the
characters recognized through decoding the received
encoded message.
IV. THEORETICAL ASPECT OF TRAINING THE PROPOSED
STATISTICAL MODEL
The proposed scheme achieves better compression
ratio with relatively low complexity by means of
computational simplicity and effective expert model,
which is used to train the statistical context. Firstly, the
prime modification is performed in the syllable selection
section of [4, 5] proposed by Lansky et al. In their
papers, only syllables were considered by defining
maximal subsequence of vowels including pseudo-vowel
© 2010 ACADEMY PUBLISHER
‘y’. Here our proposal is to consider as a
Prime Syllable. Using as a prime syllable may
dramatically reduce the total number of characters
needed to represent the message through sophisticated
encoding. Secondly, the paper [4, 5] does not clearly
express the criteria of choosing the training syllables for
the model. The term used in [4, 5] to define the criteria of
choosing syllables was simply “frequentness”. However,
we employ a new theoretical perspective on “Text
Ranking or Component Ranking” method to choose the
syllables for training our proposed model. We, in the first
step count the frequency of each vowel from the standard
Text Calgary Corpora. In the second step, we count the
number of vowels having space either at (i - 1) or (i +
1) position with itself at position (i). The reason of
adding the with-space vowels is simply to increase the
weightage of the corresponding vowels. These entries
having vowels are also inserted in the knowledgebase as
separate entity. In this step, we also count the statistics of
consonants in order to build the dictionary or multigrams
entries. In the step of extending primary
knowledgebase, proposed in papers [4, 5] the criteria was
that, the frequency would be 1:65000, which results a
knowledgebase size of more than 4000 entries initially.
For low memory devices, it is obviously difficult to
afford this amount of storage as well as to facilitate a
well suited mechanism of searching, which leads our
proposed scheme to redefine the knowledge base span as
short as possible and hence to reduce the scope of loosely
choosing the multi-grams. In the phase of choosing
multi-grams, we give emphasis on Probability
Distribution, which is computed with the help of its
neighbor characteristics.
A new text weighting or component ranking scheme
has been employed to select the multi-grams that
facilitates us to efficiently construct the knowledgebase
[11, 15]. This developed novel text weighting scheme is
employed with an aim to get the knowledgebase. The
proposed text weighting or component ranking is
obtained through the following equation:
Suppose we choose a multi-grams D consisting of
characters C1 C2 C3… Cn of length n. Thus,
∂(D
) =
C ,C ,C ,......... ,C ,
1 2 3 n
n
∑ ( ∂(D
) -1)
-∂
(D
) + λ
C C ,C ,C ,......... ,C , C ,C ,C ,......... ,C
i = 1
i 1 2 3 n−1
1 2 3 n
for i > 1 and,
∂ ( DC
) = λ(Ci ) for i = 1, where λ(Ci ) is the frequency
i
of the character Ci with assumption that each character of
the alphabet must exist in the training data. And the value
of ∂(Dφ) indicates the multi-grams index of character
Dφ.. Here, we refer the multi-gram index obtained from
∂ (D C ,C ,C ,......... ,C , ) as the resultant text-weightage
1 2 3
n
for the text C1 C2 C3… Cn .

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 53
For example, we want to have the probability distribution
of multi-grams “and”.
The steps will be:
Let ∂ ( a ) = x1 ,
∂ ( n ) = x2 ,
∂( d ) = x3 ,
∂( an ) = x4 and,
λ(C and ) = Frequency of ‘and’ in the training document = x5.
Therefore, ∂(and) = ( ∂( a ) -1) + ( ∂( n ) -1) + ( ∂ (d ) -1)
— ∂ (an) + λ (C and ) = x 1+ x 2 + x 3 — x 4 + x 5
Static coding is defined as a set S = {B1, B2, B3 ….., Bn-1 ,
Bn } of binary streams, where each element of the set B1,
B2, B3 ….., Bn-1 , Bn are uniquely identifiable providing that,
B1, B2, B3 ….., Bn-1 ,Bn are not necessarily to be of equal bit
lengths. As all the elements of the sets supports our three
basic points of concern- identifiability, uniqueness and
variability of any encoding scheme, we are interested to
use static coding. This theoretical aspect let us do not
consider updating the model throughout the compression,
because, firstly, updating a larger knowledgebase using a
very short data is not computationally affordable and
effective, secondly, a larger portion of the source code
may be saved if the update is avoided resulting faster
execution [1, 3, 11] and finally, if the training statistical
data are not permitted to be updated as well as to be
expanded, a constant and consistent memory
optimization for the overall compression process may
ensured. That is, there is no possibility to expand the
knowledgebase arbitrarily and thus there is no risk of
arising the overloaded memory problem or out-ofmemory
problem. As the knowledgebase is not updated,
the use of static coding is also perfect for the same.
V. PERFORMANCE ANALYSIS
Though it is a general idea that compression and
decompression time should have an inter-relation, the
proposed scheme demonstrated a little exception. The
points behind that may be summarized through the
following discussions.
A. Performance Analysis of Compression Process
Let the total number of training entries for the
statistical model be Ng , where Ng is a non-negative
integer and the maximum level for statistical modeling is
L. The first level of the statistical model must contain the
single characters, where the total number of distinct
character is l1. For levels 1, 2, 3, …., n-1, n the total
number of distinct multi-gram entries are l1, l2, l3, ……,
ln-1, ln respectively.
When any text is to be compressed, it is hierarchically
compared with each level of statistical model starting
from the highest order. If there is any match, the
corresponding static coding for multi-gram entry is
assigned for the text. If the multi-gram entry is not found
throughout the level, it is forwarded to the next level.
This assignment uses efficient searching procedures. Let
the code m is found at the i-th level with offset k
© 2010 ACADEMY PUBLISHER
resulting a search cost of Sm ( l j ) + km , where km < li
and, j=L, L-1,……, i-1 with respect to search space.
Here j limits from L to (i-1) instead of i because, as we
find the code in somewhere of i-th level not requiring to
search the whole element-space of the i-th level, rather
searching through an offset value k for i-th level, the
overall search-space is L to (i-1). That is why, for the
above consequences, the total searching appears: search
overhead for (i-1) number of levels with additional
search overhead of k elements. Here the term “search
overhead” stands for the search space complexity as well
as other related computational requirements like time and
power consumptions. When the code matches, it is
placed in output stream as character representation. This
step requires padding the bit-stream and then conversion
into character stream. Assume that, the process of overall
conversion for each successful entry occurs with the
overhead B. That is, for any multi-gram matching, the
required overhead is,
C
∑ −
1 =
i 1
(S 1(
l
j = L
j )) + k 1 + B 1
Similarly,
C n
=
∑ −
C 2 =
i 1
(S 2 ( l
j = L
j )) + k 2 + B 2
∑ − i 1
(S n ( l j )) + k n
j = L
+ B n
, and
In such a way if n multi-grams are identified and then
encoded, the required resultant number of operations in
compression process is:
n n i −1
n
n
T = C = (S ( l )) + k + B (1)
∑
y = 1
y
∑
∑ y j ∑ y ∑
y=
1 j = L
y=
1 y=
1
B. Performance Analysis of Decompression Process
For the decompression process, the text to be
decompressed is converted into binary stream. If the
largest code is of length cmax and the smallest code is of
length cmin then the decompression process will start the
searching with the cmax number of bits and search through
the codes up to cmin bits by reducing one bit in each step
for unsuccessful match. It is necessary to mention that,
the codes with same bit length do not essentially
comprise any specific level. So, to reveal the character
representation for each entry d if a switch of h levels are
required, where cmin ≤ h ≤ cmax where the maximum level
is p, with the matching offset for corresponding level kd,
and the assignment of the code with character
representation for each successful match requires an
overhead of Bd, then the overall requirement for
comparing through the each level settings results (=
overhead of searching through level + overhead for
searching through offset + overhead of representation).
For detecting first character the overhead is,
y

54 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
h
E 1 = ∑ ( S ′ 1
q = p
( l ′ ))
q
+
k ′
1
+
B ′
Similarly, for detecting the second character, the levelwise
overhead will be:
h
E 2 = ∑ ( S ′ 2 ( l q′
)) + k 2′
+ B 2′
q = p
Hence,
h
∑
q = p
E = ( S ′ ( l ′ )) + k ′ + B ′ (2)
n
n
q
Here, p = maximum level, and S / is a function that
denotes the search-overhead for searching in element
space provided as parameter of the function and h =
minimum level. The computation progresses through p,
p-1, p-2,……, h+2, h+1, h . Here the subscript n is used
to denote the level-wise overhead for detecting one
character representation with respect to level.
In order to detect a single multi-grams f, the total searchoverhead
with respect to search space for level-wise
calculation is,
h
( S ′ ( l ′ )) + k ′ because, we are to
∑=
q
p
f
q
start with maximum level p and then proceed
decreasingly towards the downward levels h (as
explained above). If S / is the search-overhead function,
then searching from level p to h will result
h
∑
q = p
( S
′
f
(
l ′
q
))
f
n
where f is the multi-grams,
which is being revealed. For the matching level, as only a
partial number of elements are to be searched, the offset
k is used to denote the offset.
After checking through the levels, the procedure follows
searching through the bit-wise statistics for any
unsuccessful match in level-wise statistics. If there are a
total of u bit-phases, we are to perform searching through
the search-space consisting of starting from the
maximum bit phase to the minimum bit phase in
descending order. Because of any unsuccessful match in
any bit-phase, a bit switch is performed and level wise
calculation for that level is forwarded. That is, an
g
overhead of ∑=
b
d
(E is incurred for each level-wise
b )
analysis. Consequently, the overhead of unit step will be,
g
∑
b = d
C ′ = (E ) where d and g are maximum and
1
b
minimum bit phases respectively and d ≥ g.
Substituting the value of Eb, we get,
C ′
1
=
g
∑
h
∑ ( S 1,b ′ ( l q′
)) +
g
∑ k 1,b ′ +
g
∑
b = d q = p
b = d
b = d
Similarly, we get,
© 2010 ACADEMY PUBLISHER
1
.
n
.
B ′
1,b
C 2′
=
g
∑
h
∑ ( S 2′
, b ( l q′
)) +
g
∑ k 2′
,b +
g
∑ B 2′
,b
b = d q = p
b = d
b = d
And,
C n′
=
g
∑
h
∑ ( S n′
, b ( l q′
)) +
g
∑ k n,b ′ +
g
∑
b = d q = p
b = d
b = d
B ′
Here, we use the subscript 1 with k and B in order to
denote that, the calculations are for detecting unit code
only where the calculation is performed starting from d
to g in decreasing order, that is, in the order of d, (d-1),
(d-2), …… , (g+2), (g+1), g. If we are to reveal n number
of codes, then the total overhead becomes:
T ′ =
n
∑
y = 1
C ′
y
.
As for each bit wise overhead calculation, level-wise
calculations would must be included; we may omit the
subscript notation for search overhead function for
simplicity,
n g h
n g n g
T′
= ∑ ∑ ∑( S′
y b(
lq′
)) + ∑∑k′
y,b+
∑ ∑B′
(3)
,
y,b
y = 1 b = d q = p
y = 1b
= d y = 1 b = d
TABLE 1:
COMPLEXITIES OF COMPRESSION AND DECOMPRESSION PROCESSES
T′
=
Compression Complexity of Proposed Scheme
T
=
∑ ∑
− n i 1
(
y=
1 j = L
(S ( l )) + k
y
Decompression Complexity of Proposed Scheme
n
∑
(
g
∑
h
j
∑ ( S′
y,b(
lq′
)) + ∑ k′
y,b + ∑ B′
y = 1 b = d q = p
b = d b = d
C. Performance Analysis with respect to Compression
Ratio
Compression ratio may be defined as the ratio of total
number of bits to represent the compressed information
and the total number of bits in original text.
Let the training knowledgebase contains n items. The
items may contain any of the total s symbols of the
source language. Again, the knowledgebase items vary
from one to c characters. For traditional encoding,
required overhead (i.e. total number of required bits) to
represent each character of knowledgebase entries is
log(s) in average. Therefore, if we categorize the
knowledgebase items into kn categories, (where the
categorization aspect is total number of characters in
each knowledgebase entry) and there are e1, e2, e3, …, en
elements in k1, k2, k3, …, kn categories respectively, then
g
y
+
B
y
)
g
y,b
n,b
)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 55
we may easily calculate the overhead for coding the total
knowledgebase.
For category k1, we need a total of log(s) * k1 bits is
needed to code each knowledgebase items.
As there are a total of e1 elements, total bit requirement
for coding all the elements of category k1 is,
log(s) * k1 * e1.
Here, log(s) , k1 and e1 are non-negative integer values.
That is,
Ok1 = log(s) * k1 * e1
Ok1 = k1 e1 log(s)
Where Ok1 indicates the total bit requirement for
coding all the elements of category k1 .
For category k2, we need a total of log(s) * k2 bits to
code each knowledgebase entries.
As there are e2 elements, total bit requirement for
coding all the elements of category k2 is, log(s) * k2 * e2.
Here, log(s) , k2 and e2 too are non-negative integer
values.
That is, Ok2 = log(s) * k2 * e2
Ok2 = k2 e2 log(s)
Where Ok2 indicates the total bit requirement for
coding all the elements of category k2 .
Similarly, For category kn, we can write ,
Okn = log(s) * kn * en
Okn = kn en log(s)
Where Okn indicates the total bit requirement for
coding all the elements of category k2 .
Symbolically, the total bit requirement for
representation of the knowledgebase entries is:
Ot = Ok1 + Ok2 + Ok3 + ... ... ... + Okn
= log(s)* k1 * e1 + log(s)* k2 * e2 + . . . + log(s)* kn * en
= k1 e1 log(s) + k2 e2 log(s) + . . . + kn en log(s)
= log(s) ( k1 e1 + k2 e2 + . . . + kn en )
n
That is , O = log(s)
t
∑ ( k . e
i i
i = 1
Since, 1 ≤ ei ≤ log(n) , and s being the total number of
symbol unit in source language, in worst case,
n
O =
t
log(s) . n . log n ∑ (
i = 1
k
i
)
)
(4)
(5)
Now let us consider our proposed scheme, where
components of knowledgebase entries are grouped into
© 2010 ACADEMY PUBLISHER
several levels and each entry in the level is chosen by
means of effective statistical entity. Because of using
hierarchical statistics, any element of certain level
possess greater probability of occurrence than the
element placed at any position below that. However the
levels are placed in descending order to facilitate that,
higher-gram texts or knowledgebase components are
coded in advance to any lower-gram texts irrespective of
probability distribution [11]. Here, it is noteworthy that,
though the probability distribution or statistical context
was not taken into consideration to organize the levels,
the total structure resulted an automatic statistical
distribution, because, the text-ranking scheme used to
build the knowledgebase followed hierarchical steps that
inherently inferred statistics from lower groups.
Consequently, whenever we are formulating the
statistical entries, any subgroup from its upper group will
have lower values and specific coding schemes may be
employed taking this criteria into consideration.
As we are using static coding in order to encode the
total knowledgebase and the knowledgebase is
hierarchically grouped, the resultant outcome leads our
proposed scheme into a low-bit consuming one. In our
proposed scheme, symbolically, any knowledgebase
entry varies from 1 to r characters. That is, the levels are
r, (r - 1), (r - 2), ………, 2, 1. Formation of levels start
from single characters and proceed incrementally. Any
entry in the knowledgebase containing any substring
from its successor level will have greater multi-gram
index because of being inferred from the previously
encountered entry. This aspect results in a sustainable
knowledgebase architecture if we sort the knowledgebase
in any order considering multi-gram index as the primary
criteria. Since we are calculating the multi-gram-index or
multi-gram-weight through a comprehensive text ranking
scheme we may consider the underlying text elements as
a single unit. Even though, we are considering the overall
knowledgebase a single unit, it has been clarified earlier
that, the architecture of the knowledgebase will provide
us elementary grouping facilities. This coherence model
provides us the opportunity to use static coding. If we
have a total of m entries in the knowledgebase, coding of
those values using binary stream may vary from 1 to
log(m) bits. Let it be y in an average, where 1 ≤ y ≤
log(m).
Again, as we are using multi-grams as single
component, if the multi-gram consists of even k
characters where 1 ≤ k ≤ r, it will be turned into a single
one. Consequently, the average requirement may be
specified as
O = y + y + y + .......... ... + y
p 1 2 3
n

56 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
= ∑
=
n
O y
p i
i 1
where, 1≤ yi ≤ log(m) for 1≤ i ≤ n .
For worst case O = n log(n)
p
where y = log(n) for 1 ≤i
≤ n .
i
(6)
(7)
Even if, the total number of components n is same for
both the cases, since 1 ≤ y ≤ log(n) (because here m=n,
both indicating the total number of elements) and 1 ≤ ei ≤
log(n) multiplying any value (category index k) with ei
will definitely result much more than that of yi.
Consequently, we may deduce that,
Op ≤ Ot
Here, n is total number of elements in the
knowledgebase and y (1 ≤ y ≤ log(m)) refers the total
bits needed to code component i. It is inherently clear
that, Op ≤ Ot . That is, the total number of bits needed to
encode the same source symbols using our proposed
scheme is less than that of the traditional schemes. As,
compression ratio (R) is the ratio of compressed bit and
source bits, we get that,
and,
Rt = Ot / nt
Rp = Op / nt
As, Op ≤ Ot we get that,
Rp ≤ Rt.
Even for the worst case, we get that,
O
t
= log(s)
. n . log
n
n ∑
i = 1
(
n
O = log(s)
t
. O p ∑ (
i = 1
k )
i
( 8 )
k
i
Hence we may deduce that, even for worst case,
Op ≤ Ot
The lower the value of compression ratio, the better
the compression. Consequently ,we may conclude that,
the compression efficiency of proposed scheme is better
than that of traditional dictionary based compression
schemes.
Our proposed scheme requires less space because we
have constrained the growth of the knowledgebase. The
facility to use a couple of levels will ensure greater
flexibility in memory management. As, the search space
is minimized, the computational complexity will also be
reduced. It is of no doubt that, lower computational
complexity will ensure faster performance.
© 2010 ACADEMY PUBLISHER
)
VI. EXPERIMENTAL RESULTS AND DISCUSSIONS
The performance evaluation is performed on the basis
of the file “book1”, “book2”, “paper4”, “paper5” and
“paper6” from “Calgary Corpus”. As the prime aspect of
our proposed Compression Scheme is not to compress
huge amount of text rather to compress texts with limited
size affordable by the mobile devices i.e. embedded
systems, we took blocks of texts less than five hundred
characters chosen randomly from those files ignoring
binary files and other non-text files and performed the
efficiency evaluation.
The most recent study involving compression of text
data are:
1. “Low Complex and Power Efficient Text
Compressor for Cellular and Sensor Networks” (Mode 1)
by Rein et al. [1, 2, 3] and,
2. “A modification Of Greedy Sequential Grammar
Transform based data Compression" by Islam et al [4]
We denote the above two methods as DCM-1 and
DCM-2 respectively, where DCM stands for Data
Compression Method.
The simulation was performed in a 2.0 GHz Personal
Computer with 112 MB of RAM with the object oriented
programming language Java. The average compression
ratio for three random execution results for different size
of blocks of text is as follows
Corpus Number of
Characters
Considered
TABLE 2
COMPRESSION RATIO
Compressi
on Ratio
(%) for
DCM-1
Compressi
on Ratio
(%) for
DCM-2
paper 4 108 44.01 44.03 42.94
paper 5 061 44.98 45.51 44.02
paper 6 032 45.22 45.96 44.30
book 1 191 46.84 48.11 45.97
book 2 104 43.95 46.69 45.27
Compressi
on Ratio
(%) for
proposed
scheme
The compression ratio is a metric to describe how
many compressed units are required to describe one unit
of data. The lower the presented value shows better
compression. A general observation is that higher modes
lead to better compression ratios even if the difference
with higher orders becomes smaller.
The Performance of any dictionary based or
knowledge-inferred compression scheme greatly varies
with the architecture of dictionary construction and
knowledge inference. When the test bed is considered as

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 57
the same of the knowledgebase inferring one or
dictionary forming one, the performance gain will be
higher because of the greater match with the dictionary
entries or knowledgebase components. Our proposed
scheme has worse performance because, the training
scheme that we have provided makes an effective use of
text ranking scheme for several corpora and choice of the
knowledgebase entries are unbiased towards any specific
corpus. It is the novelty of our approach which ensures a
greater distribution of ranking and makes the developed
scheme uniformly usable for text data compression. But,
other schemes which are trained with specific files of any
corpora, there are greater matching with the
knowledgebase and the hence performance evaluation
demonstrates better performance for that corpus or that
specific file than that of our proposed scheme.
The prime achievement of our proposed scheme is, it is
even capable to compress source texts consisting with an
average of only five characters. Though the compression
ratio is deteriorated for that case, it is an evolutionary
step for small text compression. This achievement may
be greatly helpful for compression of sensor network
concerned data and even in asynchronous data
transmission management for web applications having
the glimpse of real time computation.
Besides of the evaluation scheme that has been
presented earlier in this section, we analyze the
performance of the proposed scheme in terms of
compression ratio with respect to the text presented in
[3].
TITLE TEXT
The international system of units consists of a
set of the units together with a set of the
prefixes. The units of the SI can be divided into
two subsets. There are seven base units. Each
TEST_A
of these base units are dimensionally
independent. From the seven base units all
other units are derived.
TEST_B
TEST_C
However, few believed that SMS would be
used as the means of sending text messages
from a mobile to the another. One factor in the
takeup of the SMS was that operators were
slow to eliminate billing fraud which was
possibly by changing SMSC setting on
individual handsets to the SMSC's of other
operators.
Alice is a fictional character in the books of the
Alice’s adventures in the Wonderland and its
sequel through the Looking-Glass, which were
written by Charles Dodgson under the pen
name Lewis Caroll. The character is based on
Alice Liddel, a child friend of Dodgson's. The
pictures, however do not depict Alice Liddel,
as the illustrator never met her. She is seen as a
logical girl, sometimes being pedantic,
especially with Humpty Dumpty in the second
book.
© 2010 ACADEMY PUBLISHER
The performance of proposed compression scheme (in
terms of compression ratio) for the above texts in
comparison with [3] (indicated as DCM-1) is given
below.
Compression Ratio in percentage
(Lower Values indicate better performance)
54
52
50
48
46
44
42
40
TEST_A TEST_B TEST_C
DCM- 1 PROPOSED SCHEME
Figure 1. Performance Comparison for example text.
From the figure we find that the compression ratio is
lower for all the cases for the text presented in [11].
It is necessary to mention here that In order to
implement our proposed scheme no additional hardware
is necessary. Rather it is possible to use even in any lowpowered
and low-memory devices. Basically the aspect
of ensuring an affordable text compression scheme for a
greater variety of smart devices we emphasis on static
coding in lieu of on-the-fly coding. Besides the results
presented in this section, detailed theoretical analysis on
the space and time requirements, computational
complexity i.e. overall computational overhead is already
presented in section V.
VII. CONCLUSION AND RECOMMENDATION
We have presented an effective and efficient approach
of compressing short English text message for lowpowered
embedded devices. Here modified syllable
based dictionary matching and static coding have been
employed to obtain the compression. Moreover, a new
theoretical concept of choosing the multi-grams is used,
which has facilitated us to obtain mentionable

58 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
compression ratio using a small number of
knowledgebase entries than other methods consuming
less resource. The overall strategy of computational
simplicity has also ensured the reduced time complexity
for the proposed compression and decompression
process. The main aspect of our proposed scheme resides
in the text ranking based knowledge-base construction
with space integration that initiates a new arena of text
compression methodology. A consistent and relevant
mathematical analysis of the overall performance also
establishes a strong technical basis of the proposed
scheme. Moreover, the prime achievement is in the scale
of starting threshold of text compression; that we have
reduced to less than five characters. With limited
knowledge-base size, the achieved compression is of no
doubt efficient and effective. As the knowledge base is
not accepted to be grown through the continuous
applications, we may keep out the low-memory system
from the risk of expanding its knowledgebase crossing
optimal memory size and thus, the applicability of the
proposed system even in any very low memory devices is
ensured.
REFERENCES
[1] Stephan Rein, Clemens Guhmann, Frank H. P. Fitzek:
“Compression of Short Text on Embedded Systems”,
Journal of Computers : Volume 1, No: 06, September
2006.
[2] S. Rein, C. Guhmann, and F. Fitzek , “Low Complexity
Compression of Short Messages,” Proceedings of the
IEEE Data Compression Conference (DCC’06), March
2006, pp.123–132.
[3] Stephan. Rein, F. Fitzek, M. P. G. Perucci, T. Schneider
and C. Guhmann, “Low Complex and Power Efficient
Text Compressor for Cellular And Sensor Networks”, In
15 th IST Mobile and Wireless Communication Summit,
June 2006.
[4] J. Lansky , M. Zemlicka , “Compression Of Small Text
Files Using Syllables”. Proceedings of Data Compression
Conference (DCC’06) , Los Alamitos, CA, USA, 2006.
[5] J. Lansky and M. Zemlicka. “Text Compression:
Syllables”. In Annual International Workshop on
DAtabases, TExts, Specifications and Objects (DATESO),
Volume 129 of CEUR Workshop Proceedings, pp. 32–45.
2005.. CEUR-WS.
[6] Md. Rafiqul Islam, Sajib Kumar Saha, Mrinal Kanti
Baowaly. “A modification of Greedy Sequential Grammar
Transform based Universal Lossless data Compression”.
Proceedings of 9th International Conference on Computer
and Information Technology (ICCIT 06), 28-30 December,
2006, Dhaka, Bangladesh.
[7] Przemysław Skibiński, “Two-Level Dictionary Based
Compression”, Proceedings of the IEEE Data
Compression Conference (DCC’05), page 481.
[8] F. Awan and A. Mukherjee, "LIPT: A Lossless Text
Transform to improve compression", Proceedings of
International Conference on Information and Theory:
Coding and Computing, IEEE Computer Society, Las
Vegas Nevada, 2001.
[9] H. Kruse and A. Mukherjee, “Preprocessing Text to
Improve Compression Ratios”, Proceedings of Data
© 2010 ACADEMY PUBLISHER
Compression Conference, IEEE Computer Society,
Snowbird Utah, 1998, pp. 556.
[10] S. A. Ahsan Rajon, “A Study on Text Corpora for
Evaluating Data Compression Schemes: Summary of
Findings and Recommendations”, Research Report,
Computer Science and Engineering Discipline, Khulna
University, Khulna, Bangladesh, December, 2008.
[11] Md. Rafiqul Islam, S. A. Ahsan Rajon and Anonda Podder,
“Short Text Compression for Smart Devices", Proceedings
of 11th International Conference on Computer and
Information Technology (ICCIT 2008), 25-27 December,
2008, Khulna, Bangladesh, pp. 453-558.
[12] S. Rein and C. Guhmann, “Arithmetic Coding–A Short
Tutorial”, Wavelet Application Group, Technical Report,
April 2005.
[13] David Hertz: “Secure Text Communication for the Tiger
XS”. Master of Science Thesis, Department of Electrical
Engineering, Linköpings University, Linköping, Sweden.
[14] S. A. Ahsan Rajon and Anonda Podder, “Lossless
Compression of Short English Text for Low-Powered
Devices”- Undergraduate thesis, CSE Discipline, Khulna
University, Khulna, Bangladesh, March, 2008.
[15] Md. Rafiqul Islam, S. A. Ahsan Rajon, Anonda Podder,
“Lossless Compression of Short English Text for Low-
Powered Deices”, in the proceedings of International
Conference on Data Engineering and Management
(ICDEM 2008) Tiruchirappalli, Tamil Nadu, India.
February 9, 2008.
[16] Ross Arnold, and Tim Bell, “A Corpus for the evaluation
pf lossless compression algorithms”, Data Compression
Conference, pp. 201-210, IEEE Computer Society Press,
1997.
Md. Rafiqul Islam obtained Master of Science (M. S.) in
Engineering (Computers) from Azerbaijan Polytechnic Institute
(Azerbaijan Technical University at present) in 1987 and Ph.D.
in Computer Science from Universiti Teknologi Malaysia
(UTM) in 1999. His research areas include design and analysis
of algorithms and Information Security. Dr. Islam has got a
number of papers related to these areas published in national
and international journals as well as in referred conference
proceedings.
He is currently working as the Head of Computer Science
and Engineering Discipline, Khulna University, Bangladesh.
S. A. Ahsan Rajon received his B.Sc. Engineering degree
from Computer Science and Engineering Discipline, Khulna
University, Khulna in April 2008. He is now a postgraduate
student of Business Administration Discipline under
Management and Business Administration School of same
university. Rajon is also working as an Adjunct Faculty of
Computer Science and Engineering Discipline, Khulna
University, Bangladesh. Rajon has made several publications in
International conferences. His research interests include data
engineering and management, electronic commerce and
ubiquitous computing. Currently he is working on robotics.
He is a member of Institute of Engineers, Bangladesh (IEB).

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 59
Design and Analysis of an Effective Corpus for
Evaluation of Bengali Text Compression
Schemes
Md. Rafiqul Islam
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh.
Email: dmri1978@yahoo.com
S. A. Ahsan Rajon
Computer Science and Engineering Discipline, Khulna University, Khulna, Bangladesh.
Email: ahsan.rajon@gmail.com
Abstract — In this paper, we propose an effective
platform for evaluation of Bengali text compression
schemes. A novel scheme for construction of Bengali text
compression corpus has also been incorporated in this
paper. A methodical study on the formulation-approaches of
text corpus for data compression and present an effective
corpus named Ekushe-Khul for evaluating the Bengali text
compression schemes has also been presented in this paper.
To design the Bengali text compression corpus, Type to
Token Ratio has been considered as the selection criteria
with a number of secondary considerations. This paper also
presents a mathematical analysis on data compression
performance with structural aspects of corpora. A
comprehensive analysis on the evolving criteria of text
compression corpora with related issues in designing
dictionary based compression are extensively incorporated
here. The proposed corpus is effective for evaluating
compression efficiency of small and middle sized Bengali
text files.
Index Terms — Corpus, Bengali Text, Bengali Text
Compression, Dictionary Coding, Data Management,
Evaluation Platform, Compression Efficiency, Type to Token
Ratio (TTR).
I. INTRODUCTION
With a number of languages in the world, Bengali is
the only language, which has been established at a cost of
lives. However, in the establishment of Bengali as a
glittering candidate in the field of research the number of
steps or achievements is not so mention-worthy.
Enhancing Data Management techniques for Bengali text
has not yet got any robust base. Text Compression, which
is one of the important aspects of elementary data
This paper is an enhancement of the paper “On the Design of an
Effective Corpus for Evaluation of Bengali Text Compression
Schemes”, by Md. Rafiqul Islam, and S. A. Ahsan Rajon, which was
published in the proceedings of International Conference on Computer
and Information Technology (ICCIT’ 2008), 25-27 December 2008,
Khulna, Bangladesh.
Contact Author: S. A. Ahsan Rajon
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.59-68
management and text manipulation schemes, is still now
mostly based on generalized universal compression
techniques. The remaining sophisticated Bengali text
compression schemes also suffer from unavailability of
standard evaluation platform, i.e. unavailability of text
compression corpus. This paper proposes a text
compression corpus for evaluating Bengali text
compression schemes.
The construction of a data compression corpus is now
a demand-of-time. The state-of-the-art technologies of
data management have already equipped the benchmarks
and standard collection of experimental data sets for
performance evaluation [1]. In Bengali, though a highly
impressive step has been initiated in [2] for constructing a
linguistic corpus having CIIL (Central Institute of Indian
Languages) corpus [4] as a pioneer, still now any
mentionable sophisticated data compression corpus is not
available with text compression benchmarks. This paper
focuses on designing a Bengali text compression corpus
based on Type to Token Ratio (TTR) and provides the
mathematical framework for various aspects of choosing
the corpus with compression benchmarks.
The paper is targeted to design a corpus that will
facilitate the future researchers especially in the research
of Bengali text compression to have a test-bed for
evaluating their performance. For designing such a
corpus we are to derive a novel scheme of text
compression corpus formation because of having no
existing benchmark.
The paper is organized into eight sections. Section II
provides elementary concepts on corpus and text
compression schemes. The key-points for constructing a
corpus are presented in section III. Section IV is devoted
with the literature review presenting various aspects of
the existing corpora (corpuses). Section V describes the
proposed approach for formulating the corpus. A brief
description on the files comprising the proposed text
compression corpus is presented in section VI. Analysis
of the proposed corpus is depicted in section VII. Various

60 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
issues on the usability of the proposed corpus for other
languages is presented in section VIII. The paper ends
with section IX providing conclusion and
recommendation.
II. OVERVIEW OF CORPUS AND DATA COMPRESSION
Text compression is an elementary aspect of data
compression. Compression is the process of reducing the
size of a file or data by identifying and removing
redundancy in its structure [7]. Data Compression offers
an effective approach of reducing communication costs
by using available bandwidth effectively. Data
Compression technique is generally divided into two
categories; namely, Lossless Data Compression and
Lossy Data Compression. For lossless schemes, the
recovery of data should be exact. Lossless compression
algorithms are to some extent essential for all kinds of
text processing, scientific and statistical databases
applications, medical and biological image processing,
DNA and other biological data management and so on.
However, a lossy data compression technique does not
ensure the exact recovery of data. For image compression
and multimedia data compression, there is a great use of
lossy data compression. From the early 1990, certain
specific platforms of evaluating English text compression
schemes are adapted for standard researches, which are
better known as data compression corpus [1].
In general sense, a corpus is simply a collection of
texts. From the point of view of data compression,
compression-corpus is a standard collection of texts or
data for analyzing and evaluating effectiveness and
efficiency i.e. performance of compression schemes. The
mostly used data compression corpora for evaluating data
compression are namely Calgary corpus, Canterbury
corpus and Project Gutenberg. Though the corpora for
English were introduced in the early 1990, still now, for
analyzing and evaluating Bengali text compression
performance, no such complete and standard corpus is
available. In this paper, we present a novel approach for
designing data compression corpus for evaluation of
Bengali text compression scheme. Implementation of the
proposed scheme has also been focused in the developed
Ekushe-Khul corpus.
For analyzing domain specific or language specific
text-compression scheme, it is a must to have the domain
oriented or language oriented benchmark to analyze the
compression schemes. There are a number of English text
compression corpora and corresponding benchmarks
which are extensively used for evaluation of English Text
compression schemes. There are also a number of
multilingual text corpora involving English for evaluating
English text compression with respect to other languages.
But in case of Bengali there is neither any corpus for
evaluation of text compression performance nor any
benchmark for the same. A detail explanation on the
necessity of Language specific corpora and the suitability
© 2010 ACADEMY PUBLISHER
(or unsuitability) of any corpora for evaluating non-
English concerns is provided by Sarker and Roeck in [10]
with a comprehensive explanation by N. S. Dash in [12]
and by Akshar et al. in [11]. The required changes for
designing a new corpus for Bengali in order to meet new
demands of computational linguistics and other
approaches is elaborately provided in [9].
III. CHARACTERISTICS OF A DATA COMPRESSION CORPUS
The characteristics of a corpus are defined by the
purpose of creating the corpus and the evaluation arena
for which the corpus is designed for. These criteria give
rise to the necessity to field-specific corpus like data
compression corpus, information retrieval corpus, corpus
for data mining research etc. According to Arnold et al.,
there are six criteria for choosing a corpus [1]. Firstly, the
files should be chosen as representative of the files used
and expected to be used in compression method.
Secondly, the availability of the file should be ensured.
Availability of public domain materials is the third
consideration for the corpus. They also suggest that the
files should not be larger than necessary and it should be
perceived to be valid and useful. Finally, the corpus
should be actually valid and useful.
The mentioned criteria for choosing corpus have
successfully established Canterbury Corpus as one of the
extensively adapted corpus for data compression.
However, due to the rapid growth of technology,
additional criteria have been evolved in constructing the
corpus. Firstly, the criteria were specified for evaluating
the compression of middle-sized files. But, with the
advancement of embedded devices with lower processing
speed and small memory (like mobile-phones) it has
become extremely important to provide a new test-bed
for evaluation of small text compression schemes.
Frequent uses of mobile phones, short text messages and
emails have already made a transformation on the use of
texts, and hence necessity of field and size specific
corpus formation has been evolved. Taking these
limitations into concern, we provide specific files
regarding the issues with consideration of other
traditional characteristics. Moreover, the file structure i.e.
sentence structure was not considered as any important
criteria for choosing corpora. But, compression ratio
gradually varies with variations of sentence construction
and file structure. Besides these, probability of
occurrences of texts-components and frequency of those
are important criteria, which fluctuate the text
compression performance.
IV. LITERATURE REVIEW
The history of mostly adapted data compression corpus
started a long ago. Calgary Corpus, which was collected
in 1987 and was published in 1990 is considered as the
pioneer of data compression corpus. The files in Calgary
corpus [1] with their content category are listed in table I.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 61
TABLE I
FILES IN THE CALGARY CORPUS
Title Description
bib Bibliographic File
book1 Hardy: Far From the Madding crowd
book2 Witten: Principles of computer speech
geo Geophysical data
news News batch file
obj1 Compiled Code for Vax
obj2 Compiled Code for Apple Macintosh
paper1 Witten: Arithmetic Coding for compression
paper2 Witten: Computer Security
paper3 Witten: In Search of Autonomy
paper4 Cleary: Programming by example revisited
paper5 Cleary: logical implementation of arithmetic
paper6 Cleary: Compact Hash tables
pic Picture from CCITT Facsimile
progc C source code
progl Lisp source code
progp Pascal Source Code
trans Transcript of a session on terminal
The next initiative and the mostly adapted one is
Canterbury Corpus. It was developed by Arnold et al.
[1]. Canterbury corpus considered the benchmark derived
from the Calgary corpus as their corpus development
basement. In Table II, the files constituting Canterbury
corpus are presented.
TABLE II
FILES IN THE CANTERBURY CORPUS
Title Description
alice29.txt English text
ptt5 Fax images
fields.c C source code
kennedy.xls Spread-sheet Files
Ssum SPARC Executable
lcet10.txt Technical documents
plrabn12.txt English Poetry
cp.html HTML
grammar.lsp Lisp source code
xargs.1 GNU Manual pages
asyoulik.txt plays
The above files are solely for evaluation of English text
compression evaluation corpus and hence not effective or
at all applicable for evaluating Bengali text compression.
The most recent and extremely organized analysis on
Bengali news corpus is provided by Majumder et al. [2].
© 2010 ACADEMY PUBLISHER
This paper proposes a new corpus on the basis of the well
circulated newspaper prothom-alo [5] and provides
exhaustive analysis regarding TTR (Type-to-Token-
Ratio), Function Word and other morphological and
linguistic analysis. For analysis of data compression
efficiency, this corpus requires some specific
modification in the sense that, data compression involves
the additional criteria of heterogeneity and text-size.
Moreover, the news corpus proposed in [2] is extensively
tested for linguistic analysis only. In order to use this for
data compression, further text-compression related
analysis is a must. According to their methodology, they
collect html news files from prothom-alo websites and
converting it to Unicode, they categorize the total
collection into twenty seven distinct groups, which are
also available in a single file of size three hundred and
eighteen mega bytes. For most practical applications and
researches on middle-sized or small text compression, it
is rarely necessary to have larger files, rather it is
essential to have standard small and medium sized
corpus.
The fist corpus in Bengali developed by Central
Institute of Indian Languages (CIIL) in the years 1991-
1995 possesses a collection of three million Bengali
words, which provide valuable linguistic data for research
on Bengali language analysis [4]. In evaluating
compression efficiency, this impressive corpus also
suffers from the limitation of heterogeneity and lack of
embedded and small device supportability.
Besides the corpora presented there are a number of
researches available describing the features of Bengali
linguistic corpus. A number of works is also devoted to
have corpora on Bengali OCR (Optical Character
Recognition). The contribution of B. B. Chaudhury is
pioneer in this specific field of Bengali OCR formation
[13].
The necessity of devising specific and relevant criteria
for building Bengali corpora and the basic limitations as
well as pitfalls of existing non-Bengali corpora building
concepts are extensively incorporated by N. S. Dash in
[9]. Prime techniques of Bengali text corpus processing,
which are used for various linguistic activities including
far more complicated orthographical, morphological, and
lexicological concerns is also presented in [9]. Various
criteria including Concordance of Words, Lexical
Collocation, Key-Word-In-Context, Local Word
Grouping, Lemmatization of Words, Parsing Sentences
etc are widely described by N. S. Dash in [9]. Being the
overall aspect of the literature in [9] is to furnish
linguistic and information retrieval corpora, there is little
usability of the technique [9] in text compression corpora
formation.
The essentiality of domain specific corpora has
motivated us to design Bengali text compression
evaluation corpus. The unavailability of reference or
benchmark in the field has also leaded us to develop a
new concept for designing a corpus.

62 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
V. PROPOSED CORPUS FOR EVALUATION OF BENGALI
TEXT COMPRESSION SCHEMES
We propose a new corpus for evaluation of Bengali
text compression schemes namely Ekushe-Khul corpus
[8]. This corpus is intended for evaluating only short and
middle-sized text compression schemes rather than
evaluation for large text compression. One of the worthmentionable
points is, this corpus is selected by applying
a new approach, which takes both TTR and Compression
Ratio (CR) into account. The two concepts have been
adapted because we do not have any standard and
sophisticated Bengali text compression corpus available
with which we may incorporate and compare the
performance and thus perform the selection from
candidate files as done for other existing data
compression corpora.
To form the corpus we have considered ten groups-
Articles, Poems, Advertisements, Speeches, News, Sms,
Emails, Particulars, Stories And Reports. The files are of
sized 4 KB to 1800 KB. The steps of constructing the
proposed corpus are depicted in Fig 1.
COLLECT CATEGORIZED FILES
REMOVE ENGLISH AND OTHER NON-
BENGALI TEXTS OR SYMBOLS
CONVERT TO UNICODE
CALCULATE TTR FOR EACH FILE
SELECT FILES WITH
MAXIMUM TTR
IF MORE THAN ONE,
SELECT ONE WITH
MAXIMUM SOURCE FILE
SIZE AND IGNORE
OTHERS
SELECT FILES WITH
MINIMUM TTR
IF MORE THAN ONE,
SELECT ONE WITH
MAXIMUM SOURCE
FILE SIZE AND IGNORE
OTHERS
CATEGORY_TITLE_1.txt CATEGORY_TITLE_2.txt
Fig. 1: Steps for constructing the Proposed Bengali text
compression corpus.
© 2010 ACADEMY PUBLISHER
The files constituting the proposed Ekushe-Khul corpus
are shown in Table III. A detail description on the
proposed corpus selection criteria with potential issues
and concerned data compression parameters that we have
proposed for developing Bengali text corpora for
evaluation of Bengali text compression schemes is
presented in section VI.
TABLE III
FILES CONSTITUTING THE PROPOSED EKUSHEY-KHUL CORPUS.
File Name Type
Article1.txt
Article2.txt
Poem1.txt
Poem2.txt
Advertise1.txt
Advertise2.txt
Speech1.txt
Speech2.txt
News1.txt
News2.txt
SMS1.txt
SMS2.txt
Email1.txt
Email2.txt
Particulars1.txt
Particulars2.txt
Story1.txt
Story2.txt
Report1.txt
Report2.txt
Articles in Bengali
Bengali Classic Poems
Advertisements
Various Speeches
News from Newspaper
Short Bengali Message
Emails in Bengali
Particulars
Bengali Stories
Academic Reports
To collect the files, the first step was to remove
redundancies and other components (e.g. html tags,
letters and words from other languages, formatting etc.)
and convert them into Unicode as several documents
were non-standard encoded and saved as Unicode text
files. These files are then forwarded for further analysis.
For each category of files, all the files are passed to a
TTR Calculator and then Compression Ratio Evaluator
for selection of final documents. In case of Canterbury
Corpus, which is the pioneer for data compression corpus
was formed by comparing each file with the standard of
the Calgary Corpus of same group or category, and the
file demonstrating the closest match in terms of
compression performance was selected. But, as no such
standard exists, we had to take account of both the criteria
and hence selected two with the maximum and minimum
consistence.
TTR is a measurement of how many times a previously
encountered words repeat themselves before a new word
makes its appearance in the text [2]. It is note worthy that
though this measure is essential for language engineering,

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 63
its application in data compression is not mandatory. For
calculative approaches, TTR is calculated by dividing the
total number of word tokens by the total number of
distinct words. Texts with a large number of distinct
words result into lower TTR. For this case, the task of
forming a data compression dictionary appears to be
more accurate. Conversely, for texts with greater TTR i.e.
with greater number of frequent words (resulting lower
number of distinct words) provide a suitable arena for
constructing and evaluating dictionary-based and
repetition-analysis oriented text compression schemes.
This is the reason, which inspires us to provide peer-files
for each category. The files with suffix ‘2’ indicate the
file with lower TTR (suitable for text compression
fluctuation evaluation) and files with suffix ‘1’ indicate
greater TTR (suitable for dictionary-based text
compression scheme evaluation) for the respective group.
The next step involves analyzing the files in terms of
compression ratio [6] to establish logical and
implementational validity of choosing TTR as a
benchmark for constructing the corpus. Though a
number of approaches exist for compression of English
texts, (like Burrows-wheeler Transform [4], Arithmetic
Coding, Huffman Coding etc.) it is a sad tale that, still
now a little research has been dedicated for Bengali text
compression schemes. The uses of non-sophisticated text
compression schemes do not provide optimal
performance for Bengali text compression purposes. For
providing a benchmark of the proposed corpus, we adapt
the dictionary based compression scheme in [7] and
demonstrate the inter-relation of TTR with Compression
Ratio.
Unavailability of existing sophisticated and
multidimensional text Bengali text compression
approach, we had a little alternatives to compare the
performance fluctuation of our corpora. Though there are
a number of English text compression schemes available,
testing with those may never provide a comprehensive
view to out developed scheme with implemented
prototype.
VI. ASPECTS OF CORPUS SELECTION CRITERIA AND TEXT
COMPRESSION PERFORMANCE
Let n be the total number categories of collections of
texts. The categories are representative of texts bearing
distinct themes with differences in sentence construction,
sentence wording and sentence structure.
Let the categories are c1 , c 2 , c 3 , … … , c n
respectively.
For category c1 the files are
, f , f , … … , f
f1,1 1,2 1,3
1,n
with length l1,1 , l1,2
, l1,3
, … … , l1,n
, where, the
length indicates total number of characters (including
blank spaces, tabs, new lines and other punctuation
marks).
© 2010 ACADEMY PUBLISHER
If there are total of w1,1 , w 1,2 , w 1,3 , … … , w1,n
words in f1,1 , f 1,2 , f1,3
, … … , f 1,n respectively
with number of distinct words be
, d , d , … …,
d for the same.
d1,1 1,2 1,3 1,n
Here, for any file fi, j, where i denotes category index
and j denotes file index, and wi ,j ≥ di,j. Assume Ti. j be
the Type to Token Ratio (TTR) for file fi, j.
w
That is,
i , j
T = .
i , j
d
i , j
For any two files p, q in category i , with TTR Ti,p
and T1,q, if Ti, p ≥ Ti, q indicates that, Ti, p contains larger
number of distinct words than that of Ti, q.
Applying this rule, TTRs are counted for the files.
Then the file with maximum and minimum TTRs are
selected. For the file with maximum TTR, if the elements
comprise the set Sd and St for distinct words and total
words respectively, then for ideal case, set Sd − St would
tend towards Φ (empty or null) set.
Though TTR has been considered as a means of
forming linguistic corpora and Natural Language
Processing corpora including Information Retrieval
corpora, taking it into concerns, it is possible to design
Data Compression corpora too. Here, we provide an
analysis has been presented on the effectiveness of using
TTR as a criteria of constructing Corpora.
The basic relation between TTR and word-distribution
is:
� If TTR is larger (greater) then, the number
(frequency) of matching word is also greater;
consequently, number of distinct words is smaller.
� If TTR is smaller (lower) then, the number
(frequency) of matching word is smaller;
consequently, number of distinct word is larger.
For designing dictionary based data compression
approaches, TTR plays a great role. For constructing the
dictionary, the prime concern is to analyze the probability
of occurrence of each character or syllable or substring or
words (or any combination of the mentioned units). The
frequencies in which they occur are an important aspect
of figuring the probability. As the total number of distinct
words is greater in files with lower TTR value, the
detection of the probability of syllables may be easier as
the word distribution is approximately flat in those files.
For building word-based dictionaries for data
compression, files with higher TTR values are more
effective as the recurring words may be very easily
pointed out. The fluctuation may be easily identified by
comparing the performance of the compression scheme
using the peer files.
Let us consider a tertiary dictionary based text
compression approach DCA, which uses a dictionary of n
items. The total items consist of c characters, s syllables
and w substrings, which save (in average) sc, ss, and sw
bits each. Traditionally, sw ≥ ss ≥ sc as words or
substrings are generally considered as a collection of
syllable and/or symbols and syllables are considered as a

64 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
collection of characters, but the indices or pointers in
most designs require consecutive values spanning against
a threshold value. For a file with greater TTR ( i.e.
Greater number of frequent words and lower number of
distinct words), the probability of encoding using wordbased
coding is high. Consequently, the bit-savings will
also be high as the saving weightage is greater for words
in compassion with other units. Let the replacement of
number of distinct words be rw. Here, rw theoretically
tends towards the total occurrence in average.
Therefore, the total savings in this phase is
S
1
=
rw
∑
y = 1
(sw
y
×
fw
y
)
where fw is the frequency of corresponding distinct
word. That is, fw indicates the frequency of occurrence
of distinct word-indexed at y and ideally, fw → tw.
The number of distinct syllable rs in a collection of
highly distinct words should be lower. If the weightage is
considered as double as distinct words, the total savings
would be-
S 2
=
rs
∑
y = 1
(ss
y
×
fs
y
)
where fs is the frequency of the syllable.
The rest of the characters should be encoded separately
resulting savings of
S 3
=
rc
∑
y = 1
(sc
y
×
fc
y
)
If we encode a total of p characters separately, which
saves sp bits for each character and if we encode q words
separately with an average of d characters per word,
whenever even d = p, the savings will be much more
greater for word based bit-savings, since the indexing of
the constructed code base will be approximately
sequential and hence, each single character and each
single word based encoding will occupy about same
length.
Thus, the total saving is,
S = S + S + S
S
=
1
rw
∑
y = 1
2
y
3
(sw × fw ) +
y
rs
∑
y = 1
(ss
y
× fs ) +
y
rc
∑
y = 1
(sc × fc ) (1)
For a file with lower TTR, (i.e. greater number of
distinct words and lower number of frequent words) there
will have non-frequent occurrence of matching words and
syllables. This will result extensive application and usage
of character based encoding, as the primary unit of
encoding is character. That is, r w′
< rw and fw′ → 1 .
Similarly, r s′
< rs and Therefore, if there are total of
tw words in the file, then, fs′ → 1 .
But, because of remaining large number of character
units, r c′
→ tc′
.
© 2010 ACADEMY PUBLISHER
y
y
If there a total of t w′
words in the file with lower TTR
values, then,
∑ ′
S
′
1 =
rw
y = 1
and w′
→ tw
(sw
y
×
fw
′
y
f ′
As low frequent words would be encountered, as much
frequent syllables would be faced. However, a minor
portion of the syllables may be coded in form of words.
That is, we will have a larger number of character units
and syllable units ( t s′
and t c′
) left to be coded.
For syllable based coding,
∑ ′ rs
S
′
2 = (ss y × fs′
y )
y = 1
After the two steps, a greater number of characters are
left to be coded with distinct characters. Typically this
value t c′
may be t c′
< ts′
and for usual cases, t c′
< tw′
.
Let these steps save,
∑ ′ rc
S
′
3 = (sc y × fc
′ y )
y = 1
S ′ = S ′ + S ′ + S ′
S ′ =
rw′
∑
y = 1
1
(sw
y
2 3
rs′
× fw′
) + (ss
y
∑
y = 1
y
)
× fs′
) +
y
rc′
∑
y = 1
(sc
y
× fc′
) (2)
It has been stated that, the ratio of bit-saving for word
based , syllable based and character based encoding is
4:2:1 (i.e. for the ratios a : b : c we may write, a > b >
c). The ratio is taken as standard because, for character
based dictionary oriented encoding, single character or
symbol is considered as the primary unit. Consequently,
substrings and syllables ranges from two to any higher
value. The number of characters that comprises any word
may ideally be considered in the range of greater than
four whereas short length words may be thought of as the
subgroup of substrings. Hence, it can be easily remarked
that, for the above consequences,
(i) S ≥
′
1 S1
; since the number of word-matching is
greater in S1 and the difference is a multiplier of four.
(ii) S ≥
′
2 S 2 ; since the number of syllable-matching
is greater in S2 and the difference is a multiplier of two.
(iii) S ≤
′
3 S 3 ; since the number of matching-character
is greater in S3 and the difference is an unit multiplier.
From the above three consequences, we may easily
point that, though S
′
3 ≥ S 3 , it is an average case that,
′
S will cross the saturation point or threshold value
3
comprising of (4 + 2 = 6) times higher amplitude.
Consequently, we may easily conclude: S > S ′ .
It may be also proved the same as follows:
From the third clause we may write,
S α = S
′ ,
3 + 3
y

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 65
where α is a constant bit-factor.
From clause (i) and (ii), we can write
S S ≥ S
′
+ S
′
.
1 + 2 1 2
From (i), (ii), we may deduce that (by adding the same
value S3 / in both sides),
′
≥
′
+
′
1 + S 2 + S 3 S1
S 2 + S 3
S
Being a constant factor
′
(3)
α = 1, 2, 3, ........etc(number
of bits) , we may write (3) as,
S S + S + α ≥ S
′
+ S
′
+ S
′
(4)
1 + 2 3
1 2 3
It has been presented that,
S = S1
+ S 2 + S 3
and,
S ′ = S1′
+ S 2′
+ S 3′
.
Hence, from (4),
S + α ≥ S
′
.
Recall from earlier explanation that, S and S / are the
aggregation of bit savings for matching characters,
syllables and words. Again, as the ratio of bit saving for
word-based, syllable based and character based encoding
is considered 4:2:1, consideration of linear transformation
factor l indicates that, S will rush towards 6l where
α tends words l. for the average cases.
In better cases, as all the characters are likely to be precoded
with either syllable based coding or word based
coding section. After coding with word based and
syllable based coding there will be little portion of source
text left for character based coding. That is, the total
saving is probable to be derived from word based saving.
Consequently, S 1 + S 2 → S , whereas, character
based saving will tend towards zero. S 3 → 0 .
Again, being S + =
′
3 α S 3 ; α too will tend towards
zero. That is, α → 0 .
Again, from clause (i), we get,
S = S ′ + α
(5)
1
1
Similarly, from clause (ii),
2
2
1
S = S ′ + α
(6)
2
We get from (iii), S 3 = S 3′
− α 3
(7)
Where 1 , α 2 , α 3
α are non negative integer values
because they all are equalizing factors in terms of bits.
It has already been stated that, S 3 → 0 .
Consequently, ′ 3 − α 3
S will only tends towards zero
if and only if, α 3 → 0 .
That is, it can be clearly deduced that,
α 1 2 3
+ α − α = c ≥ 0 .
As a result, we may write,
© 2010 ACADEMY PUBLISHER
α 1 2 3
+ α − α ≥ 0 .
It has been previously stated that, α 3 → 0 .
Consequently, α1 + α 2 ≥ α 3
(8)
Now, (5) + (6) + (7) results-
′ ′
S 1 + S 2 + S 3 = S 1 + α 1 + S 2 + α 2
′
+ S 3 − α 3
⇒ S + S + S
′
= S + S
′ ′
+ S + α + α − α
1
2
3
1
⇒
′ ′ ′
S 1 + S 2 + S 3 = S 1 + S 2 + S 3 + c [Using (8)]
⇒
′ ′ ′
S 1 + S 2 + S 3 ≥ S 1 + S 2 + S 3
⇒ S ≥ S
′ [Since c is non-negative]
That is, the performance of dictionary based data
compression would be better for files with larger TTR
Values and the result will deteriorate for lower TTR
valued files. This is the aspect, which motivates us to
provide peer files of each category to recognize the
fluctuation of performance for dictionary based data
compression.
VII. ANALYSIS AND DISCUSSIONS ON THE PROPOSED
CORPUS
In the previous sections, overview of the corpus has
been presented. In this section, we present statistics
regarding TTR and Compression ratio of the files of each
groups. The experimental results demonstrate that, not
only in expressing the effectiveness of any corpus, TTR
may also be employed as a criteria for constructing data
compressing corpus with the great scope of presenting the
performance fluctuation of dictionary based (i.e.
repetition analysis oriented) text compression scheme for
best case and worse case analysis. The maximum TTR for
each group is provided in Table IV.
2
TABLE IV
MAXIMUM TTR FOR EACH GROUP OF FILES OF PROPOSED CORPUS
File Name TTR
Article1 2.944
Poem1 1.451
Advertise1 1.170
Speech1 5.846
News1 3.087
SMS1 1.138
Email1 2.681
Particulars1 3.510
Story1 4.862
Report1 3.364
3
1
2
3

66 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
The minimum Type-to-Token Ratio (TTR) for each
group is provided in Table V.
TABLE V
MINIMUM TTR FOR EACH GROUP OF FILES OF PROPOSED CORPUS
File Name TTR
Article2 1.042
Poem2 1.132
Advertise2 1.012
Speech2 2.164
News2 1.611
SMS2 1.103
Email2 1.812
Particulars2 1.489
Story2 2.002
Report2 2.157
We also analyze the compression ratio for the files
using the dictionary-based approach for Bengali text
compression provided in [7] and find the following
statistics-
6.5
6
5.5
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Article
Relation between TTR and Compression Ratio
Poem
Advertise
Speech
News
SMS
Email
Particulars
Story
Report
MAXIMUM TTR MINIMUM TTR
CR OF MAX TTR FILE CR OF MIN TTR FILE
Figure. 2: Relation between Type to Token Ratio and
Compression Ratio.
© 2010 ACADEMY PUBLISHER
Fig 2, has presented a comparative analysis on
Compression Ratio with TTR in terms of applicability of
dictionary based data compression techniques.
Compression Ratio (CR) is a metric of data compression
efficiency. Compression ratio indicates the number of bits
required to describe one byte in compressed form. The
lower the compression ratio, the better is the
compression. In order to evaluate the performance
fluctuation of dictionary based Bengali text compression
schemes, we propose to use the peer files with
assumption that, the best case would be obtained from the
files with greater TTR values and the worse case will
occur for lower TTR files. The construction of the
dictionary may really be facilitated from the files with
larger TTR.
As, there is no existing corpus for evaluation of
Bengali Text Compression Scheme, it is not possible to
incorporate any comparative analysis with respect to any
parameter. Because of the same in terms of compression
benchmarks, there is no way to integrate the performance
(effectiveness) analysis of proposed corpus. Basically,
this is the reason which leads us to develop a new scheme
for designing any corpus without any reference
benchmarks. The proposed scheme is also to some extent
a novel approach for automatic creation and evaluation of
suitability (or unsuitability) of any corpora.
VIII. APPLICABILITY OF PROPOSED CORPUS FORMATION
SCHEME FOR OTHER LANGUAGES
The proposed method of building a text corpus for
evaluation of Bengali text compression scheme is also
applicable for non-Bengali data compression corpus
formation. As per the presented scheme for building the
corpus, we analyze various statistical analysis of the
source text to decide whether the text will be included or
not irrespective of the text-structure and other linguistic
features. This text-structure independence feature
provides a greater flexibility in applying the proposed
concept for non-Bengali text.
We have already presented the theoretical background
of choosing the peer files in view of dictionary based
compression scheme. Here the same has been described
in terms of traditional character based encoding scheme.
For a character based encoding scheme i.e. a singlegram
based coding approach, the main strategy is to
define code-words for each character and then replacing
each character with corresponding code-word. The codewords
are considered as binary stream and in practical
cases the encoding scheme takes static coding into
account. Static coding is an encoding mechanism where
source components are encoded with non-conflicting
variable length binary stream. In the simplest case, the
length of each binary stream to be used for representing
each character is determined through their frequency
distribution. Often this length determination is obtained

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 67
through probabilistic distribution based component
ranking schemes.
Let there are n unit source symbols in the source
language. That is, the total number of letters in the source
language be n. If TTR is larger (greater) then the number
(frequency) of matching word is also greater;
consequently, number of distinct words is smaller.
Whenever there will greater number of matching words,
it may be assumed that, for average cases, the number of
matching characters will also be greater. This may not be
the same if the non-repeating words or low-repeating
words individually contain unusually large set of
repeating characters and consequently, even though the
frequency of matching character is greater, sum of the
characters present in the non-matching words is greater.
That is, for a n character-set based language if the
considering text consist of m characters, and there are tmw
matching words where each of the matching word
contains an average of acw characters, the total pool of
matching character is tmw * acw . Being n character-set
based language, this pool of matching character may
contain at best n distinct characters. That is, the TCR
(Type to Character Ratio in analogy with TTR, Type to
Token ratio) of the file will be a multiplier of u = 1/n .
Consequently, for any file with greater TTR, will have a
large multiplier of u. As for practical cases we may
assume that the average fluctuation of each word-length
from average number of characters per word is
approximately zero, we may consider the word length
versus character distribution of the source text as a linear
distribution. As a result, it may assume that, whenever
there will have greater number of frequently occurring
words, it implies frequently occurring characters.
Whenever static coding is used, the code-words are
defined according to the frequency of the occurring
elements. That is, frequently appearing elements will be
assigned lower bit consuming codes in order to minimize
the total overhead (in terms of bit consumption) whereas,
the non-frequently occurring elements may be coded with
greater threshold value. The same may be expressed in
terms of Type to Character Ratio (TCR) that, elements
with larger TCR will be assigned low overhead and
elements with lower TCR will be assigned with grater
overhead in comparison with lower TCR elements. If we
want to deploy such scheme for employing in Bengali
text compression scheme, it is an effective idea to build
such a corpus which will contain reversible components,
which presents the fluctuation of TCR clearly. That is, to
provide a sound evaluation of the corpus, it may
inherently express that, for a suitable corpus
taw ∝ tac
and,
t
t
t
aw
dw
∝
t
t
taw ac ∝
dw
t
n
ac
dc
© 2010 ACADEMY PUBLISHER
Where,
taw = Total number of words appeared in the source
text.
tdw = Total number of distinct words in the source
text.
tac = Total number of characters appeared in the
source text.
taw = Total number of distinct characters in the
source text.
n = Total number of symbols in the source language.
This criteria demonstrates an important and to some
extent essential aspect of forming a data compression
corpus that will facilitate of an effective and efficient
design and implementation scheme of text compression
approaches irrespective of language. Though the main
concern or uses of data compression corpus that has been
adapted traditionally is only evaluation of compression
scheme, the evolving nature of data management has
motivated (and necessarily forced) to develop corpora for
being considered as a knowledgebase as well as test-base
for text compression scheme. The analysis presented in
this paper establishes the criteria to be taken into
consideration for building any dictionary based text
compression scheme irrespective of adapting single-gram
dictionary based text compression or multi-gram
dictionary based compression.
VIII. CONCLUSION
We have proposed a new Corpus for evaluation of
Bengali text compression schemes named Ekushey-Khul.
The methodology is also robust to some extent that, it
takes both linguistic and compression-ratio into account.
It is also a contribution that, we specify new criteria for
choosing text compression corpus, which includes
consideration of expected device and the changed attitude
of text varying with communication framework for which
the compression is intended. Though it is an inaugurating
step towards forming a Bengali text compression
evaluation corpus, further improvement may be achieved
by integrating other types and modes of document like
technical documents, Bengali dictionary, and Bengali
spreadsheets etc. with consideration of more training
files.
REFERENCES
[1] Ross Arnold, and Tim Bell, “A Corpus for the evaluation
of lossless compression algorithms”, Data Compression
Conference, pp. 201-210, IEEE Computer Society Press,
1997.
[2] Khair Md. Yeasir Arafat Majumder, Md. Zahurul Islam,
and Majuzmder Khan, “Analysis of and Observations from
a Bangla News Corpus”, Proceedings of 9th International
Conference on Computer and Information technology
ICCIT 2006, pp. 520-525, 2006.
[3] Mat Powel, “Evaluating Lossless Compression
Algorithms”, February, 2001.

68 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
[4] N. S. Dash, “Corpus Linguistics and Language
Technology”, 2005.
[5] Official Web site of Prothom-Alo, www.prothom-alo.com
[6] M. Burrows and D. J .Wheeler. “A block sorting lossless
data compression algorithm”. Technical report, Digital
Equipment Corporation, Palo Alto, CA, 1994.
[7] S. A. Ahsan Rajon, “A study on Bengali Text Compression
Schemes”, Research Report, Khulna University, Khulna,
June 2008.
[8] Md. Rafiqul Islam, and S. A. Ahsan Rajon, “On the
Design of an Effective Corpus for Evaluation of Bengali
Text Compression Schemes”, Proceedings of 11th
International Conference on Computer and Information
Technology (ICCIT 2008), 25-27 December, 2008, Khulna,
Bangladesh, pp. 236-241.
[9] Niladri Sekhar Dash, “Some Techniques Used for
Processing Bengali Corpus to Meet New Demands of
Linguistics and Language Technology”, SKASE Journal of
Theoretical Linguistics. 2007, vol. 4, no. 2 ISSN 1336-
782X.
[10] Avik Sarkar, Anne De Roeck , A Framework for
Evaluating the Suitability of Non-English Corpora for
Language Engineering [online].
[11] Akshar Bharathi, Rajeev Sangal and Sushma M Bendre:
Some Observations Regarding Corpora of Indian
Languages. Proc. of Int. Conf. on Knowledge-Based
Computer Systems (KBCS-98), 17-19 Dec 1998, NCST,
Mumbai.
[12] Niladri Sekhar Dash and Bidyut Baran Chaudhuri , Why
do we need to develop corpora in indian languages.
[online]
[13] Upal Garain, B. B. Chaudhuri, “A Complete printed
Bangla OCR System”, Pattern Recognition Journal, Vol.
31, No. 5, pp. 531 – 549, Elseiver Science Limited, 1998.
© 2010 ACADEMY PUBLISHER
Md. Rafiqul Islam obtained Master of Science (M. S.) in
Engineering (Computers) from Azerbaijan Polytechnic Institute
(Azerbaijan Technical University at present) in 1987 and Ph.D.
in Computer Science from Universiti Teknologi Malaysia
(UTM) in 1999. His research areas include design and analysis
of algorithms and Information Security. Dr. Islam has got a
number of papers related to these areas published in national
and international journals as well as in referred conference
proceedings.
He is currently working as the Head of Computer Science
and Engineering Discipline, Khulna University, Bangladesh
S. A. Ahsan Rajon is an Adjunct Faculty of Computer
Science and Engineering Discipline, Khulna University,
Khulna. He has completed his graduation from the same
discipline in April 2008. He is also pursuing M.B.A. from
Business Administration Discipline under Management and
Business Administration School of same university. Rajon has
made several publications in International conferences. His
research interest includes data engineering and management,
electronic commerce and ubiquitous computing. Currently he is
working on robotics.
He is a member of Institute of Engineers, Bangladesh (IEB).

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 69
A Corpus-based Evaluation of a Domain-specific
Text to Knowledge Mapping Prototype
Rushdi Shams
Department of Computer Science and Engineering, Khulna University of Engineering & Technology (KUET),
Bangladesh
Email: rushdecoder@yahoo.com
Adel Elsayed
M3C Research Lab, the University of Bolton, United Kingdom
Email: a.elsayed@bolton.ac.uk
Quazi Mah- Zereen Akter
Department of Computer Science & Engineering, University of Dhaka, Bangladesh
Email: mahzereen@yahoo.com
Abstract— The aim of this paper is to evaluate a Text to
Knowledge Mapping (TKM) Prototype. The prototype is
domain-specific, the purpose of which is to map
instructional text onto a knowledge domain. The context of
the knowledge domain is DC electrical circuit. During
development, the prototype has been tested with a limited
data set from the domain. The prototype reached a stage
where it needs to be evaluated with a representative
linguistic data set called corpus. A corpus is a collection of
text drawn from typical sources which can be used as a test
data set to evaluate NLP systems. As there is no available
corpus for the domain, we developed and annotated a
representative corpus. The evaluation of the prototype
considers two of its major components- lexical components
and knowledge model. Evaluation on lexical components
enriches the lexical resources of the prototype like
vocabulary and grammar structures. This leads the
prototype to parse a reasonable amount of sentences in the
corpus. While dealing with the lexicon was straight forward,
the identification and extraction of appropriate semantic
relations was much more involved. It was necessary,
therefore, to manually develop a conceptual structure for
the domain to formulate a domain-specific framework of
semantic relations. The framework of semantic relations-
that has resulted from this study consisted of 55 relations,
out of which 42 have inverse relations. We also conducted
rhetorical analysis on the corpus to prove its
representativeness in conveying semantic. Finally, we
conducted a topical and discourse analysis on the corpus to
analyze the coverage of discourse by the prototype.
Index Terms— Corpus, Knowledge Representation,
Ontology, Lexical Components, Knowledge Model,
Conceptual Structure, Semantic Relations, Discourse
Analysis, Topical Analysis
I. INTRODUCTION
Text to Knowledge Mapping (TKM) Prototype [1] is a
domain-specific NLP system, the purpose of which is to
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.69-80
parse instructional text and to model it with its predefined
ontology. During development, the prototype has
been tested with a limited data set from the domain
instructional text on DC electrical circuit. The prototype
reached a stage where its lexical components and
knowledge model need to be evaluated with a
representative linguistic data set, a corpus- a collection of
text drawn from typical sources. Information retrieval
during parsing, activation of concepts and relating them
with predicate and semantic relations contribute to map
and model domain-specific text on its knowledge domain.
Therefore, the usability of the TKM prototype as a
specialized knowledge representation tool for the domain
depends on the evaluation of its lexical components like
vocabulary and grammar structures, knowledge model
like ontology and coverage of discourse.
An important precondition to evaluate NLP systems is
the availability of a suitable set of language data called
corpus as test and reference material [2]. With an
extensive web-based search, we did not find any corpus
for the domain DC electrical circuit. Therefore, we need
to develop a representative corpus to evaluate the
prototype because a representative corpus reflects the
way language is used in the domain [3]. A usable corpus
requires various annotations according to the scope and
type of evaluation. As we intend to evaluate both the
lexical components and knowledge model of the TKM
prototype, the corpus should be annotated with
information like Parts of Speech (POS) tagging, phrasal
structure annotations, and stem word tagging. These
annotations can lead us to adjust the lexical components
of the prototype according to the qualitative and
quantitative layers [1] [4] of its knowledge model.
Thereafter, evaluation on knowledge representation of the
prototype demands both development of domain-specific
ontology and a generic framework of semantic relations
in the domain. The evaluation helps developing a

70 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
representative knowledge representation tool for the
domain DC electrical circuit.
In this paper, we proposed a stochastic development
procedure of a domain-specific representative corpus that
is used to evaluate two major components of the TKM
prototype. We presented detail procedure of corpus-based
evaluation of an NLP system- that includes enriching the
lexicon and morphological database, testing the parsing
ability of the prototype, and the adjustment of the lexical
components according to the linguistic information in the
corpus. We also developed ontology according to the
human conceptualization. As successful knowledge
representation depends on predicate and semantic
relations in the text, we developed frameworks for
semantic relations with which any NLP system can read
and realize text in the domain. We evaluated the coverage
of discourse by the TKM prototype with a topical and
discourse analysis on the corpus.
The remainder of this paper is organized as follows. In
Section II, corpus-based evaluations of various NLP
systems have been discussed. Section III describes the
proposed procedure of representative corpus development
and annotations. Section IV describes the evaluation of
lexical components of the TKM prototype such as the
vocabulary and grammar structure. Section V contains
the outline of developing an ontology and framework for
semantic relations. The section also includes the
rhetorical and topical analysis. Section VI concludes the
paper.
II. RELATED WORK
A text based domain-specific NLP system can be
evaluated according to the type, context or discourse of
text from the domain although no established agreement
has been developed on test sets and training sets [5].
Corpus is not restricted today only for researches on
linguistics [6]; it is now becoming the principal resource
to evaluate such domain-specific NLP systems. Many
NLP systems like Saarbrucker Message Extraction
System (SMES) [8] have been tested with a corpus as
proper evaluation depends on a representative test set of
data like corpus [7]. Corpus contains structured and
variable but representative text. A corpus is said
representative if the findings from it can be generalized to
language or a particular aspect of language as a whole
[3]. Corpus-based evaluations like MORPHIX [9] and
MORPHIX++ [7] showed that the evaluation with a
representative corpus results in proper adjustments.
MORPHIX++ was tested with a corpus and systematic
inspection revealed some necessary adjustments like
missing lexical entries, discrepant morphology
incomplete or erroneous single words.
NLP systems use either pre-defined or customized
grammar rules. For instance, the lexical components of
the TKM prototype use Combinatory Categorical
Grammar (CCG) [10]. The prototype follows some
specific clausal and phrasal structures according to CCG.
As it follows a particular grammar, we need to adjust the
grammar and phrasal structures according to the
structures of text from the domain. For example, TKM
© 2010 ACADEMY PUBLISHER
prototype, on its early test, was able to parse simple
sentences only [35]. This becomes a drawback if majority
of text in the domain is written in compound and complex
sentences. Therefore, necessary adjustment on CCG can
let the prototype parse compound and complex sentences
as well. In addition, NLP systems may recognize specific
clausal and phrasal structure which maybe absent in
domain-specific text. For example, if an NLP system uses
grammars that handle one subject and one object, both
parsing and knowledge extraction from domain-specific
text becomes difficult if majority of the text contains
more than one subject and one object. These linguistic
properties of domain-specific text bring in the issue of
adjustment. The lexicographical resources of such
systems can be increased by analyzing linguistic patterns
in domain-specific corpus. Statistical data like frequency
of words, number of simple, complex or compound
sentences, number of subject and object present in the
sentences assist to adjust the lexical components of the
systems. The grammar structure MORPHIX++ supported
was not efficient in its early days. It was adjusted and
extended according to the corpus used as its test suite.
The text in the corpus sometimes conveys ambiguity to
a knowledge mapping prototype if its knowledge model
differs from human cognition. For a sentence a resistor is
both a circuit component and a diagrammatic
representation, the role of a resistor is a component in
physical connection or a component in diagram. To
differentiate between them, the machine has to
conceptualize the domain like human. We need semantic
relations in text to conceptualize the domain. If a
knowledge model is developed with domain-specific
semantic relations, then machine identifies the proper role
played by a concept in the domain. Semantic relations for
a large domain can be obtained by developing conceptual
structure of the domain with concept maps as it represents
both textual and semantic relations graphically [11].
A team at Information Sciences Institute of University
of Southern California was working on computer-based
authoring. They suffered for an unavailability of a theory
of discourse structure. Responding to this, Rhetorical
Structure Theory (RST) was developed out of studies of
edited or carefully prepared text from a wide variety of
sources. It now has a status in linguistics that is
independent of its computational uses [29]. RST is an
approach to the study of text organization which
conceptualizes in relational terms a domain within the
semantic stratum [30]. After the formulation of RST in
the 1980s, it becomes an emerging area of research for
computational linguistics. It eventually draws the
attention of researchers in natural language processing.
Discourse analysis helps understanding the behaviour
of a domain-specific NLP system in its discourse. Corpus
is a strong source of discourse analysis as linguistic and
semantic relations confined in it play important role to
manifest, adjust and extend systems to attune with its
discourse. Researches like [31], [32], and [33]
incorporated corpus in discourse analysis where
emphases were given on finding linguistic relations,
manual annotation and correlation of discourse structures.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 71
III. CORPUS DEVELOPMENT
In this section, we will discuss regarding the
development approach of a domain-specific corpus, proof
or its representativeness, and its annotation procedure.
A. Development Approach
As we did not find any corpus for the domain DC
electrical circuit with extensive web searches, we
initiated WebBootCaT [12] to develop a representative
corpus. We developed five corpora using the
WebBootCaT and analyzed them by comparing the
number of distinct domain-specific terms and number of
distinct words present. The significant difference between
these two numbers and inconsistency on the size of the
corpus in Figure 1 state that web-based tools are not
usable to develop domain-specific corpora.
Figure 1. Inconsistency of WebBootCat to develop domain-specific corpus.
Therefore, we decided to develop the corpus manually
and collected text from 141 web resources containing
1,029 sentences and 18,834 words. During the
development, we left the non textual information (e.g.,
equations and diagrams) as the TKM prototype operates
only on text.
B. Representativeness of the Corpus
The representativeness of the corpus can be justified
with a notion of saturation or closure described by [13].
© 2010 ACADEMY PUBLISHER
At the lexical level, saturation can be tested by dividing
the corpus into equal sections in terms of number of
words or any other parameters. If another section of the
identical size is added, the number of new items in the
new section should be approximately the same as in other
sections [14].
Figure 2. Representativeness of the corpus with technical terms, verbs, prepositions and coordinators

72 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
To find out the representativeness for the corpus, it has
been segmented into 15 samples. Each sample is
comprised of 1,267 words on average. We plotted the
cumulative frequency of the most frequent technical
terms in the samples.
Figure 2 depicts that the presence of the domainspecific
technical terms becomes stationary after a few
samples. This is one of the criteria showing the
representativeness of the corpus. After a certain point, no
matter how much text we add to the corpus, the
frequencies of the terms are becoming stationary.
Similarly, we counted the frequency of non-technical
words in the corpus and grouped them according to their
parts of speech. Statistics on verbs, prepositions and coordinators
in Figure 2 show that the corpus has been
saturated after sample 11.
We also counted the frequency of types of sentences in
the corpus. As the domain contains instructional text and
most of which are simple sentences, it needs to be
reflected on the corpus as well. Figure 3 shows that
majority of the text is simple sentence (in percentage).
Figure 3. Sentence structure in the corpus
C. Corpus Annotation
To annotate the corpus with POS tags, Cognitive
Computation Group POS tagger [15] has been used as it
works on the basis of learning techniques like Sparse
Techniques on Winnows (SNOW). The corpus is
annotated with nine parts of speech include noun,
pronoun, verb, adverb, adjective, preposition,
coordinator, determiner, and modal. The phrasal structure
of the corpus has been annotated by the slash-notation
grammar rules defined by CCG. We developed an XML
version of the corpus with seven tags.
IV. EVALUATION OF LEXICAL COMPONENTS
The evaluation of vocabulary and the grammar
structure of the prototype are illustrated in this section.
This section also refers to the efficiency in parsing and
richness of lexical entries of the prototype.
A. Evaluation of Vocabulary
The lexicon of the prototype is mapped on the unique
words of the corpus. The words present both in the
morphology and in corpus are called the vocabulary of
the prototype. Initially, only five percent of the
vocabulary was covered by the prototype (Table I).
© 2010 ACADEMY PUBLISHER
TABLE I.
PRELIMINARY VOCABULARY COVERAGE OF THE TKM PROTOTYPE
Words in Morphology Unique Words in Vocabulary
and in Corpus
the Corpus Coverage
101 1,902 5%
MORPHIX++, a second generation NLP system,
covered 91 percent of word in the corpus developed to
evaluate it. The reason behind this difference is the
augmentation of the vocabulary of MORPHIX++ ran
parallel with the development of the system where the
main focus in case of TKM prototype was to develop an
operational system first rather than increasing its
vocabulary.
We used the POS tags of the corpus to populate the
lexicon. We retrieved every distinct word for each
distinct POS from the corpus and we simply added it if
that word was absent in the lexicon. The number of added
entries into the lexicon is shown in Table II. On
completion of the process, the vocabulary of the
prototype covers 90 percent of the corpus (Table III).
TABLE II.
AUGMENTATION OF LEXICAL ENTRIES IN THE TKM PROTOTYPE
POS Augmented Entries
Determiner 19
Coordinator 5
Noun and Pronoun 2,094
Adjective 364
Preposition 71
Adverb 177
Verb 264
TABLE III.
VOCABULARY OF THE TKM PROTOTYPE
Words in Morphology Unique Words in Vocabulary
and in Corpus the Corpus
Coverage
1,783 1,902 90%
B. Evaluation of Grammar
The TKM prototype struggles to parse modals or
auxiliary verb because CCG does not provide any
specification to categorize modals into finite and nonfinite
[16]. We defined grammar formalisms for modals
and adjusted the lexicon that increased the ability of the
prototype to parse modals.
CCG does not have any mechanism for phrasal
structures like adjective–adjective–noun although
researches showed that numerous adjectives can be
placed before a noun [17]. Except the regular adjectives,
we defined grammar formalisms for noun equivalents
(e.g., two common types of circuits), participle equivalent
(e.g., the connected wire), gerund equivalents (e.g., the
conducting material), and adverb equivalents (e.g., the
above circuit is series circuit) of adjectives that increased
the rate of parsing adjectives.
CCG is unable to parse sentences that start and end
with a prepositional phrase [18]. For example, in series
circuit, the current is a single current- this sentence is not

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 73
parsed by CCG. In contrast, the current is a single
current in series circuit- is sometimes parsed by CCG.
The lexicon the prototype is using has nine different types
of prepositions. Sometimes, it is difficult to even identify
regular prepositions. For instance, the sum of potential
differences in a circuit adds up to zero voltage- Though
in regular grammars, up is not treated as adverbs- these
are called particles where prepositions have no objects
and require specific verbs with them (e.g., throw out, add
up). The parsing ability of the prototype increased as we
defined grammar rules for such prepositions.
Complementizer, although it is a form of preposition, it
is not recognized by CCG. Adverbs, on the other hand,
have a strong coverage by CCG. In many cases, adverbs
sit at the end of the sentence- CCG does not provide any
category to define these adverbs although it has fully
featured adverb categories for other two positions of an
adverb in sentence- adverbs that start a sentence or that
sit in the middle of a sentence. These issues have been
resolved by adding new grammar rules.
The lexicon has two categories for coordinatorssitting
at the beginning of a sentence (e.g., since, as) and
relating two clauses (e.g., and, or). CCG defined that they
can be in the middle of two noun phrases only with
np\np/np but the sentence series and parallel circuits are
the types of circuits has the category n\n/n rather than
np\np/np. CCG handles adverbs and conjunctions well
but it seriously lags in handling sentences having similar
verbs as in the sentence the sum of current flowing into
the junction is eventually equal to the sum of current
flowing out of the junction. The identical verbs flowing
(gerund) appear twice with another verb (be) is
concerning. Moreover, a verb has to be present in a
sentence to form predicate argument structure but we
discovered that there are sentences which do not have any
verbs- the bigger the resistance, the smaller the current.
Gerund of verb is known as noun. Gerund is formed by
placing ing at the end of the verb. For example, current
flowing into a junction is equal to the current flowing out
of the junction- in this sentence, flowing is a gerund.
Gerunds are not treated as nouns in CCG. In other words,
gerunds, if treated as nouns in CCG, the sentence
struggles to be parsed.
After creating grammar rules and phrasal structures
and adding them into the lexicon and morphology of the
prototype, the parsing ability of the prototype increased to
31 percent (Table IV). Although the prototype was tested
with a limited dataset, it was unable to parse any sentence
from the corpus before the evaluation.
TABLE IV.
AUGMENTATION OF LEXICAL ENTRIES IN THE TKM PROTOTYPE
State of the
Prototype
Total
Sentences
Parsed
Sentences
Efficiency
Preliminary 1,029 0 0%
Evaluated 981 300 31%
We analyzed the 300 sentences parsed by the prototype
and figured out the number of subject, object and verb
they consist. In Figure 4, we see that the prototype works
well when the number of subjects and objects in a
© 2010 ACADEMY PUBLISHER
sentence do not exceed two and when the number of
verbs does not exceed one.
The inefficiency of the prototype to parse sentence is
due to the absence of phrasal structures (hence, the
categories). 69 percent of the sentences in the corpus
have phrasal structures that are not supported by the CCG
structure. It should be noted that the prototype fails to
parse sentences even for absence of just one category. For
example, One simple DC circuit consists of a voltage
source (battery or voltaic cell) connected to a resistor –
this sentence is not parsed by the prototype for the
absence of category of conjunction or (np\n/np) and for
the category of verb connected (s\np/pp). In the corpus,
these absent categories are identified so that modification
of the lexicon becomes easier.
Figure 4. Number of subjects, objects, and verbs in the
sentences parsed by TKM prototype.
The inefficiency of the prototype to parse sentence is
due to the absence of phrasal structures (hence, the
categories). 69 percent of the sentences in the corpus
have phrasal structures that are not supported by the CCG
structure. It should be noted that the prototype fails to
parse sentences even for absence of just one category. For
example, One simple DC circuit consists of a voltage
source (battery or voltaic cell) connected to a resistor –
this sentence is not parsed by the prototype for the
absence of category of conjunction or (np\n/np) and for
the category of verb connected (s\np/pp). In the corpus,
these absent categories are identified so that modification
of the lexicon becomes easier.
V. EVALUATION OF KNOWLEDGE MODEL
In this section, we will discuss the procedure of
developing a domain-specific ontology and framework
for semantic relations. The results of rhetorical, topical,
and discourse analysis are also outlined in this section.
A. Ontology and Framework for Semantic Relations
From the 300 parsed sentences, the prototype is able to
map only 10 percent of the sentences effectively on its
pre-built ontology. We investigated the ontology and
found that it was not developed according to a
representative data set like our corpus. We decided to
develop ontology for the domain-specific corpus that
helps to adjust the knowledge model of the TKM
prototype.
We developed the ontology in a similar way human
conceptualizes a domain. In conjunction with the

74 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
development of the ontology for the domain, we
developed a framework for semantic relations. The
framework is built upon the framework proposed by
FACTOTUM thesaurus [19] [20]. These semantic
relations help to represent hierarchical knowledge apart
from predicate information.
We conceptualized every sentence in the corpus
manually. The outcome of the conceptualization led us to
develop concepts and relations among them and
graphically represented them as concept maps with Cmap
Tools [21].
To illustrate this procedure, for the sentence One
simple DC circuit consists of a voltage source (battery or
voltaic cell) connected to a resistor, we firstly
conceptualized the sentence in the following manner-
1. DC circuit has voltage source as its component.
2. Battery and voltaic cell are voltage sources.
3. Battery and voltaic cell have similarity.
4. Voltage source can be connected to resistor.
5. DC circuit has resistor as its component.
6. As they all are satisfying the properties of a circuit,
DC circuit is a type of circuit.
We used this information to develop base level concept
maps that represent the predicate relations in the text. To
develop higher level concept maps, we require to group
concepts and to find relations among the groups. For this
particular sentence, we defined groups named circuit and
circuit component. We assigned DC Circuit and Circuit
to the group Circuit and the rest of the concepts to the
group circuit component. We can also find a relation
between these two groups- circuit is made of circuit
components. For a sentence Resistors in the diagram are
in parallel- the concept resistor would be assigned to
group of concepts called Diagrammatic Notation rather
than Circuit Components. This process of grouping the
concepts from the base level concept maps and finding
relations among them produced four levels of concept
maps for the corpus. The conceptual structure of the
domain is comprised of all these concept maps resulted
from human conceptualization at four different levels.
The predicate relations in the sentence are as follows-
1. DC Circuit Have Component Voltage source
2. Battery Type Of voltage source
3. Voltaic cell Type Of voltage source
4. Battery Is Voltaic Cell
5. Voltage Source Connected To Resistor
6. Battery Connected To Resistor
7. Voltaic Cell Connected To Resistor
8. DC Circuit Have Component Resistor
9. DC Circuit Type of Circuit
These relations are then analyzed to initiate developing
the framework for the semantic relations in the text. The
analysis provides us the following semantic relations-
1. Relation which describes parts that are physically
related (e.g., Have Component)
2. Relation which describes hyponymy (e.g., Type Of),
and synonymy (e.g., Is) that are similar
3. Relation which describes hierarchy or class (e.g.,
Type Of)
© 2010 ACADEMY PUBLISHER
4. Relation which describes spatial relations
(specifically location of objects) (e.g., Connected To)
As we represent knowledge by conceptualization
followed by mapping linguistic information on
knowledge model, it will allow the prototype to map
knowledge from the text onto the ontology efficiently.
For example, the prototype now can provide the user
knowledge like voltage source is a physical part of the
DC circuit- which is not stated in the sentence literally
but semantically.
As we developed conceptual structure for the corpus
with the Cmap Tool, the total number of concepts and
relations increases but number of new concepts and
relations decreases. On completion, the number of
concepts and relations are plotted against the corpus size.
Figure 5 shows the cumulative increment of the number
of concepts and relations. We see a plateau showing that
the number of concepts and relations are becoming
stationary.
Figure 5. Graph to show that the number of concepts and
relations in the corpus is becoming stationary.
We also plotted number of new concepts and relations
against the corpus size (Figure 6). The plateau in Figure 6
shows that the number of new concepts and relations are
becoming stationary. These two observations led us to a
decision that if we put semantic relations in the corpus
into a framework, then it will be representative.
Figure 6. Graph to show that the number of new concepts
and relations in the corpus is becoming stationary.
We found 97 predicate relations and 166 concepts in
the corpus and we developed Tier 2 of our framework
(Table V) to support these relations. Afterwards, we

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 75
grouped level 0 concepts and relations to produce level 1
of concept maps. As we came across new predicate
relations, we created Tier 1of our framework to support
the semantic relations in Tier 2. These two tiers of
semantic relations comprise the domain-specific
framework for semantic relations and can be supportive
to all the predicate relations of the domain. In essence,
the level 0 concept maps have the predicate relations and
TABLE V.
FRAMEWORK FOR SEMANTIC RELATIONS IN THE CORPUS
the semantics conveyed by them are supported by
relations in Tier 2. Predicate relations in level 1 and level
2 concept maps are supported by Tier 1 semantic
relations. The ontology along with the concept maps is
depicted in [34].
Relation Category
Tier 1 Semantic
Relations
Tier 2 Semantic Relations
Predicate Relations Inverse Predicate Relations
Hierarchy Have type Type of
Physically Related
Parts
Constituent Material
Have component
Make, Produce
Component of
Made of, Produced by
Take place between, Direction of
Location of Objects Connected to, Flows
Spatial Relations
through, Have direction
Location of Activities
Transfer, Find, Divide,
Commence from
Transferred by, Found by,
Divided by, End to
Affect, Cause, Vary in, Affected by, Caused by,
Effect/ Partial Cause Resist, Force, Limit, Resisted by, Forced by,
Opposite to, Related to Limited by
Causally/
Functionally Related
Production/ Generation
Destruction
Manifestation
Produce
Collide, Melt
Represent
Produced by
Collided by, Melted by
Represented by
Conversion
Convert, Convertible to Converted by, Convertible
from
Predicate Relations
Carry, Measure, Supply, Carried by, Measured by,
Instrumental
Function/ Usage
Functions
Share, Depend on,
Protect, Absorb
Supplied by, Shared by,
Depended by, Protected by,
Absorbed by
Use Use, Do not use Used by, Not used by
Human Role Deal with Dealt by
Topic Govern Governed by
Representation
Represent, Characterize Represented by,
Characterized by
Conceptually Related
Have state, Have unit, State of, Unit of, Source of,
Property
Have source, Have
Magnitude, Have
Terminal
Magnitude of, Terminal of
Similarity
Synonymy
Hyponymy
Is, Referred to
Have type
Is
Type of
Quantitative Relations Numerical Relations
Proportional, Inverse proportional to, Gain, Lose, Do
not gain, Do not lose
Instantiation Have instance Instance of
Extension Have Extension Extension of
B. Rhetorical Analysis
To find the stereotypical relations in the domain, RST
proposed by Mann & Thompson [22] is used as a
descriptive tool. Research work like Rosner & Stede [23]
and Vander Linden [24] also used this framework for the
rhetorical analysis in their corpus. We used a framework
based on the work of Hunter [25] who outlines the
structural model of content of information for second
language learning materials proposed within the frame of
machine-mediated communication [26]. This framework
© 2010 ACADEMY PUBLISHER
defined text structures, textual expressions and
information structures within domain-specific text.
One common characteristic of expository text is that
they use text structures. Text structures refer to the
semantic and syntactic organizational arrangements used
to present written information. Text structure used in the
analysis includes introduction, background,
methodologies, results, observations, and conclusions.
Textual Expressions are relations that describe the
nature of a sentence at phrase level. It eventually outlines
the type of the sentence. These all are mononuclear
relations- the relations do not depend on the semantic of

76 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
the adjacent sentences. We used the following textual
expressions for the analysis- common knowledge, cite,
report, explanation, claim, evaluation inference, and
decision.
Information structures are used at both phrase level
and sentence level in the analysis. We analyzed the
meaning of the sentence and its impact on other
juxtaposed sentences with relations like description,
classification, comparison, sequence, cause-effect, and
contrast.
We used RSTTool [27] to annotate the corpus with the
rhetorical relations. The procedure shows that the corpus
has 2,701 relations grouped into 19 rhetorical relations
(Table VI).
TABLE VI.
RHETORICAL RELATIONS IN THE CORPUS
Higher means of relations background (13 percent) and
observations (7 percent) are significant as the corpus
contains instructional text and instructional text mostly
describes background and observation of events [28]. The
analysis also shows that most of the text are descriptive
(30 percent) and presented as report (20 percent) thus
proved the representativeness of the corpus in case of
containing semantic relations. Qualitative layer of the
prototype deals with the causal relationship between
concepts and a significant amount of cause-effect relation
(2.14 percent) is of particular interest for us to deal with.
We found that the prototype is able to map about 70
percent of the causal relations in the text.
Rhetorical Structures Rhetorical Relations Appearance Mean
Text Structures Introduction 117 4.33%
Background 356 13.19%
Methodologies 60 2.22%
Results 51 1.89%
Observations 180 6.66%
Conclusions 42 1.55%
Textual Expressions Common Knowledge 74 2.74%
Report 545 20.18%
Explanation 192 7.11%
Claim 85 3.15%
Evaluation 3 0.11%
Inference 13 0.48%
Decision 59 2.18%
Information Structures Description 817 30.25%
Classification 13 0.48%
Comparison 25 0.93%
Sequence 34 1.26%
Cause-effect 58 2.14%
Contrast 17 0.63%
Total 2,701 100%
C.Topical Analysis
We intend to analyze the topical progression of the
corpus as the prototype both handled and failed to handle
text on various topics. The analysis will help us to
determine the context and discourse awareness of the
prototype. The prototype is not developed to parse and
map text of any particular topic or context and it should
represent the whole domain. Since we have the
representative corpus, the topical analysis of the
prototype can help us understand the topical coverage of
its context and discourse.
First, we annotated the corpus with three types of
topical progressions- parallel progression, sequential
progression, and extend parallel progression. As the topic
in the text progresses onwards, we indented the text of the
corpus according to the type of progression it belongs to.
For example, indentation is the starting topic,
indentation is the sequential topic originated from
, indentation is the sequential topic originated
from , and indentation is the extended parallel
topic of (Figure 7). On completion, we found six
indentations of topical progression in the corpus.
Putting more resistors in the parallel circuit decreases the total resistance because the electricity has additional branches to flow along and so the
total current flowing increases.
This is very useful because it means that we can switch the lamp on and off without affecting the other lamps.
The brightness of the lamp does not change as other lamps in parallel are switched on or off.
For this reason, lamps are always connected in parallel.
© 2010 ACADEMY PUBLISHER
Figure 7. Annotation of the corpus with topical progression.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 77
Second, we counted total number of sentences in each
indentation of the corpus. As we expected, indentation 1
covers the most of the corpus and indentation 6 has the
least number of sentences. We also counted number of
sentences TKM prototype handled in each indentation to
find its topical coverage. The more the topic progresses
away from the context, the possibility of not
understanding the context increases but the prototype
TABLE VII.
TOPICAL ANALYSIS OF THE CORPUS AND TKM PROTOTYPE
showed that even if the topic is six indentations away
from the original context, it can represent the knowledge
(Table VII). The prototype efficiently handled language
and knowledge on topics that are four, five, and six
indentations away from the original context with 38, 36,
and 33 percent coverage, respectively. However, topics
nearer to the starting context are covered relatively low
with 25 and 26 percent.
Indentation Number of Sentences Corpus Coverage Number of Sentences handled by Topical Coverage by
the Prototype
the Prototype
1 641 66% 197 31%
2 259 22% 65 25%
3 86 7% 22 26%
4 26 3% 10 38%
5 11 1% 4 36%
6 4 1% 2 33%
D.Discourse Analysis
The discourse of the prototype contains high level
concepts developed during the progress of ontology. High
level concepts are those that are related with Tier 1 and
Tier 2 semantic relations of our framework and convey
knowledge rather than predicate information. According
to our research, the domain has 12 high level concepts
shown in Table VIII. The table is organized in
descending order according to the number of concepts in
discourse. We semi-automatically analyzed the corpus
and found that the high level concepts of the ontology are
present 4,120 times in the corpus- this is the discourse of
the prototype. Moreover, we also found that the high
level concepts of the domain are present 969 times in the
sentences that the prototype can handle- which is the
discourse covered by the prototype. If we divide the
TABLE VIII.
DISCOURSE ANALYSIS OF THE TKM PROTOTYPE
discourse coverage of prototype by the total number of
concepts in the discourse, then we will find the discourse
covered by the prototype. In this case, Table VIII shows
that the discourse coverage of the TKM Prototype is 24
percent.
If we consider individual high level concepts, then
Units and Measuring Instruments are the areas of
discourse the prototype covers, mostly, with 48 percent of
coverage. Rules is next to them with 28 percent of
coverage. The prototype covers only 15 percent of the
discourse of Electrical Process though the discourse is
significant in the domain. Environmental Factors, a
narrower high level concept, is next to it in case of less
discourse coverage by the prototype.
High Level Concepts Coverage of Prototype Concepts in Discourse Difference with Discourse
Discourse
Coverage Deviation
(1) Electrical Quantity 335 1433 1098 24% 77%
(2) Circuit Components 154 685 531 23% 78%
(3) Diagrammatic Notation 94 482 388 20% 81%
(4) Electrical Process 63 442 379 15% 86%
(5) Electrical Device 83 313 230 27% 74%
(6) Units 100 211 111 48% 53%
(7) Atomic Level 31 161 130 20% 81%
(8) Circuits 30 140 110 22% 79%
(9) Environmental Factors 20 110 90 19% 82%
(10) Measuring Instrument 46 96 50 48% 53%
(11) Rules 13 47 34 28% 73%
(12) Materials 4 21 17 20% 81%
Total 969 4,120 3,151 24% 77%
We also analyzed the deviation of the prototype from
the discourse. First, we measured the difference of the
coverage of the prototype and coverage in discourse.
Then, we measured the deviation- the difference with
© 2010 ACADEMY PUBLISHER
discourse divided by the concepts in discourse. This
deviation is the measure of unawareness in discourse-
how much of the discourse the prototype failed to pursuit.
The data show that the prototype is strong to represent
knowledge from the discourse of Units and Measuring

78 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Instrument (both 53 percent). The prototype has the
overall deviation from the discourse of 77 percent- means
its discourse awareness is 23 percent.
We plotted the presence of high level concepts in the
discourse and the coverage of the discourse by the
prototype in Figure 8. The difference between the
coverage of discourse by the prototype and the discourse
itself is depicted with vertical lines. From Figure 8 and
Table VIII, we see that the difference is proportional to
each other from Electrical Quantity to Electrical
Processes and then a sudden rise in case of Electrical
Device and Units indicates that most of the simple
sentences in the corpus are situated in this area. Another
smooth maintenance of difference between the discourse
and the coverage of discourse is manifested from
concepts Atomic Level to Environmental Factors. The
prototype shows efficiency in knowledge representation
in Measuring Instruments that indicates to the possibility
of having understandable knowledge for the prototype
lies in this discourse with suitable linguistic and semantic
arrangement.
Figure 8. Difference between the discourse and the
coverage of discourse by the prototype.
VI. CONCLUSIONS
In this paper, we presented a corpus-based evaluation
of lexical components and knowledge model of a
domain-specific Text to Knowledge Mapping prototype.
We developed a domain-specific corpus and proved its
representativeness in linguistic elements with stochastic
approach and its soundness in semantic features with
rhetorical analysis. The representative corpus, with
enriched multimodality, can be used as a reference in text
summarization, for context and discourse analysis, and
for developing ontology. The linguistic resources of the
corpus have been used to evaluate and adjust lexical
components of the prototype like vocabulary and
grammar. This evaluation led the prototype to parse
reasonable amount of domain-specific text. During
evaluation on knowledge model, we developed a domainspecific
ontology and a framework for semantic relations
associated with it. We conducted topical and discourse
analysis on the prototype to see its context awareness and
the performance of the prototype is satisfactory.
However, limited conceptual acquisition of the prototype
© 2010 ACADEMY PUBLISHER
refers to limited knowledge representation and demands a
framework for domain-specific linguistic relations.
Using the domain-specific corpus, a generic corpus
parsing and lexical component analysis tool is developed
[36] that extracts lexical information from any XML
corpus and store the information in database. The corpus
also contributed in domain-specific text summarization
and the result of summarization was satisfactory [37].
REFERENCES
[1] W. Ou and A. Elsayed, “A knowledge-based approach for
semantic extraction”, International Conference on
Business Knowledge Management, Macao, 2006.
[2] T. Declerck, J. Klein, and G. Neumann, “Evaluation of the
NLP components of an information extraction system for
German”, Proceedings of the first international
Conference on Language Resources and Evaluation
(LREC), Granada, 1998, pp. 293-297.
[3] D. Evans, “Corpus building and investigation for the
humanities”, Available:
http://www.humcorp.bham.ac.uk/humcorp/information/cor
pusintro/Unit1.pdf [28 January 2009].
[4] K. D. Forbus, “Qualitative process theory: Twelve years
after”, Artificial Intelligence, vol. 59, no. 1-2, 1993, pp.
115-123.
[5] M. Palmer and T. Finin, “Workshop on the evaluation of
natural language processing systems”, Computational
Linguistics, vol. 16, no. 3, 1993, pp. 175-181.
[6] T. McEnery and A. Wilson, Corpus Linguistics,
Edinburgh: Edinburgh University Press, United Kingdom,
1996.
[7] J. Klein, T. Declerck, and G. Neumann, “Evaluation of the
syntactic analysis component of an information extraction
system for German”, Proceedings of the 1st International
Conference on Language Resources and Evaluation,
Granada, Spain, 1998, pp. 293-297.
[8] G. Neumann, R. Backofen, J. Baur, M. Becker, and C.
Braun, “An information extraction core system for real
world German text processing”, Proceedings of the 5th
Conference on Applied Natural Language Processing
(ANLP), USA, 1997, pp. 209-216.
[9] W. Finkler and G. Neumann, “Morphix: A fast Realization
of a classification–based approach to morphology”,
Proceedings der 4. Österreichischen Artificial Intelligence
Tagung, Wiener Workshop Wissensbasierte
Sprachverarbeitung, Berlin, Germany, 1988, pp. 11-19.
[10] S. Clark, M. Steedman, and J. R. Curran, “Objectextraction
and question-parsing using CCG”, Available:
http://www.iccs.inf.ed.ac.uk/~stevec/papers/emnlp04.pdf
[28 January 2009].
[11] N. Nathan and E. Kozminsky, “Text concept mapping: The
contribution of mapping characteristics to learning from
texts”, Proceeding of the 1 st International Conference on
Concept Mapping (CMC2004), Pamplona, Spain, 2004.
[12] M. Baroni, A. Kilgarriff, J. Pomikálek, and P. Rychlý,
“WebBootCaT: A web tool for instant corpora”,
Proceeding of the EuraLex Conference, Italy, 2006, pp.
123-132.
[13] T. McEnery, R. Xia, and Y. Tono, Corpus-Based
Language Studies: An Advanced Resource Book, London:
Routledge, 2006.
[14] D. Biber, “Representativeness in corpus design”, Literary
and Linguistic Computing, vol. 8, no. 4, 1993, pp. 243-257.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 79
[15] University of Illinois at Urbana-Champaign, “The SNoW
learning architecture”, Available:
http://l2r.cs.uiuc.edu/~danr/snow.html [28 January 2009].
[16] A. S. Hornby, Guide to Patterns and Usage in English, 2 nd
Edition, Oxford University Press, Delhi, 1995, pp. 1-2.
[17] M. A. Covington, Natural Language Processing for
Prolog Programmers, 1 st Edition, Prentice Hall, 1993, pp.
88-90.
[18] M. Steedman and J. Baldridge, “Combinatory categorical
grammar”, Unpublished Tutorial Paper, 1993.
[19] Z. Chen, Y. Perl, M. Halper , J. Geller, and H. Gu,
“Partitioning the UMLS semantic network”, IEEE
Transactions on Information Technology in Biomedicine,
vol. 6, no. 2, 2002.
[20] Micra, Inc., “FACTOTUM”, Available:
http://www.micra.com/factotum/ [25 January 2009].
[21] A.J.Cañas, G. Hill, R. Carff, N. Suri, J. Lott, G. Gómez,
T.C. Eskridge, M. Arroyo, and R. Carvajal, “CmapTools:
A knowledge modelling and sharing environment”,
Proceeding of the 1 st International Conference on Concept
Mapping, Pamplona, Spain, 2004.
[22] W. C. Mann and S. A.Thompson, Rhetorical Structure
Theory: Toward a functional theory of text organization,
Text 8 (3), 1988, pp. 243-281.
[23] D. Rosner and M. Stede, “Customizing RST for the
automatic production of technical manuals”, Aspects of
Automated Natural Language Generation, Lecture Notes
in Artificial Intelligence, Trento, Italy, 1992, pp. 199-214.
[24] K. V. Linden, “Speaking of actions: Choosing rhetorical
status and grammatical form in instructional text
generation”, Ph. D. thesis, University of Colorado, 1993.
[25] L. Hunter, “Dimensions of media object
comprehensibility”, 7th IEEE International Conference on
Advanced Learning Technologies (ICALT 2007), Niigata,
Japan, 2007, pp. 925-926.
[26] A. Elsayed, “Machine-mediated communication: The
technology”, 6 th IEEE International Conference on
Advanced Learning Technologies (ICALT 2006), Kerkrade,
The Netherlands, 2006, pp. 1194-1195.
[27] D. Marcu, “RST annotation tool”, Available:
http://www.isi.edu/licensed-sw/RSTTool/ [27 January
2009]
[28] L. Kosseim and G. Lapalme, “Choosing rhetorical
structures to plan instructional texts”, Computational
Intelligence, vol. 16, no. 3, 2000, pp. 408-445.
[29] W. C. Mann and M. Taboada, “Rhetorical Structure
Theory: Looking back and moving ahead”, Discourse
Studies, vol. 8, no. 3, 2006, pp. 423-459.
[30] V. Stuart-smith, “The hierarchical organization of text as
conceptualized by rhetorical structure theory: A systemic
functional perspective”, Australian Journal of Linguistics,
vol. 27, no. 1, 2007, pp. 41-61.
[31] R. Williams and E. Reiter, “A corpus analysis of discourse
relations for natural language generation”, Proceedings of
Corpus Linguistics, Lacaster, United Kingdom, 2003, pp.
899-908.
[32] D. Marcu, M. Romera, and E. Amorrortu, “Experiments in
constructing a corpus of discourse trees”, University of
Maryland, 1999, pp. 48-57.
[33] B. Webber, “D-LTAG: Extending lexicalized TAG to
discourse”, Journal of Cognitive Science, vol. 28, 2004, pp.
751-779.
[34] R. Shams and A. Elsayed, “Development of a conceptual
structure for a domain-specific corpus”, 3 rd International
Conference on Concept Maps (CMC 2008), Estonia and
Finland, 2008.
© 2010 ACADEMY PUBLISHER
[35] R. Shams and A. Elsayed, “A Corpus-based evaluation of
lexical components of a domain-specific text to knowledge
mapping prototype”, 2008 International Conference on
Computer and Information Technology (ICCIT), Khulna,
Bangladesh, 2008, pp. 242-247.
[36] S. Chowdhury, A. S. Shawon, and R. Shams, “CorParse: A
corpus parsing and analysis tool”, Unpublished Tutorial
Paper, 2008.
[37] A. Hossain, S. R. Akter, M. Gope, M. M. A. Hashem, and
R. Shams, “A corpus-dependent text summarization and
presentation using statistical methods”, Undergraduate
Thesis, Khulna University of Engineering & Technology,
2009.
Rushdi Shams was born in Khulna, Bangladesh on January
3, 1985. He pursued his M.Sc. in Information Technology from
the University of Bolton, United Kingdom in 2007 and his B.Sc.
(Engineering) in Computer Science and Engineering from
Khulna University of Engineering & Technology (KUET),
Bangladesh in 2006. He has major fields of study like intelligent
systems, internet security, data warehouse, IT management and
artificial intelligence.
He is currently a Lecturer in the Department of Computer
Science and Engineering, KUET. Formerly, he was a Research
Intern at M3C Laboratory, University of Bolton. So far, he has
publications in the area of Ad hoc Networks and Knowledge
Processing. He supervised undergraduate theses on diversified
fields like knowledge processing, wireless sensor networks,
corpus linguistics, and web engineering. Currently, he is
developing frameworks for web 3.0 and acquisition and
machine representation of commonsense knowledge. His
current research interests are Knowledge and Language
Processing, Computational Linguistics, and Wireless Networks.
Mr. Shams is an Associate Member of Institute of Engineers,
Bangladesh (IEB).
Adel Elsayed pursued his Ph.D on Applied Optimal Control
from Loughborough University, UK in 1985. His M.Sc is in
Electronic Control Engineering from Salford University, UK in
1975 preceded by a B.Sc. in Communications Engineering from
Libya University, Libya in 1972. He has a diversified field of
study from signal processing to multimodal communication and
from communications to intelligent systems and humancomputer
interaction.
He joined the University of Bolton towards the end of 2001.
His main objective was to establish a research base by building
on his work on multimodal communication, a new line of
research that he has started few years earlier. His research at
Bolton has started as he set up the “Active Presentation
Technology” Lab (APT Lab). Since then, he diversified into
investigating the underpinning knowledge structures that
support human-machine communication. This led to researching
information structures and their applications. Consequently, the
scope of his work grew out of the narrow area of active
presentation technology into the wider scope of Machine-
Mediated Communication, hence the new name of the research
lab. He has esteemed numbers of conference and journal
publications. Besides these, his research interests are cognitive
tools, knowledge and language processing, and speech
technology.

80 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Dr. Elsayed established M3C as international workshop
attached to well known IEEE International Conference on
Advanced Learning Technologies (ICALT). He acts as guest
editor for special issues of many well known journals as well as
reviewer in number of international conferences.
Quazi Mah- Zereen Akter was born on November 25, 1985
in Dhaka, Bangladesh. She is now completing her M.Sc. in
Computer Science and Engineering from the University of
Dhaka, Bangladesh. She completed her B.Sc. (Engineering) in
Computer Science and Engineering from the University of
Dhaka in 2008. Bioinformatics, artificial intelligence, computer
vision, distributed database systems, digital system design, and
VLSI are her major fields of study.
She is currently working on her Masters thesis with
reversible logic and computational intelligence. Besides
Intelligent Systems, her research interests are Bioinformatics,
Management Information Systems, Reversible Logic, and Logic
Simplification and Minimization.
© 2010 ACADEMY PUBLISHER

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 81
Implementation of Low Density Parity Check
Decoders using a New High Level Design
Methodology
Syed Mahfuzul Aziz and Minh Duc Pham
School of Electrical & Information Engineering, University of South Australia, Mawson Lakes, Australia
Email: {Mahfuz.Aziz, Minh.Pham}@unisa.edu.au
Abstract—Low density parity check (LDPC) codes are
error-correcting codes that offer huge advantages in terms
of coding gain, throughput and power dissipation. Error
correction algorithms are often implemented in hardware
for fast processing to meet the real-time needs of
communication systems. However hardware
implementation of LDPC decoders using traditional
hardware description language (HDL) based approach is a
complex and time consuming task. This paper presents an
efficient high level approach to designing LDPC decoders
using a collection of high level modelling tools. The
proposed new methodology supports programmable logic
design starting from high level modelling all the way up to
FPGA implementation. The methodology has been used to
design and implement representative LDPC decoders. A
comprehensive testing strategy has been developed to test
the designed decoders at various levels. The simulation and
implementation results presented in this paper prove the
validity and productivity of the new high level design
approach.
Index Terms—Error correction coding, digital systems,
digital communication, logic design, FPGA.
I. INTRODUCTION
Information passing through a practical
communication channel may be corrupted in transit by
noise present in the channel [1]. Therefore it is of
paramount importance for communication systems to
have adequate means for the detection and correction of
errors in the information received over communication
channels. Turbo codes and LDPC (low density parity
check) codes are most commonly used for error detection
and correction nowadays [2]. Both of these codes provide
coding gains [3] close to Shannon’s limit [4]. LDPC
codes however outperform turbo codes in terms of coding
gain for large SNR [5, 6]. LDPC code of length 1 million
with a coding rate (ratio of information bits to the sum of
information and parity bits) of 0.5 and BER of 10-6
provides a capacity which is only 0.13db from Shannon’s
limit [5]. Further advantages of using LDPC codes are
Project number: ITEE-09/CGD-12, University of South Australia.
Corresponding author: Dr Syed Mahfuzul Aziz, Email:
mahfuz.aziz@unisa.edu.au
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.81-90
given in [7]. These include less computational complexity
as compared to turbo codes, ability to pipeline the
decoder to increase throughput at the cost of registers and
some latency, and less number of iterations than turbo
codes. Increased number of iterations reduces the
throughput and increases power dissipation [5, 8]. The
studies conducted in [7] and [8] indicate that lesser
number of iterations in LDPC codes helps achieving
higher throughput and decreases power dissipation.
Another factor which contributes to increasing the
throughput of LDPC decoder is the degree of parallelism
which is adjustable [9, 10]. Another reason of using
LDPC codes is that it is relatively easier to implement
than turbo codes [11].
Despite all these advantages of LDPC codes the
random parity check matrix makes the wiring between
variable and check nodes complex, especially for large
matrices. This leads to increased routing congestion in the
decoder and eventually the size of the LDPC decoder
increases and speed decreases [12, 13]. Complexity of
practical implementation makes high throughput very
difficult to achieve [14, 15]. Moreover, designing LDPC
decoders in VHDL (Very high speed integrated circuit
Hardware Description Language) becomes a cumbersome
task when the size of the design increases. Huge amount
of time and effort are required to model such large
designs in VHDL [16]. It results in decrease in
productivity. It becomes a nightmare for the designer to
write VHDL code for thousands of connections and to
make the changes required. Therefore, hardware
implementation of LDPC decoder remains a challenge.
This paper examines high level modelling and
synthesis techniques of LDPC decoders using emerging
industry tools. It compares the high level approaches with
traditional hardware description language based
approaches in terms of modelling complexity, efforts and
time. It also compares the results obtained for a
representative LDPC decoder design using the high level
approaches and using a traditional HDL based approach.
II. DESIGN APPROACH
This paper investigates a new high-level modelling and
synthesis methodology for LDPC decoder using state of

82 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
the art tools. As opposed to using only hand written
VHDL codes for the entire design, this research examines
high level design methods using Simulink, involving a
combination of blocks designed with predefined library
components and with embedded Matlab codes. A VHDL
(VHSIC Hardware Description Language) code is
generated automatically from the high level Simulink
model using Mathworks’s Simulink HDL coder (version
1.2). The entire design process is captured in Fig. 1.
Simulink HDL coder gives the flexibility of integrating
different design approaches, namely Matlab programs,
Simulink models and VHDL codes. Simulink and Matlab
programs provide a much higher level of abstraction than
VHDL. Therefore our design approach utilizes as many
Simulink library blocks and Matlab blocks as possible to
design various modules of the decoder provided these
modules can be successfully processed by the HDL coder
for automatic generation of VHDL codes. All the
modules are then integrated in a top level Simulink model
to produce the overall decoder model. VHDL code is
generated automatically using Simulink HDL coder along
with an optional test bench. This test bench can be used
in ModelSim (a HDL simulator) to check correctness of
the design. The auto-generated VHDL code can also be
used in Altera’s Quartus II or Xilinx ISE (Integrated
Software Environment) for synthesis and implementation
on desired FPGA (Field Programmable Gate Array).
Automatic HDL code generation will lead to reduction in
design efforts thereby increasing design productivity.
Simulink
library blocks
Simulation in
Simulink
MATLAB
Programs
Simulation in
MATLAB
Simulink HDL Coder (Automatic
Code Generation)
Simulation, Analysis & Synthesis
Implementation on FPGA
Figure 1. Proposed design flow.
III. OVERVIEW OF LDPC CODES AND DECODING
ALGORITHM
Low density parity check (LDPC) codes are a class of
linear block codes which are used for error detection and
correction [17]. LDPC codes were first discovered by
Gallager in 1962. In 1981, Tanner worked on LDPC
codes and came up with important tanner graphs or
bipartite graphs [18]. LDPC codes could not be
implemented at the time of invention because the
technology was not advanced, though they were studied
again after almost three decades because they are more
© 2010 ACADEMY PUBLISHER
HDL Coder
VHDL
Programs
Simulation in
Quartus/Xilinx
advantageous than any other error-correcting codes.
LDPC codes are extensively used in standards such as
10Gigabit Ethernet (10GBaseT) & digital video
broadcasting (DVB-S2) [19].
There are different algorithms which could be used for
decoding purposes. We used the min-sum decoding
algorithm [14, 20], which is a special case of sum-product
algorithm [18, 21, 22]. Sum-product algorithm reduces
the computational complexity and makes the decoder
numerically stable [18]. Assume that the messages from
the host communication system are represented by I and
are passed on to the decoder for error correction. The
LDPC decoder consists of a number of variable nodes (v)
and check nodes (c). The operation of the min-sum
algorithm can be summarised as follows [14]:
A. Variable Node Operation
A variable node performs the operation given in (1)
and passes the outputs to check nodes.
L cv = ∑ R + I
mv
m∈
M ( v)
\ c
v
(1)
where, Iv is the input to variable node v, also known as
Log Likelihood ratio (LLR), Lcv is the output of variable
node v going to check node c, M(v)\c denotes the set of
check nodes connected to variable node v excluding the
check node c, Rmv is the output of check nodes going to
variable node v.
B. Check Node Operation
A check node receives messages from variable nodes
and performs the operation given by (2):
R
cv
=
∏
n∈N
( c)
\ v
sign(
L ) × | L | (2)
cv
min
n∈N
( c)
\ v
where, Rcv is the output of check node c going to variable
node v.
C. Parity Check
Every check node also checks whether the parity
condition is satisfied by looking at the sign of the
messages coming from the variable nodes. Until all the
parity checks from all the check nodes are satisfied the
messages are sent back to variable nodes and the variable
nodes do the operation specified in part (A), otherwise
the decoder stops the process.
Min-sum decoding algorithm uses soft decision.
However, hard decision is taken on the new LLR (Iv). If
the new LLR is negative then the output bit would be a 1
otherwise a 0. Interested readers can find the details of
the sum-product and min-sum decoder algorithms in [18,
20-22].
IV. REPRESENTATIVE DECODER DESIGNS
LDPC codes can be represented by an M x N sparse
matrix, usually called H matrix. The H matrix contains
mostly zeros and a small number of 1s [9, 10]. It can also
cv

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 83
be represented by a graph called bipartite or tanner graph.
Tanner graphs contain variable and check nodes. The
decoder prototype designed in this research is based on
the matrix shown in Fig. 2. It is a very small matrix
compared to the real matrices. This 10x5 matrix can be
transformed into a tanner graph having 10 variable nodes
and 5 check nodes as is shown in Fig. 3. The signals go
from the variable nodes to the check nodes and from the
check nodes back to variable nodes. Typically different
sets of wires are used for the signals going from variable
nodes to check nodes and vice versa. The process is
iterative and goes until all the parity checks are satisfied.
The messages in our prototype designs are 4-bit long.
1 1 1 1 0 1 1 0 0 0
0 0 1 1 1 1 1 1 0 0
0 1 0 1 0 1 0 1 1 1
1 0 1 0 1 0 0 1 1 1
1 1 0 0 1 0 1 0 1 1
V1
V2
V3
V4
V5
V6
V7
V8
V9
V10
© 2010 ACADEMY PUBLISHER
Figure 2. A 10x5 sparse matrix.
Figure 3. Tanner graph.
C1
C2
C3
C4
C5
Figure 4. Check node design in Simulink.
A. Design 1
The approach taken in the first decoder design was to
use a combination of embedded Matlab blocks and
Simulink library blocks. No handwritten VHDL blocks
were used. A generic Simulink model was first developed
for each of a variable node and a check node.
Check node design in Simulink: The check node has
been designed in the same way as the variable node. The
Simulink model of the check node is shown in Fig. 4. It
uses blocks from Simulink library (absolute and bitwise
XOR). The check node finds the minimum of all the
inputs and performs the parity checks. It contains a
control block, which controls its operation.
Variable node design in Simulink: The variable node,
shown in Fig. 5, has been designed in Simulink using the
basic add block from the Simulink library. The control
block controls the operation of the variable node. The
variable node has four inputs, the first three come from
various check nodes and the fourth one is the raw LLR,
which is an external input supplied by the host
communication system. It performs the operation given in
(1) and passes the outputs to the check nodes to which it
is connected.
Once the Simulink models are developed for the check
node and the variable node, the VHDL codes for each of
these can be generated automatically using the HDL
Coder tool according to the design flow presented in Fig.
1. The functionality of each of these models can be
verified at both Simulink and VHDL levels. Each VHDL
model can also be synthesized for a target FPGA.
10x5 LDPC decoder: A LDPC decoder with 10
variable and 5 check nodes has been designed using
instances of the variable and check node components
described above. The various nodes are interconnected in
a way that corresponds to the Tanner graph of Fig. 3.

84 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
B. Design 2
Design 2 uses a combination of embedded Matlab
blocks, Simulink library blocks and VHDL/Link for
ModelSim blocks. ‘Link for ModelSim’ is a utility that
enables modules coded in VHDL to be embedded in
Simulink models. Design 2 is different from the first
design in that it uses serial communication of messages
between variable and check nodes. This is achieved using
SIPO (serial-in-parallel-out) and PISO (parallel-in-serialout)
registers at each input and output port in all the
nodes. This greatly simplifies the interconnections by
reducing the number of wires four times at the cost of
some extra registers. The variable and check node
components are same and are connected in the same way
as in Design 1. The PISO and SIPO components are
coded in VHDL and are used in the Simulink model with
the help of the ‘Link for ModelSim’ utility (to link the
VHDL blocks with ModelSim for simulation and code
generation purposes). The way PISO and SIPO are used
in a variable node is shown in Fig. 6. Check nodes have
been modified in exactly the same manner. The inputs
and outputs of all the variable and check nodes are now
1-bit long. The effects of using a VHDL block in a
Simulink design are discussed in the next section. The
HDL coder does not generate the VHDL code of the Link
for ModelSim blocks. The VHDL code for these blocks
needs to be added before synthesizing the code.
SIPO1
SIPO2
SIPO3
Variable
Node
PISO1
PISO2
PISO3
Figure 5. Variable node design in Simulink.
New
Variable
Node with
SIPO &
PISO
Figure 6. A variable node with PISO and SIPO.
© 2010 ACADEMY PUBLISHER
V. RESULTS AND ANALYSIS
A. Convergence Test
The convergence characteristics of the LDPC decoders
presented in this paper have been tested using VHDL
testbench and compared with that of a Matlab code for a
functionally equivalent decoder. For this purpose the
VHDL code automatically generated from the Simulink
models was used. Fig. 7 shows the spread of the number
of iterations for Design 2 by using plots of (mean number
of iterations + standard deviation) and (mean number of
iterations − standard deviation) versus SNR (Eb/No). It is
clear that convergence is achieved in 6 iterations. This is
consistent with the convergence result of a functionally
equivalent decoder shown in Fig. 8, designed using
Matlab code [23].
Figure 7. Spread of the number of iterations for Design 2 obtained
from VHDL testbench.
Figure 8. Spread of the number of iterations for a functionally
equivalent Matlab code.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 85
B. Algorithm Performance
The performance of our LDPC algorithm has been
evaluated by simulating the decoder over AWGN channel
and plotting the Bit-Error-Rate (BER) against Signal-to-
Noise-Ratio (SNR) [24, 25]. Fig. 9 shows the BER
performance of Design 2 obtained from simulation of the
high-level Simulink model. Fig. 9 also shows the BER
plot of the unencoded BPSK channel. Clearly our decoder
demonstrates increasing gain in BER with increasing
SNR compared to the unencoded BER. This proves that
the proposed high-level design method can deliver
competitive designs with the desired gain in BER
performance [24, 25].
Figure 9. Performance simulation of the LDPC decoder.
C. Synthesis
Synthesis results for the high-level designs are given
below and are compared with their ‘VHDL only’
counterparts. The results were obtained using Altera’s
Quartus II software with Cyclone II EP2C70F672C6 as
the target FPGA device. Table 1 compares the resources
used by our Design 1 and its maximum operating
frequency (Fmax) with the same decoder designed solely
using hand coded VHDL. Our design uses 2.1% of the
FPGA’s logical elements, only 120 registers and has a
maximum frequency of 42.96 MHz. Clearly our Simulink
design uses much less resources and is faster than the
VHDL-only design. The higher register count in the
VHDL design is due to the presence of dedicated
registers for latching variable and check node outputs,
which were removed from the Simulink design. Table 2
compares our Design 2 with its VHDL-only counterpart.
It uses 3.5% of the logic elements, which is higher than
the amount used by the VHDL-only design (2.8%).
TABLE I.
DESIGN 1 COMPARED WITH VHDL-ONLY COUNTERPART
Design 1 VHDL only design
Logical elements 1432 (2.1%) 1492 (2.2%)
Combinational functions 1432 (2.1%) 1492 (2.2%)
Total registers 120 (0.2%) 686 (1%)
Maximum frequency (Fmax) 42.96MHz 37.3 MHz
© 2010 ACADEMY PUBLISHER
TABLE II.
DESIGN 2 COMPARED WITH VHDL-ONLY COUNTERPART
Design 2 VHDL only design
Logical elements 2416 (3.5%) 1916 (2.8%)
Combinational functions 2416 (3.5%) 1916 (2.8%)
Total registers 480 (0.7%) 881 (1.3%)
Maximum frequency (Fmax) 67.20MHz 67.72MHz
The reason for the higher number of logic elements in
our design is that the control unit contains handshaking
circuitry to facilitate communication of large amount of
data between the PC and the FPGA board for testing.
However, the VHDL-only design has a simple control unit
and does not include any such handshaking circuitry. Our
Design 2 uses nearly half the number of registers and
achieves nearly the same maximum frequency (Fmax).
The time required by our high-level approach for
successful modelling, simulation and synthesis of the
LDPC designs was almost a quarter of that required by
the hand coding method.
D. Behavioral Simulation of VHDL Model
Fig. 10 shows the functional simulation result for the
VHDL model of Design 2 generated from its top level
Simulink model. A set of raw LLRs (6, 6, 2, 4, 7, 4, -2, 6,
4, 7) are applied to the variable nodes. The signals
end_o_vn & end_o_cn are the control signals. The parity
becomes 1 in the third iteration when all the parity checks
are satisfied and the decoder stops the iterations. The
corrected LLRs output by the variable nodes are 7, 7, 7,
7, 7, 7, -8, 4, 5, 7.
E. Hardware Implementation and Testing
After fully simulating the Simulink as well as the
VHDL models of Design 1 and Design 2, both designs
were implemented on a Xilinx Spartan 2E FPGA. The
FPGA platform we used is shown in Fig. 11. It contains
three separate modules: the USB communication module
for communicating with the PC, the main FPGA module
housing the Xilinx Spartan IIE, and the I/O module. The
LLRs are generated by a MATLAB program and are
stored in a text file on the PC. A LabVIEW program
running on the PC sends these LLRs to the decoder via
the USB module and receives the decoded LLRs back
along with the parity information. The decoded LLRs are
written into a separate file by the LabVIEW program and
analysed for correctness by a MATLAB program by
comparing with the LLRs generated by simulation of the
Simulink and/or VHDL models. The PC controls the
operation of the decoder on the FPGA through a number
of handshaking signals as is shown in Fig. 12, for
example start/stop, max_iteration etc. The I/O module has
been used to display useful runtime information, for
example the number of iterations completed by the
decoder for each LLR. This enables us to obtain a visual
indication that the LDPC decoder is operating. The
performance results obtained from the implemented
decoders are presented in the next sub-section along with
the performance of Simulink and VHDL models.

86 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Figure 11. The FPGA platform used to implement the decoders.
LLR_In
Start Stop
Clock
LDPC Decoder
in FPGA
Figure 12. Block diagram of the LDPC decoder implemented on the
Xilinx Spartan 2E FPGA device.
F. Comprehensive Testing Strategy
The Matlab and Simulink environment used in the high
level design methodology make it very easy and fast to
build a full test system for testing the whole design. In
this section we present a comprehensive testing strategy
for LDPC decoders whereby we can test the decoder
outputs from three different environments
simultaneously. Fig. 13 shows the proposed testing
strategy. The test system allows the design to be
simultaneously tested in Simulink, VHDL (ModelSim)
and on the FPGA.
© 2010 ACADEMY PUBLISHER
Figure 10. Behavioral simulation results of decoder Design 2.
Max_iteration
Reset
Parity
LLR_Out
Number_of
_Iterations
The LDPC Encode Test Data module generates a
sequence of LDPC encoded test data and sends these to
the Simulink simulation model, VHDL testbench and
FPGA at the same time. After decoding is done in the
three environments, the data are sent back to the Test
Data Analysis module. This module analyses the decoded
data from the three environments and compares the parity
information for correctness. We have used this scheme to
validate the decoder designs presented in this paper.
LDPC
Encode
Test Data
LDPC
Decoder
in
Simulink
LDPC
Decoder
in
VHDL
LDPC
Decoder
in
FPGA
Figure 13. Structure of the testing system.
LDPC
Test Data
Analysis
Fig. 14 shows the performance plot (BER) obtained
from the FPGA and compares it with the BER obtained
from the Simulink model. The performance plots from
the VHDL testbench and FPGA are also compared in Fig.
15. These figures show a close match among the
performance plots obtained from three different
environments and therefore prove that the design has
been correctly implemented and run on the FPGA device.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 87
Figure 14. Bit-Error-Rates obtained from FPGA implementation
compared with Simulink simulation.
Figure 15. Bit-Error-Rates obtained from FPGA implementation
compared with VHDL testbench.
VI. ANALYSIS OF THE HIGH LEVEL DESIGN METHOD
The high-level design methodology presented in this
paper reduces design complexity, effort and time. For
example, to design the variable and check nodes of Fig. 4
and Fig. 5 completely in VHDL a designer has to write
behaviours of quite a few modules in VHDL and
manually do the port mapping for the top level design. As
the design gets larger and larger it becomes very difficult
to code everything in VHDL. It not only takes huge effort
and time, but also managing complex designs and reusing
the designs are often quite difficult. A great deal of
expertise and experience in VHDL is also required. For a
large LDPC decoder with hundreds of nodes, the
interconnections among the variable and check nodes
could easily become a designer’s nightmare if the entire
design has to be coded in VHDL. An alternative, intuitive
and highly efficient design method has been a long
standing desire of the engineers engaged in the design of
complex digital systems such as LDPC decoders.
© 2010 ACADEMY PUBLISHER
The high level methodology we have presented in this
paper offers a very attractive alternative. The designs can
be done intuitively in Simulink using predefined
Simulink library blocks and blocks made from high-level
Matlab code. The design complexity, effort and time are
reduced drastically because the need for manually writing
complex VHDL code is almost eliminated. Our estimates
have shown that the time required to successfully design,
simulate and synthesise a LDPC decoder using hand
coded VHDL is almost four times that required by the
proposed high-level methodology. Our decoder Design 2
had 1069 lines of VHDL code generated automatically by
HDL Coder from the top level Simulink model as
opposed to only 225 lines of hand written VHDL code.
The main reason for the large code produced by HDL
Coder is that it used a very large number of internal
signals to generate the VHDL description of the decoder.
This is something we did not have any control over. Of
course our decider Design 2 had additional handshaking
circuitry to facilitate communication between the PC and
the FPGA board for testing purpose. This contributed to
the larger code to some extent. However, we did not
optimise the auto-generated VHDL code. Yet our decoder
designs compare favourably with the hand coded designs
(see Tables 1 and 2).
Another important aspect of the proposed high-level
design methodology is that design reuse requires much
less effort because changes are made either at block level
in Simulink or in high-level Matlab code. In addition
Matlab is a software programming language that is used
much more widely than hardware description languages
like VHDL and Verilog. Therefore designers without
specific skills in hardware languages are able to design
complex digital systems without much problem. Even
software engineers and algorithm developers are able to
quickly implement and test their high-level designs due to
the ability to automatically generate HDL descriptions
from the Simulink models. This will surely offer great
flexibility and efficiency in the design and reuse of
complex LPDC decoders. There are some other benefits
of the high level design methodology:
• Different parts of the model can be enveloped using the
‘create subsystem’ property of Simulink. It reduces the
design complexity for large designs.
• User created Simulink library blocks provide great
flexibility because the revisions made to the user
defined library seamlessly propagate through the entire
model.
• HDL Coder can generate either VHDL or Verilog
descriptions from the high-level Simulink models,
allowing greater flexibility in the choice of the
hardware description language.
In the high-level design methodology, it is also easy
and fast to build a complete test system. The Matlab and
Simulink library provides a lot of powerful tools for
generating and analysing test data such as graphical plot,
scripts, etc.
The design methodology we have presented requires
the use of some emerging design tools and library
functions, such as Simulink hardware library and

88 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
embedded Matlab blocks, Link for ModelSim and HDL
Coder tools. Because these are very recent developments
the library of Simulink blocks to support the functions a
designer needs is rather limited at the present time.
Similarly the capability of HDL Coder is limited to
conversion of a few commonly used Simulink blocks and
a few embedded Matlab blocks. Some of the specific
limitations are discussed below.
A. Current Limitations
Some common blocks in the Simulink library are not
currently supported by the HDL Coder for automatic
generation of VHDL code, e.g. flip-flops. Although we
could build the PISO and SIPO registers easily in
Simulink using the flip flops from its library the registers
were not converted to VHDL by the HDL Coder. A
careful selection of supported Matlab functions and
Simulink library blocks may help addressing this type of
problems, but not necessarily always. Some other
limitations we experienced are listed below:
• Multiple instances of some modules (components) are
used in the LDPC decoder, and in fact in most modular
designs utilising a hierarchical design methodology.
For example, in our LDPC decoder multiple instances
of the check and variable nodes, and PISO and SIPO
registers are used. The HDL Coder dumps the full
behaviour of each component as many times as it is
instantiated in the design. This produces redundant
instances of component behaviour in the generated
VHDL code. It is necessary for the designer to edit the
auto-generated code.
• There are some functions which are supported by the
HDL Coder, but it produces the output in a particular
data type. For example, the sign function of Matlab
gives output only in 8-bit integer (int8) format. It is not
possible to change these default data types in the
current version. The only option is to manually edit the
auto-generated HDL code.
• Link for ModelSim: This utility enables blocks
designed in VHDL to be included in Simulink models.
However the ports of the blocks that use Link for
ModelSim get interchanged while simulating in
Simulink. HDL code cannot be generated for the
Simulink model in this situation. The code can be
generated only after correcting the design. This makes
the design process difficult and time consuming. The
other drawback of blocks utilising Link for ModelSim
is that these blocks may make changes in the autogenerated
code. In case of the LDPC decoder these
blocks changed the signals of type signed to unsigned.
Once again this requires manually editing the autogenerated
VHDL code.
VII. CONCLUSIONS
In this paper a new high-level intuitive design
methodology based on Simulink for modelling, synthesis
and implementation of LDPC decoders has been
presented. It utilises the higher level of abstraction
offered by the Simulink modelling environment. The
© 2010 ACADEMY PUBLISHER
modelling, simulation and synthesis process utilises a
combination of emerging design tools and associated
library functions. These include the Simulink HDL Coder
and ‘Link for ModelSim’ tools, and embedded Matlab
and Simulink hardware library blocks. Two versions of
10x5 LDPC decoders have been designed, simulated,
synthesised and successfully implemented on a Xilinx
Spartan 2E FPGA device. A comprehensive testing
strategy has been adopted to test the decoders at all
levels, from the high level Simulink model through
VHDL all the way up to hardware implementation. Our
testing strategy supports simultaneous testing of the
decoders at the three levels which is useful for real-time
debugging. The proposed high level design methodology
facilitates the creation of such testing strategy very
quickly because the test data generated at the high level is
used for testing at all three levels. This helps to reduce
the time of evaluating and testing the design as well as
ensures that the final design is efficient and error-free.
The performance figures on Bit-Error-Rates obtained at
the three levels compare favourably with those of the
LDPC decoders reported in the literature.
The proposed high level design methodology offers
great advantages in terms of design complexity, effort
and time compared to a HDL-only design method. Our
Design 1, completed using the new methodology, uses
less FPGA resources and achieves higher Fmax compared
to its HDL-only designs. Our Design 2 achieves nearly
the same Fmax as its VHDL-only counterpart, but uses
slightly higher number of logic elements due to the
inclusion of additional handshaking circuitry required for
testing the design on FPGA. Given the significant
reduction in design effort and time, the above results
make the proposed design methodology a very attractive
one. We envisage that with further enrichment of the
Simulink Library blocks, and enhancements of HDL
Coder and Link for ModelSim tools, design
methodologies similar to the one presented in this paper
will eventually replace tedious HDL-based design
approach. This will pave the way for cost effective design
and reuse of a new generation of complex high
performance LDPC decoders.
ACKNOWLEDGEMENT
This work has been supported in part by a research
grant from the Division of IT, Engineering and the
Environment of the University of South Australia
(UniSA). The authors wish to thank Mr Sunil Sharma for
his initial investigations into developing variable and
check node models using Simulink and related tools. The
authors also acknowledge Prof Bill Cowley of UniSA’s
Institute of Telecommunications Research for his useful
suggestions and critical feedback. The authors understand
that Prof Cowley’s time has been partially supported
through a research grant from Sir Ross and Sir Keith
Smith fund. Finally the authors would like to thank Dr
Mark Ho of the School of Electrical and Information
Engineering of UniSA for his suggestions on the
performance simulations of the decoder.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 89
REFERENCES
[1] R. G. Gallager, Low-Density Parity-Check Codes.
Cambridge, Mass: Monogram, 1963.
[2] S. Johnson, Introducing Low-Density Parity-Check
Codes. Australia: University of Newcastle, 2006.
[3] S. Lin and D. J. Costello, Error Control Coding:
Fundamentals and Applications. New Jersey: Prentice
Hall, 2004.
[4] B. Reiffen, “Sequential Decoding for Discrete Input
Memoryless Channels,” IRE Trans. Inf. Theory, vol. 8,
no. 3, pp. 208-220, April 1962.
[5] A. J. Blanksby and C. J. Howland, “A 690- mW 1-Gb/s
1024-b, rate-1/2 low-density parity check code decoder,”
IEEE J. Solid State Circuits, vol. 37, no. 3, pp. 404–412,
2002.
[6] T. J. Richardson, M. A. Shokrollahi, and R. L. Urbanke,
“Design of capacity-approaching irregular low-density
parity-check codes,” IEEE Trans. Inf. Theory, vol. 47, pp.
619-637, February 2001.
[7] J. Nguyen, B. Nikolic, and E. Yeo, Design of a low
density parity check iterative decoder. University of
California, Berkley: EECS, College of Engineering, 2002.
[8] S. Hong and W. Stark, “Design and implementation of a
low complexity VLSI turbo-code decoder architecture for
low energy mobile wireless communications,” J. VLSI
Signal Processing, vol. 24, pp. 43-57, 2000.
[9] M. Karkooti, P. Radosavljevic, and J. R. Cavallaro,
“Configurable, High Throughput, Irregular LDPC
Decoder Architecture: Tradeoff Analysis and
Implementation,” Proc. Int. Conf. Application Specific
Systems, Architectures and Processors, pp. 360-367,
September 2006.
[10] M. Karkooti, P. Radosavljevic, and J. R. Cavallaro,
“Configurable LDPC Decoder Architectures for Regular
and Irregular Codes”, J. Signal Processing Systems, vol.
53, pp. 73-88, October 2008.
[11] Y. Lei, L. Hui, and R. C. J. Shi, “Code Construction and
FPGA Implementation of a low-error-floor multi-rate
low-density Parity-check code decoder,” IEEE Trans.
Circuits & Systems I, vol. 53, pp. 892-904, April 2006.
[12] A. Darabiha, C. A. Carusone, R. F. Kschischang, and E.
S. Rogers, “Multi-Gbit/sec Low Density Parity Check
Decoders with Reduced Interconnect Complexity,” Proc.
IEEE Int. Symp. Circuits & Systems, vol. 5, pp. 5194-
5197, 2005.
[13] A. Darabiha, C. A. Carusone, and R. F. Kschischang,
“Block-Interlaced LDPC Decoders With Reduced
Interconnect Complexity,” IEEE Trans. Circuits &
Systems I I, vol. 55, pp. 74-78, January 2008.
[14] J. Sha, M. Gao, Z. Zhang, Li Li, and Z. Wang, “An FPGA
implementation of array LDPC decoder,” Proc. IEEE
Asia Pacific Conf. Circ. & Systems, pp. 1675-1678,
December 2006.
[15] Mauro Cocco, “A scalable architecture of LDPC
Decoding,” Proc. Design, Automation & Test in Europe
Conf., vol. 3, pp. 88-93, February 2004.
[16] J. A. Wicks and J. R. Armstrong, “Efficiency ratings for
VHDL behavioral models,” IEEE Proc. Southeasrcon’98,
pp. 401-404, April 1998.
[17] M. Eroz, F.W. Sun, and L.N. Lee, “DVB-S2 low density
parity check codes with near Shannon limit performance,”
Int. J. Satellite Communications and Networking, pp.
269-279, 2004.
[18] W. E. Ryan, An Introduction to LDPC Codes. USA: The
University of Arizona, 2003.
© 2010 ACADEMY PUBLISHER
[19] T. Mohsenin and B. M. Baas, “Split-Row: A reduced
complexity, high throughput LDPC decoder architecture,”
Proc. Int. Conf. on Computer Design (ICCD 2006), San
Jose, CA, pp. 220-225, 1-4 Oct 2007.
[20] Z. Jianguang, F. Zarkeshvari, and A. H. Banihashemi, “On
implementation of min-sum algorithm and its
modifications for decoding low-density Parity-check
(LDPC) codes”, IEEE Trans. on Communications, vol. 53,
no. 4, pp. 549-554, April 2005.
[21] F. R. Kschischang, B. J. Frey, and H. Loeliger, “Factor
graphs and the sum-product algorithm,” IEEE Trans. on
Inf. Theory, vol. 47, no. 2, pp. 498-519, 2001.
[22] Sae-Young Chung, T. J. Richardson, and R. L. Urbanke,
“Analysis of Sum-Product Decoding of Low-Density
Parity-Check Codes Using a Gaussian Approximation,”
IEEE Trans. Info. Theory, vol. 47, no. 2, pp 657-670, 2001.
[23] S. M. Aziz and S. Sharma, “New Methodologies for High
Level Modeling and Synthesis of Low Density Parity
Check Decoders“, Proc. 11 th Int. Conf. on Computers and
IT (ICCIT 2008), Khulna, pp. 276-281, 24-27 December
2008.
[24] D. Sridhara and T. E. Fuja, “LDPC Codes Over Rings for
PSK Modulation,” IEEE Trans. Inf. Theory, vol. 51, no. 9,
pp. 3209-3220, September 2005.
[25] J. K. S. Lee and J. Thorpe, “Memory-Efficient Decoding of
LDPC Codes,” Proc. IEEE Int. Symp. on Information
Theory (ISIT 2005), Adelaide, Australia, pp. 459-463, 4-9
November 2005.
Syed Mahfuzul Aziz received
Bachelor and Masters Degrees, both in
electrical & electronic engineering,
from Bangladesh University of
Engineering & Technology (BUET) in
1984 and 1986 respectively. He
received a Ph.D. degree in electronic
engineering from the University of
Kent (UK) in 1993 and a Graduate
Certificate in higher education from
Queensland University of Technology, Australia in 2002.
He was a Professor in BUET until 1999, and led the
development of the teaching and research programs in
integrated circuit (IC) design in Bangladesh. He joined the
University of South Australia in 1999, where he is currently an
associate professor and the inaugural academic director of first
year engineering program. In 1996, he was a visiting scholar at
the University of Texas at Austin when he spent time at Crystal
Semiconductor Corporation designing advanced CMOS
integrated circuits. He was a visiting professor at the National
Institute of Applied Science Toulouse, France in 2006, where he
has collaborations in the area of nanoscale CMOS technology
modelling and integration with educational IC design tools. He
has been involved in numerous industry projects in Australia
and overseas, and has attracted funding from reputed research
organisations such as the Australian Defence Science and
Technology Organisation (DSTO), and the Pork CRC
(Cooperative Research Centre), Australia. He has authored
eighty five refereed research papers. His research interests
include digital CMOS IC design and testability, modelling and
FPGA implementation of high performance processing systems,
biomedical engineering and engineering education.

90 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Dr Aziz is a senior member of IEEE and a member of
Engineers Australia. He has received numerous professional
awards. These include: an Excellent Achievement Award in
Networking and Internet System Development (1998) from the
Centre of the International Co-operation for Computerisation,
Japan; the International Network for Engineering Education and
Research Achievement Award (2007); a Citation for outstanding
contributions to student learning from both the Australian
Learning & Teaching Council (2007) and the Australasian
Association for Engineering Education (AaeE–2007); an AaeE
Award for Teaching Excellence - Highly Commended (2008).
He has served as member of the program committees of many
international conferences. He reviews papers for the IEEE
Transactions on Computer and Electronics Letters, UK.
Recently he has been appointed a reviewer of the National
Priorities Research Program, a flagship funding scheme of the
Qatar National Research Fund.
© 2010 ACADEMY PUBLISHER
Minh Duc Pham received B.S.
degree in electronic engineering from
HCM National University of
Technology, Vietnam in 2003 and M.S.
degree in microsystems technology from
the University of South Australia in
2008. He has been working as an
ASIC/FPGA engineer since 2003 for
Arrive Technologies Inc, a fab-less
silicon supplier of Disruptive Next
Generation Solutions for PDH, SONET, SDH and Ethernet
Internetworking. Mr Pham is currently working as a Research
Assistant in the School of Electrical and Information
Engineering of the University of South Australia. His research
interests are in the fields of VLSI implementation of
communication systems such as SoC for next generation
networking, automation in VLSI design, forward error
correction and coding theory.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 91
A Formal Model for Abstracting the Interaction of
Web Services
Li Bao
Institute of Software Engineering, Dalian Maritime University, Dalian, China
Email: ebond@163.com
Weishi Zhang and Xiong Xie
Institute of Software Engineering, Dalian Maritime University, Dalian, China
Email: {teesiv, xxyj}@dlmu.edu.cn
Abstract—This paper addresses the problems of modeling
the interaction of Web services when they are composed
together. Many subtle errors such as message not received
and deadlock may occur due to uncontrolled concurrency of
Web services. A model called IMWSC (Interaction Module
for Web Service Composition, IMWSC for short) is
proposed. The proposed model is used to abstract and
analyze the interaction of web services. IMWSC is given a
formal semantics by means of CCS (Calculus of
Communicating System, CCS for short), which is a kind of
process algebra that can be used to model concurrent
systems. The application of this model is further
investigated in a case study. Some important points related
to verify the correctness of interaction of Web service are
discussed.
Index Terms—Web Service, Interaction, Formal Method,
IMWSC
I. INTRODUCTION
In order to survive the massive competition created by
the new online economy, many organizations are rushing
to put their core business competencies on the Internet as
a collection of web services for more automation and
global visibility [1] . The concept of web service has
become recently very popular. Web services are software
applications which can be used through a network
(intranet or Internet) via the exchange of messages based
on XML standards [2] . It has become a vehicle of web
services rather than just a repository of information.
The ability to efficiently and effectively share services
on the Web is a critical step towards the development of
the new online economy driven by the Business-to-
Business (B2B) e-commerce [1] . Existing enterprises
would form alliances and integrate their services to share
costs, skills, and resources in offering a value-added
service to form what is known as composite service.
A composite web service is a system that consists of
several conceptually autonomous but cooperating units.
In order to establish a long-running service composition,
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.91-98
many languages and tools emerged, which provide
different schemas to glue service operations properly.
Service composition approaches can be generally divided
into two categories [3, 4] : business flow based approach
and semantic based approach. Some famous projects on
web service are based on business flow [24] , such as
eFlow [5] , METEOR-S [6] , SELF-SERV [7] ; Semantics based
approach composes services based on ontology and relies
on the use of AI planning techniques to automatically
search, orchestrate, compose and execute services.
Representative projects on web service research that is
based on Semantics are: WebDG [8] , SWORD [9] , SHOP [10] .
From a software engineering viewpoint, the
construction of new services by the static or dynamic
composition of existing services raises exciting new
perspectives which can significantly impact the way
industrial applications will be developed in the future —
but they also raise a number of challenges. Among them
is the essential problem of guaranteeing the correct
interaction of independent, communicating software
pieces [2] .
One legitimate question is therefore whether or not the
correct and reliable interaction of web services can be
guaranteed to a great extent by introducing the formal
description techniques. Our investigations suggest a
positive answer. This paper addresses the problem of
formally modeling the interaction of web services when
they are composed together, be it in a dynamic or static
way. A model for abstracting and analyzing one scenario
of the interaction process of web services called IMWSC
is proposed. After the interaction of web service is
described in an abstract way, available supporting tool
can be used to determine whether or not this interaction
process satisfies the desired properties which are
expressed in a kind of modal logic.
This paper is structured as follows. Section 2 discusses
the related work. In Section 3, we present IMWSC.
Section 4 defines the semantics of IMWSC. The
application of IMWSC is investigated in a case study in
Section 5. And the conclusion and future work are drawn
up in Section 6.

92 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
II. Related Work
Petri nets are a formal model for concurrency. Since
the semantics of Petri nets is formally defined, by
mapping each BPEL process to a Petri net a formal model
of BPEL can be obtained which allows the verification
techniques and tools developed for Petri nets to be
exploited in the context of BPEL processes. Many works
such as [11, 12, 21, 22] introduce the Petri net based
method for describing and verifying web service.
In [21], Schmidt and Stahl discuss a mapping from
BPEL to Petri nets by giving several examples. Each
BPEL construct is mapped into a Petri net pattern.
In [22], Schlingloff, Martens and Schmidt also
consider the usability problem. They show that usability
can be expressed in alternating-time temporal logic. As a
consequence, model checking algorithms for this logic
can be exploited to check for usability.
As research aiming at facilitating web services
integration and verification, WS-Net introduced in [11] is
an executable architectural description language
incorporating the semantics of Colored Petri-net with the
style and understandability of object-oriented concepts.
In [12], Tao provide a web service composition model
which is based on a kind of advanced Petri-net, OOPN
(Object Oriented Petri Net). A web service can be
mapped to an OOPN system based on this model and
different OOPN system can be integrated together into a
composite service via message passing.
A process algebra is a rather small concurrent language
that abstracts from many details and focuses on particular
features. There are several relevant publications [1, 13, 14]
for process algebra based methods.
Gwen Salaün, Lucas Bordeaux present an overview of
the applicability of process algebras in the context of web
services in [1].
Authors present a framework for the design and the
verification of WSs using process algebras and their tools
in [13].
Li Bao, Weishi Zhang present a CCS based method for
describing and verifying the behaviour of web service in
[14].
III. Defining IMWSC
A. Initiative of IMWSC
For the Petri net based methods, one major defect is
that the number of the places and the transitions described
in a Petri net is too large. Researchers often map each
element in a web service composition language to an
element in Petri net and do not restrict the number of the
places and the transitions in a Petri net. If the number of
the places and the transitions described in a Petri net is
not restricted, the designers will meet a condition of state
explosion, which is very difficult to be dealt with; another
major defect for the Petri net based methods is the lack of
the description of interaction process of web services.
The Petri net based methods often put their emphasis on
describing the workflow inside a web service, and do not
present the complicate interaction process of web services.
© 2010 ACADEMY PUBLISHER
For the process algebra based methods, one major
defect is that some kinds of complex structure of web
service composition can not been defined by using these
methods; another major defect for the process algebra
based methods is that the lack of rigorous translation
mechanism between the element of web service
composition language and the element of process algebra.
These methods often give simple corresponding relations
and translation rules. These relations and rules can not
guarantee the correct reservation of the information
related to behavior and are apt to lead to the loss of
information. We adopt a kind of hierarchically refined
description method to define the interaction process of
web services, i.e., we divide the interaction process of
web services into smaller parts, which is defined as
Interaction Module for Web Service Composition
(IMWSC in short). For each of these parts, a scenario of
the interaction of web services is defined. These smaller
parts, i.e., modules, have a common property that the
outcome of each module is determinate, in other words,
each of the module has only one terminative state. This
important property suggests these modules can be
composed. Therefore, by mapping each module to a
transition in a Petri net, modules which describe the
scenarios of the interaction of web services are strictly
composed. However, for the limit of the length, we only
introduce the definition and properties of a module, i.e.
IMWSC model, method about how to compose these
modules will be introduced in further work.
Instead of composing activities of web service, we
compose modules. The merit of our approach is that it
can effectively reduce the number of the objects to be
analysis, such that the interaction process of web service
is described more concisely, as well as the state explosion
can be avoided. At the same time, web service
composition with complex structure can be described by
composing these modules, while the process algebra
based methods can not achieve.
Another benefit of our approach is the introducing of
the semantic of IMWSC. The semantic of IMWSC
comprises three parts: semantic domain, semantic range,
and valuation function. A process calculus CCS (Calculus
of Communicating Systems, CCS in short) [15, 16] is
introduced as a semantic range, and then valuation
functions are defined that translate an IMWSC(semantic
domain) into a process term. Since the valuation
functions are rigorously defined, the correct reservation
of the information related to behavior can be guaranteed,
such that the loss of information can be avoided.
B. Formal Definition of IMWSC
A web service is a software application which can be
used through a network (intranet or Internet). For a web
service, the basic functional unit is operation. The process
of web service invocation is actually the process of
operation invocation. For IMWSC, the invocation of
operation is modeled by Activity. For a better control of
the structure of activities, we introduce a set of processes,
i.e. Proc, as the basic control unit. A process, i.e. an
element in Proc, is a linear concatenation of activities. If

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 93
the output data of one operation opr1
is the input data of
another operation opr2
, we consider that there is a
corresponding relation between and opr . For
opr1 2
IMWSC, we introduce the binary relation a to
represent this kind of relation. Symbol L is introduced
into IMWSC to record the interaction history of web
services. We described the interaction process of web
services in a scenario in the way defined by IMWSC, in
other words, the definition of an instance of IMWSC is
the definition of the interaction process of web services in
a scenario.
Definition 1. (IMWSC) Formally, an IMWSC is a
septuple ,
where:
● Service denotes a set of web services;
● Proc is a set of processes ;
● Activity is a set of activities ;
● L is a set of sequences of activities;
● Message is a set of messages that are exchanged by
services;
R ⊆ Activity×Activity is a binary relation;
● a
● F is a sextuple < , , , , ,
f mA
>, where:
f → { c,
b }
R
f pT f pS f pU f aP f aT
─ pT : Proc is a mapping that
describes the type of each process ( composite
or basic ) ;
─ f pS : Proc → Service is a mapping that
describes the type of each process ( composite
or atomic ) ;
─ f pU : Proc → Proc is a mapping that
associates a process with a composite process;
─ f aP : Activity → Proc is a mapping that
associates each activity with a process ;
─ f aT : Activity → {ii, io, ei, eo, ex} is a
─
mapping that describes the type of each
activity ( internal input, internal output,
environmental input, environmental output,
execute );
f mA : Message → Activity is a mapping that
associates each message with an Activity.
f con
( proc)
{ a | a ∈ Activity ∧ (a)
f aP
We let = =
proc} for p ∈Proc
∧ f pT ( p)
= b ; Let c < ⊆ Activity
× Activity be an partial order relation over Activity,
defined as: < c = { ( a1 , a2
) | Activity
a a1 2 ∈ ,
∧ f aP ( a1)
= f aP ( a2
) ∧ ( a1
happens earlier than
a2
)}; An element proc in Proc is constructed by the
following grammar:
© 2010 ACADEMY PUBLISHER
proc = α | || proc | proc p proc ,
1 2
proc 1
2
where: α ∈ Activity; proc1, proc2∈Proc.
● 1 || proc2
is a new process that performs
proc1 and proc2 independently;
proc
● proc1 p proc2
is a new process that performs
proc1 and proc2 sequentially.
Fig. 1 presents an illustration of the structure of
IMWSC. In Fig.1, a service is visualized by a circle;
interaction of services is visualized by a pair of parallel
arrows (with opposite directions); the interaction process
Definition, i.e., the definition of an instance of IMWSC,
is visualized by a rectangle.
Figure 1. Structure of IMWSC
C. The Necessary Condition for the Correctness of
IMWSC
The fundamental requirement for a correct interaction
process of services is that each input of a service shall be
met by another service. Thus the basic requirement that
guarantees the correctness of IMWSC is:
for any activity a1 ∈Activity,
if it is an input activity,
then, there shall be another activity a2
∈ Activity, such
that a1Ra a2
. The necessary condition for the correctness
of IMWSC can also be defined as the following
predicative formula:
∀ a ∈ Activity (( f a ii
) (
aT ( ) = ∨ f aT ( a)
= io →
∃a' ∈ Activity ( aR a'
∨ )))
a 'R
a
a a
IV. Formal Semantics of IMWSC
Formal semantic descriptions of a model are the basis
for proving properties of this model. Moreover, they
provide precise documentation of model design and
standards for implementations, and (sometimes) they can
be used for generation of prototype implementations.
The formal semantics of IMWSC comprises three parts:
semantic domain, semantic range, and valuation function.
A process calculus CCS (Calculus of Communicating
Systems, CCS for short) is introduced as a semantic range,

94 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
and then valuation functions are defined that translate an
IMWSC (semantic domain) into a process term.
A. Basic Syntax of CCS
Let A be a countably infinite collection of names, and
the set = { a | a ∈ A} be the set of complementary
names (or co-names for short). Let L = A U be a set of
labels, and Act = L U {τ } be the set of actions, where
τ denote the activities which are not externally visible.
Let K be a countably infinite collection of process names.
The collection of CCS expressions is given by the
following grammar:
P, Q :: = K | α . P | P | P | Q | P [ f ] |
P \ L
where:
● K is a process name in K ;
● α is an action in Act ;
● I is an index set;
● f : Act → Act is a relabelling function satisfying
the following constraints:
─ f ( τ ) = τ and
─ f ( a)
= f ( a)
for each label a ;
● L is a set of labels.
B. Operational Semantics of CCS
CCS is formalized using axiomatic and operational
semantics. To formally capture the understanding of the
semantics of the language CCS, the collection of
inference rules are therefore introduced as follows (a
transition P Q holds for CCS expressions P, Q if,
and only if, it can be proven using these rules) :
α
→
For a detailed introduction to the syntax and
operational semantics of CCS, readers are referred to [17,
18].
∑ ∈I
i
C. Defining Valuation Functions
The valuation functions of IMWSC, and their
corresponding semantic domains, semantic ranges are
given in Tab. 1 ( symbols IMWSC denotes an IMWSC
instance; P denotes process term in CCS; A denotes the
set of atomic processes; Activity denotes a set of activities;
Act denotes the set of actions in CCS).
© 2010 ACADEMY PUBLISHER
i
where:
TABLE I
Valuation Functions and their Domains and Ranges
● fm ( procr ) = fc (proc1) | fc (proc2) | ⋅⋅⋅ | fc (procn),
where , proc ∈ Service ( 1 ≤ i ≤ n)
, and
procr i
( ) = Ø
s r proc
f ∧ f s proci
) = procr
( ;
● ( ) =
iff = a ;
c i proc f ) ( a i proc f ) ( f pT proci
● ( ) = iff = c ;
c i proc f ) ( r i proc f ) ( f pT proci
● ( ) = , where
a i proc f ) ( ) ( ) ( f e a1
⋅ f e a2
L f e an
a ∈ Activity ∧ f ( a ) = proc , and
i
a < a < L <
1
c
2
c
● f iff ii ei ;
e ( ai
) = ! ai
f aT ( ai
) = ∨ f aT ( ai
) =
e
c
● f iff io eo ;
e ( ai
) = ? ai
f aT ( ai
) = ∨ f aT ( ai
) =
● f ( ) ( ) | ( ) iff
r proci
= f c proc1
f c proc2
f pU ( proc1)
= proci
∧ f pU proc ) = proci
|| proc proc = ;
proc i
1
a
i
n
2
;
i
( 2 ∧
● f ( ) ( ). ( ) iff
r proci
= f c proc1
f c proc2
f pU ( proc1)
= proci
∧ f pU proc ) = proci
= proc p proc .
proc i
1
2
( 2 ∧
By means of the valuation functions defined in Tab. 1,
an algorithm aiming at translating an IMWSC instance to
CCS terms can be developed:
Algorithm. IMWSC_Instance_to_CCS
INPUT: IMWSC Instance
OUTPUT: The corresponding CCS terms
Process Trans_fm (IMWSC Instance)
{
1. Str Exp = Empty ;
2. For each p in Proc ;
3. Exp= Exp | Trans_fc ( p ) ;
4. Return Exp ;

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 95
}
Process Trans_fc ( process p ∈Proc
)
{
1. If ( process Type = basic )
{ Return Trans_fa ( p ) } ;
2. Else { Return Trans_fr ( p ) };
}
Process Trans_fa ( process pi ∈Proc)
{
}
1. Str name = getName( p ) ;
2. SET name = NIL ;
a i of process
3. For each activity in Activity
4. If ( activityType = output )
name = ! getName( i ). name ;
5. Else If (activityType = input )
name = ? getName( i ). name.
6. RETURN name ;
Process Trans_fr (process pi ∈Proc)
{
1. Str Exp = Empty ;
}
2. For each subService j of process
3. If ( compositionType of pi
is parallel)
a
a
p
u i
Exp = Exp | Trans_fc ( j ) ;
4. Else Exp = Exp. Trans_fc ( u j ) ;
5. Return Exp ;
If the IMWSC instance to be translated comprises m
basic processes, and n = max num(
p )} , where
num(
pi
)
p
O ( m×
n)
u
pi
returns the number of the activities contained
in process , the complexity of above algorithm will
be .
i
{ i
V. Case Study: Application of IMWSC to a Concrete
Scenario
A. Abstracting the Interaction of Web Services
We will investigate the application of IMWSC in a
simple scenario. There are three services involved in this
scenario:
― The Client Service, which need to find out some
useful information (for convenience, client here is
considered as a service);
― The Response Service, which is responsible for
dealing with information inquiry requests;
― The Information Service, which acts as a database
and providing the useful information.
© 2010 ACADEMY PUBLISHER
The business process of this scenario is introduced
briefly as follows:
1. The Response Service receives a request from the
Client Service which need to find out some useful
information;
2. The Response Service contacts the Information
Service and relay the information inquiry request;
3. The Response Service answers the questions to the
Client Service.
Fig. 3 presents an illustration of the structure of this
scenario, where
● A service is visualized by a rectangle (with round
angles);
● A state of a service is visualized by a circle (the
initial and the terminative states of a service are
visualized by icons , respectively);
● A transition between states is visualized by an arrow
(with curve line), from the source state to the target
state ;
● The supply channels of services in this scenario is
visualized by a pair of parallel arrows (with opposite
directions).
Figure 2. A Scenario of Interaction of Services
By applying IMWSC, the interaction process of
services in this scenario is described as follows:
_____________________________________________
f con (Client) = { cReq, cAsk, cInquiry, cInfo };
f con (Reponse) = { rReq, rAsk, rInquiry, rAnswer };
f (InfoS) = { iAnswer, iInfo }.
con
f aT (cReq) = ii; f aT (cAsk) = io; f aT (cInquiry) = ii;
f aT (cInfo) = io; f aT (rReq) = io; f aT (rAsk) = ii;
f aT (rInquiry) = io; f aT (rAnswer) = ii;
f aT (iReq) = io; f eT (iInfo) = ii; f aT (iAnswer) = io.
cReq < c cAsk < c cInquiry < c cInfo;

96 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
< c < c < c
rReq rAsk rInquiry rAnswer;
<
iAnswer iInfo.
c
< rReq, cReq >∈ ; < cAsk, rAsk >∈ R ;
Ra a
R ∈ a
< rInquiry, cInquiry >∈ ; < cInfo, iInfo >
a
R
R ;
< iAnswer, rAnswer >∈ a ;
______________________________________________
By means of the semantics of IMWSC defined in
Section 4, the corresponding CCS terms translated are as
follows:
Client = ! Req. ? Ask. ! Inquiry. ? Info. nil ;
Response = ? Req. ! Ask. ? Inquiry. ! Answer. nil;
InfoS = ? Answer. ! Info. nil;
Scenario = ( Client | Response | InfoS ) / { req, ask,
info, Inquiry, Answer }
B. Verifying the Interaction of Web Services
CCS is an effective modeling language which has
available supporting tool CWB-NC (Concurrency
Workbench of the New Century, CWB-NC for short) [20] .
We use this tool to reason on and verify the behavior of
an instance of IMWSC.
Using the supporting tool of CCS, i.e., CWB-NC, aims
at assist the design and verification of a system. Applying
CCS in the design phase of a system is helpful to show
explicitly the interaction of the components that compose
this system; after the model of a system has been
constructed, modal μ − calculus [23] can be used to reason
on the system behavior. For a detailed introduction to
modal logic, readers are referred to, for example, [19, 23].
One type of verification supported by the tool is
reachability analysis. Here, as in each type of verification,
our first step in using the tool is to write a description of
the system supported by CWB-NC. The description is
then parsed by the tool and checked for syntactic
correctness. We then give a logical formula describing a
“bad state” that the system should never reach. Given
such a formula and system description, CWB-NC
explores every possible state the system may reach during
execution sequence and checks to see if a bad state is
reachable. If a bad state is detected, a description of the
execution sequence leading to the state is reported to the
user. Many bugs such as deadlock and critical section
violation may be found using this approach [20] .
Correct termination is one of the main properties a
proper web service should satisfy. We use can_terminate
to define the state of termination of a system. And the
explanation for this state is as follows:
can_terminate is true of a system if it will reach a
terminative state. We express this property the system
should have in modal −
μ calculus:
prop can_terminate =
min X = [−]ff \/ X
© 2010 ACADEMY PUBLISHER
Reachability analysis is actually a special case of a
more general type of verification called model checking.
In the model checking approach a system is again
described using a design language and a property the
system should have is formulated as a logical formula [20] .
Another type of verification supported by CWB-NC
involves using a design language for defining both
systems and specifications. Here the specification
describes a system behavior more abstractly than the
system description [20] . A relation, i.e., Observational
equivalence needs to be introduced before we conduct
this type of verification.
Observational equivalence is useful in verification as
they lay the conceptual basis for deciding that the
behavior of two web services can be considered to be the
same. They can also be used as a tool for reducing
verification effort by replacing a process by a smaller (in
size), but equivalent one. The bisimulation equivalence
between two processes is a relation between their
evolutions such that for each evolution of one of the
services there is a corresponding evolution of the other
service such that the evolutions are observationally
equivalent and lead to processes which are again
bisimilar. This characterization of the behavior of web
services using the notion of bisimulation helps service
designer optimize composite services by, e.g., changing
their component web services with equivalent ones.
Another motivation is customization of services. To
enhance competitiveness a service providers may modify
their service for customers' convenience and this
customized service must conform to the original one.
Formally, the relation of observational equivalence is
defined as:
Definition 1 [Weak Transitions] [23] :
● q q'
iff , ;
ε
⇒ q q q qn
q'
= → → → = L n ≥ 0
0
● q q'
iff ;
τ
⇒ q q'
ε
⇒
● q q'
iff , (
α
ε α
⇒ q⇒
q →q
ε
⇒ q'
α ≠ τ ).
Definition 2[Observational Equivalence] [23] :
1
τ
Let S ⊆ Q× Q . The relation S is a weak
bisimulation relation if whenever 1 2 then: q S q
● q1 q'1
implies for some such
that ;
α
→ q2 q'2
α
⇒ 2 ' q
q'1
S q'2
● q2 q'2
implies for some such that
.
α
→ q1 q'1
α
⇒ 1 ' q
q'1
S q'2
q 1 and q2 are observationally equivalent, if 1 2
for some weak bisimulation relation , written
q S q
q .
1
2
τ
τ
S 1 q ≈ 2

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 97
In this scenario, Client is considered as a service which
interacts with the composition of the services Response
and InfoS.
The behaviour of the composition of the services
Response and InfoS can be described in two ways:
1. The system description of the composition of
services Response and InfoS is:
Response = ? Req. ! Ask. ? Inquiry. ! Answer. nil;
InfoS = ? Answer. ! Info. nil;
Info_Response = ( Response | InfoS ) / {answer} ;
2. The specification of this composition is:
Spe = ? Req. ! Ask. ? Inquiry. ! Info. nil
Command ‘ eq -S obseq ’ of CWB-NC tool can be
used to examine whether or not two processes are
observationally equivalent. By executing this command,
we know that processes Info_Response and Spe are
observationally equivalent.
VI. Conclusions and Future Work
Formal description and verification of the interaction
of web services is an important research field. After the
description and verification of a practical application of
web service, we come to a conclusion that IMWSC has
very good capability in abstracting, simulating, and
analyzing a scenario of the interaction process of web
services, which will facilitate the correct implementation.
Currently many service composition methods do not
take into account abstracting and analyzing the interactive
features of services in a composition. Therefore it is apt
to make mistakes when using these methods. Our work is
an attempt to abstract and verify the interaction process of
web services which will make the composition process
more reliable.
Further work will involve defining the way IMWSC
instances are composed. An instance of IMWSC model
defined only one scenario of the interaction process of
web services. To model the complete interaction process
of web services, there is a need for composing the
instances of IMWSC model. Since Petri nets are a well
known formal model that is capable of defining the
composition process, we plan to compose the instances
by using Petri net. In our further work, we will present
the fixed point property of IMWSC model. The fixed
point property indicated the outcome of each instance of
IMWSC model is determinate, in other words, each of the
module has only one terminative state. This property lays
the mathematical foundation for mapping a module to a
transition in a Petri net.
ACKNOWLEDGMENT
This research is supported by the National Natural
Science Foundation of China under Grant No.60573087.
© 2010 ACADEMY PUBLISHER
REFERENCES
[1] Rachid H., Boualem B.. A Petri net based model for Web
service composition. Proc. of the 14th Australian Database
Conference on Database Technologies, Adelaide, South
Australia, 2003: 191-200.
[2] Salaün G., Bordeaux L.. Describing and reasoning on Web
services using process algebra. Proc. of the 2nd IEEE
International Conference on Web Services, San Diego,
California, USA, 2004: 43-51.
[3] Schahram Dustdar, Wolfgang Schreiner. A survey on web
services composition. Int. J. Web and Grid Services, 1(1):
1-30, 2005.
[4] Muhammad Adeel Talib. Modeling the Flow in Dynamic
Web Services Composition. Information Technology
Journal, 3 (2): 184-187, 2004.
[5] Casati F, Sayal M, and Shan M C. Developing e-services
for composing e-services. Proc. of the 13th International
Conference on Advanced Information Systems
Engineering (CAiSE2001), Interlaken, Switzerland, 2001:
171-186.
[6] Patil A, Oundhakar S, Sheth A, Verma K. METEOR-S
Web service annotation framework [A]. Proc. of the 13th
International World Wide Web Conference
(WWW2004)[C], New York, USA, May 2004: 553 – 562.
[7] Benatallah B, Sheng Q Z, Dumas M. The Self-Serv
Environment for Web Services Composition[J], IEEE
Internet Computing, 2003, 7(1): 40 – 48.
[8] Medjahed B, Bouguettaya A, Elmagarmid A K.
Composing Web services on the Semantic Web[J], VLDB
Journal, 2003, 12(4): 333 – 351.
[9] Shankar R P, Armando F. SWORD: A developer toolkit
for Web service composition, Proc. of the 11th
International World Wide Web Conference, Honolulu,
Hawaii, USA, May 2002, 786 – 810
[10] E. Sirin, B. Parsia, D. Wu, J. Hendler, and D. Nau. HTN
Planning for Web Service Composition using SHOP.
Journal of Web Semantics, 1(4):377-396, 2004.
[11] Zhang Jia , Chung Jen Yao , Chang C. K. . WS-Net: A
petri-net based specification model for web services. Proc.
of the 2nd IEEE International Conference on Web
Services,San Diego,California,USA, 2004:420-427.
[12] Tao X. F. Formalizing Web service and modeling Web
service based on object oriented Petri net. Lecture Notes in
Computer Science. Berlin, Heidelberg. Springer-Verlag,
2004.
[13] A. Ferrara. Web Services: a Process Algebra Approach”,
Proc. of the 2nd Int. Conf. on Service-Oriented Computing
(ICSOC'04), ACM Press, 2004: 242-251.
[14] Li Bao, Weishi Zhang, Xiuguo Zhang. Describing and
Verifying Web Service Using CCS. In Proc. of the 7th
International Conference on Parallel and Distributed
Computing, Applications and Technologies, 2006: 421-426.
[15] R. Milner. Communication and Concurrency. Prentice Hall,
1989.
[16] R. Milner. A calculi of Communication and Concurrency.
Spinger-Verlag, 1980.
[17] Colin Fidge. A comparative introduction to CSP, CCS and
LOTOS. Technical Report, Department of Computer
Science, University of Queensland, Brisbane, Australia,
April 1994. Available at:
http://sky.fit.qut.edu.au/~fidgec/Publications/fidge94g.pdf.
[18] Luca Aceto, Kim G. An Introduction to Milner’s CCS.
Larsen. Technical Report, Department of Computer
Science, Aalborg University, Aalborg, Denmark, March
2005. Available at:

98 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
http://www.dimi.uniud.it/miculan/Didattica/MFGC04/intro
2ccs.pdf.
[19] C. Stirling. Modal logics for communicating systems.
Theoretical Computer Science, 49(2-3): 311-347, 1987.
[20] R. Cleaveland and S. Sims. The NCSU concurrency
workbench. In R. Alur and T. Henzinger, editors,
Proceedings of the 8th International Conference on
Computer Aided Verification, volume 1102 of Lecture
Notes in Computer Science, New Brunswick, NJ, USA,
1996: 394-397.
[21] K. Schmidt and C. Stahl. A Petri net semantic for
BPEL4WS - validation and application. Proc. of the 11th
Workshop on Algorithms and Tools for Petri Nets,
Paderborn, Germany, 2004: 1-6.
[22] B. Schlinglo, A. Martens, and K. Schmidt. Modeling and
model checking web services. Proc. of the 2nd
Li Bao received the BS degree in
computer science from Dalian
Nationality University, China, in
2003, and the MS degree in computer
science from Dalian Maritime
University, China, in 1996. From
2006 to date, he works as a PhD
candidate in the Institute of Software
Engineering, Dalian Maritime
University, China. His research interests include
distributed computing, software engineering, and formal
description techniques.
© 2010 ACADEMY PUBLISHER
Weishi Zhang received the BS
degree in computer science from
Xi’an Jiaotong University, China, in
1984, and the MS degree in
computer science from the Chinese
Academy of Science, China, in 1986.
He received the PhD degree in
computer science from the
University of Munich, Germany, in
International Workshop on Logic and Communication in
Multi-Agent Systems. volume 126 of Electronic Notes in
Theoretical Computer Science, Nancy, France, August
2004: 3-26.
[23] M. Hennessy and R. Milner, Algebraic laws for
nondeterminism and concurrency, J. Assoc. Comput.
Mach., 32: 137-161, 1985.
[24] Hull R., Su J. (2005) Tools for Composite Web Services:
A Short Overview. SIGMOD Record, 34 (2): 86-95, 2005.
1996. From 1986 to1990, he was an assistant researcher
at the Shenyang Institute of Computing, Chinese
Academy of Science, China. From 1990 to 1992, he was
a visiting scholar at Passau University, Germany. From
1992 to 1997, he was an assistant professor at the
University of Munich, Germany. In 1997, he joined the
Department of Computer Science, Dalian Maritime
University, China, where he is currently a professor of
computer science. His research interests include
distributed computing, software engineering, software
architecture, formal specification techniques, and
program semantics models.
Xiong Xie works as a PhD candidate
in the Institute of Software
Engineering, Dalian Maritime
University, China. Her research
interests include distributed
computing, software engineering,
and formal description techniques

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 99
Performance Evaluation of Elliptic Curve
Projective Coordinates with Parallel GF(p) Field
Operations and Side-Channel Atomicity
Turki F. Al-Somani
Computer Engineering Department, Umm Al-Qura University, P.O. Box: 715 Makkah 21955, Saudi Arabia
Email: tfsomani@uqu.edu.sa
Abstract—This paper presents performance analysis and
evaluation of elliptic curve projective coordinates with
parallel field operations over GF(p). Side-channel atomicity
has been used in these comparisons. The field computations
of point operations are segmented into atomic blocks that
are indistinguishable from each other to resist against
simple power analysis attacks. These atomic blocks are
executed in parallel using 2, 3 and 4 multipliers.
Comparisons between the Homogeneous, Jacobian and
Edwards coordinate systems using parallel field operations
over GF(p) are presented. Results show that Edwards
coordinate system outperforms both the Homogeneous and
Jacobian coordinate systems and gives better area-time
(AT) and area-time 2 (AT 2 ) complexities.
Index Terms— elliptic curve cryptosystems, projective
coordinate systems, Edwards coordinates, side-channel
atomicity.
I. INTRODUCTION
Elliptic Curve Cryptosystems (ECCs) have been recently
attracting increased attention [1]. The ability to use
smaller key sizes and the computationally more efficient
ECC algorithms compared to those used in earlier public
key cryptosystems such as RSA [2] and ElGamal [3] are
two main reasons why ECCs are becoming more popular.
They are considered particularly suitable for
implementation on smart cards or mobile devices.
Because of the physical characteristics of such devices
and their use in potentially hostile environments, Side
Channel Attacks (SCA) [4 - 8] on such devices are
considered serious threats. Two main types of SCAs have
gained considerable attention: simple power analysis
(SPA) attacks and differential power analysis (DPA)
attacks. An SPA attack uses only a single observation of
the power consumption, whereas a DPA attack uses many
observations of the power consumption together with
statistical tools.
SCA seek to break the security of these devices
through observing their power consumption trace or
computations timing. Careless or naive implementation of
Manuscript received December 13, 2008; revised April 11, 2009;
accepted April 27, 2009.
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.99-109
cryptosystems allows side channel attacks to infer the
secret key or obtain partial information about it. Thus,
designers of cryptosystems seek to introduce algorithms
and designs that are not only efficient, but also side
channel attack resistant [9].
The primary operation of ECCs is scalar
multiplication. Scalar multiplication in the group of
points of an elliptic curve is analogous to exponentiation
in the multiplicative group of integers modulo a fixed
integer m. The scalar multiplication operation, denoted as
kP, where k is an integer and P is a point on the elliptic
curve, represents the addition of k copies of point P.
Scalar multiplication is computed by a series of point
doubling and point addition operations of the point P
depending on the bit sequence representing the scalar
multiplier k. Several scalar multiplication algorithms have
been proposed in the literature. A good survey is
conducted by Hankerson et. al. in [10].
Several countermeasures against SCA have been
proposed in the literature. Chevallier-Mames et al. [11]
proposed side-channel atomicity as an efficient
countermeasure against only SPA attacks. Side-channel
atomicity involves almost no computational overhead to
resist against SPA attacks. It splits the elliptic curve point
operations into atomic blocks that are indistinguishable
from each other. Hence, side-channel atomicity is
considered to be an inexpensive countermeasure that does
not leak any data regarding the operation being
performed [11 - 13].
The group operations in an affine coordinate system
involve finite field inversion, which is a very costly
operation, particularly over prime fields. Projective
coordinate systems are used to eliminate the need for
performing inversion. Several projective coordinate
systems have been proposed in the literature including the
Homogeneous, Jacobian and Edwards coordinate systems
[9][14][15].
The selection of a projective coordinate is based on
the number of arithmetic operations, mainly
multiplications. This is to be expected due to the
sequential nature of these architectures where a single
multiplier is used. For high performance
implementations, such sequential architectures are too
slow to meet the demand of increasing number of
operations. One solution for meeting this requirement is

100 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
to exploit the inherent parallelism within the elliptic
curve point operations in projective coordinate [16 - 19].
The performance of these projective coordinates varies
when parallel field multipliers are used. This is because
of the nature of their critical paths. This paper
investigates and compares the performance of the
Homogeneous, Jacobian and Edwards coordinate systems
with side-channel atomicity when parallel field
multipliers are employed. The rest of this paper is
organized as follows. Section II gives a brief introduction
to ECCs. Section III introduces projective coordinate
systems. Section IV shows how the point operations of
the projective coordinate systems are segmented into
atomic blocks and how they are executed in parallel.
Section V shows the performance evaluation of the
selected projective coordinate systems using parallel field
multipliers. Finally, Section VI concludes the paper.
II. ELLIPTIC CURVE PRELIMINARIES
The elliptic curve cryptosystem (ECC), which was
originally proposed by Niel Koblitz and Victor Miller in
1985, is seen as a serious alternative to RSA because the
key size of ECC is much shorter than that of RSA and
ElGamal. To date, no significant breakthroughs have
been made in determining weaknesses in the EC
algorithm, which is based on the discrete logarithm
problem over points on an elliptic curve. The fact that the
problem appears so difficult to crack means that key sizes
can be reduced considerably, even exponentially. This
makes ECC a serious challenger to RSA and ElGamal.
Extensive research has been done on the underlying
math, security strength, and efficient implementation of
ECCs [20]. Among the different fields that can underlie
elliptic curves, prime fields GF(p) and binary fields
GF(2 m ) have been shown to be best suited for
cryptographic applications. An elliptic curve E over the
finite field GF(p) defined by the parameters a, b ∈ GF(p)
with p > 3, consists of the set of points P = (x, y), where
x, y ∈ GF(p), that satisfy the equation:
y 2 = x 3 + ax + b,
where a, b ∈ GF(p) and 4a 3 + 27b 2 ≠ 0 mod p together
with the additive identity of the group point O known as
the “point at infinity”.
Scalar multiplication (kP) is the primary operation of
ECCs Several scalar multiplication algorithms have been
proposed in the literature [10]. Computing kP can be
done with the straightforward double-and-add algorithm,
the so-called binary algorithm, based on the binary
expression of k = (km-1,…,k0) where km-1 is the most
significant bit of the multiplier k. The double-and-add
scalar multiplication algorithm is the most
straightforward scalar multiplication algorithm. It
inspects the bits of the scalar multiplier k, and if the
inspected bit ki = 0, only point doubling is performed. If,
however, the inspected bit ki = 1, both point doubling and
addition are performed. The double-and-add algorithm
requires (m-1) point doublings and an average of (m/2)
point additions [10].
Non-adjacent form (NAF) reduces the average number
of point additions to (m/3) [21]. In NAF, signed-digit
© 2010 ACADEMY PUBLISHER
representations are used such that the scalar multiplier’s
coefficient ki ∈ {0, ±1}. NAF has the property that no
two consecutive coefficients are nonzero. NAF also has
the property that every positive integer k has a unique
NAF encoding, denoted NAF(k).
III. PROJECTIVE COORDINATE SYSTEMS
Projective coordinate systems are used to eliminate the
need for performing inversion. Several projective
coordinate systems have been proposed in the literature
[9][14][15], including the Homogeneous, Jacobian and
Edwards coordinate systems. For the Homogeneous, so
called projective, coordinate system, an elliptic curve
point P takes the form (x, y) = (X/Z, Y/Z), while for the
Jacobian coordinate system, P takes the form (x, y) =
(X/Z 2 , Y/Z 3 ) [9].
Let P1, P2 and P3 be three different points on the
elliptic curve over GF(p), where P1=(X1, Y1, Z1), P2=(X2,
Y2, Z2=1) and P3=(X3, Y3, Z3). Point addition with the
Homogenous coordinate systems can be computed as:
A=Y2Z1, B=X2Z1−X1, C=A 2 Z1−B 3 −2B 2 X1, X3=BC,
Y3=A(B 2 X1−C)−B 3 Y1, Z3=B 3 Z1. Point doubling, on the
other hand, can be computed as: A=aZ1 2 +3X1 2 , B=Y1Z1,
C=X1Y1B, D=A 2 −8C, X3=2BD, Y3=A(4C−D)−8Y1 2 B 2 ,
Z3=8B 3 .
With the Jacobian coordinate system, point addition
can be computed as: A=X1, B=X2Z1 2 , C=Y1, D=Y2Z1 3 ,
E=B−A, F=D−C, X3=F 2 –(E3+2AE2),
Y3=F(AE 2 −X3)−CE 3 , Z3=Z1E. Point doubling, on the
other hand, can be computed as: A=4X1Y1 2 , B=3X1 2 +aZ1 4 ,
X3=B 2 −2A, Y3=B(A−X3)−8Y1 4 , Z3=2Y1Z1.
Recently, Edwards showed in [14] that all elliptic
curves over prime fields could be transformed to the
shape: x 2 + y 2 = c 2 (1 + x 2 y 2 ), with (0, c) as neutral
element and with the surprisingly simple and symmetric
addition law of two points P1 = (x1, y1) and P2 = (x2, y2)
as:
P1 + P2 → ((x1y2+x2y1)/(c(1+x1x2y1y2)),(y1y2x1x2)/(c(1-x1x2y1y2))).
To capture a larger class of elliptic curves over the
original field, the notion of Edwards form have been
modified in [15] to include all curves x 2 + y 2 = c 2 (1 +
dx 2 y 2 ) where cd(1−dc 4 ) ≠ 0.
Point addition with the Edwards coordinate systems
can be computed as: B=Z1 2 Z1, C=X1X2, D=Y1Y2, E=G–
(C+D), F=dCD, G=(X1+Y1)(X2+Y2), X3=Z1E(B–F),
Z3=(B–F)(B+F), Y3=Z1(D–C)(B+F). Point doubling, on
the other hand, can be computed as: A=X1+Y1, B=A 2 ,
C=X1 2 , D=Y1 2 , E=C+D, F=B–E, H=Z1 2 , I=2H, J=E–I,
X3=FJ, Z3=EJ, Y3=E(C–D).
IV. THE PROPOSED METHODOLOGY
Since field multiplications and squarings are the
dominant operation in elliptic curve point operations in
projective coordinates that require much higher
computation time than field additions and subtractions,
the emphasis in this paper is to perform comparisons
between projective coordinate systems when parallel
multiplications or squarings are performed at the same

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 101
time. Furthermore, the field computations of point
operations are segmented into atomic blocks that are
indistinguishable from each other to resist against SPA
attacks, which is called side-channel atomicity [11]. The
approach adopted in this paper is:
1. Analyzing the dataflow of point operations for
each projective coordinate system in the
following manner:
a. Find the critical path which has the lowest
number of field multiplications.
b. Find the maximum number of multipliers
that are needed to meet this critical path.
2. Segmenting the field computations of point
operations for each as follows:
a. An atomic block contains at most one field
multiplication, two field additions, and one
field subtraction.
b. A Field squaring is performed by a
multiplier instead of using a special
hardware unit for squaring.
3. Varying the number of parallel multipliers from
two to the number of multipliers specified by the
critical path to find the following:
a. The best schedule of each dataflow using
the specified number of multipliers.
b. The area-time (AT) and area-time 2 (AT 2 )
complexities.
Table I shows the field arithmetic operations of the
selected projective coordinate systems according to the
presented formulas in Section III. In Table I, αis and βjs,
represent multiplications/squarings and
additions/subtractions respectively. For example, the first
possible multiplication for point addition in the
Homogenous coordinate system (Y2 × Z1) is represented
by α1. The second possible field addition for point
doubling in Edwards coordinate system (C + D), as
another example, is represented by β2. The data
dependencies between the αis and βjs in point operations
for the Homogenous, Jacobian and coordinate systems
are depicted in Fig. 1, 2 and 3 respectively.
In Table II, the αis and βjs are grouped in atomic
blocks. Table II shows the atomic blocks for point
doubling and point addition, denoted by ∆ and Γ
respectively. An empty field operations within an atomic
block are marked by “*”. In Table II, for example, the
atomic block ∆1 of point doubling in the Jacobian
coordinate system contains the on field multiplication α1,
one field addition β1 and two empty slots. The atomic
block Γ7 of point addition in the Homogenous coordinate
system, as another example, contains one field
multiplication α7 and three field additions β2, β3 and β4.
Let the unit of time be the required time to execute
an atomic block. In Table II, point addition requires 11
time units for the three selected projective coordinate
systems. Point doubling, on the other hand, requires 13,
10 and 7 for the Homogenous, Jacobian and Edwards
coordinate systems respectively.
© 2010 ACADEMY PUBLISHER
Table III, IV and V show the scheduling of the atomic
blocks of the Homogenous, Jacobian and Edwards
coordinate systems respectively on parallel multipliers
according to the proposed methodology early in this
section. In Table III, IV and V, the first column shows the
number of multipliers. The second column shows the
required time units to perform point operations using
parallel multipliers. The utilizations of the parallel
multipliers depends on the number of multipliers and the
critical path of the projective coordinate system. Adding
more multipliers, on the other hand, does not imply better
performance. For example, the number of the required
time units to perform point addition using the Jacobian
projective coordinate is the same when three or four
multipliers.
V. RESULTS & PERFORMANCE ANALYSIS
The lower bound on the area-time cost of a given
design is usually employed as a performance metric
(area) x (time) 2α , 0 ≤ α ≤ 1, where the choice of α
determines the relative importance of area and time [22].
Such lower bounds have been obtained for several
problems, e.g., discrete Fourier transform, matrix
multiplication, binary addition, and others [22]. Once the
lower bound on the chosen performance metric is known,
designers attempt to devise algorithms and designs which
are optimal for a range of area and time values. Even
though a design might be optimal for a certain range of
area and time values, it is nevertheless of interest to
obtain designs for minimum values of time, i.e.,
maximum speed performance, as well as designs for
minimum area. In order to make a more meaningful
comparison between the selected projective coordinate
systems with parallel multipliers, both the AT and AT 2
measures are evaluated.
Table IV shows the AT and AT 2 measures for the
selected projective coordinate systems with m = 160 bits.
In Table IV, the Area (A) is the number of multipliers.
The Time (T), on the other hand, is calculated using the
NAF binary algorithm as:
T = m(DBL) + m/3(ADD),
where DBL and ADD are the required time units for
performing point doubling and addition respectively in
Tables III, IV and V. For example, T = 160 ×(4) + ×
160 × (6) = 960 time units for Edwards coordinate system
with two parallel multipliers. Another example with the
Jacobian coordinate system with three multipliers gives:
T = 160 ×(5) + × 160 × (5) = 1066.66667 time units.
Fig. 4 and Fig. 5 depict the comparisons results of
Table IV for AT and AT 2 respectively. The results show
that the Edwards coordinate system provides the best AT
and AT 2 results. A key observation is that the Edwards
coordinate system provides better AT and AT 2 using only
two multipliers when compared to the other two
coordinate systems with four multipliers, which makes
the Edwards coordinate system more attractive.
Despite that the Jacobian coordinate system provides
better performance than the Homogenous coordinate
system with sequential designs [23], the results show that

102 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
the Homogenous and the Jacobian coordinate systems
provide the same AT and AT 2 when three multipliers are
used. The results also show that the Homogenous
coordinate system provides better AT and AT 2 than the
Jacobian coordinate system when four multipliers are
used. This is because of the nature of the critical path of
the Homogenous coordinate system that allows for more
parallelism when four multipliers are employed.
VI. CONCLUSION
In this paper, the performance of the Homogeneous,
Jacobian and Edwards coordinate systems with sidechannel
atomicity have been analyzed when parallel
GF(p) field multipliers are used. The point operations of
the selected projective coordinate systems have been
segmented into atomic blocks. These atomic block are
executed in parallel using 2, 3 and 4 multipliers. An
atomic block can contain at most one field multiplication,
two field additions, and one field subtraction. A Field
squaring is performed by a multiplier instead of using a
special hardware unit for squaring.
The AT and AT 2 performance metric have been
evaluated for each of the selected projective coordinate
systems. The results show that the Edwards coordinate
system provides the best AT and AT 2 as compared to the
other two coordinate systems. The results also show that
the Homogenous coordinate system provides better
performance than the Jacobian coordinate systems when
four multipliers are used.
ACKNOWLEDGMENT
The author would like also to acknowledge the support
of Umm Al-Qura University (UQU).
REFERENCES
[1] N. Koblitz, “Elliptic curve cryptosystems,” Mathematics of
Computation, Vol. 48, pp. 203-209, 1987.
[2] R. Rivest, A. Shamir, and L. Adleman, “A method for obtaining
digital signatures and public key cryptosystems,”
Communications of the ACM, Vol. 21, No.2, pp. 120-126, 1978.
[3] T. ElGamal, “A Public-Key Cryptosystem and a Signature
Scheme Based on Discrete Logarithms,” Advances in
Cryptology: Proceedings of CRYPTO 84, Springer Verlag, pp.
10-18, 1985.
[4] P. Kocher, J. Jaffe, and B. Jun, “Differential power analysis,”
CRYPTO '99, LNCS 1666, pp. 388-397, 1999.
[5] P. Kocher, “Timing Attacks on Implementations of Diffe-
Hellman, RSA, DSS, and Other Systems,” CRYPTO '96, LNCS
1109, pp. 104-113, 1996.
[6] P. Fouque and F. Valette, “The doubling attack – why upwards is
better than downwards,” Cryptographic Hardware and Embedded
Systems – CHES’03, LNCS 2779, Springer-Verlag, pp.269–280,
2003
[7] L. Goubin, “A refined power-analysis attack on elliptic curve
cryptosystems,” Public Key Cryptography – PKC’03, LNCS
2567, Springer-Verlag, pp.199–210, 2003.
[8] T. Akishita, and T. Takagi, “Zero-value point attacks on elliptic
curve cryptosystem,” Information Security Conference – ISC’03,
LNCS 2851, Springer-Verlag, pp.218–233, 2003.
[9] H. Cohen, G. Frey, R. M. Avanzi, C. Doche, T. Lange, K.
Nguyen and F. Vercauteren, “Handbook of Elliptic and
Hyperelliptic Curve Cryptography,” Discrete Mathematics and
Its Applications, vol. 34, Chapman & Hall/CRC, 2005.
© 2010 ACADEMY PUBLISHER
[10] D. Hankerson, A. J. Menezes, and S. Vanstone, “Guide to Elliptic
Curve Cryptography,” Springer-Verlag, 2004.
[11] B. Chevallier-Mames, M. Ciet, and M. Joye, “Low-cost solutions
for preventing simple side-channel analysis: side-channel
atomicity,” IEEE Trans. Computers, Vol. 53, No. 6, pp. 760-768,
2004.
[12] P. Mishra, , “Pipelined computation of scalar multiplication in
elliptic curve cryptosystems (extended version),” IEEE Trans. On
Computers, Vol. 55, No. 8, pp. 1000-1010, 2006.
[13] T.F. Al-Somani, “Overlapped parallel computations of scalar
multiplication with resistance against Side Channel Attacks,” Int.
J. Information and Computer Security, Vol. 2, No. 3, pp.261–
278, 2008.
[14] H. M. Edwards, “A normal form for elliptic curves,” Bulletin of
the American Mathematical Society 44, pp. 393–422, 2007.
[15] D. J. Bernstein and T. Lange, “Faster addition and doubling on
elliptic curves,”, Advances in Cryptology – ASIACRYPT 2007,
LNCS 4833, Springer-Verlag, pp.29-50, 2007.
[16] A. Gutub and M. K. Ibrahim, “High Radix Parallel Architecture
For GF(p) Elliptic Curve Processor,” IEEE Conference on
Acoustics, Speech, and Signal Processing, ICASSP 2003, Pages:
625- 628, Hong Kong, April 6-10, 2003.
[17] A. Gutub, “Fast 160-Bits GF(p) Elliptic Curve Crypto Hardware
of High-Radix Scalable Multipliers,” International Arab Journal
of Information Technology (IAJIT), Vol. 3, No. 4, Pages: 342-
349, October 2006.
[18] A. Gutub, M. K. Ibrahim and T. Al-Somani, “Parallelizing GF(P)
Elliptic Curve Cryptography Computations for Security and
Speed,” IEEE International Symposium on Signal Processing and
its Applications in conjunction with the International Conference
on Information Sciences, Signal Processing and their
Applications (ISSPA), Sharjah, United Arab Emirates, February
12-15,2007.
[19] A. Gutub, “Efficient Utilization of Scalable Multipliers in
Parallel to Compute GF(p) Elliptic Curve Cryptographic
Operations,” Kuwait Journal of Science & Engineering (KJSE),
Vol . 34, No. 2, Pages: 165-182, December 2007.
[20] A. Menezes, “Elliptic Curve Public Key Cryptosystems,” Kluwer
Academic Publishers, 1993.
[21] M. Joye, and C. Tymen, “Compact Encoding of Non-Adjacent
Forms with Applications to Elliptic Curve Cryptography,” Public
Key Cryptography, LNCS 1992, Springer-Verlag, pp. 353-364,
2001.
[22] D. Thompson, “A complexity theory for VLSI,” Ph.D.
dissertation, Carnegie Mellon University, Dep. Computer
Science, 1980.
[23] I. Blake, G. Seroussi and N. Smart, “Elliptic Curve in
Cryptography,” Cambridge University Press, New York, 1999.
Turki F. Al-Somani received his B.Sc. and M.Sc. degrees in
Electrical and Computer Engineering from King Abdul-Aziz
University, Saudi Arabia in 1997 and 2000, respectively. He
obtained his PhD degree from King Fahd University of
Petroleum and Minerals (KFUPM), Saudi Arabia in 2006.
Currently, he is an assistant professor at the Computer
Engineering Department in Umm Al-Qura University (UQU),
Saudi Arabia. His research interests include computer
arithmetic, System-on-Chip designs, security and
cryptosystems, theories of information, application specific
processors and FPGAs. He published several journal and
conference papers in the areas of his research.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 103
TABLE I
FIELD ARITHMETIC OPERATIONS OF THE SELECTED PROJECTIVE COORDINATE SYSTEMS
Homogeneous Coordinate System Jacobian Coordinate System Edwards Coordinate System (with c = 1)
Mixed Addition Doubling Mixed Addition Doubling Mixed Addition Doubling
α1 A = Y2 × Z1 α1 Z1 2 = Z1 × Z1 α1 Z1 2 = Z1 × Z1 α1 Y1 2 = Y1 × Y1 α1 B = Z1 × Z1 β1 A = X1 + Y1
α2 X2 × Z1 α2 X1 × Y1 α2 Y2 × Z1 β1 2Y1 2 = Y1 2 +
Y1 2
α2 C = X1 × X2 α1 C = X1 × X1
β1 B = X2Z1– X1 α3 B = Y1 × Z1 α3 B = X2 × Z1 2 α2 X1 2 = X1 × X1 α3 D = Y1 × Y2 α2 D = Y1 × Y1
α3 B 2 = B × B α4 X1 2 = X1 × X1 β1 E = B – A β2 2X1 2 = X1 2 +
X1 2
β1 X1 + Y1 β2 E = C + D
α4 A 2 = A × A β1 2X1 2 = X1 2 +
X1 2
α5 B 3 = B 2 × B β2 3X1 2 = 2X1 2 +
X1 2
α4
D = Y2 Z1 ×
Z1 2
β3
3X1 2 = 2X1 2 +
X1 2
β2 X2 + Y2 β3 C – D
β2 F = D – C α3 Z1 2 = Z1 × Z1 α4 G = (X1 + Y1) ×
(X2 + Y2)
α3
B = A × A
α6 A 2 × Z1 α5 a × Z1 2 α5 E 2 = E × E α4 Y1 × Z1 β3 C + D β4 F = B – E
α7 B 2 × X1 β3 A = a4 Z1 2 +
3X1 2
α6 F 2 = F × F β4 Z3 = Y1Z1 + β4 D – C α4 H = Z1 × Z1
Y1Z1
β2 2B 2 X1 = B 2 X1 +
B 2 α6 C = X1Y1 × B α7 Z3 = Z1 × E α5 X1 × Y1
X1
2 β5 E = G – (C + β5 I = H + H
D)
β3 (B 3 + 2B 2 X1) β4 2C = C + C α8 E 3 = E 2 × E β5 2X1Y1 2 =
X1Y1 2 + X1Y1 2
α5 C × D β6 J = E – I
β4
C = A 2 Z1 – (B 3
+ 2B 2 X1)
β5 4C = 2C + 2C α9 A × E 2 β6 A = 2X1Y1 2 +
2X1Y1 2
α8 X3 = B × C β6 8C = 4C + 4C β3 2AE 2 = AE 2 +
AE 2
α9 Z3 = B 3 × Z1 α7 B 2 = B × B β4 E 3 +2AE 2 α6 4Y1 4 = 2Y1 2 ×
2Y1 2
β5 (B 2 X1– C) β7 2B 2 = B 2 + B 2 β5 X3 = F 2 – (E 3
α10 A × (B 2 X1– C) β8 4B 2 = 2B 2 +
2B 2
α11 B 3 × Y1 β9 8B 2 = 4B 2 +
4B 2
β6
Y3 = A ×
(B 2 X1– C) –
B 3 Y1
© 2010 ACADEMY PUBLISHER
+2AE 2 )
α8 Y1 2 = Y1 × Y1 β6 Y3 = F (AE 2 –
X3) – CE 3
α6 Z1 × (D – C) α5 X3 = F × J
β7 2A= A + A α7 Z1 × E α6 Z3 = E × J
β8
8Y1 4 = 4Y1 4 +
4Y1 4
α8 F = d × CD α7 Y3 = E × (C – D)
β6 B – F
α10 C × E 3 α7 Z1 4 = Z1 2 × Z1 2 β7 B + F
α11 F × (AE 2 – X3) α8 a × Z1 4 α9 X3 = Z1E × (B
β9
B = 3X1 2 +
a4Z1 4
α10
– F)
Z3 = (B – F) ×
(B + F)
α9 A 2 = A × A α9 B 2 = B × B α11 Y3 = Z1(D – C)
× (B + F)
β10 D = A 2 – 8C β10 X3 = B 2 – 2A
β11 4C – D β11 A– X3
α10 B × (A– X3)
α10 Z3 = 8B 2 × B
α11 Y1 2 × –8B 2 β12 Y3 = B(A– X3)
– 8Y1 4
α12 B × D
β12 X3 = BD + BD
α13 A × (4C – D)
β13 Y3 = A(4C –
D) – 8Y1 2 B 2

TABLE II
104 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
© 2010 ACADEMY PUBLISHER
POINT OPERATIONS IN ATOMIC BLOCKS
Homogeneous Coordinate System Jacobian Coordinate System Edwards Coordinate System (with c = 1)
Mixed Addition Doubling Mixed Addition Doubling Mixed Addition Doubling
Γ1 α1 ∆1 α1 Γ1 α1 ∆1 α1 Γ1 α1 ∆1 β1
* * * β1 * α1
* * * * * *
* * * * * *
Γ2 α2 ∆2 α2 Γ2 α2 ∆2 α2 Γ2 α2 ∆2 α2
β1 * * β2 * *
* * * β3 * β2
* * * * * β3
Γ3 α3 ∆3 α3 Γ3 α3 ∆3 α3 Γ3 α3 ∆3 α3
* * β1 * β3 *
* * * * β4 *
* * * * * β4
Γ4 α4 ∆4 α4 Γ4 α4 ∆4 α4 Γ4 β1 ∆4 α4
* β1 β2 β4 β2 β5
* β2 * * α4 *
* * * * β5 β6
Γ5 α5 ∆5 α5 Γ5 α5 ∆5 α5 Γ5 α5 ∆5 α5
* * * β5 * *
* * * β6 * *
* β3 * β7 * *
Γ6 α6 ∆6 α6 Γ6 α6 ∆6 α6 Γ6 α6 ∆6 α6
* β4 * β8 * *
* β5 * * * *
* β6 * * * *
Γ7 α7 ∆7 α7 Γ7 α7 ∆7 α7 Γ7 α7 ∆7 α7
β2 β7 * * * *
β3 β8 * * * *
β4 β9 * * * *
Γ8 α8 ∆8 α8 Γ8 α8 ∆8 α8 Γ8 α8
* β10 * β9 β6
* β11 * * β7
* * * * *
Γ9 α9 ∆9 α9 Γ9 α9 ∆9 α9 Γ9 α9
* * β3 β10 *
* * β4 * *
* * β5 * *
Γ10 β5 ∆10 α10 Γ10 α10 ∆10 α10 Γ10 α10
α10 * * β11 *
* * * * *
* * * * *
Γ11 α11 ∆11 α11 Γ11 α11 Γ11 α11
β6 * β6 *
* * * *
* * * *
∆12 α12
β12
*
*
∆13 α13
β13
*
*

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 105
TABLE III
POINT OPERATIONS FOR THE HOMOGENEOUS COORDINATE SYSTEM WITH PARALLEL MULTIPLIERS
No. of
Multipliers
Homogeneous Coordinate System
Time Mixed Addition Doubling
Mul1 Mul2 Mul1 Mul2
2 1 Γ1 Γ2 ∆1 ∆2
2 Γ3 Γ4 ∆3 ∆4
3 Γ5 Γ6 ∆5 ∆6
4 Γ7 Γ8 ∆7 ∆8
5 Γ9 Γ10 ∆9 ∆10
6 Γ11 ∆11 ∆12
7 ∆13
Mul1 Mul2 Mul3 Mul1 Mul2 Mul3
3 1 Γ1 Γ2 ∆1 ∆2 ∆3
2 Γ3 Γ4 ∆4 ∆5 ∆6
3 Γ5 Γ6 Γ7 ∆7 ∆8 ∆9
4 Γ8 Γ9 Γ10 ∆10 ∆11 ∆12
5 Γ11 ∆13
Mul1 Mul2 Mul3 Mul4 Mul1 Mul2 Mul3 Mul4
4 1 Γ1 Γ2 ∆1 ∆2 ∆3 ∆4
2 Γ3 Γ4 ∆5 ∆6 ∆7 ∆8
3 Γ5 Γ6 Γ7 ∆9 ∆10 ∆11
4 Γ8 Γ9 Γ10 Γ11 ∆12 ∆13
TABLE IV
POINT OPERATIONS FOR THE JACOBIAN COORDINATE SYSTEM WITH PARALLEL MULTIPLIERS
No. of
Multipliers
© 2010 ACADEMY PUBLISHER
Jacobian Coordinate System
Time Mixed Addition Doubling
Mul1 Mul2 Mul1 Mul2
2 1 Γ1 Γ2 ∆1 ∆2
2 Γ3 Γ4 ∆3 ∆4
3 Γ5 Γ6 ∆5 ∆6
4 Γ7 Γ8 ∆7 ∆8
5 Γ9 Γ10 ∆9
6 Γ11 ∆10
Mul1 Mul2 Mul3 Mul1 Mul2 Mul3
3 1 Γ1 Γ2 ∆1 ∆2 ∆3
2 Γ3 Γ4 ∆4 ∆5 ∆6
3 Γ5 Γ6 Γ7 ∆7 ∆8
4 Γ8 Γ9 ∆9
5 Γ10 Γ11 ∆10
Mul1 Mul2 Mul3 Mul4 Mul1 Mul2 Mul3 Mul4
4 1 Γ1 Γ2 ∆1 ∆2 ∆3 ∆4
2 Γ3 Γ4 ∆5 ∆6 ∆7
3 Γ5 Γ6 Γ7 ∆8
4 Γ8 Γ9 ∆9
5 Γ10 Γ11 ∆10

106 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
TABLE V
POINT OPERATIONS FOR THE EDWARDS COORDINATE SYSTEM WITH PARALLEL MULTIPLIERS
No. of
Multipliers
Edwards Coordinate System (with c = 1)
Time Mixed Addition Doubling
Mul1 Mul2 Mul1 Mul2
2 1 Γ1 Γ2 ∆1 ∆2
2 Γ3 Γ4 ∆3 ∆4
3 Γ5 Γ6 ∆5 ∆6
4 Γ7 Γ8 ∆7
5 Γ9
6 Γ10 Γ11
Mul1 Mul2 Mul3 Mul1 Mul2 Mul3
3 1 Γ1 Γ2 Γ3 ∆1 ∆2
2 Γ4 Γ5 Γ6 ∆3 ∆4
3 Γ7 Γ8 ∆5 ∆6 ∆7
4 Γ9 Γ10 Γ11
Mul1 Mul2 Mul3 Mul4 Mul1 Mul2 Mul3 Mul4
4 1 Γ1 Γ2 Γ3 Γ4 ∆1 ∆2 ∆3 ∆4
2 Γ5 Γ6 Γ7 ∆5 ∆6 ∆7
3 Γ8
4 Γ9 Γ10 Γ11
TABLE VI
AT & AT 2 COMPARISONS (with m = 160 bits)
Projective Coordinate System Jacobian Coordinate System Edwards Coordinate System
Area (A) = No. of Multipliers Time (T) AT AT 2 Time (T) AT AT 2 Time (T) AT AT 2
2 1440 2880 4147200 1280 2560 3276800 960 1920 1843200
3 1066.6667 3200 3413333.3 1066.6667 3200 3413333.3 693.33333 2080 1442133
4 853.33333 3413.3333 2912711.1 1066.6667 4266.6667 4551111.1 533.33333 2133.333 1137778
© 2010 ACADEMY PUBLISHER

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 107
α2
β1
α3
© 2010 ACADEMY PUBLISHER
α1
α4
α5 α6 α7
β2
β3
β4
β5
α8 α9 α10 α11
β6
α1 α2 α3 α4
α5 α6 α7
(a) Point Addition (b) Point Doubling
Figure1. The data dependency graph of the Homogenous coordinate system.
β3
α9
β10
β11
α12
β12
β4
β5
β6
α13
β13
β7
β8
β9
α10
β1
β2
α8
α11

108 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
α2
α3
β1
α1
α4
α5 α6 α7
α8
α10
© 2010 ACADEMY PUBLISHER
β2
α9
β3
β4
β5
α11
β6
α1 α2 α3 α4
β1
β2
β3
α5 α6 α7
(a) Point Addition (b) Point Doubling
Figure 2. The data dependency graph of the Jacobian coordinate system.
β5
β6
β7
β8
α8
β9
α9
β10
β11
α10
β12
β4

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 109
© 2010 ACADEMY PUBLISHER
(a) Point Addition (b) Point Doubling
Figure 3. The data dependency graph of the Edwards coordinate system.
Figure 4. Area x Time (AT) Comparisons.
Figure 5. Area x Time 2 (AT 2 ) Comparisons.

110 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
VPRS-Based Knowledge Discovery Approach in
Incomplete Information System
Shibao SUN 1,2
1. Electronic Information Engineering College, Henan University of Science and Technology, Luoyang
Henan 471003, China
2. National Lab. Of Software Development Environment, Beijing University of Aeronautics & Astronautics,
Beijing, 100191, China
Email: sunshibao@126.com
Ruijuan ZHENG 1 , Qingtao WU 1 , and Tianrui LI 3
1. Electronic Information Engineering College, Henan University of Science and Technology, Luoyang
471003, China
3. School of Information Science and Technology, Southwest Jiaotong University, Chengdu, 610031,China.
Email: rjwo@163.com, wqt8921@126.com, trli30@gmail.com
Abstract—Through changing the equivalence relation in the
incomplete information system, a new variable precision
rough set model and an approach for knowledge reduction
are proposed. To overcome no monotonic property of the
lower approximation, a cumulative variable precision rough
set model is explored, and the basic properties of cumulative
lower and upper approximation operators are investigated.
The example proves that the cumulative variable precision
rough set model has wide range of applications and better
result than variable precision rough set model.
Index Terms—Variable precision rough set, Incomplete
information system, Decision table, Cumulative
approximation
I. INTRODUCTION
Rough set theory (RST) has been proposed by
Pawlak [1] as a tool to conceptualize, organize and
analyze various types of data in knowledge discovery.
This method is especially useful for dealing with
uncertain and vague knowledge in information systems.
Many examples about applications of the rough set
method to process control, economics, medical diagnosis,
biochemistry, environmental science, biology, chemistry,
psychology, conflict analysis and other fields can be
found in [2,3]. However, the classical rough set theory is
based on an equivalence relation and can not be applied
in many real situations. Therefore, many extended RST
models, e.g. binary relation based rough sets [4], covering
This work is supported by National Natural Science Foundation of China
(Grant No. 60873108), the Special Research Funding to Doctoral Subject
of Higher Education Institutes in P.R. China (Grant No. 20060613007),
Supported by Natural Science Foundation of Henan Province of
China(Grant No. 072300410210), Natural Science Research Foundation of
Henan University of Science and Technology (Grant No.09001172).
Corresponding author:
Email: sunshibao@126.com (Shibao SUN).
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.110-116
based rough sets [5,6], and fuzzy rough sets [7,8] have
been proposed. In order to solve classification problems
with uncertain data and no functional relationship
between attributes and relax the rigid boundary definition
of the classical rough set model to improve the model
suitability, the variable precision rough set (VPRS) model
was proposed by Ziarko [9] in 1993. It is an effective
mathematical tool with an error-tolerance capability to
handle uncertainty problem. Basically, the VPRS is an
extension of classical rough set theory [1-3], allowing for
a partial classification. By setting a confidence threshold,
β (0 ≤ β < 0.5) , the VPRS can allow noise data or
remove error data [10]. Recently the VPRS model has
been widely applied in many fields [11].
The key issues of VPRS model mainly concentrates
on generalization of models and development of
reduction approaches under the equivalence relation. For
example, β -reduct [12], β lower (upper) distribution
reduction [13] and reduction based on structure [14], etc,
are reduction approaches under the equivalence relation.
However, in many practical problems, the equivalence
relation of objects is difficult to construct, or the
equivalence relation of objects essentially does not exist.
In this case, we need to generalize the VPRS model. The
ideas of generalization are from two aspects. One is to
generalize approximated objects from a crisp set to a
fuzzy set [15]; The other is to generalize the relation on
the universe from the equivalence relation to the fuzzy
relation [15], binary relation [16], or covering relation
[17,18]. The idea of the VPRS was introduced to fuzzy
rough set and the theory and application of fuzzy rough
set were discussed in [15]. The equivalence relation was
generalized to a binary relation R on the universe U in
the VPRS model, so that a generalized VPRS model was
obtained [16]. Covering rough set model [19] has been
obtained when the equivalence relation on the universe
was generalized to cover on the universe in rough set
model. The equivalence relation was generalized to cover

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 111
on universe U in the VPRS model and two kinds of
variable precision covering rough set models were
obtained in [17,18]. The definition of the variable
precision rough fuzzy set model under the equivalence
relation was given in [20].
The classical rough set approach requires the data
table to be complete, i.e., without missing values. In
practice, however, the data table is often incomplete. To
deal with these cases, Greco, et al [21] proposed an
extension of the rough set methodology to the analysis of
incomplete data tables. The extended indiscernible
relation between two objects is considered as a
directional statement where a subject is compared to a
referent object. It requires that the referent object has no
missing values. The extended rough set approach
maintains all good characteristics of its original version.
It also boils down to the original approach when there is
no missing value. The rules induced from the rough
approximations defined according to the extended
relation verify a suitable property: they are robust in a
sense that each rule is supported by at least one object
with no missing value on the condition attributes
represented in the rule. Obviously, these ideas can be
used to the VPRS model.
The classical and the generalized VPRS approach
based on indiscernible relations also require the data table
to be complete. In this paper, two kinds of VPRS
approaches for dealing with incomplete dada table are
proposed. The paper is organized as follows. In Section 2,
a general view of VPRS approach and incomplete
information system are given. In Section 3, based on the
extended indiscernible relation, we propose a new VPRS
model and an approach for knowledge reduction in the
incomplete information system. In Section 4, a
cumulative VPRS model in the incomplete information
system is discussed. They are based on the cumulative β
lower (upper) approximation of X . In Section 5, we
present an illustrative example which is intended to
explain the concepts introduced in Section 3 and Section
4. The paper ends with conclusions and further research
topics in Section 6.
II. PRELIMINARIES AND NOTATIONS
Definition 1 [1]. An information system is the 4tuple
S = ( U, Q, V, f)
, where U is a non-empty finite
set of objects (universe), Q= { q1, q2, L , qm}
is a finite
set of attributes, V q is the domain of the attribute q ,
V = U V and f : U× Q→ V is a total function such
q∈Q q
that f ( xq , ) ∈ Vq
for each q∈ Q , x ∈ U , called an
information function. If Q= CU { d}
and
CI { d } =∅,
then S = ( U, CU { d}, V, f)
is called
a decision table, where d is a decision attribute.
© 2010 ACADEMY PUBLISHER
To every (non-empty) subset of attributes P⊆ C is
associated an indiscernible relation on U , denoted by
R P :
RP = {( xy , ) ∈ U× U: f( xq , ) = f( yq , ) ∀q∈ P}.
(1)
If ( x, y) ∈ RP,
it is said that the objects x and y are P -
indiscernible. Clearly, the indiscernible relation thus
defined is an equivalence relation (reflexive, symmetric
and transitive). The family of all the equivalence classes
of the relation R p is denoted by U / R p and the
equivalence class containing an element x ∈ U by
[ x] = { y∈U :( x, y) ∈ Rp}
. The equivalence classes
of the relation R p are called P -elementary sets. If
P = C , the C -elementary sets are called atoms.
Definition 2 [9]. Let X and Y be subsets of nonempty
finite universe U , if every e∈ X then e∈ Y ,
we call Y contain X . It is described as Y ⊇ X . Let
⎧1
− | X IY
|/| X |, |X|>0,
cXY ( , ) = ⎨ (2)
⎩0,
|X|=0,
where | X | is cardinality of set X . cXY ( , ) is called
the relative error ratio for X with regard to Y .
Definition 3 [9]. Let S be a decision table, X a
nonempty subset of U , 0≤ < 0.5 and ∅≠P⊆ C .
The β lower approximation and the β upper
approximation of X in S are defined, respectively, by:
P ( X) = { x∈U: c([ x], X)
≤ β}.
(3)
β
β
Pβ( X) = { x∈ U: c([ x], X)
< 1 − β}.
(4)
The elements of Pβ ( X)
are those objects x ∈ U
which belong to the equivalence classes generated by the
indiscernible relation R p , contained in X with the error
β
ratio β ; the elements of P ( X)
are all and only those
objects x ∈ U which belong to the equivalence classes
generated by the indiscernible relation R p , contained in
X with the error ratio 1− β .
Definition 4 [21]. An information system is called an
incomplete information system if there exists x ∈ U and
a∈ C that satisfy that the value f ( xa , ) is unknown,
denoted as “*”. It assumes here that at least one of the
states of x in terms of P is certain where P⊆ C , i.e.
∃a∈ P such that f ( xa , ) is known. Thus,
V = VC UVd U {*} .
Definition 5 [21]. ∀x, y∈ U , object y is called
subject and object x , referent. Subject y is
indiscernible with referent x , with respect to condition
attributes from P⊆ C (denoted as yI Px
), if for every

112 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
q∈ P the following conditions satisfy: (1) f( x, q) ≠ * ,
(2) f( x, q) = f( y, q) or f( y, q)
= * , where “*”
denotes a missing value.
The above definition means that the referent object
considered for indiscernible with respect to P should
have no missing values on attributes from set P . The
binary relation I P is not necessarily reflexive and also
not necessarily symmetric. However, I P is transitive.
For each P⊆ C , let us define a set of objects
having no missing values on attributes from P :
U x U f x q for eachq P
P
= { ∈ : ( , ) ≠* ∈ }. (5)
III. VARIABLE PRECISION ROUGH SET MODEL IN THE
INCOMPLETE INFORMATION SYSTEM
Definition 6. Let S be an incomplete information
system, X a nonempty subset of U , 0≤ β < 0.5 and
P C
∅≠ ⊆ . The β lower approximation and the β
upper approximation of X in S are defined,
respectively, by:
I
P X x U c I x X β
β
I
( ) = { ∈ : ( ( ), ) ≤ }. (6)
P P
Pβ( X) = { x∈ UP: c( IP( x), X)
< 1 − β}.
(7)
I
The elements of Pβ ( X)
are those objects x ∈ U
which belong to the equivalence classes generated by the
indiscernible relation I P , contained in X with the error
I
ratio β ; the elements of Pβ ( X)
are all and only those
objects x ∈ U which belong to the equivalence classes
generated by the indiscernible relation I P , contained in
X with the error ratio 1− β .
The β boundary of X in S , denoted by
I
BN ( X)
, is:
β
I
BNβ ( X ) = { x∈ UP: β < c( IP( x), X ) < 1 − β}.
(8)
The β negative domain of X in S , denoted by
NEGβ ( X ) , is:
I
NEGβ ( X ) = { x∈UP: c( IP( x), X ) ≥1 − β}.
(9)
Corollary 1. When β = 0 , the VPRS model defined
above is equivalent to rough set model in incomplete
information system [21].
Proof. In formula (10), cI ( P ( x), X) ≤ β is
equivalent to cI ( P ( x), X) ≤ 0 , so
1 ≤| I ( x) I X |/| I ( x)|
, such that I ( x) ⊆ X , that
P P
I
is to say, P ( X)
β
is equivalent to PX ( ) .
I
Analogously, Pβ ( X)
is equivalent to PX ( ) .
© 2010 ACADEMY PUBLISHER
P
Corollary 2. If an information system is complete,
the VPRS model defined above is equivalent to the
classical VPRS model.
Proof. ∀x, y∈ U , P⊆ C , yI Px
, then for every
q∈ P , we have (1) f( x, q) ≠ * , (2)
f( x, q) = f( y, q) or f( y, q)
= * . In a complete
information system, we have f ( xq , ) = f( yq , ) , so
IP ( x ) is equal to [ x ] , such that formula (6) is
equivalent to formula (3) and formula (7) is equivalent to
formula (4).
Theorem 1. ∀X ⊆ U
where (~ X = UX - ) .
I I
, P (~ X) = NEG ( X)
β β
,
Proof. From Definition 6,
I | IP( x) I (~ X)|
Pβ(~ X) = { x∈UP:1 − ≤ β}
| I ( x)|
| IP( x) I (~ X)|
= { x∈UP: ≥1 −β}
| IP( x)|
| IP( x) I X |
= { x∈UP: ≤ β}
| IP( x)|
| IP( x) I X |
I
= { x∈UP:1− ≥1 −β}
= NEG ( X ) .
β
| I ( x)|
P
Theorem 2. Let S be an incomplete information
system, X , Y are two nonempty subsets of U ,
0≤ β < 0.5 and ∅ ≠ P⊆ C . The rough
approximations defined as above satisfy the following
properties:
I
β ⊆
I
β ;
I
β
I
I
β =
I
β = ;
I I
β β
⊆ ⇒
I
β ⊆
I
β
I
β U ⊇
I
β U
I
β
;
I
β I ⊆
I
β I
I
β
;
I
β U ⊇
I
β U
I
β ;
I
β I ⊆
I
β I
I
β ;
I
β =
I
β ;
I
β =
I
β
;
① P ( X) P ( X)
② P ( ∅ ) = P β ( ∅ ) =∅; P ( U) P ( U) U
③ X ⊆Y ⇒ P ( X) ⊆ P ( Y)
;
④ X Y P ( X) P ( Y)
⑤ P ( X Y) P ( X) P ( Y)
⑥ P ( X Y) P ( X) P ( Y)
⑦ P ( X Y) P ( X) P ( Y)
⑧ P ( X Y) P ( X) P ( Y)
⑨ P (~ X) ~ P ( X)
⑩ P (~ X) ~ P ( X)
Proof. ① Because of 0 β 0.5
I
β (
I
) such that x satisfies Pβ ( X)
.
P X
P
≤ < , x satisfies

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 113
I
② Because of 0≤ β < 0.5 and X =∅, we have
I
P ( ∅ ) =∅ and P ( )
β
β ∅ =∅ . Therefore,
I
I
= β
β = .
P ( U) P ( U) U
X Y
I
③ ∀x∈ P ( X)
, cI ( ( x), X) β
β
P
≤ , when
⊆ , we have cI ( ( x), Y) ≤ β , that is to say,
x P ( Y)
I
∈ .
β
P
I I
④ Similar to ③, we have P ( X) ⊆ P ( Y)
.
β β
⑤ From ③, we have
I I I
P ( X UY) ⊇ P ( X) U P ( Y)
.
β β β
⑥ From ③, we have
I I I
P ( X IY) ⊆ P ( X) I P ( Y)
.
β β β
⑦ From ④, we have
I I I
P ( X UY) ⊇ P ( X) U P ( Y)
.
β β β
⑧ From ④, we have
I I I
P ( X IY) ⊆ P ( X) I P ( Y)
.
β β β
⑨ From Theorem 1,
I | IP( x) I X |
Pβ(~ X) = { x∈UP:1− ≥1 −β}
| IP( x)|
| IP( x) I X |
I
= ~{ x∈UP:1− < 1 −β}
= ~ P β ( X)
.
| I ( x)|
P
I I
⑩ Similar to ③, we have P (~ X) ~ P ( X)
β
= .
The following ratio defines a β accuracy of the
approximation of X ⊆ U , X ≠∅, by means of the
attributes from P⊆ C :
I
| P ( X)|
β
α P ( X ) = I .
(10)
| P ( X)|
Obviously, 0 ( ) 1 α ≤ < .
P X
Another ratio defines a β quality of the
approximation of X by means of the attributes from
P⊆ C :
I
| P ( X)|
β λ P ( X ) = .
(11)
| X |
The quality λ ( ) represents the relative frequency of
P X
the objects with error ratio β correctly classified by
means of the attributes from P .
A primary use of rough set theory is to reduce the
number of attributes in databases thereby improving the
performance of applications in a number of aspects
including speed, storage, and accuracy. For a data set
with discrete attribute values, this can be done by
reducing the number of redundant attributes and find a
© 2010 ACADEMY PUBLISHER
β
β
subset of the original attributes that are the most
informative.
Definition 7. ( β dependability) Suppose that
S = ( U, CU { d}, V, f)
is an incomplete information
system, β dependability is defined as follows:
γ ( Cd , , β) = | posCd ( , , β)|/|
U|.
(12)
where
I
pos( C, d, β ) = U C ( Y ).
(13)
Y∈U / d
Definition 8. ( β approximation reduction) Suppose
that S = ( U, CU { d}, V, f)
is an incomplete
information system, X ⊆ U , a conditional attribute
subset A⊆ C is called a β approximation reduction if
and only if it satisfies: ① γ ( Ad , , β) = γ( Cd , , β)
and
② there does not exist a conditional attribute subset
B ⊆ A , such that γ ( Bd , , β) = γ( Cd , , β)
.
Based on Definition 8, through removing
superfluous attributes, we can obtain a reductive database.
IV. CUMULATIVE VARIABLE PRECISION ROUGH SET
MODEL IN THE INCOMPLETE INFORMATION SYSTEM
Let us observe that a very useful property of the
lower approximation within the classical rough set theory
is that if an object x ∈ U belongs to the lower
approximation of X with respect to P⊆ C , then x
belongs also to the lower approximation of X with
respect to R ⊆ C when P⊆ R (this is a kind of
monotonic property). However, formula (6) does not
satisfy this property of the lower approximation, because
it is possible that f( x, q) ≠ * for all q∈ P but
f( x, q ) = * for some q∈R− P . This is quite
problematic for some key concepts of the variable
precision rough set theory, like β accuracy and β
quality of approximation, and β dependability.
Therefore, another definition of the lower
approximation should be considered. Then the concepts
of β accuracy, β quality , and β dependability of
approximation can be still valid in the case of missing
values.
Definition 9. Given X ⊆ U and P⊆ C ,
* I
P ( X) = U P ( X).
(14)
Then
β
R⊆P β
*
Pβ ( X)
is called as the cumulative β lower
approximation of X because it includes all the objects
belonging to all β lower approximations of X , where
R ⊆ P .
It can be shown that another type of the indiscernible
*
relation, denoted by I , permits a direct definition of the
P
cumulative β lower approximation in a usual way. For
β

114 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
each x, y∈ U and for each P⊆ C ,
*
yI x means that
P
f( x, q) = f( y, q) or f( x, q)
= *
and/or
( , ) *
∈ . Let
f y q = for every q P
* *
I ( x) = { y∈ U : yI x}
for each x ∈ U and for each
P P
*
P⊆ C . I is reflexive and symmetric but not transitive.
P
We can prove that Definition 9 is equivalent to the
following definition:
* * *
Pβ( X) = { x∈UP: c( I ( x), X)
≤ β.
(15)
P
where
*
U = { x∈U: f( x, q) ≠* for at least one q∈ P}.
P
Using the indiscernible relation
I , we can define
the cumulative β upper approximation of X ,
*
β
*
Pβ ( X)
*
P
*
P β
* *
complementary to
P ( X) = { x∈ U : c( I ( x), X)
< 1 − }. (16)
For each X ⊆ U , let X = X I U . Let us remark
*
P
P P
*
that x ∈ U if and only if there exists R ≠∅ such that
P
R ⊆ P and x ∈ U R .
Rough approximations
*
Pβ ( X)
and
*
Pβ ( X)
satisfies the following properties:
① For each X ⊆ U and for each P⊆ C :
* *
Pβ( X) ⊆ Pβ( X)
;
② For each X ⊆ U and for each P⊆ C :
* *
P
*
β
Pβ ( X) = U −P ( U − X)
;
③ For each X ⊆ U and for each PR , ⊆ C , if
P⊆ R , then
*
β
*
Pβ( X) ⊆ R ( X)
. Furthermore, if
* *
* *
UP= UR,
then Pβ( X) ⊇ Rβ( X)
.
Due to the property of monotonic, when augmenting
attributes set P , we get a lower approximation of X
that is at least of the same cardinality. Thus, we can
define analogously for the case of missing values the
following key concepts of the variable precision rough
sets theory: the cumulative β accuracy of approximation
*
of X (denoted as ( ) X α ), the cumulative β quality
P
*
λ ( ) of approximation of X (denoted as ( ) X λ ),
P X
and the cumulative β dependability (denoted as
*
γ ( Cd , , β ) ). These concepts have the same definitions
as those given in Sections 3 but they use rough
*
approximation
Pβ ( X)
.
*
Pβ ( X)
and
© 2010 ACADEMY PUBLISHER
V. AN EXAMPLE
P
The illustrative example presented in this section is
to explain the concepts introduced in Section 3 and
Section 4. The director of the school wants to make a
global evaluation to some students. This evaluation
should be based on the level in Mathematics, Physics and
Literature. However, not all the students have passed all
three exams and, therefore, there are some missing values.
The director made the examples of evaluation as shown
in Table 1.
TABLE I
STUDENT EVALUATIONS WITH MISSING VALUES
Stud
ent
Mathem
atics
Physics Literature
Global
evaluation
1 medium bad bad bad
2 good medium * good
3 medium * medium bad
4 * medium medium good
5 * good bad bad
6 good medium bad good
For β = 0.35 , The lower and upper
approximations can be calculated from Table 1:
Let C = { Mathematics, Physics, Literature}
be
condition attributes and { Global evaluation} be
decision attribute. Let bad = {1,3,5} and
good = {2, 4,6} .
U = {1, 6} , I (1) = {1} , I (6) = {2,6} ,
C
I
C
I
C β ( bad)
= {1} , C β ( bad)
= {1} , C β ( good)
= {6} ,
I
C β ( good)
= {6} , γ ( Cd , , β ) = 1/3.
Let L = { Literature}
, such that
γ ( Ld , , β ) = 1/3 . It is easy to validate that L is
reduction of condition attribute set C .
For β = 0.35 , the cumulative lower and upper
approximations can be calculated from Table 1:
*
*
U = {1,2,3,4,5,6} , I (1) = {1} ,
C
*
I (2) = {2, 4,6} ,
C
C
I
C
*
I (3) = {3, 4} ,
*
I (4) = {2,3,4} , * I (5) = {5} , * I (6) = {2,6} ,
C
C
*
C β ( bad)
= {1,5} ,
*
C ( good)
{2,4,6}
β = , *
*
C
C
C β ( bad)
= {1,3,4,5} ,
C β ( good)
= {2,3,4,6} ,
*
γ ( Cd , , β ) = 5/6.
Let L = { Literature}
and P= { Physics}
, such
that γ ( Ld , , β ) = 5/6 and γ ( Pd , , β ) = 5/6 . It is
easy to validate that L and P are cumulative β
approximation reductions of condition attribute set C .
From this example, we can see that the cumulative
variable precision rough set model better reflects the
rough set’s essence.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 115
VI. CONCLUSIONS
In incomplete information system, a new VPRS
model was obtained through defining the relation of
objects. To overcome no monotonic property of the
proposed VPRS model, a cumulative data relation was
defined and a cumulative VPRS model was established.
However, these models were limited in applications on a
small database with incomplete information. Moreover,
this paper only presented the basic reduction approach for
a decision table. In our future work, we will focus on the
development of reduction algorithms and extracting
minimal exact rules from the large decision table. How to
deal with fuzzy data in incomplete information system
will also be one of our future research work.
ACKNOWLEDGMENT
The authors thank the editor and anonymous referees
for their constructive comments and suggestions, which
have improved the quality of this paper.
REFERENCES
[1] Z. Pawlak, A. Skowron, Rudiments of rough sets.
Information Sciences, vol. 177, no. 1, pp. 3–27, 2007.
[2] Z. Pawlak, A. Skowron, Rough sets: some extensions.
Information Sciences, vol. 177, no. 1, pp. 28–40, 2007.
[3] Z. Pawlak, A. Skowron, Rough sets and boolean reasoning.
Information Sciences, vol. 177, no. 1, pp. 41–73, 2007.
[4] W. Zhu, F.-Y. Wang, Binary relation based rough sets, in:
IEEE FSKD 2006, LNAI, vol. 4223, pp. 276–285, 2006.
[5] W. Zhu, Topological approaches to covering rough sets,
Information Sciences, vol. 177, no. 6, pp. 1499–1508,
2007.
[6] W. Zhu, F.-Y. Wang, On three types of covering rough sets,
IEEE Transactions on Knowledge and Data Engineering,
vol. 19, no. 8, 2007.
[7] K.Y. Qin, Z. Pei, On the topological properties of fuzzy
rough sets, Fuzzy Sets and Systems, vol. 151, no. 3, pp.
601–613, 2005.
[8] W.Z. Wu, W.X. Zhang, Constructive and axiomatic
approaches of fuzzy approximation operators, Information
Sciences, vol. 159, no. 3–4, pp. 233–254, 2004.
[9] W. Ziarko, Variable precision rough set model, Journal of
computer system science, vol. 46, no. 1, pp. 39–59, 1993.
[10] D. Slezak, W. Ziarko, The investigation of the Bayesian
rough set model, International Journal of Approximation
Reason, vol. 40, pp. 81–91, 2005.
[11] Z. Tao, B. D. Xu, D. W. Wang, R. Li, Rough Rules Mining
Approach Based on Variable Precision Rough Set Theory,
Information and Control,vol. 33, no. 1, pp. 18-22, 2004
(in Chinese).
[12] M. Beynon, Reducts within the variable precision rough
sets model: A further investigation, European journal of
operational research, vol. 134, pp. 592-605, 2001.
[13] W. X. Zhang, Y. Liang, W. Z. Wu, Information System
and Knowledge Discovery, Science Press, Beijing, China,
2003 (in Chinese).
[14] M. Inuiguchi, Structure-Based Approaches to Attribute
Reduction in Variable Precision Rough Set Models,
Proceeding of 2005 IEEE ICGC, pp. 34-39, 2005.
[15] A. Mieszkowicz-Rolka, L. Rolka, Remarks on
approximation quality in variable precision fuzzy rough
sets model Rough Sets and Current Trends in Computing:
© 2010 ACADEMY PUBLISHER
4th International Conference, Springer-Verlag, pp. 402-
411, 2004.
[16] Z. T. Gong, B. Z. Sun, Y. B Shao, D. G. Chen, Variable
precision rough set model based on general relations,
Journal of Lanzhou University (Natural Sciences), vol. 41,
no. 6, pp. 110-114, 2005 (in Chinese).
[17] Y. J. Zhang, Y. P. Wang, Covering rough set model based
on variable precision, Journal of Liaoning Institute of
Technology, vol. 26, no. 4, pp. 274-276, 2006 (in Chinese).
[18] S. B. Sun, R. X. Liu, K. Y. Qin, Comparison of Variable
Precision Covering Rough Set Models, Computer
engineering, vol. 34, no. 7, pp. 10-13, 2008 (in Chinese).
[19] W. Zhu,F. Y. Wang, Reduction and axiomization of
covering generalized rough sets, Information Sciences, vol.
152, pp. 217-230, 2003.
[20] A. Mieszkowicz-Rolka, L. Rolka, Fuzziness in Information
Systems, Electronic Notes in Theoretical Computer
Science, vol. 82, no. 4, pp. 1-10, 2003.
[21] X. B. Yang, J. Y. Yang, C. Wu, D. J. Yu, Dominancebased
rough set approach and knowledge reductions in
incomplete ordered information system, Information
Sciences, vol. 178, pp. 1219–1234, 2008.
Shibao SUN was born in Henan, China on 07/10/1970. He
received B.S. degree in computer application from Henan
University, Kaifeng, China in 1994, the M.S. degree in
computer application from Zhengzhou University, Zhengzhou,
China in 2004, the D.E. degree in traffic information
engineering and control from Southwest Jiaotong University,
Chengdu, China in 2008.
In 2008, he was a Post-Doc of Computer Science and
Technology at the National Lab. Of Software Development
Environment, Beijing University of Aeronautics & Astronautics,
Beijing, China. He is currently an Associate Professor in
Electronic Information Engineering College, Henan University
of Science and Technology, Luoyang, China. His research
interests can be summarized as intelligent information
processing, machine learning. Particularly, he is currently
interested in rough sets theory and approaches, as well as their
applications in information processing.
Ruijuan ZHENG was born in Henan Province, China, on
01/20/80. She received B.S. degree in Computer Science and
Technology from Henan University, Kaifeng, China in 2003,
the M.S. and D.E. in bio-inspired computer network security
theory and technology from Harbin Engineering University,
Harbin, China in 2006 and 2008, respectively.
She works in Electronic Information and Engineering
College, Henan University of Science and Technology now. Her
research interests involve Computer immunology, Bio-inspired
dependability, etc.
She is to join the China Computer Federation.
Qingtao WU was born in Jiangxi, China on 03/13/1975. He
received M.S. degrees in computer science from Henan
University of Science and Technology, Luoyang, China in 2003,
the D.E. degree in computer applications from East China
University of Science and Technology, Shanghai, China in 2006.
Currently he is working as Associate Professor and Head in
Dept. of Computer Science, Henan University of Science and
Technology, Luoyang, China. His area of interest includes
network & information security, software formal methods.
Tianrui LI was born in Fujian, China on 06/14/1969. He
received B.S. and M.S. in mathematics from Southwest Jiaotong
University, Chengdu, China in 1992 and 1995, respectively. He

116 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
received D.E. degree in traffic information engineering and
control from Southwest Jiaotong University, Chengdu, China in
2002.
In 2005-2006, he was a Post-Doc of Computer Science and
Engineering at the Belgian Nuclear Research Centre, Belgium.
In 2008, he was a visiting professor for three months in Hasselt
University, Belgium. He is currently a Professor in School of
Information Science and Technology of Southwest Jiaotong
University, Chengdu, China. His research interests can be
summarized as developing effective and efficient data analysis
techniques for novel data intensive applications. Particularly, he
is currently interested in various techniques of data mining,
granular computing and mathematical modeling, as well as their
applications in traffic.
Prof. Li has served as ISKE2007, ISKE2008, ISKE2009
program chairs, IEEE GrC 2009 program vice chair and
RSKT2008, FLINS2010 organizing chair, etc. and has been a
reviewer for several leading academic journals.
© 2010 ACADEMY PUBLISHER

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 117
NaXi Pictographs Input Method and WEFT
Abstract—NaXi Pictograph , which is the only hieroglyph in
use is important for researching into the evolution of the
character. In the past, processing NaXi Pictograph by hand,
that is very inefficient. The NaXi Pictograph information
processing modules is developed, such as pictograph Outline
Font lib, the input method module, the embedded module
and so on. A method of NaXi Pictographs Outline font
extraction is proposed, which has been testified better in the
Linux. As the specialty of NaXi Pictographs,two input
mothods -- NaXi pictographs pinyin and graphic primitive
input method are advanced, by evaluating ,which is found
that the second is better than the first in precision. For
display in web, The web embedding fonts technology of
NaXi Pictographs is brought forward, and The web
embedding fonts technology of NaXi pictographs makes
Internet clients browse NaXi pictographs web without
downloading NaXi pictographs font. The development of the
NaXi Pictograph Information Processing System is
significant to research into NaXi pictographs and apply it.
Index Terms—NaXi pictographs, information processing,
outlines font, IME (Input Method Editors), embedded font,
WEFT
I. INTRODUCTION
NaXi pictograph belongs to the NaXi language of yi
language branch of Tibetan-Burmese languages which
was created by the ancestor of NaXi. It’s credited as “the
only living ancient hieroglyph” and still being used in
writing lection and composition and in the field of
communication, therefore, NaXi pictograph has a special
role in the history of human character. Since hieroglyph
appeared in the early stage of the development of
characters, through a deep study of it we can get
something about the evolutionary history of human
characters and human culture which will make a great
contribution to research on the origin of modern
characters. Many experts and scholars at home and
aboard have been working on NaXi pictograph for a long
time, among which Harvard and Yunnan Academy of
Social Sciences lead a dominant role. But the problem
that a lot of literature and ancient books can’t be
processed efficiently makes it urgent to realizing an
information processing system for NaXi pictograph.
The traditional hand-drawing method of processing
NaXi pictograph has low efficiency and can’t guarantee
the problem can be solved by the output of computer’s
*Corresponding author: Jing-ying Zhao Email: dalianzjy@gmail.com
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.117-124
Hai Guo
Dalian Nationalities University, Dalian, China
Email: guohai@dlnu.edu.cn
Jing-ying Zhao*
Dalian Nationalities University, Dalian, China
Email: dalianzjy@gmail.com
standard font, therefore, our project designs and develops
the NaXi pictograph information processing system for
the first time which ends the processing history of NaXi
pictograph without computerization.
II. NAXI PICTOGRAPH PROCESSING SUMMARIZATION
The Naxi Pictograph input platform is developed for
special groups, including units of printing and publication,
Naxi pictograph research institutes and art designing
companies, so it’s not a universal software platform.
These reasons make it definite that our development must
meet the requirement of different customers. The analysis
of application level of this project can be summarized as
Table I.
According to the analysis above, our project has
designed various Naxi pictograph input processing
systems which can meet specific requirement of different
customers. For professional customer doing publication
and printing, we developed Naxi pictograph standard
edition with TrueType and PostScript outline fonts, and
Four input methods including internal code input method,
Latin input method, graphic primitive Method and English
input method, which can fully satisfy the needs of
printing and typesetting. As for the Naxi pictograph
research institutes, we developed Naxi pictograph
TrueType outline font and Naxi Pinyin input method.
Considering that the artistic designing units just need the
font style outline, their system only have Naxi pictograph
TrueType outline fonts and Naxi pictograph English input
method. At last, we developed web embedding fonts and
PDF document creating technology of Naxi pictograph
TABLE I.
THE ANALYSIS OF CUSTOMER REQUIREMENT OF NAXI PICTOGRAPH
INPUT PLATFORM
Regular
customer requirement
customer group high accuracy, fast input, output, various
convenient input method
Printing
Publication
general output accuracy, advanced input method
research various of output font styles, input method easily
institutes
to learned
artistic offer web pages and electronic document
designing
browsing
ordinary user customer requirement

118 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
for ordinary customers.
III. THE PRINCIPLE OF NAXI PICTOGRAPH INFORMATION
PROCESSING
A. The principle of NaXi Pictograph Outline Fonts
Since outline font can be zoomed in and out at will
without distortion, it is widely applied in printing,
typesetting, character processing and art creating, and
Windows we uses everyday is displayed with outline font
too. This type of font is composed of Bézier curve, and
the TrueType font uses quadratic Bézier curve, the
PostScript font uses cubic Bézier curve, so the
reducibility of PostScript font is better than the TrueType
font. A Bézier curve is controlled by three points, as is
shown in Fig. 1, when the middle control point changes
the whole curve changes too. If an outline word consists
of n+1 points (P0, P1 … Pn), it needs m Bézier curves to
form the font. Equation (1) shows the detail.
m =
n
∑
i=
0
⎛ i ⎞
⎜ ⎟t
⎝n
⎠
i
( 1−
t)
n−i
A PostScript outline is formed of a group of cubic Bézier
curves which can be described in (2), (3):
3
2
x = a * t + b * t + c * t + d
X
y
X
3
2
y = a * t + b * t + c * t + d
y
A TrueType outline is formed of a group of quadratic
Bézier curves and every curve is defined by three control
points. For a quadratic Bézier curve’s three control points
are (Ax, Ay),(Bx, By)and(Cx, Cy)described in (4)
and (5) :
P
X
=
© 2010 ACADEMY PUBLISHER
X
y
P
2
2
( 1 − t)
AX
+ 2t(
1 − t)
B X + t C X
i
X
y
(1)
(2)
(3)
(4)
P
y
=
Figure 1. Bézier Curve
2
2
( 1 − t)
Ay
+ 2t(
1 − t)
B y + t C y
By modifying the parameter t from 0 to 1, all values of
p defined by A, B and C can be generated, from which
the quadratic Bézier curve is obtained.
Font is defined as a combination of a set of outline
curves and every curve includes three or more points. The
illustration using Bézier outline curve of NaXi pictograph
“deer” is shown in Fig. 2. The outline is constituted of a
set of instructions which pictured the character’s exterior.
All the parameter information of outline fonts is stored
in a NaXi pictograph table “glyf”. Because of the high
complexity and similarity of Naxi pictograph, the
precision of outline curve delineation is greatly higher
than the dot matrix font delineation. Based on the two
outline font technique, we developed NaXi pictograph
TrueType font and PostScript font which basically met
the requirement of processing basic model of NaXi
pictograph with computer.
B. The Feature Extraction Method of NaXi Pictograph
Outline Font
The outline font is filled with a series of outline
curves and delineating the outline accurately is the key
point for the success of a font. It can be seen from
Figure 2. the Outline Curve delineation of NaXi Pictograph “deer”
(5)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 119
Figure3 that the NaXi pictograph is complicated, various
and also with many strokes which make it difficult for
truly reducing the outline of NaXi pictograph. According
to these characteristics, our project proposes a dual-mode
transformation algorithm for extracting the NaXi
pictograph outline points and uses some rules to check
whether a pixel point is in the outline.
Rule1: if the center of a pixel point is in the outline, the
point is lightened and becomes a part of the outline curve.
Rule2: if an outline line passes the center of a pixel,
the pixel is lightened.
Rule3: if a scanning line located in the center of two
adjacent pixels (horizontal or vertical) intersects with the
on-Transition and the off-Transition and the points on
this line haven’t lightened by rule1 and rule2, the left
endpoint is lightened when this line is horizontal, while
the right endpoint is lighted when it’s vertical.
Rule4: rule3 can be used only in the case that two
contour surfaces still intersect the scanning line in twoway,
but this doesn’t mean these pixels are “stubs”. By
checking the scanning line-segment formed a square with
crossing scan line-segment, whether they intersect each
other through two contour surfaces can be verified. There
is the possibility which is small but exists that more than
one contour intersect with discontinuity point, so it’s
necessary to control some characters outline using gridfitting.
The method discussed above is used in the secondary
development of Pgaedit based on Linux which effectively
prevents the generation of discontinuity point, and the
accuracy of outline point extraction of NaXi pictograph
reaches 99.99%.
IV NAXI PICTOGRAPH FONT DEVELOPMENT
A computer font is an electronic data file containing a
set of glyphs, characters, or symbols such as dingbats.
Although the term font first referred to a set of metal type
sorts in one style and size, since the 1990s most fonts are
digital, used on computers.There are three basic kinds of
computer font file data formats:
� Bitmap fonts consist of a series of dots or pixels
representing the image of each glyph in each face
and size.
� Outline fonts (also called vector fonts) use Bézier
curves, drawing instructions and mathematical
formulas to describe each glyph, which make the
character outlines scalable to any size.
� Stroke fonts use a series of specified lines and
additional information to define the profile, or size
and shape of the line in a specific face, which
together describe the appearance of the glyph.
Bitmap fonts are faster and easier to use in computer
code, but inflexible, requiring a separate font for each
size. Outline and stroke fonts can be resized using a
single font and substituting different measurements for
components of each glyph, but are somewhat more
complicated to use than bitmap fonts as they require
additional computer code to render the outline to a bitmap
for display on screen or in print.
© 2010 ACADEMY PUBLISHER
The difference between bitmap fonts and outline fonts
is similar to the difference between bitmap and vector
image file formats. Bitmap fonts are like image formats
such as Windows Bitmap (.bmp), Portable Network
Graphics (.png) and Tagged Image Format (.tif or .tiff),
which store the image data as a grid of pixels, in some
cases with compression. Outline or stroke image formats
such as Windows Metafile format (.wmf) and Scalable
Vector Graphics format (.svg), store instructions in the
form of lines and curves of how to draw the image rather
than storing the image itself.
A bitmap image can be displayed in a different size
only with some distortion, but renders quickly; outline or
stroke image formats are resizable but take more time to
render as pixels must be drawn from scratch each time
they are displayed.
Fonts are designed and created using font editors.
Fonts specifically designed for the computer screen and
not printing are known as screenfonts.
B. NaXi pictograph font library
Font library is the basis and also the core of the system,
accurate font and correct code set solid foundation for
developing the complete system. The NaXi pictograph
font library includes NaXi pictograph TrueType outline
font and NaXi pictograph PostScript font. TrueType
outline font is composed of quadratic Bézier curves and
PostScript is built with cubic Bézier curves. Since a
quadratic Bézier curve can be transformed to a cubic
Bézier curve, we develop TrueType outline font first and
then realize the transformation by expert tools.
The developement of NaXi Pictographs Outline Font is
mostly going by to design character draft, to scan the
font,to digitize the font、to modify and coordinate the
font、to examine the quality of character、to conformity
character storeroom. After calculation,draft is scanned as
lattice character storeroom in high precision by scanner,at
the same time the character storeroom code is given.To
simulate and conformity digitallization is according to the
method of double pattern switch ,which tries its best to
shift lattice graph to numeral infomation as real as
manuscript automatically (curve outline).The outline
point,line,angle and location can be rectified through
parameter controling,which present extraordinarily
important in the process of producing NaXi Pictographs
with complex font and large diverse manner.
The outline font development can be divided into four
steps: font and script designing, digital fitting, modifying
and font generating. Our project uses the original script of
A dictionary of NaXi pictograph sound-indication which
is transformed to standard bit map by scanning, and then
the describing points are extracted for digital fitting, after
that the NaXi pictograph outline font is integrated and
created. This kind of technique can guarantee the veracity
of original script and the generated outline font has the
features of excellent reducibility and vector property as
well..
V NAXI PICTOGRAPH INPUT METHOD BASED ON IMM-
IME

120 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
A. IMM-IME Introduction
As Windows is the most widely used operation system at
present, this paper focuses on the input system for NaXi
pictograph based on Windows. The input method based
on Windows transforms standard ASCII string into
NaXi word or string using some particular coding rule.
With different application program the user can not
design a transformation program himself, due to which
the task of inputting NaXi pictograph should be taken by
the Windows system administration.
As shown in Fig.4, at first, the keyboard event of NaXi
pictograph input system is received by the Windows file
use.exe, then use.exe transfers the event to the Input
Method Manager (IMM), after that the IMM conveys the
event to the input method editor which translates the
keyboard event to its corresponding NaXi character (or
string) with reference to user’s encoding dictionary, when
this is done the translated event is propagated back to
use.exe and then to the executing application, until now
the whole input process of NaXi pictograph is finished[4-
5].
B. The Basic Structure of NaXi Pictograph IMM-IME
The IMM-IME structure provides various input
methods for applications, each tread of an application can
keep an active input window. The processing order of
other messages won’t be disturbed by inserting the NaXi
pictograph message to message circle. The head file
TABLE II. NAXI PINYIN CODE
IPA Latin letter IPA Latin letter IPA Latin letter IPA Latin letter
p‘ p p
b b bb f(ɯ) f(w)
t d t’ t d dd n n
l l k g k‘ k g gg
Ŋ ng h h ʈɕ j ʈɕ‘ q
dʐ jj ȵ ni ɕ(ʑ) x(y) ʐ r
ʂ sh dʐ rh tʂ zh tʂ‘ ch
z ss s s dz zz ts‘ c
ts z i i u u y iu
a a o o ә e v v
ɯ ee әr er e ei ӕ ai
ie iei iӕ iai ia ia iә ie
uei ui uӕ uai ua ua uә ue
Figure 3. the Input Method Principle of NaXi Pictograph.
© 2010 ACADEMY PUBLISHER
immdev.h should be included for using these new
features. The detailed working principle of NaXi
pictograph Pinyin input method is shown in Fig. 3.
C. NaXi pictograph pinyin input method
Current input method includes two types: Pinyin and font.
The font input method can be divided by strokes which
needs the user has the ability to write. Unfortunately,
writing NaXi pictograph is so hard for the ordinary user
that makes the shape code unsuitable for NaXi pictograph
being inputted to the computer. The phonetic code input
method just requires the user know how to read the
character, therefore, Pinyin input method is more suitable
for NaXi pictograph.
Early dictionaries of NaXi pictograph sound-indication
employ International Phonetic Alphabet to mark NaXi
character, while the computer uses Latin letters as input
code, the conversion between NaXi Phonetic and Latin
Pinyin becomes necessary. Table II lists the mapping
between International Phonetic Alphabet and Latin
Pinyin in detail. When designing the input method of
NaXi pictograph one can encounter the problem of too
long encoding due to the characteristics of NaXi
pronunciation. For instance, the NaXi character �
can be coded as “ssoxiqssoddassa”, it needs fifteen
English characters to map this single one character which
will result in very low efficiency when input an article.
Research shows that the initials repetition phenomenon is
common in NaXi pictograph pronounced coding, through
the method of designing code with simplified initials the
Pinyin input method of NaXi pictograph can greatly
reduce the coding length and improve the coding
efficiency. Let’s take the character � as an example,
its pronounced coding is “ssoxiqssoddassa”, after the
simplification of pronunciation it becomes to
soxiqsodasa, the length reduces to eleven from fifteen.
NaXi pictograph consists of 2120 characters and the
average coding length is twelve bits which is reduced to
eight bits after the simplification of pronunciation, so the
input speed is also increased.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 121
D. NaXi pictograph graphic primitive input method
1) the Structure Coding Method of NaXi
Pictograph:The NaXi Pictograph has four common
structures, those are undivided whole structure codedby b,
up and lower structure coded by s, right and left structure
coded by z, surounding structure coded by b. The NaXi
Pictograph character of undivided whole structure are
such as�,�,�,�,�and so on, the upper and
lower structure,�,�,�,�,�,�; right and left
structure, �,�,�,�,�surounding structure,
�,�,�,�,�. After the coarse segmentation and
coding above, the fine granularity coding is introduced.
2) the Graphic Primitive Coding Method of NaXi
Pictograph:The NaXi pictograph character hasn’t radical
component such as Chinese characters, so the
representation method using Graphic Primitive is
proposed in this paper,which inspired by syntactic pattern
recognition. The basic elements of graph, graphic
TABLE III.
GRAPHIC PRIMITIVE CODE OF NAXI PICTOGRAPH
Graphic Primitive code
point a
line b
circle c
circular curve f
left oblique line g
right oblique line h
vertical line i
vertical curve g
oval curve k
rectangle l
right oblique line h
primitives, are point, line, circle, circular curve, left
oblique line. right oblique line, vertical line, vertical
© 2010 ACADEMY PUBLISHER
curve, oval curve and rectangle. As showed in Table III,
the basic graphic primitives are coded. Different with
Chinese and other characters, many NaXi pictograph
characters contains digit, for example � (dice), the
number of dots, and � (fire), the number of vertical
curve, and � (treasure),the number of circle. So the
quantity is coded, one is 'y', two is 'e', three is 's', four is 'f',
five is 'w', six is 'l', seven is 'q', eight is 'b', nine is 'j', the
number greater than nine is 'd'. The user interface of
NaXi primitive input method is shown in Fig. 4.
E. Evaluation of NaXi Pictograph Input Method
In order to evaluate the efficiency of NaXi proposed by
this paper, we develop an evaluating system for NaXi
pictograph input method. Its flow procedure is shown in
Fig. 6. The evaluating system employs the Windows
message mechanism to automatically converse the NaXi
Pinyin text to NaXi pictograph text under the conditions
of activating the NaXi input method with and without
optimization. After getting the evaluation result, we can
calculate the conversion accuracy.
In the experiment, we choose five NaXi texts to the
evaluating system, the results demonstrate that the NaXi
graphic primitive input method achieved higher average
conversion accuracy than the Pinyin input method. Table
IV shows the details.
VI WEB EMBEDDING APPLICATION
Aiming at the ordinary customers’ requirement, our
project also made some study on the application of NaXi
pictograph and developed Web embedding and PDF
embedding technology.
A. The Classification of Web Embedding
The lattice font has been gradually eliminated at the
Figure 4. NaXi Pictograph Primitive Input Method.

122 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Conversion
Accuracy
Pinyin
Method
graphic
primitive
Method
beginning of 90s in last century with the updating of
operating system, the new outlines font has replaced it.
The mainstream outlines font includes:
①PostScript Type 1 font, brought forward by Adobe
corporation years ago, applied in publication typesetting
system, belongs to the first generation of outlines font.
② TrueType font, brought forward by Apple and
Microsoft. Because it has so many merits, it has been
used in a variety of operating system on Mac and Pc,
belongs to the second generation of outlines font.
③ OpenType font, Put forward by Adobe and
Microsoft as a new generation of outlines font standard. It
fuses the merits of Type 1 and TrueType, belongs to the
third generation of outlines font.
The TrueType font embedding technology of NaXi
pictographs can be divided into embedded OpenType and
True Doc. Put forward by Microsoft, Embedded
OpenType technology can compresse TrueType font into
Eot file, and then embeds it into HTML web pages.
While TrueDoc brought forward by Netscape and
Bitstream can compress TrueType font into TrueDoc file
and then embed it into HTML web pages.
B. The Principal of Web Embedding Technology of NaXi
Pictographs
Web embedding technology of NaXi pictographs uses
Figure 5. NaXi Pictograph Input Method.
© 2010 ACADEMY PUBLISHER
TABLE IV.
COMPARISON OF THE NAXI PINYIN INPUT METHODS
text1 text2 text3 text4 text5 Average
Accuracy
99.1% 98.1% 98.4% 96.1% 97.8% 98.1%
99.6% 99.1% 98.9% 97.1% 98.1% 98.6%
CSS2 calling compressed OpenType font files to embed
NaXi pictographs into web page. Downloading
compressed fonts of NaXi pictographs downloaded to
clients’ temporary directories lets clients browse web
pages with NaXi pictographs
correctly without installing NaXi pictographs font..
The principle is that a client sends requirement for
browsing, and then HTTP server sends HTML files to
client browser. Client sends the information of web pages
contained NaXi pictographs to server through CSS2, the
server send this information to EOT database. EOT calls
the TrueType font files stored in HTTP server, then
HTTP server sends this inmessage back to the browser.
The NaXi pictographs will be deleted autonomously after
the client closes the browser. In this way fonts copyright
can be better protected from piracy. By calling CSS2
several times in a web page, not only NaXi pictographs
but Chinese, Mongolian, Tibetan and other minority
languages can be displayed in the same web page.
C. Realization of Web Embedding Fonts Technology of
NaXi Pictographs
1) The environment of developing web embedding fonts
of NaXi pictographs:The development environment is an
environment of making web pages, EOT files, and CSS
tables, it is mainly composed of DreamWeaver, WETF
(Web Embedding Fonts Tool), Font Creator and etc. Free
BSD installing Apache is adopted as the server. Firstly
this structure is totally compatible with Microsoft IIS
system. Secondly, it provides more useful functions,
faster operating speed and better stability than Microsoft
IIS system.
2) The application of CSS2 in web embedding fonts of
NaXi pictographs:pictographs TrueType font into web
pages possible. From the way that CSS is inserted, there
are three kinds of CSS: inline mode list, embedded mode
list and exterior mode list, and the inline mode list and
embedded mode list are broadly used in web embedding
fonts technology of NaXi pictographs. The key sentence
of CSS is @font-face, which has defined the name, type,
thickness and other information of a planted font.
3) The generation of planted font database of NaXi
pictographs: The substance in creating font database is to
compress NaXi pictographs TrueType fonts into
OpenType fonts. There are many ways in generating
NaXi pictographs planted databases, this paper adopts
Microsoft WETF ( Web Embedding Fonts Tool). Before
planting the validity of NaXi pictographs TrueType fonts
in the development environment are checked by WEFT.
WEFT displays the font validity in graph。.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 123
After the success of font-checking, add the needed
NaXi pictographs into EOT files, then calling NaXi
pictographs font while browsing will not be a problem.
4) The embedding of NaXi pictographs font:You can
input NaXi pictographs by using NaXi pictographs input
method developed by the information system lab, then
create a web page containing NaXi pictographs with
Dreamweaver. Then insert codes as follows between
and :
Embedding function @font-face provides with four
parameters: you can use font-family to name this font in
the current webpage, this project is defined as NaXi; fontstyle
can be anyone of normal, italic or oblique,
commonly defined as normal; font-weight can be normal,
bold, bolder, lighter or other legal thickness value;
FontURL is a URL pointing to OpenType files, normally
adopts absolute pathway. After saving the NaXi
pictographs webpage, you can test it by transfering it to
the HTTP server. The testing Figure can be seen in Fig. 6.
Up to now, the preliminary process of planting NaXi
pictographs is officially finished.
VII CONCLUSION
A complete NaXi pictograph information processing
platform is developed in this paper. It ends the history of
© 2010 ACADEMY PUBLISHER
Figure 6. NaXi Pictograph WEFT Web.
NaXi pictograph processing without computer, provides
valuable reference for creating other minority languages
information processing system and also plays a
significant role in promoting the computerization process
of minority characters in China.
ACKNOWLEDGMENT
This work is supported by National Natural Science
Foundation of China (No. 60803096).
REFERENCES
[1] G. Eason, B. Noble, and I. N. Sneddon, “On certain
integrals of Lipschitz-Hankel type involving products of
Bessel functions,” Phil. Trans. Roy. Soc. London, vol.
A247, pp. 529–551, April. 1955.
[2] Joseph Francis Charles Rock,A Nakhi-English
encyclopedic dictionary, I.M.E.O, Rome,1963.
[3] Liu Yongkui, Guo Hai, Lu Guiyan,Li Hongyan,"Input
technology and information processing of NaXi
pictograph",Journal of Computational Information
Systems,vol.3, no.1, p361-368, February, 2007.
[4] Guo Hai, Che Wengang,Nie Juan,Li Bin etc,"Web
embedding fonts technology of NaXi pictographs",Jisuanji
Gongcheng/Computer Engineering, vol.31, no.17, pp.203-
204+207, Sep. 2005.
[5] Guo Hai, Zhao Jing-ying,"Development of the NaXi
Pictographs Information Processing System",Control &
Automation,vol.22, no. 22, pp.122-124, 2006.
[6] Xie Qian, Jiang Li,Wu Jian,etc,"Research on Chinese
Linux input method engine standard",Jisuanji Yanjiu yu
Fazhan/Computer Research and Development, vol.43,
no.11, pp.1965-1971,November, 2006.
[7] Tseng Chun-Han, Chen Chia-Ping,"Chinese input method
based on reduced Mandarin phonetic alphabet", in Proc.
INTERSPEECH 2006 and 9th International Conference on
Spoken Language Processing, 2006, pp.733-736.

124 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
[8] Tanimoto Yoshio,Nanba Kuniharu,Rokumyo Yasuhiko,
etc,"Evaluation system of suitable computer input device
for patients", Proceedings of the Third Workshop - 2005
IEEE Intelligent Data Acquisition and Advanced
Computing Systems, 2007, pp. 369-373.
Hai Guo received the M.Sc. degree in pattern recognition
and intelligent system from the Kun Ming University of science
and technology, China, in 2004.
He has been teaching at Dalian Nationalities University
since graduation. For many years, he has done research in
© 2010 ACADEMY PUBLISHER
image processing, pattern recognition. He has published
more than 10 papers.
Jing-ying Zhao received the M.Sc. degree in computer
application from the Chang Chun University of science and
technology, China, in 2003.
She has been teaching at Dalian Nationalities University
since 2003. For many years, she has done research in image
processing, pattern recognition. He has published more than 10
papers

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 125
A Systematic Decision Criterion for Urban
Agglomeration: Methodology for Causality
Measurement at Provincial Scale
Yaobin Liu
Research Center of Central China Economic Development, Nanchang University, Nanchang, P.R.China
Email: liuyaobin2003@163.com
Li Wan
Research Center of Central China Economic Development, Nanchang University, Nanchang, P.R.China
Email: Sandy_gf@163.com
Abstract—The paper employs the synergism theory and grey
relative technique to analyze the inner nonlinear
relationships among factors that attribute to process of
urban agglomeration, and builds a synergic model to reveal
the causality running urban agglomeration system. The grey
relative analysis shows that the urban agglomeration system
is an open and dissipative system, and it is affected by many
factors, which are not just simple linear relative but
relevant, non-uniform and irreversible. Meanwhile, the
synergic model reveals that the key parameters that
dominated the process of urban agglomeration in Jiangxi
province are the rate of industrialization, percentage of
urban fixed asset investment in the total fixed asset
investment and the highway mileage on per unit area.
Obviously, the simulated results are available in practice, so
the nonlinear systematical methods can be applied to
analyze the causality mechanism in process of urban
agglomeration.
Index Terms—Urban agglomeration; causality
measurement; grey relative technique; synergic model
I. INTRODUCTION
Urban agglomeration is made up of cities with
different sizes to be linked by traffic network in a given
area, and it is an inevitable result when urbanization
reaches a certain level [1]. Today, while the global
economic development is integrated, the urban
agglomeration has become a carrier, by which a country
or region can participate in international competition. For
example, in the United States, three urban agglomerations
including the New York area, Los Angeles area and the
Great Lakes concentrated about 46% of the national
population and 67% of the GDP. So, it is vital for a
nation and area to nurture and build a stronger and
competitive urban agglomeration for participating in
international competition. How to nurture and build the
urban agglomerations, rapidly? This is not only an actual
Manuscript received January 1, 2009; revised June 1, 2009; accepted
July 1, 2009.
Corresponding author. Tel.: +86791 3163385; fax: + 867918304401
E-mail address: liuyaobin2003@163.com (Y.B.Liu).
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.125-130
problem for city planners and urban managers, but also
an important academic issue for the theoretical scholars.
Although many scholars pay more attention to the city
cluster in the developed areas of western countries, the
modern researches have been extended to the developing
areas including Africa and Asia countries. Here,
mechanism and pattern of urban agglomeration are still a
focus and topic in urban planning field. For example,
Naude and Krugell(2003)analyzed the spatial
development of urban agglomeration in South Africa, and
thought that the size of the primate city in Johannesmoreeast
Rand might be relatively too large[2]. Kanemoto et
al.(2005)took Tokyo metropolitan as an example, and
analyzed the rational size of Tokyo as the primate city[3].
Qin et al.(2006) analyzed formation mechanism and
spatial pattern of urban agglomeration in central Jilin of
China[4].
Most modern scholars of urban development
acknowledge that transnational processes are having an
increasingly important influence on the evolution of
urban agglomeration. An early observation was the
recognition of an emerging system of world cities[5], a
kind of urban elite which is shaped in part by the new
international division of labor. These urban
agglomerations are also thought to be controlling and
coordinating global finance, producer and business
services[6]. The view of world cities as the “key nodes”
of the international urban system is a widely held one,
underpinned in particular by rapid advances in the
development of information technology and
telecommunications. However, because the surrounding
metropolitan area has experienced profound changes of
spatial organization, with suburbanization bringing the
most radical reorganization of metropolitan
space[7][8][9], the growing role of suburbanization in
metropolitan development is not unique as other major
cities in post-communist countries follow a similar path
[10][11][12][13][14][15][16][17].Suburbanization should
thus be considered as one of the crucial topics in the
study of urban agglomeration in post communist cities.
Consequently, a variety of explanation attribute to urban
agglomerations at present.

126 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
When we carefully review the relevant operation
mechanism of urban system, we may find that urban
agglomeration being a complicatedly non-linear systems,
has a lot of “disturbance” and “fluctuation” in orderly
appearance. The “disturbance” and “fluctuation” can
make the self-organized process changed and promoted,
and then the system can change from disordered to
ordered (or from junior orderly to senior orderly). The
process of urban agglomeration also can appear a
“Lorenz Butterfly Effect” phenomenon when the system
inclines to recession or collapse[18]. Therefore, we have
to depend on non-linear mathematical methods to analyze
the interaction mechanism of formation and evolution of
urban agglomeration, by which we can find the
quantitative relationship among the key factors and reveal
the causality in the system. However, there is still a lack
of studies on the integrated systems from the holistic
region point of view, thus we take Jiangxi province in
China as a case, and use the grey relative technique to
analyze the main factors of urban agglomeration
evolution and develop a synergic method to reveal the
causality. The objective of this study is to investigate the
key factors of urban agglomeration in Jiangsu province.
Improved understanding of its interactive mechanism will
certainly help planning efforts for future urbanization
management.
II. METHODOLOGY AND DATE
A. Grey relative technique
Grey system theory thinks that the world is the
material world, and it is also the information world. In the
information world, the white system can be understood
and the black system can not be known relatively and
temporarily, only the grey system can partly be explained
eternally and absolutely. The grey system theory is just
an analytical method which generally is used to study
relationships with incomplete information. The main
characteristic of the grey system theory is to compare the
data arrays and to calculate the grey relational degrees
between each element of the system with incomplete
information. Hereinto, the grey relative technique is an
effective way to reveal the grey relation. So-called grey
relative technique is to find the links among the elements
through the quantitative analysis. Thus, we can reveal the
incomplete information world by the grey relative
technique. The basic idea of grey relative technique is to
judge similar of its close link by grey relational degrees
[19]. To calculate the grey relational degree rij, the first
step must calculate the grey relational coefficient between
main-array and sub-array. The grey relational coefficient
is defined as:
minmin xi( k) − xj( k) + ξ max max xi( k) −xj(
k)
j k j k
rij() k =
xk () − xk () + ξ maxmax xk () −xk
()
i j i j
j k
Where rij(k) is the grey relational coefficient of the
main-array and sub-array at time k; xi(k) is the mainarray;
xj(k) is the sub-array and ξ the differentiation
© 2010 ACADEMY PUBLISHER
(1)
coefficient, which is used to increase the most prominent
differences between two arrays. Generally, the value is
assumed to be 0.5. The second step is to calculate the
grey relational degree between the main-array and subarray.
It is defined as:
n 1
r r( x ( k), x ( k))
= ∑ (2)
ij i j
n k = 1
Where rij is the grey relational degree of sub-array j
and main-array i; n is the length of the arrays.
B. Synergic model
The formation of urban agglomeration is the result of
nonlinear links among the elements of system, and its
evolution results from their synergy, cooperation and
competition. Therefore, there exists a relatively stable
form in process of urban agglomeration, in which the
form changes from the disorder state to orderly another.
Besides, the process of urban agglomeration is also an
interminable developing period with temporal-spatial
structure on the whole. In order to reveal the inner
mechanism and evolution causality, we develop a
synergic model to analyze the nonlinear function and find
out theirs quantitative relationship[20].
If the dynamic system of urban agglomeration have n
state variable, the x 1 , 2 x ,…, x m are the date sequences
during 1-m. Because the variables have different units,
they can not directly be applied for the synergic model
formerly, and the variables have to be normalized.
Supposed normalized variables
(0)
be x ()( t i = 1,2, L m)
, and the corresponding
i
sequence after first accumulation is
variation rate is
(1)
dxi () t
dt
(1)
x i , thus the
. The variation rate is
determined by two aspects. The first aspect results from
the inner synergic effects among the elements of system,
and it is self-development result. If the self-development
ax t , the restrain items is just
(1) ()
ii i
items is
(1) 2
− bii ( xi ( t))
. The second one lies in the exterior
synergic effects among the subsystems, and it is so-called
(1)
the synergic function. If the synergic items is ax ij i () t ,
(1) 2
the restrain items is just − b ( x ( t))
. Thus, the total
variation rate
dx
(1)
i
ij i
(1)
xi () t can be expressed as follows:
m
(1) (1) 2 (1)
ii i ( ii ij)( i ) ij j
j
ax b b x ax
dt + =− + +∑ (3)
If bii + bij = bi,
we can get the following nonlinear
equation:

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 127
dx
dt
(1)
i
m
(1) (1) 2 (1)
ii i i i ij j
j
∑ (4)
+ ax =− b( x ) + ax ( i≠ j)
Because the evolution of system is from one critical
point to another, the fluctuation is only a disturbance.
Correspondingly, the stability mechanism of system
always makes the fluctuation to decay or even disappear.
Considering the synergic model of urban agglomeration
that we build is to find out the dominant driving
parameters, they can lead to formation and evolution of
urban agglomeration. In practical, we will give the
parameters an important action to strengthen the synergic
function between systems. Beside, the purpose of our
work help the systems promote a higher orderly
development and not forecast its trend when the system
happens to change or transition. Therefore, we neglect the
role of the fluctuation when we build model.
The parameters aii, aij and bi are indicators, by which
we measure the intensity of synergy and competitive
among the elements. These parameters can be solved
through using the nonlinear differential equations. To
analyze the status equations of the model, we can found
which is the positive relaxation coefficient among the
state variables? And which is the negative relaxation
coefficient in the status equation? When the state
variables with the negative relaxation coefficient are
eliminate with adiabatic approximation approach, the
remained state variables are all the slow ones of this
system, and they is so-called order parameter. On the
other hand, we can also use numerical method, i.e.
Runge-Kutta approach with four times differences, to
gain the numerical solution of the equation, and then
eliminate the rapid state variables of system by adiabatic
approximation approach. At last, the remaining slow state
variables are just the order parameters of system. Because
the
x t is obtained by first accumulation and
(1) ()
i
normalization in the equation. After the numerical
solution of the equation is reduced, the simulated values
of state variables in line with synergic development of
urban agglomeration system can be obtained by Runge-
Kutta approach, subsequently.
C. Date collection
Jiangxi province is located in Southeast China,
bordering Fujian, Guangdong, Zhejiang, Hunan, Hubei
and Anhui provinces. It represents about 1.8% of the
surface of the P.R. of China and 3.5% of its population.
In Jiangxi province, agricultural production dominates its
economy. The share of agriculture in GDP was 17.93% in
2005, i.e. 5.32% higher than the average level of the
whole country. Its GDP per capita in 2006 was equal to
9440.62 Yuan (US$8.19), i.e. 67.24% of the national
average[21].
The data of non-agricultural population and GDP of
provincial cities and the other major cities of central
region origin from “China city statistical yearbook
(2007)”which is drawn up by urban social-economic
survey department of National statistics bureau[22]. The
social-economic data of major cities of Jiangxi province
© 2010 ACADEMY PUBLISHER
origin from “Statistical yearbook of Jiangxi province
(2007)”, which is edited by statistics bureau of Jiangxi
province[23]. The distance between every city of Jiangxi
province and the other cities is measured by mileage of
railway of National railway ministry, these data origin
from http://qq.ip138.com/train/ , specially, when there is
direct train between the two cities, we adopted their
nearest distance. Taking into account that the
considerable number of cities doesn’t have any direct
train, we use the shortest distance of transit based on the
accessibility.
III. RESULT AND ANANLYSIS
A. Factors analysis
For the evolution of urban agglomeration is interaction
process among variables, in order to find out their
behavior characterize, we have to inspect the relationship
between elements of urban agglomeration system.
According to qualitative analysis, it can be found that the
dominant driving force of formation and evolution of
urban agglomeration in Jiangxi province comes from six
aspects, i.e. industrialization, migration and transferring
of industries, foreign investment, pressure of competition
in the central area, construction of traffic network and
macro-control of government[24]. Because the
relationship among the variable in the internal dynamic
systems is relatively complex, and they are affected by
some man-made policy, therefore, it is very difficult to
directly quantify the action characteristics of urban
agglomeration system. Thus, we can use the indirect
variables to character the dynamic systems. First of all,
we selected six variables as the original series during
2000-2006. The six variables are respectively the rate of
industrialization, percentage of the tertiary industry in
GDP, percentage of the actual utilized foreign investment
in GDP, percentage of urban fixed asset investment in
total fixed assets investment, the highway mileage on per
unit area, and percentage of financial income in GDP
(Table 1). We can use grey relative technique to calculate
the grey relational degrees of the six variables, and
then obtain cross table of the dynamic elements during
2000-2006 in Jiangxi province (Table 2). Through grey
relative analysis, we can have four results from table 2.
The first result shows that the grey relative degrees are all
above 0.45, which indicates the interactive intensity and
the mutual constrains of the various elements in urban
agglomeration are significant. The second result shows
that the grey relative degrees of the variables are
different, and they are not simple linear relative but
relevant, non-uniform and irreversible, which indicates
the system is complex. The third result shows that there is
a higher grey relative degree between percentage of the
actual utilized foreign investment in GDP and the other
variables, which indicates that the ability of attract
foreign investment is more important than that of other
elements in process of urban agglomeration in Jiangxi
province. At the same time, the rate of industrialization,
percentage of the tertiary industry in GDP, percentage of
urban fixed asset investment in total fixed assets

128 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
investment, the highway mileage on per unit area and
percentage of financial income in GDP pay an important
role in the ability of attracting foreign investment. The
last aspect shows that there exists a higher relationship
between percentage of urban fixed asset investment in
total fixed assets investment and the other variables,
which indicates that the city construction investments
have an important impact on the evolution of urban
agglomeration. However, the rate of industrialization,
percentage of the tertiary industry in GDP, financial
income and the highway mileage on per unit area are also
play a vital role in urban fixed asset investment.
Table 1. The data of action characteristics of urban agglomeration during 2000-2006 in Jiangxi province,China
items variables times
2000 2001 2002 2003 2004 2005 2006
The rate of industrialization (%)
Percentage of the tertiary industry in GDP (%)
Percentage of the actual utilized foreign investment in
GDP (%)
x1 x2
x3
27.2
40.8
1.021
27.7
40.6
1.637
28.7
39.6
3.993
30.8
37.2
5.169
33.0
35.5
5.344
35.9
34.8
5.375
38.7
33.5
5.408
Percentage of urban fixed asset investment in total x4 84.4 84.6 85.1 85.8 86.4 87.7 88.5
fixed assets investment (%)
The highway mileage on per unit area(km/km 2 ) x 5 0.361 0.364 0.368 0.369 0.371 0.373 0.768
Percentage of financial income in GDP (%) x6 0.01 0.01 0.01 0.01 0.01 0.01 0.01
Table 2. The cross table of factors of urban agglomeration in Jiangxi province,China
x 1 x 2 x 3 x 4 x 5 x 6
x 1 1 0.66 0.65 0.76 0.59 0.62
x 2 0.62 1 0.74 0.72 0.55 0.67
x 3 0.65 0.77 1 0.64 0.67 0.90
x4 0.71 0.75 0.60 1 0.58 0.65
x5 0.56 0.57 0.64 0.57 1 0.72
x 6 0.46 0.45 0.47 0.46 0.43 1
B. Causality analysis
Supposed that the rate of industrialization, percentage
of the tertiary industry in GDP, percentage of the actual
utilized foreign investment in GDP, percentage of urban
fixed asset investment in total fixed assets investment, the
highway mileage on per unit area and percentage of
financial income in GDP are the state variables in urban
⎧dx
⎪
⎪
dt
⎪dx
⎪
dt
⎪
⎪dx
⎪ dt
⎨
⎪dx4
⎪ dt
⎪
⎪dx5
⎪ dt
⎪
⎪dx6
⎩⎪
dt
(1)
1
(1)
2
(1)
3
(1)
(1)
(1)
agglomeration, respectively, and they can be denoted as
x t ,
(0)
x2 () t ,,
(0)
x3 () t ,
(0)
x4 () t ,
(0)
x5 () t and
x t . With the software of MATLAB applied, we
(0)
1 ()
(0)
6 ()
can get the parameters of system by using the synergic
model. And then the parameters are put into Equa.(4), we
can get the state equations of system:
=− 0.5870x + 0.2503( x ( t)) + 0.3474x + 0.5219x + 0.4146x + 0.3292x + 0.4158x
(1) (1) 2 (1) (1) (1) (1) (1)
1 1 2 3 4 5 6
= 0.3037x −0.7274x − 0.0529( x ) + 0.4679x + 0.5047x + 0.4053x + 0.3922x
(1) (1) (1) 2 (1) (1) (1) (1)
1 2 2 3 4 5 6
=− 0.7287x + 0.3509x + 0.3726x + 0.0724( x ) + 0.4635x + 0.5064x + 0.4826x
(1) (1) (1) (1) 2 (1) (1) (1)
1 2 3 3 4 5 6
= 0.3275x + 0.3062x + 0.2272x − 0.7271x + 0.0807( x ( t)) + 0.4627x + 0.5243x
(1) (1) (1) (1) (1) 2 (1) (1)
1 2 3 4 4 5 6
= 0.4276x
+ 0.3907x + 0.1278x + 0.4347x −0.5264x − 0.0902( x ( t)) +−0.3526x
(1) (1) (1) (1) (1) (1) 2 (1)
1 2 3 4 5 5 6
=− 0.5264x + 0.3281x + 0.5071x + 0.2229x + 0.3281x + 0.3209x −0.4563(
x ( t))
(1) (1) (1) (1) (1) (1) (1) 2
1 2 3 4 5 6 6
Using Runge-Kutta approach with four times
differences, we can get the numerical solution as the
indicate
(1)
x% i ()( t i = 2, L ,6) . After the numerical
solution of the equation is reduced, the simulated values
% . Through
of state variables can be got, i.e. (0) xi () t
© 2010 ACADEMY PUBLISHER
analyzing the state equations of this model, we can get
some results. Firstly, except that the coefficients of the
first state variable (the rate of industrialization), the third
state variable (percentage of urban fixed asset investment
in total fixed assets investment) and the fourth state
variable (the highway mileage on per unit area) are
(5)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 129
positive, the coefficients of other three variables
including percentage of the tertiary industry in GDP,
percentage of the actual utilized foreign investment in
GDP and percentage of financial income in GDP are
negative. Secondly, with the adiabatic approximation
approach employed, we can eliminate the three state
variables, and the remaining state variables which are the
rate of industrialization, percentage of urban fixed asset
investment in total fixed assets investment and the
highway mileage on per unit area are just the slow state
variables, which can be called order parameter. By
contrasting the cumulative curve of initial values and that
of forecasted values (Fig.1 and Fig.2), it can be seem that
the simulated rule of dynamic evolution of urban
agglomeration is consistent with the actual situation, so
the model is success.
The synergic model reveals that the dominant
parameters of urban agglomeration in Jiangxi province
are the rate of industrialization, percentage of urban fixed
asset investment in total fixed assets investment and the
highway mileage on per unit areas, which are key driving
factor in formation and evolution of urban agglomeration.
Therefore, the degree of industrialization, the competition
ability of the central urban agglomeration and the
construction of transport network play an important role
in all driving factor. Therefore, we should pay more
attention to the new industrialization for developing
urban economics in the process of urban agglomeration.
At the same time, the expenditure and maintenance funds
in urban construction and the construction of transport
network take a vital impact on developing of urban
agglomeration in Jiangxi province. It is showed that
urban construction and transport links are an important
foundation for economic development. If there is no a
guarantee for good urban environment in economic
growth, the healthy economic development can not be
achieved. Therefore, we should pay more attention to
planning on the whole urban construction in practice, and
raise funds for construction and accelerate the
infrastructure of transports in Jiangxi province.
x(0)
3
2. 5
2
1. 5
1
0. 5
0
x1 x2 x3 x4
x5 x6
1 2 3 4 5 6 7 t
Figure 1. Cumulative curve of initial values
© 2010 ACADEMY PUBLISHER
x(0)/
3
2. 5
2
1. 5
1
0. 5
0
x1 x2 x3 x4
x5 x6
1 2 3 4 5 6 7 t
Figure 2. Simulation curve of forecasted values
IV. CONCLUSIONS
There is still a lack of studies on the integrated urban
agglomeration systems from the holistic region point of
view, thus we take Jiangxi province in China as a case,
and use the grey relative technique to analyze the main
factors of urban agglomeration evolution and use
synergic method to build a nonlinear differential model
that can reveal the causality in the process of urban
agglomeration and get the parameters of system by
simulation. The results by grey relative analysis shows
that urban agglomeration system is an open and
dissipative system, and it affected by many factors, which
are not simple linear relative but relevant, non-uniform
and irreversible. And the synergic model further reveals
that the dominant parameter of urban agglomeration in
Jiangxi province are the rate of industrialization,
percentage of urban fixed asset investment in total fixed
assets investment and the highway mileage on per unit
area, which are key driving factor in formation and
evolution of urban agglomeration. Therefore, the
promoting of industrialization, the enhancing of
competitive ability of the central urban agglomeration
and the boosting construction of transport network are
vital in process of urban agglomeration.
The studied result shows that the integrated model
constituted of the grey relative technique, the synergic
model and numerical analysis can to the degree reveal the
evolution direction of urban agglomeration, and hold the
internal mechanism of its formation and evolution.
Therefore, the nonlinear systematical methods can be
applied to analyze the mechanism of urban agglomeration
development. As an applied research, when we use the
other variables to describe the action characteristics in the
dynamic systems, the method is simple but it can reflect
the uncertainty and subjectivity. At the same time, only a
series of assumptions is considered do we build the
synergic model by nonlinear differential grey system
when. Because the fluctuation is not considered, the
universal practicability can not be ensured when it is
applied. Therefore, we should amend the model in
accordance with the specific issues in practice.
ACKNOWLEDGEMENT
This research was supported by funding from

130 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Provincial Education Department of Jiangxi (No.
GJJ08014) and National Social Science Foundation of
China (NSSC) (No.07CJL031).
REFERENCES
[1] S.M. Yao, Y.M. Zhu, Z.G.Chen, The Urban
Agglomerations of China, 3rd ed. Hefei: Science and
Technology University of China Publishing House,2006,
pp.23-28. (In Chinese)
[2] W.A.Naude, W. F.Krugell, “Are South Africa's cities too
small?” Cities, 2003,20(3),pp.175-180.
[3] Y,Kanemoto, T.Kitagawa, H.Saito, et al., Estimating
urban agglomeration economies for Japanese
metropolitan areas: is Tokyo too large? Tokyo: Graduate
School of Economics, University of Tokyo,2005,pp.45-
48.
[4] G. Qin, P.Y.Zhang, B.Jiao, “Formation mechanism and
spatial pattern of urban agglomeration in central Jilin of
China”, Chinese Geographical Science,2006,
16(2),pp.154-159.
[5] P.Hall, The World Cities, 3rd eds. London: Weidenfeld
and Nicolson, 1966, pp.124-128.
[6] N.Thrift, “Globalization, regulation, urbanisation: the case
of the Netherlands”, Urban Studies,1986,31, pp.365-380.
[7] L.Sykora, “Changes in the internal spatial structure of
post-communist Prague”, Geographical Journal,1999, 49
(1),pp. 79-89.
[8] M.Ourdnicek, “Differential suburban development in
Prague urban region”, Geografiska Annaler: Series B,
Human Geography, 2007,89 (2),pp.111-126.
[9] L.Sykora, and M.Ourednicek, “Sprawling postcommunist
metropolis: commercial and residential suburbanization in
Prague and Brno, the Czech Republic”, in Dust, M.,
Razin, E. and Vozquez, C. (eds): Employment
Deconcentration in European Metropolitan Areas:
Market Forces versus Planning Regulations,
Geographical Journal Library 91, New York: Springer,
2007,pp. 209-233.
[10] K.Leetmaa, and T.Tammaru,“Suburbanization in
countries in transition: destinations of suburbanizer , s in
the Tallinn metropolitan area”, Geografiska Annaler:
Series B, Human Geography ,2007,89 (2), pp.127-146.
[11] H.Kok, and Z.Kovacs, “The process of suburbanization in
the agglomeration of Budapest”, Netherlands Journal of
Housing and Built Environment,1999, 14 (2), pp.119-141.
[12] H.Nuissl, and D.Rink,”The production of urban sprawl in
eastern Germany as a phenomenon of post-socialist
transformation”, Cities,2005,22 (2),pp.123-134.
[13] N.Pichler-Milanovi, “Ljubljana: from beloved city of the
nation to Cenral European capital”, in Hamilton, F.E.I.,
Dimitrowska, K. and Pichler-Milanovi, N. (eds):
Transformation of Cities in Central and Eastern Europe:
Towards Globalization, Tokyo: United Nations University
Press, 2005, pp. 318–363.
[14] U.Sailer-Fliege, “Characteristics of post-socialist urban
transformation in East Central Europe”, GeoJournal, 1999,
49 (1),pp.7-16.
[15] T.Tammaru, “Suburbanization, employment change, and
commuting in the Tallinn metropolitan area”,
Environment and Planning A, 2005,37 (9), pp.1669-1687.
[16] J.Timar, and M.Varadi, “The uneven development of suburbanization
during transition in Hungary”, European
Urban and Regional Studies,2001, 8 (4), pp.349-360.
© 2010 ACADEMY PUBLISHER
[17] I.Tosics, “Post-socialist Budapest: the invasion of market
forces and the response of public leadership”, in
Hamilton, F.E.I., Dimitrowska, K. and Pichler-Milanovi,
N. (eds): Transformation of Cities in Central and Eastern
Europe: Towards Globalization, Tokyo: United Nations
University Press,2005, pp. 248–280.
[18] X.F.Pan, “Urban construction and economic development
of collaborative system through using the model of grey
correlative analysis”, Science and Technology
Management Research, 2005, 10, pp.43-45. (In Chinese)
[19] L.Fu, Grey system theory and application, Beijing:
Science and Technology Literature Publishing House ,
1992,pp.212-216. (In Chinese)
[20] K.X.Bi, H,Y.Wang “The Gray Cognate Analyzing and
Model Structuring of Synergy Development System of
Universities Science Research and Subject Construction”,
Systems Engineering-theory & Practice, ,2003,4, pp.140-
143. (In Chinese)
[21] State Statistical Bureau of China, China Statistical
yearbook(2007), Beijing: China Statistical Press,
2007,pp.243-246.
[22] Urban social-economic survey department of National
statistics bureau. China City Statistical yearbook(2007),
Beijing: China Statistical Press, 2007,pp.142-126.
[23] statistics bureau of Jiangxi province, Statistical yearbook
of Jiangxi province (2007), Beijing: China Statistical
Press, 2007,pp.67-69.
[24] Y.B. Liu, W.J.Song, , L.Wan, “Formation of urban
agglomeration and dynamic mechanism analysis in
Jiangxi province”, Group of Economic Research, 2007,30,
pp.28-29. (In Chinese)
Yaobin Liu received an academic degree in Management
Science and Engineer. His PhD was on urban management
relation of function and development. He received his PhD
(2005) from China University of Mining and Technology.
Following two post-doctoral years at Huazhong University of
Science and Technology, he served on Research Center of
Central China Economic Development of Nanchang
University until 2005. He initiated and led a number of projects
with the urban agglomerations in the Central China since
1999. It started with geography mapping, through several
conservation development plans to a conservation master plan
for the whole urban agglomeration. He developed a main
interest in urban management and planning.
Li Wan is currently a PhD candidate in the Research
Center of Central China Economic Development of
Nanchang University. He received a BS in Management
Science from Nanchang University and MS in Human
Geography from Nanchang University. His research focuses on
urban agglomeration assessment and strategic planning
assessment (SEA), with special emphasis on applying
management science principles into SEA on spatial plan. He
had also been involved in several programs in urban functional
planning and integrated urban agglomeration management
when he worked as an assistant researcher in Urban Planning
Center of Jiangxi university of Science and Technology in 2002
and 2003.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 131
Application of Improved Fuzzy Controller for
Smart Structure
Jingjun Zhang
Department of Science Research, Hebei University of Engineering, Handan, China
Email: santt88@163.com
Liya Cao, Weize Yuan, 2 Ruizhen Gao, and Jingtao Li
College of Civil Engineering, Hebei University of Engineering, Handan, China
2 College of Mechanical and Electrical Engineering, Hebei University of Engineering, Handan, China
Email: {yaer305@163.com, yuanweize520@126.com, 217taozi@163.com}
Abstract—In order to reduce the vibration of smart
structures, this paper gets the optimal location of
piezoelectric patch by the D-optimal design principle, and
then uses the fuzzy logic to control the smart structures
vibration. The fuzzy IF-THEN rules are established on
analysis of the motion traits of cantilever beam. The fuzzy
logic controller (FLC) designs base on using the
displacement and velocity of the cantilever beams tips as
the inputs, the control force on cantilever beams as the
output. This new method improves calculation efficiency
and reduces calculation complexity. Besides that, the paper
establishes parameter self-adjustment factor in fuzzy
controller by s-function to make the fuzzy logic control
more effective. The simulation results with Matlab
illustrate that the proposed method has a better control
performance than existing methods.
Index Terms—smart structures, optimal location, fuzzy IF-
THEN rules, fuzzy logic controller, parameter selfadjustment,
s-function
I. INTRODUCTION
In 1985 Bailey and et al. [1] performed experimental
research on active vibration control using surface-bonded
piezoelectric polymer actuators on a cantilevered beam.
Their experiment has greatly inspired the active vibration
control related field. Recently, much research has been
developed in the field of smart materials and structures.
Piezoelectric is a kind of smart material, due to the
following two characteristics: the first is direct and
inverse piezoelectric effects and the second is the ability
to be used as the sensor or the actuator in active vibration
control systems. Vibration control of smart structures is
very important because of the lightly damped of the
materials which were used. The placement of
piezoelectric patches plays an important role in the
design procedure of the active structures. The researchers
have focused on development of the optimal
piezoelectric patch location. Guo and et al [2] presented
a global optimization of sensor locations based on the
damage detection method for structural health
monitoring systems. Martin Kögl and et al [3] presented
a novel approach to the design of piezoelectric plates and
shell actuators using topology optimization. In this
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.131-138
approach, the optimization problem consists of
distributing the piezoelectric actuators in such a way as
to achieve a maximum output displacement in a given
direction at a given point of the structure. Cao and et al
[4] use the element sensitivities of singular values to
identify optimal locations for actuators.
There have been many performance criterion
presented, such as controllability and observability of
control system measures, dissipation energy measures
and system stability measures. However, in order to
make use of the above-mentioned measures, a flexible
structure state space equation should be modeled by the
given location of piezoelectric patches. The D-optimal
design principle is an optimization method presented by
Bayard and et al [5] which suggests that the maximum
determinant of Fisher Information Matrix Criteria is
chosen as the optimization function and then simplified
to determine an optimal principle for the best location for
piezoelectric elements.
In the regions of civil engineering and the spaceflight
engineering, the structures always accompany with
complicated kinetic characteristics and uncertain factors.
On the other hand, a robot system is a highly nonlinear
and heavily coupled mechanical system. The
mathematical model of such system usually consists of a
set of linear or nonlinear difference equations derived by
using some form of approximation and simulation. In
1983, Brown and Yao [6] used the fuzzy theory to the
engineering structures at first time. In 1986, Juang and
Elton [7] adopted fuzzy logic to estimate the density of
earthquake on the extent of damage for the constructions.
Battaini [8] designed the fuzzy logic control about mass
systems and experimented. Symans and Kelly [9] applied
fuzzy logic control strategy to control the system of
control. Based on the virtues of fuzzy logic control,
H.Park [10] established the approximate model of the
driver, the sensor and the fuzzy logic controller to solve
the problems of vibrations for flexible structure. The
result indicated that the fuzzy logic control had the
stronger robust and self-adaptive for the linear and nonlinear
system. R.Y.Y.Lee [11] carried on the similar
experiments of designing the ordinary fuzzy logic
controller and self-buildup controller for the non-linear

132 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
piezoelectric driver. The simulated result showed that the
fuzzy logic control has excellent suppression
effectiveness on the vibration of non-linear flexible. This
paper presents an optimization method using the Doptimal
design principle simplified to determine an
optimal principle for the best location of piezoelectric
elements, and then designs a fuzzy logic controller to
control the beam’s vibration. The simulation results
show that this proposed method has more superiority
than the others.
II. MODELING AND THEORY
In this section, we consider beam or plate type
structures bonded with rectangular shaped piezoelectric
sensors and actuators. The sensors and actuators are
symmetrically collocated on both sides of the same
position of the structure. The research of reference [12]
is verified that the symmetrical collocation can avoid
observation spillover and control spillover induced by
modal truncation, and ensures the controlled system is
minimum phase system. The finite element method can
analyze the arbitrary geometry models and the
anisotropic properties of the piezoelectric materials.
Considering the piezoelectric effect, special finite
elements with a degree of volts have been developed.
These elements have become available in some
commercial finite element software such as ANSYS. In
this paper, the solid45 3-d solid elements are used to
model the host structures and the solid5 3-d solid
elements are used to model the piezoelectric elements for
analysis the low modals of flexible structure are
extracted using ANSYS software.
III. D-OPTIMAL DESIGN PRINCIPLE
2dy
Actually, sensors are used to estimate state parameter.
Based on the theory of mathematical statistics, the
determinant of Fisher Information Matrix det(F) has
inverse ratio to low bound of variance of parameter
unbiased estimation. Due to the symmetrical collocation
of the piezoelectric patches, if the locations of sensors
have been confirmed, the actuators will be at the same
locations as the sensors. For a lightly damped structure,
the D-optimal design principle can be simplified to
decouple the problems of placement of actuators/sensors
and input control. The principle can be written as:
max(det( F ))
(1)
m
i k j
n
2dx
Figure 1. Piezoelectric patch finite element
© 2010 ACADEMY PUBLISHER
M
∑
k=
1
Subject to m B ∈ β , } : { Bm = β β k = m
where β is the location selection matrix; Bm is the set of
possible locations. The objective function can also be
written as:
N M
T 2
( γ k φi
)
S ( m)
= max ∑ log( ∑ β ) (2)
k
β
2
i= 1 k = 1 ϕ k
where N and M are the number of modals and the
possible locations of sensors, m is the number of the
sensors; β is composed of 0 and 1, if k
β k = 1 , this
denotes that a sensor is located on the location, in
contrast, if β k = 0 , this denotes that the location has no
a sensor; γ k is a vector coefficient related to kth
location of sensor; φ i is a ith normalized mode shape
vector; ϕ k is covariance of sensor signal noise, it can be
defined as 1.
The physical sense of (2) can be regard as finding the
location of maximum charges or volt output. Therefore,
T
γ k φi
is equivalent to the output charges of sensors. If
we suppose that the sensor area is enough small relative
compared to the beam (plate), the sensor charge can be
written as
q = D × A = A × ( Dx
+ D y )
(3)
2
∂ ω t p
Dx
= d31E
peε
x = d31E
pe 2
∂x
2
(4)
2
∂ ω t p
D y = d 31E
peε
y = d 31E
pe 2
∂y
2
(5)
Where Dx is the electric displacement generated by the x
axis strain, Dy is the electric displacement generated by
the y axis strain, and A is the area of sensor. For the
beam structure, electric displacement is generated by
unidirectional strain. d31 is the piezoelectric strain
constant. Ep and t p are the Young modulus and the
thickness of the piezoelectric patch. ω is the deflection
of structure. Dimensions of the piezoelectric patch and
the finite element are shown in Fig. 1.
Based on second-order difference scheme, k
x 2
2
∂ ω
∂
2
∂ ω
and k can be expressed as:
(6)
y 2
∂
2
∂ ω 2 ω(
i)
+ ω(
j)
− 2ω(
k)
d
2 k ≈ xω
=
2
∂x
( dx)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 133
2
∂ ω 2 ω(
m)
+ ω(
n)
− 2ω(
k)
d 2 k ≈ yω
=
2
∂y
( dy)
(7)
Substitute (4) (5) (6) (7) into (3) to yield
q = d31E
pe Aλk
(8)
where
t p ω(
i)
+ ω(
k)
− 2ω(
j)
ω(
m)
+ ω(
n)
− 2ω(
j)
λk
= (
+
)
2
2
2 ( dx)
( dy)
;
The deflection of structureω can be expressed by mode
shapeφ i
N
∑
i=
1
ω = η φ
(9)
where η is ith modal generalized coordinate, substitute
i
N
∑
i=1
T
T
(9) into (8), assuming γ φ = q , γ k can be written
k
as
T
γ k =
1
i
η i Ad 31 E pe λ (10)
k
2
where
t i p 1
λk
= φi
( [ 0K1
2 i
2 ( dx)
− 2 k 1 j K0]
+
;
1
[ 0K1
2 m
( dy)
− 2 k 1n
K0])
φ i is a ith normalized mode shape vector. Substitute (10)
into (2), objective function can be simplified as:
N
i
S ( m)
= max ( η β λ )
β
i
M
∑∑
i
i=
1 k = 1
i
i
k
k
(11)
Note (11), we conclude that the objective function is
composed of all step of mode shapes with the
coefficientη i .
With the vibration generating force, the equation of
motion in modal coordinates can be written as:
2 1
& η&i ( t)
+ 2ξ
iω
& iη
i ( t)
+ ω i η i ( t)
= f ( t)
M i
(12)
where η& & i , η& i andηi represent modal acceleration, velocity
and displacement, respectively, ω i and ξ i are the
natural frequency and damping ratio of the ith mode, due
to flexible structure and absence of interior damping,
ξ i = 0 .
For a different force, every mode shape has a different
proportion in the vibration of a structure. Assuming that
the vibration generating force is taken as a unit impulse,
structure vibrates according to the lower mode shape and
the other mode shapes can be ignored. Assuming that the
vibration generating force is taken as sine force with a
frequency of θ, the vibration of structure has relation to
© 2010 ACADEMY PUBLISHER
the frequency of the force rather than uniquely according
to one mode shape. For the above-mentioned, in order to
develop the performance of the vibration, the adhered
patches must control every mode shapes. So, the
piezoelectric patches should be placed on the maximum
of all the mode stains. Unitary mode stain can be written
as:
i i
λ k = λk
/ max( φi
)
(13)
Integrating equation (13) and (11), a new objective
function can be written as
∑∑
*
i
S ( m)
= max ( β λ )
β
N
M
i=
1 k=
1
(14)
Accordingly, the performance criterion of the sensor
locations can be obtained using mode shapes of
structures.
IV. FUZZY LOGIC CONTROLLER
A. Modeling the piezoelectric structure with the finite
element
The piezoelectric material is PVDF of β . It lays on
the above and below surfaces that acting as the
piezoelectric sensors and actuators respectively. The
sensor and actuator are symmetrically collocated on the
structure in the same position. Considering the coupling
effect, the equations which use the limit element could
be written as:
M u&
& ( t)
+ Cu&
( t)
+ Ku(
t)
= F + U (15)
Where M, K and C is the whole mass, the whole
stiffness, the whole damping respective, while u (t)
、
u& (t)
、 u& & (t)
is the displacement, velocity and
acceleration, F is the external force ; U is the force
produced by the piezoelectricity. In the process of
modeling, the influence of the force which produces by
on account of the piezoelectric material is neglected. So
the model is approximate. The figure of the active
vibration control of intelligent structure is shown in
Fig.2.
Figure 2. Active of vibration piezoelectric structure
k
k

134 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
B. Modeling the relation of fuzzy logic control
The field of fuzzy system and control has been making
a big progress motivated by the practical success in
industrial process control. Fuzzy systems can be used in
as closed-loop controllers. In this case the fuzzy system
measures the outputs of the process and takes control
actions on the process continuously. The fuzzy logic
controller uses a form of quantification of imprecise
information (input fuzzy sets) to generate by an inference
scheme, which is based on a knowledge base of control
force to be applied on the system.
The advantage of this quantification is that the fuzzy
sets can be represented by a unique linguistic expression
such as small, medium, and large etc. The linguistic
representation of a fuzzy set is known as a term, and a
collection of such terms defines a term-set, or library of
fuzzy sets. Fuzzy control converts a linguistic control
strategy usually based on expert knowledge into an
automation control strategy. There are three functions
required to be performed by fuzzy logic controller before
the controller can generate the desired output signals.
The first step is to fuzzify each input. This can be
realized by associating each input with a set of fuzzy
variables. In order to give semantics of a fuzzy variable a
numerical sense, a membership function is assigned with
each variable. The logical controller is made of four
main components: (1) Fuzzifier; (2) Knowledge base
containing fuzzy IF-THEN rules and membership
functions; (3) Fuzzy reasoning; and (4) Defuzzifer
interface [13, 14].
In this paper, fuzzy logic controller is designed as the
double-input, single-out (DISO) system: The inputs are
the displacement and the velocity of the tip of cantilever
beams, and the output is the control force on cantilever
beams. In this fuzzifier, the displacement is defined from
-5 to 5(-5,-4,-3,-2,-1,0, 1,2,3,4,5),the velocity is defined
from -2 to 2(-2,-1,0, 1,2),the control force is defined
from -7 to 7 (-7,-6,-5,-4,-3,-2,-1,0,1,2,3,4,5,6,7). Two
types of membership functions commonly adopted in
fuzzy logic control are triangle and trapezoidal shape.
We use these two type membership functions. In this
paper, compared with other methods, the method of
Mom was more effective. Accordingly, in this paper, a
way of establish fuzzy system is proposed as following:
(1) At first, the scope of the displacement and the
velocity are the maximal response of when received step
response.
(2) Plot the scopes of displacement’s and the force of
control’s out NB,NM,NS,ZO,PS,PM,PB; Then plot the
scopes of velocity ’s out N and P.
(3) According to the fuzzy rule of L.A.Zadeh’s[15], we
get the process of fuzzy illation.
At last, we use the way of Mom method to calculate in
order to obtain the result.
C. The rule of fuzzy control and the fuzzy controller
The fuzzy rule shows the fuzzy relation between the
input and output. The inputs and output are connected
with this relationship. In this paper, the displacement of
the tip of cantilever beam is chosen for the one input, the
© 2010 ACADEMY PUBLISHER
velocity is the other. In tradition method, the inputs
usually are the velocity and the rate of the velocity. In
this way, the time of calculation has been improved.
Figure 3. The basic configuration of the fuzzy system
Basing on the control rules, the signal is translated to
the driver. The function of fuzzy logic controller is
making the inputs fuzz up. In other words, it is the fuzzy
control that executes the process of fuzzed. The fuzzy
control’s basis was the rule database which was
composed of several rules. The final purpose of the fuzzy
logic controller was to make the fuzzy rule come true,
Fig. 3. The method of control vibration is minimized the
response of displacement of cantilever beam.
The function of the fuzzy logic controller is to provide
a force to prevent the vibration of the beam. On analysis
of the motion traits of cantilever beam, the rules were
obtained as fellows:
(1) If the displacement is PS and the velocity is PB, the
tip of beam is up far from of equilibrium positionH. So it
needs to add the downwardH force of NB and makes the
tip of beam close to the reference valuesH.
(2) If the displacement is PB and the velocity is PS, the
tip of beam is upward to the maximum displacementH. So
it needs to add the downwardH force of NS and make the
tip of beam close to the reference valuesH.
(3) If the displacement is PB and the velocity is NS, the
tip of beam is downward close to the equilibrium
positionH. Adding the downwardH force of NS is required
and makes the tip of beam close to the reference valuesH.
(4) If the displacement is PS and the velocity is NB, the
tip of beam is downward close to the equilibrium
positionH. Adding the downwardH force of NB is required
and makes the tip of beam close to the reference valuesH.
(5) If the displacement is NS and the velocity was NB.
The tip of beam is down far from of equilibrium positionH.
The upwardH force of PB should be added and makes the
end of beam close to the reference valuesH.
(6) If the displacement is NB and the velocity was NS.
The tip of beam is downward to the maximum
displacementH. The upwardH force of PS should be added
and makes the end of beam close to the reference valuesH.
(7) If the displacement is NB and the velocity is PS, the
tip of beam is upward close to the equilibrium positionH.
Making the tip of beam close to the reference valuesH
could be obtained by adding the P upwardH force.
(8) If the displacement is NS and the velocity is PB, the
tip of beam is upward to the maximum displacementH.
Making the tip of beam close to the reference valuesH
could be obtained by adding the PB upwardH force.
Fuzzy IF-THEN rule base is obtained by the analysis,
the author of this study, with many trial-and-errors.
(Table I). Fuzzy IF-THEN rule is the center of control

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 135
system. The fuzzy rule base is not invariable, it could be
modify in practice. The system of fuzzy logic control is
in Fig. 4.
TABLE I.
FUZZY IF-THEN RULE BASE
NB
NM
Figure 4. The system of fuzzy logic control
V. PARAMETER SELF-ADJUSTMENT
Base on the fuzzy logic control system, with big
weighting coefficient to be used against bad errors, and
small weighting coefficient of change rate to be used for
slight errors. The principle of establish the parameter
self-adjustment is using the different parameter selfadjustment
factor to implement the fuzzy control rules.
In this paper, the effective means of designing fuzzy
logic controller with fuzzy logic toolbox of Matlab is
introduced. Self-parameter is realized by compiling sfunction.
The organic combination of Matlab and
Simulink makes the design and simulation of parameter
self-adjustment fuzzy logic control system be easily and
effectively. These show the means is easy and elastic. It
can promote working efficiency of designers. The sfunction
is used to adjust the parameter because the
blocks in Matlab are not enough. The source programmer
of s-function is:
function[sys,x0,str,ts]=fpids(t,x,u,flag,m,AH,AL,Escope,
ECscope,Uscope)
ke=m/Escope;
kc=m/ECscope;
ku=Uscope/m;
switch flag,
case 0,
[sys,x0,str,ts]=mdlInitializeSizes(ke,kc,ku,m,AH,AL,
Escope,ECscope,Uscope);
case 3,
sys=mdlOutputs(t,x,u,ke,kc,ku,m,AH,AL,Escope,
ECscope,Uscope);
case {1,2,4,9}
NS
NO
PS
PM
NB ZO PM PB ZO NB NM ZO
NS PS PM ZO ZO ZO NM NS
ZO ZO ZO ZO ZO ZO ZO ZO
PS PS PM ZO ZO ZO NM NS
PB ZO PM PB ZO NB NM ZO
© 2010 ACADEMY PUBLISHER
PB
sys=[];
otherwise
error({'Unhandled flag=',num2str(flag)});
end
function[sys,x0,str,ts]=mdlInitializeSizes(ke,kc,ku,m,AH,
AL,Escope,ECscope,Uscope)
sys=[0,0,1,2,0,0,0];
function[sys]=mdlOutputs(t,x,u,ke,kc,ku,m,AH,AL,
Escope,ECscope,Uscope)
if u(1)>=Escope
u(1)=Escope;
end
if u(1)=Escope
u(2)=Escope;
end
if u(2)Uscope
result=Uscope;
end
if result

136 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
position response of the smart structure system with
fuzzy logic control is shown in Fig.11. Fuzzy Logic
Controllers is designed to control nonlinear vibration of a
smart structure and the effects of the controllers over the
system are examined. Fuzzy logic controller using in
such system which has nonlinear vibrations gives a good
result as tip position control of a smart structure. Settling
time of system is approximately 9s, and the displacement
of the tip beam is decreased from3.5mm to 2mm, so this
method has practical value in preventing vibration of
displacement. According to these results, suitable
performance of fuzzy logic controller is determined for
tip position control of a smart structure system. The
fuzzy logic controller designed is established properly
and this new controller can be used for such kind of
system.
Length
(mm)
strain quantity of the modal
Figure 5. Configurations of the cantilevered plate
5
4
3
2
1
0
Width
(mm)
TABLE II.
AMETER OF THE BEAM
Thickness
(mm)
Density
Kg/m3
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
number of t he nodes
Figure 6. The strain quantity of the node
Flexibility
modulusH(Pa)
350 25 4 2900 6.8e+10
TABLE III.
THE PARAMETER OF THE PIEZOELECTIC PATCHES
Length Width Thickness Density Flexibility
(mm) (mm) (mm) Kg/m3 modulusH(Pa)
35 25 1 8300 3.8e+9
………………
1 3 ………………
© 2010 ACADEMY PUBLISHER
39
Figure 7. The fuzzy logic control system
Figure 8. Rule viewer
Figure 9. Control surface of Mom

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 137
Figure 10. Tip position response of the beam
Figure 11. Tip position response of the beam with fuzzy logic
controller
According to the result in Fig. 10, the system has an
obvious oscillation phenomenon at the time of 4s. In
order to avoid this phenomenon, we use parameter selfadjustment
which is compiled by s-function to improve
the control effect. Where m is 6, AH is 0.8, AL is 0.2,
Escope is 0.005, ECscope is 0.1, Uscope is 0.008. The
result was shown in Fig. 12. Settling time of system is
approximately 8s; the displacement of the tip beam is
decreased from 0.2mm to 0.043mm. In addition, this
parameter self-adjustment method eliminates the
oscillation phenomenon basically.
© 2010 ACADEMY PUBLISHER
Figure 12. Tip position response of the beam with fuzzy logic
controller
VII. CONCLUSION
This new fuzzy logic controller which designs in this
paper improves calculation efficiency and reduces
calculation complexity. More importantly, it has the
obvious effect to reduce the displacement of the beams
tips. Besides that, this paper uses a new method to design
a fuzzy logic controller and compiles parameter selfadjustment
factor with s-function. The simulation results
show that this new fuzzy logic control has more
superiority than the ordinary’s, the performance is
satisfactory.
ACKNOWLEDGMENT
This work was financially supported by Natural
Science Foundation of Hebei Province under Grant
No.E2008000731, and Scientific Research Project of
Education Department of Hebei Province under Grant
No.2006107.
REFERENCES
[1] Bailey, T. and Hubbard, J. E., 1985, Distributed
piezoelectric polymer active vibration control of a
cantilever beam, AIAA J. Guidance, Control Dynam, 8:
605-11.
[2] Guo, H. Y. Zhang, L. L. and Zhou, J. X., 2004, Optimal
placement of sensors for structural health monitoring
using improved genetic algorithms, Smart Materials and
Structure, 13:528-534.
[3] Martin Kögl and Emílio, C. N., Silva, 2004, Topology
optimization of smart structures: design of piezoelectric
plate and shell actuators, Smart Materials and Structure,
14:387-399.
[4] Zongjie Cao, Bangchun Wen and Guangwei Meng, 2000,
Topological optimization of placements of piezoelectric
actuator of intelligent structure, Journal of Northeastern
University (Natural Science), 21(4): 383-385.

138 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
[5] Bayard D. S., Hadaegh, F. Y. and Meldrum, D. R., 1988,
Optimal experiment design for identification of large
space structures, Automatica, 24(3): 357-364.
[6] Brown C.B., Yao J.T.P. (1983). “Fuzzy sets and Structural
Engineering”. Journal of Structural Engineering, ASCE,
No. 109, pp.1211-1225.
[7] Juang C., Elton D.J. (1986). “Fuzzy Logic for Estimation
of Earthquake”. Intensity Based on Building Damage
Record Civil Engineering System. No.3, pp.187-197.
[8] Battaini M., Casciati F., Faravelli L. (1998) “Fuzzy
Control of Structural Vibration. An Active Mass System
Driven by a Fuzzy controller”. Earthquake Engineering
and Structural Dynamics, No.27, pp.1267-1276.
[9] Symans M.D. and Kelly A.S. (1999). “Fuzzy Logic
Control of Bridge Structures Using Intelligent Semi
Active Seismic Isolation Systems”. Earthquake
Engng.Struct.Dyn,vol.28, pp.37-60.
[10] Park H., R.Agarwal, K.Nho (2000). “Fuzzy Logic Control
of Structure Vibration of Beams”. Aerospace Aerospace
Sciences Meeting and Exhibit, A IAA-0172.
[11] Lee R.Y.Y., Abdel-Motagaly K., Mei C. (1998). “Fuzzy
Logic Control for Nonlinear Free Vibration of Composite
Plates with Piezoactuators”. CAIAA, pp. 20-23.
[12] Ding Wenjing, 1994, Current main topics on active
vibration control, Advances in mechanics 24(2): 173-180.
[13] L. Faravelli, T. Yao, Use of Adaptive Networks in Fuzzy
of Civil Structures, Micricimput, Cilvil ENG. 11:67-
76,1996.
[14] J.Yen, R. Langari. Fuzzy Logic-Intelligence Control and
Information. Prentice Hall, Englewood Cliffs, NJ, 1998.
[15] Zadeh L.A. (1965). “Fuzzy set”. Informat. Control, vol. 8,
pp. 338-353.
Jingjun Zhang, Male, born in Yucheng City, Henan
Province, in 1963. Received his B.Sc. degree in Engineering
Mechanics from Lanzhou University, Lanzhou, China, in 1985,
and the M.Sc. and PhD degrees in Mechanics from Jilin
University, Changchun, China, in 1993 and 1996 respectively.
He is currently a Professor in the College of Civil
Engineering, Hebei University of Engineering, Handan, China.
He is also the Chief of Department of Science Research in
Hebei University of Engineering, China. His research interests
include the area of flexible spacecraft structural control using
smart structure technologies and piezoelectric materials. He has
also developed research programs in attitude control of flexible
spacecraft, smart structures, and space robotics. Recently, he is
conducting research in spacecraft attitude dynamics, structural
dynamics, spacecraft testing and optimization design. More
than 70 papers have been published in academic journals in
domestic and abroad.
Prof. Zhang is an Associate Fellow of the Journal of China
Coal Society.
Liya Cao, Female, born in Handan City, Hebei Province,
in 1982. She is currently a graduate student in the College of
Civil Engineering, Hebei University of Engineering, Handan,
China. Her main research interests focus on intelligent structure
and the active vibration control with fuzzy logic.
Weize Yuan, Male, born in Baoding City, Hebei Province,
in 1981. Received his B.Sc. degree in College of Civil
Engineering from Hebei University of Engineering, Handan,
China, in 2006. He is currently a graduate student in the
College of Civil Engineering, Hebei University of Engineering,
Handan, China. His main research interests focus on the
© 2010 ACADEMY PUBLISHER
problem of obtaining the piezoelectric actuator’s optimal
location and size of the smart structure.
Ruizhen Gao, Male, born in Baoding City, Hebei
Province, in 1979. Received his B.Sc. degree in College
of Civil Engineering from Hebei University of
Engineering, Handan, China, in 2005. He is currently a
teacher in the College of Mechanical and Electrical
Engineering, Hebei University of Engineering, Handan,
China. His main research interests focus on GA Genetic
Algorithm.
Jingtao Li, Male, born in Shijiazhuang City, Hebei
Province, in 1981. Received his B.Sc. degree in College
of Natural Science from Hebei University of Engineering,
Handan, China, in 2005. He is currently a graduate
student in the College of Civil Engineering, Hebei
University of Engineering, Handan, China. His main
research interests focus on topological optimization.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 139
Numerical Simulation of Snow Drifting Disaster
on Embankment Project
Shouyun Liang, Xiangxian Ma and Haifeng Zhang
Key Laboratory of Mechanics on Disaster and Environment in Western China (Lanzhou University) /Ministry of
Education, Lanzhou, P. R. China
Email: liangsy@lzu.edu.cn, maxxan04@lzu.cn, hai_feng2008@yahoo.com.cn
Abstract—Snow drifting is a typical natural disaster, and an
increasing importance has been attached to its theory
research and field observation for the sake of cold region
engineering. Numerical simulation, as an efficient method
of studying snow drifting, promises to be widely applied in
this field. In this paper, the finite element method is used to
simulate the wind velocity field. Selected upwind,
downwind and middle part of a road as key comparison
points, a quantitative analysis constructed of the influence
of subgrade width, height and slope ratio on the wind
velocity of embankment. The results showed that, under
certain circumstances, the velocity of snow drifting
demonstrated different function types of increase with the
height and slope ratio of subgrade enhancing, and the
influence of the height is greater than the slop ratio of
subgrade, while the velocity of wind decreases with the
width of subgrade increasing; the numerical values
approximately agreed with the field observed results, but
the numerical simulation is more sensitive to various forms
of embankment. The numerical results can offer references
for engineering construction when the region in which wind
velocity is slower than the threshold velocity as snowpack
area.
Index Terms — embankment project, snow drifting
disaster, numerical simulation, ANSY
I. INTRODUCTION
Snow drifting is a typical natural disaster. The
frequent occurrence of snow drifting has directly brought
serious loss to industrial and agricultural production and
people’s life and property, baffling sight, blocking
traffic, breaking electricity, leading industry and
agriculture to go out of production and other hazards [1].
Therefore, an increasing importance has been attached to
its theory research and field observation [2]. Sato et al.
have brought forward turbulence model and other
theories based on Eurler view and multiphase flow and
Prandtl mixed length; Liston et al. have applied simple
k-ε turbulence model to simulate accumulation process
of snow according to the way in which the control unit is
filled in by the information of the transporting capacities
of particles of snow. In the study of road snow drifting
disaster, foreign scholars mainly employ CFD to
simulate the influence of snow fence on snow drifting
Manuscript received January 23, 2009; revised February 4, 2009;
accepted February 20, 2009.
Corresponding author, E-mail: liangsy@lzu.edu.cn
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.139-143
distribution, and by taking account of suspension and
saltation analyze the law of flow field transformation and
snow distribution [3]. By using FLOW-3D, Thordarson
[4] has simulated distribution and change of volume of
the snow near road flow field and within the unit, and
thereby simulated the movement of snow drifting. Xi and
Ying et al. [5-6] have by employing bilateral flow theory
simulated the flow field of snow drifting disaster through
FLUENT software, and concluded the qualitative change
of different subgrade flow field. Hu et al. [7] has applied
ANSYS software to simulate the change of wind velocity
field in order to discuss the impact of snow-sand
blockings on subgrade snow, and thereby put forward
sound setting and form of snow-sand blockings. Among
these informed simulation research of snow drifting,
there exists a lack of a qualitative study of the impact of
blockings on flow field, and there runs short of
considering the relation between engineering structures
and snow drifting disasters. Taking the embankment
project for example, the author of this paper has
simulated its variation of wind velocity field through
ANSYS, deeply analyzed every factor of embankment
sections with different models, the quantitative
relationship of wind velocity field, and the combined
effects of each factor on wind velocity field, and
achieved the objective of indirectly reflecting the
graveness of snow drifting disasters.
II. MODEL DESIGN
According to related theory of aerodynamics, there
existed a certain degree of relevance between
accumulated snow and wind speed, furthermore,
particles of snow bore a better following performance,
therefore, the author has drew on the variation of wind
velocity field to reflect indirectly the change of snow. A
simulation experiment has constructed based on Fortran
CFD of ANSYS, by adopting simple algorithm, N-S
equation and k-e model [8-10]. And this conception has
been proved feasible by the contrast of wind velocity
field observations (Feb, 29th 2008) of Jinghe-yining (JY)
railway embankment section (DK176+848) under
construction with its simulation values (Fig.1). Besides,
some simulation conditions of embankment sections
have been given by referring to standards concerning
railway and civil engineering techniques (Table I).
Seeing that the wind vertical with road surface influences
accumulated snow the most, the author has only taken
account of the condition when the wind is vertical with

140 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
road surface, leaving out account of the condition when
the wind is oblique crossing and horizontal.
Figure 1. Comparison of wind speed of simulation and field
measured of JY railway embankment section
(DK176+848.555)
Condi-
tion
1
TABLE I.
PARAMETERS BASED ON SIMULATION
Embankment
Width (m)
9, 12, 15, 18,
24.5
2 9
Embankment
height (m)
Slope
ratio
III. ANALYSES OF SIMULATION RESULTS
Velocity
(m/s)
4 1:1.5 9.4
2, 4, 8, 12,
13, 14, 15,
16, 17
3 9 4
1:1.5 9.4
1:1,
1:1.25,
1:1.5,
1:1.75
A. Impact of embankment width on wind velocity field
For subgrade with different width, their wind velocity
fields are similar, in both of which wind velocity of the
area with a certain distance to the foot of upwind slope is
blocked, and thus declines drastically. The wind velocity
reaches its minimum at the foot of the upwind slope,
with airflow subsequently climbing the slope it increases
to its maximum at the shoulder, and it reduces slightly in
between, and then increases again at the shoulder of
downwind slope (Fig. 2). The wind velocity of the area
one meter from the foot of the downwind slope reduces
almost to zero, thereafter a whirlpool comes into being,
and wind velocity declines more or less, and then it
would retrieve its former speed. Fig. 3 illustrates varied
wind velocity (0.5 m above the surface) of road surface
of different subgrade width, based on the key points of
shoulder of upwind slope, in between, and shoulder of
downwind slope. Regression analysis showed that the
wind velocity of subgrade represents different function
changes with the increase of subgrade width at the three
points:
At shoulder of upwind slope:
V= -0.1114w +16.228 (R² = 0.9903) (1)
In between:
V = 0.0021w 2 -0.0906w+14.134
(R² = 0.9907) (2)
At shoulder of downwind slope:
V = 0.0052w 2 -0.2199w+12.261
(R² = 0.9926) (3)
© 2010 ACADEMY PUBLISHER
9.4
Figure 2. The contour of velocity with different embankment
width (fig A-D: 9 m,12 m,15 m,24.5 m)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 141
Figure 3. The relationship between changes of subgrade width
and wind speed of road ( 0.5 m above pavement)
B. Impact of embankment height on wind velocity field
For subgrade with different height, their wind velocity
field bears similarity, and even resembles with wind
velocity field of subgrade with different width (Fig.4).
Fig. 5 illustrates varied wind velocity (0.5 m above the
surface) of road surface of different subgrade height,
based on the key points of shoulder of upwind slope, in
between, and shoulder of downwind slope. Regression
analysis represented that the wind velocity of subgrade
first increases and then decreases with the increase of
subgrade height at the three points. With the 15 m being
the dividing line, when the height is less than 15 m, the
wind velocity has good relativity to quadratic curves.
The law presents itself as follows:
At shoulder of upwind slope:
V = -0.0522 h 2 +1.58127 h +10.52638
(R² = 0.9581) (4)
In between:
V=-0.0439h 2 +1.1205h+7.1518
(R² = 0.9843) (5)
At shoulder of downwind slope:
V = -0.042h 2 +1.4331h+9.1942
(R² = 0.9994) (6)
When the height is more than 15 m, wind velocity
begins to decline, first considerably and then gradually.
The relativity of fitting curve reduces.
V = -0.0886h 2 +2.2006h+7.2217
(R² = 0.7877) (7)
V = -0.0731h 2 +1.7991h+5.2242
(R² = 0.7495) (8)
V = -0.1013h 2 +2.5052h+8.0499
(R² = 0.7913) (9)
C. Impact of slope ratio on wind velocity field
For subgrade with different slope ratio, the wind
velocity fields distribute generally similar to above
mentioned two situations (Fig. 6). Fig. 7 illustrates the
wind velocity (0.5 m above the surface) of road surface
with different slope ratio. We can see that, the wind
velocity represents a linear decrease with the increase of
slope ratio at the three points, in which the variation of
© 2010 ACADEMY PUBLISHER
wind velocity at shoulder of upwind slope can be
formulated as:
Figure 4. The contour of velocity with different height of road
(fig A-D: 2 m, 8 m, 15 m, 16 m)

142 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Figure 5. The relationship between changes of subgrade height
and wind speed of road (0.5 m above pavement)
V=-105.89tanα 3 +235.44tanα 2 -167.33tanα+52.271
(R² = 1) (10)
D. Comprehensive Analysis
In order to simultaneously study the combined effects
of subgrade width, height, and slope ratio on wind
velocity field of road surface, the author has adopted
linear regression to calculate correlation coefficient of
each factor with SPSS under a confidence level of 0.95.
At shoulder of upwind slope:
V=-0.089w+1.034h+3.795 tanα+8.557 (11)
In between:
V=-0.052w+2.101h+2.875tanα+10.007 (12)
At shoulder of downwind slope:
V=-0.117w+3.24h+4.018tanα+5.34 (13)
From the formulation, we can see the wind velocity is
negatively related to the subgrade width, positively
related to the other two factors. All taking into
consideration, the impact of subgrade height on wind
velocity field is much more significant than that of the
slope ratio.
IV. CONCLUSION AND DISCUSSION
Though embankment project just serves as a relatively
ideal subgrade model for snow drifting disaster, we still
can gain constructive knowledge from above research
with regard to the relation between subgrade model and
snow drifting disaster:
(1) With subgrade width increasing, the wind velocity
of road declines in different functions: it changes the
most drastically at the shoulder of upwind slope; next is
at the shoulder of downwind slope; the last is the
between part. For this reason any increasing or
decreasing of subgrade width is not a significant factor
leading to snow disaster on road. With subgrade height
and slope ratio increasing, the wind velocity of road
increases in different functions. However, the impact of
subgrade height on wind velocity field is far bigger than
that of slope ratio, yet when subgrade height is more
than15 m. the wind velocity of road reduces instead. This
dividing line of 15 m coincides with that of h=15.1 m in
© 2010 ACADEMY PUBLISHER
Wang’s (2001) wind tunnel experiment [1]. Thus, there
exists an appropriate subgrade height from the
perspective of snow on road.
Figure 6. The contour of velocity with different slope of ratio(fig
A-D: 1:1,1:1.25,1:1.5,1:1.75)

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 143
Figure 7. The relationship between changes of slope ratio
and wind speed of road (0.5 m above pavement)
pavement)
(2) The conception of considering sugrade width in
this simulation research has consulted relevant railway
and civil engineering standards, so a study of its wind
velocity field can provide some guidance to practical
practice. Comparison shows that the simulated values
agree well with the observations. The simulated values
seem more sensitive to the change of subgrade design
parameters, whereas the observations would be
frequently affected by the wild weather, terrain and other
factors. This renders the system more complicated. If we
regard the area in which wind speed is less than the
starting speed in the figure of simulated wind velocity
field as snow area, this research can offer references for
prevention and control of snow drifting disaster in
engineering aspect.
(3) This research is based only on three representative
points of a road and the wind speed of 0.5 m above the
road, therefore, the investigation of wind velocity
variation on a road remains to be carried further in depth
by taking more data on points and heights, improving the
precision and depth of research.
© 2010 ACADEMY PUBLISHER
ACKNOWLEDGMENT
We are grateful to X. Wang, Y. Yin, X. Chang and Z.
Shao for their technical support in the field experiment.
We thank T. Wang for his valuable discussions and
suggestions. This work was supported by a grant from
China Railway First Survey and Design Institute Group
Ltd.
REFERENCES
[1] Z. Wang, “Research on snow drift and its hazard control
engineering in china,” Lanzhou University Press, May.
2001. (in Chinese)
[2] L. Zhang, “Research on Snowdrifts and Snow Disaster
Simulation Technology of road,” Jilin University
Postgraduate Dissertations, Jun. 2005. (in Chinese)
[3] S. Alhajraf. “Computation fluid dynamic modelling of
drifting particles at porous fences,” Environmental
Modelling & Software, vol.19, No.2, Feb. 2004, pp.
163-170.
[4] S. Thordarson, H. Norem, “Simulation of two-dimensional
wind flow and snow drifting application for roads: part I,”
Snow Engineering, vol.7, No.4, Jun. 2004, pp. 246-265.
[5] L. Ying, J. Li, X. Zhang, J. Xi, “Study on numerical
simulation for snow fence height,” Journal of Changchun
University of Science and Technology, vol.29, No.3, Sep.
2006, pp. 54-57. (in Chinese)
[6] J. Xi, J.Li, G. Zhu, G. Zhang, “Hydrom echaical
mechanism of road snow drift deposit and its depth
model,” Journal of Jilin University (Engineering and
Technology Edition), vol.36, No. supp.2, Sep. 2006, pp.
152-156. (in Chinese)
[7] P. Hu, C. Zheng, “The anslysis on the wind speed field of
the snow protection facilities and sand control facilities,”
Journal of Chongqing Jiaotong University (Natural
Sciences), vol.24, No.3, Jun. 2005, pp. 63-68. (in Chinese)
[8] R. Essery, L. Long, J. Pomeroy, “A distributed model of
blowing snow over complex terrain,” Hydrological
Processes, vol.13, No.2, Feb. 2004, pp. 2423-2438.
[9] P. Bartelt, M, Lehning, “A physical SNOWPACK model
for the Swiss avalanche warning Part I: Numerical
model,” Cold Regions Science and Technology, vol.35,
No.3, Nov. 2002, pp. 123-145.
[10] J. Judith, M. Lehning, A. Vrouwe, “Field measurements of
SNOW-DRIFT threshold and mass fluxes, and related
model simulations,” Boundary-Layer Meteorology,
vol.113, No.8, Apr. 2004, pp. 347-368.

144 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
The Simulation of Extraterrestrial Solar Radiation
Based on SOTER in Zhangpu Sample Plot and
Fujian Province
Zhi-qiang Chen
College of Geographical Sciences, Fujian Normal University, Fuzhou, China
soiltuqiang061@163.com
Jian-fei Chen
School of Geographical Sciences, Guangzhou University, Guangzhou, China
cjf@gzhu.edu.cn
Abstract—The study establishes DEMs and the computer
models of daily extraterrestrial radiation in Zhangpu
sample plot and Fujian province and annually
extraterrestrial radiation in Fujian province using GIS base
on SOTER. The results indicate that the daily
extraterrestrial radiation is mainly 17-18MJ/m 2 in Zhangpu,
the one is mainly 14-18MJ/m 2 and the annually
extraterrestrial radiation is mainly 5000~7000MJ/m 2 in
Fujian province. The influence of topographic feature is
significant, in general the solar radiation in the chine is
bigger than the one in the valley, and the one in the sunny
slope is bigger than in the shady slope, the high radiation
value appears primarily in the sunny slope with the chine as
demarcation line. The method can extract daily and
annually extraterrestrial radiation with high precision and
speediness at small and moderate scales which don’t take
cloud into consideration. The orientation of future research
is to improve precision by RS images and combine with
SOTER.
Index Terms—DEM, GIS, solar radiation model, Zhangpu
sample plot, Fujian province
I. INTRODUCTION
Solar radiation is the main energy of physical,
biological and chemical processes on earth surface, is the
necessary parameter in many models [1-3]. In the study of
resource and environment in regional spatial scope, local
terrain impact is enormous and must be considered [2].
Because of the complexity of computer models of
extraterrestrial solar radiation considering terrain
conditions, difficulty of terrain parameters acquisition and
lack of appropriate calculation platforms, for a long time,
it is neglected or simplified [1, 4]. GIS has provided the
strong technical support for these work [2, 5]. Solar
energy study and solar energy heat utilization are usually
conducted in clear-day and extraterrestrial solar radiation
models have broad scope [6], so the study took
Manuscript received Febuary 1, 2009; revised Febuary 15, 2009;
accepted March 15, 2009.
Corresponding author: Jian-fei Chen
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.144-149
extraterrestrial solar radiation as object, took ArcGIS as
platform in an experimental plot of Zhangpu sample plot
as small scale and Fujian province as moderate scale,
calculated the solar radiation spatial distribution law
caused by hypsography under the same solar radiation in
the complex topographic area of the subtropical zone and
the model’s applicability at small and moderate scales.
The results can provide important information for solar
radiation spatial-temporal distribution in the small and
moderate areas, agricultural regionalization, effective
utilization of solar energy and precision agriculture [1].
II. STUDY AREA
The study area is located in the northeast of Zhangpu
county in Fujian province at 24°10′ ~ 24°15′N and
117°48′~118°00′E including Futan, Maping, Qianting
and part of Baizhuhu farm, is about 8236 hm 2 . Annual
mean temperature is about 21℃ and annual precipitation
is about 900mm, it’s clear in dry and humid season;
coastal plains, platform and hill appear in turn from the
sea to the land;there is no big river with weak runoff
regulation and poor hydrological condition; because of
the long term artificial activity, the original vegetation
has been destroyed and the secondary vegetation is
dominated by grass with thin forest; the stratigraphies
include Neogene Fotan Group, fluvial or marine or
Aeolian or proluvial deposits of Pleistocene and
Holocene in Quaternary, intrusive rock is early
Yanshanian biotite granite, Neogene Fotan Group is
basalt; the major soil types may be divided into the
following six orders: Ferrisols, Vertisols, Camblsols,
Ferralosols, Entisols and Anthrosols [7].
Fujian province lies between 23°32′ and 28°22′ of
north latitude with a area of 121,400 km 2 , is located on
the southeastern coast of China characterized by
subtropical climate, complicated landforms and multiple
mountainous and hilly terrains which account for 87.5%
of the total area of the province, leaving the remaining
10% for plains. It is always called the “Mountainous

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 145
Region of Southeast” with complex natural conditions. It
appears middle subtropical climate in the north mountain
and south subtropical climate in the south coast areas.
The annual mean temperature is around 15-21℃ and the
accumulated temperature (≥10℃) is 5500-7500℃, the
annual precipitation is around 1000-2000 mm and
increases gradually from the southeast coast towards the
northwest inland. The predominant natural vegetations
are the south subtropical monsoon rain forest and the
evergreen broad-leaved forest. Fujian has a history of
over 3000 years and its long term agro-economic
activities have brought about many types of land
utilization, among which that of forest land accounts for
66.8%, that of cultivated land 12.7%, that of garden plots,
3.8%, herbage land 0.1%, water areas 4.4%, construction
sites 3.1%, idle land 9.1% [8].
III. METHODS
A. Basic Astronomical Parameters
The method comes from the reference [9] in the
paper. Some basic parameters should be established when
calculating solar radiation somewhere on the earth
including solar altitudinal angle, solar aspect angle, solar
declination, sunrise & sunset time and so on [9, 10].
Eo=1.000109+0.03349cosθ+0.001472sinθ+0.000768co
s2θ+0.000079sin2θ (1)
Where Eo is the earth orbit correction factor, θ is day
angle, θ = 2πt / 365.2422, t = N-No, N is number of day,
that is date's sequence number in a year, No = 79.6764 +
0.2422 * (year-1985) - INT((year - 1985)/4).
Declination is the angle distance from the celestial
equator to the sun along right ascension circle in
equatorial coordinate system, it is positive when the sun is
in the north of the equator and on the contrary it is
negative, ranging from 0 to ±23.44. Hour angle describes
the sun’s movement in 24h, it is 0 at noon by local real
solar time, positive in the morning and negative in the
afternoon, 15 degrees per hour. Their calculation formulae
respectively are:
δ = 0.006894 - 0.399512cosθ + 0.072075sinθ -
0.006799cos2θ + 0.00089sin2θ (2)
ω = (So + Fo/60 - 12) * 15° (3)
ω1 = arcos(- tanδ * tanψ) (4)
ω2 = - ω1 (5)
Where δ is declination, ω is hour angle, So and Fo are
hour and minute by real solar time, ω1 is sunset hour
angle, ω2 is sunrise hour angle.
Solar altitudinal angle is the angle between solar ray
and horizontal surface, solar aspect angle is the angle
between projection of solar ray on horizontal surface and
local meridian, due north is 0 degree. They are influenced
by latitude, declination and hour angle. The calculation
formulae are:
H = arcsin(sinδsinψ + cosδcosψcosω) (6)
A = arccos[(sinhsinψ - sinδ)/coshcosψ] (7)
© 2010 ACADEMY PUBLISHER
Where H is solar altitudinal angle, A is solar aspect
angle, ψ is latitude.
B. Terrain Feature Calculation
When solar altitudinal angle and solar aspect angle are
decided, slope, aspect and terrain shield are the important
factors calculating extraterrestrial solar radiation
especially in mountain area. Terrain shield means that:
comparing the biggest horizon angle and corresponding
solar altitudinal angle within a certain solar aspect angle
range, if the former is bigger than solar altitudinal angle,
solar radiation is 0 in shadow range or else solar radiation
should be calculated. When GIS was not mature, slope,
aspect and terrain shield were calculated using manual
methods or complex algorithms. With the development of
GIS technology, most of GIS software have the extraction
function of terrain feature such as DRIVE SLOPE,
DRIVE ASPECT and HILLSHADE in ArcGIS;
contribute to the calculation of influence of terrain on
solar radiation [1].
C. Extraterrestrial Solar Radiation Model
The solar radiation received somewhere on the earth is
influenced by many factors such as terrain (elevation,
slope, aspect, shadowing), soil, vegetation and weather
conditions (atmospheric turbidity, cloud amount) and so
on. Because of objective conditions, slope, aspect and
elevation only were considered in the paper. The actual
solar radiation can be calculated using revised formulae if
there are more detailed ground data and weather
conditions data [2]. The extraterrestrial solar radiation
formula is:
n
W = IoTEo / 2π * ∑ [µsinδ(ωr,j - ωs,j) + νcosδ(sinωr,j
i= 1
-sinωs,j) - sinαsinβcosδ(cosωr, j- cosωs,j)] (8)
Where Io is solar constant, T is day length, n is azimuth
division number, empirical value is 36, ωr,j and ωs,j are
sunrise and sunset hour angles at every differential period
respectively, α is slope, β is aspect.
µ = sinψcosα - cosψsinαcosβ (9)
ν = sinψsinαcosβ + cosψcosα (10)
Annually extraterrestrial radiation can be obtained by
adding daily extraterrestrial radiation in a year.
VI. EXTRATERRESTRIAL SOLAR RADIATION
SIMULATIONS
DEM is the basis of terrain basic feature parameters
such as elevation, slope and aspect and so on. The scanned
1:10000 contour maps in Zhangpu and 1:250000 contour
maps in Fujian province to get DEM come from Soil and
terrain digital database (SOTER). SOTER is a cross or
edge database and can manage attribute database, spatial
database and models better [11]. SOTER is one of the
global databases [12]. The International Society of Soil
Science (ISSS) has developed a methodology for SOTER
since 1986 whose main objective is to establish a
computerized database storing attributes on topography,

146 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
soils, climate, vegetation and land use/cover, which is
linked to a Geographic Information System (GIS),
whereby each type of information or combination of
attributes can be displayed as a separate map or overlay,
or in tabular form. The database can be used for improved
mapping and monitoring of changes in world soils and
terrain resources, and for the development of an
information system capable of delivering accurate, useful
and timely information to decision-makers and
policy-makers [13]. The SOTER has been carried out in
many countries and areas including China, Argentina,
Brazil, Cuba, Mexico, Uruguay, Venezuela, Hungary in
many fields such as soil erosion, land evaluation, soil
fertility, food security and so on with powerful function
and advantages [14-18].
The scanned maps in Zhangpu and Fujian province
were digitized whose contour distances are 5m and 50m
respectively. DEM resolution should be set to generate
DEM using contour maps. Based on Z.L. Li’s study, the
relationship between DEM resolution and contour
distance on contour maps is:
D = K * Ci * cosα (11)
Where Ci is contour distance, α is average slope, K is
constant (it is between 1.5 and 2.0 taking terrain feature
into account otherwise it is between 1.0 and 1.5), D is
DEM resolution. Because terrain feature is often ignored
in DEM application, according to the formula above,
DEM accuracy is equal to the one of 1:10000 contour map
if DEM resolution is between 7 m and 42 m. The DEM
resolutions in Zhangpu and in Fujian province were set to
15m and 100m respectively [10]. The digital slope model
and digital aspect model can be acquired after DEM
establishment, at the same time the following indicators
can be calculated: sunrise and sunset angles, discrete
number of available sunshine angle, sunrise and sunset
hour angles and the ones at every differential period,
corresponding solar altitudinal angle and solar aspect
angle. The extraterrestrial solar radiation can be derived
by substituting all these indicators into the formulae in
Zhangpu sample plot and Fujian province.
Percentage of the Total Area(%)
70
60
50
40
30
20
10
© 2010 ACADEMY PUBLISHER
0
Simulation date was March 21, 2008 that was spring
equinox. The sun's meridian altitude was h = 90 – ψ + δ at
noon, sunshine time was about 12 hours, the daily
extraterrestrial solar radiation was calculated using DEM
according to the formula, that was the digital solar
radiation grid model corresponding to DEM. The daily
extraterrestrial solar radiation is mainly in 17-18MJ/m 2 in
Zhangpu (Fig 1 and Fig 2) and 16-18MJ/m 2 in Fujian
province (Fig 3 and Fig 5). Simulation year was 2008 and
the annually extraterrestrial solar radiation was calculated
which is mainly in 5000~6000MJ/m 2 and 6000~7000
MJ/m 2 with the minimum 1101.4 MJ/m 2 and the
maximum 6366.6MJ/m 2 in Fujian province (Fig 4 and Fig
6).
There is significant difference in spatial distribution
diversity of the daily extraterrestrial solar radiation in
Zhangpu sample plot and Fujian province and annually
extraterrestrial solar radiation in Fujian province with
strong spatial autocorrelation from the different
geographical locations. In general the solar radiation value
in the chine is bigger than the one in the valley, and the
one in the sunny slope is bigger than in the shady slope.
The high radiation value appears primarily in the sunny
slope with the chine as demarcation line.
Figure1. Spatial distribution of daily extraterrestrial solar radiation
in Zhangpu.
0-14 14-15 15-16 16-17 17-18 18-19
Daily Extraterrestrial Solar Radiation(MJ/m 2 )
Figure2. Statistic of daily extraterrestrial solar radiation in Zhangpu.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 147
Figure3. Spatial distribution of daily extraterrestrial radiation in
Fujian province.
Percentage of the Total Area(%)
Percentage of the Total Area(%)
70
60
50
40
30
20
10
0
70
60
50
40
30
20
10
© 2010 ACADEMY PUBLISHER
0
Figure4. Spatial distribution of annually extraterrestrial radiation in
Fujian province.
0-10 10-12 12-14 14-16 16-18 18-20
Daily Extraterrestrial Solar Radiation(MJ/m 2 )
Figure5. Statistic of daily extraterrestrial solar radiation in Fujian province.
1000 - 3000 3000 - 4000 4000 - 5000 5000 - 6000 6000 - 7000
Annually Extraterrestrial Solar Radiation(MJ/m 2 )
Figure6. Statistic of annually extraterrestrial radiation in Fujian province.

148 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
V. COMBINATION OF EXTRATERRESTRIAL SOLAR
RADIATION AND SOTER
Nowadays, with the development of science and
technology and fast growing economy, the accelerated
industrialization and urbanization, difficulties
encountered by researchers are how to manage, store and
standardized various data efficiently, how to dig out
valuable data to a great extent, and how to share the data
resources. The single solar radiation could not meet the
needs of users with the development of science and
technology and one of the way out is combination with
SOTER.
SOTER formulates a suit of regulations with
systematization and standardization for the database,
including not only the databases about topography and
soil, but also climate, land use/cover, vegetation and
reference file assistant databases. Underlying the SOTER
methodology is the identification of areas of land with
distinctive, often repetitive, pattern of landform,
lithology, surface form, slope, parent material, and soils.
The major differentiating criteria are applied in a
step-by-step manner, each step leading to a closer
identification of the land area under consideration. In this
way a SOTER unit can be defined progressively into
terrain, terrain component and soil component (Fig 7).
Tracts of land distinguished in this manner are named
SOTER units. Each SOTER unit represents one unique
combination of terrain and soil characteristics [19].
Terrain
Terrain
1:M
component
Soil
1:M
component
M:1
M:1
Terrain
component
data
1:M
Profile Horizon
Figure7. SOTER unit structure (1:M=one to many, M:1=many to one
relations).
There are two types of data in a SOTER database:
geometric data and attribute data. The geometric
component indicates the location and topology of SOTER
units, while the attribute part describes the non-spatial
characteristics. The geometric data is stored and handled
by GIS software, while the attribute data is stored
separately in a set of files, managed by a relational
database system. There are 6 attribute files and 118
attributes. A unique code is set up in both the geometric
and attribute databases to link these two types of
information for each SOTER unit [19].
The climate data were based on point observations
only and the link with the soils and terrain information
© 2010 ACADEMY PUBLISHER
exists by means of the geographical locations of these
points. Relating to 72 climate stations in Fujian province,
some climate data were interpolated using ArcGIS such
as average temperature, accumulated temperature,
monthly precipitation, minimum and maximum
temperatures, relative humidity and evaporation besides
radiation.
VI. DISCUSSIONS
The study integrated solar radiation model and DEM
based on GIS, took Zhangpu sample plot in the southeast
coastal hilly and platform area of Fujian province as small
scale and Fujian province as moderate scale, gave full
consideration to the integration of all kinds of terrain
factors, explored the effects of slope, aspect and so on to
the spatial distribution diversity of extraterrestrial solar
radiation at two scales. The method can extract daily
extraterrestrial and annually solar radiation with high
precision and speediness at small and moderate scales, but
there is certain error in the model which mainly comes
from two aspects. On the one hand, cloud was not taken
into consideration, while cloud is one of the most import
parameters of solar radiation energy transmission, so there
is certain deviation between calculation results and actual
solar radiation value [20,21]. On the other hand, DEM is
the important data in the whole model, its quality (such as
DEM accuracy, resolution etc.) determines the final
simulation results [1].
The methods taking cloud effect into consideration are
various in solar radiation models which seldom contain
space characteristics because there are no conventional
observation data of cloud spatial distribution, even if
there are observation data it is hard to simulate the
geometric relation among sun, cloud and ground. Only
using RS could effects of cloud be took into
consideration in the spatial model of solar radiation [22].
The study is going to use ASTER images to extract cloud
amount and ground state for improving simulation
precision and meeting the needs of general solar energy
engineering and science research.
The single solar radiation could not meet the needs of
users with the development of science and technology.
The SOTER database contains more sufficient data and
can be used for a wide range of applications at different
scales. At the same time, only combining with SOTER,
can the climate data develop deeply and widely and the
superiority of simulation of extraterrestrial solar radiation
based on DEM can be shown. The approach may be used
to support strategic decision-making seeking to optimize
land use/cover, prioritize research, and guide
conservation planning.
ACKNOWLEDGMENT
This work was supported in part by a grant from
foundations: National Natural Science Foundation of
China, No. 40371054; The Key Project of the Ministry of
Education of China, No. 206107; Project of Science and
Technology Department of Fujian Province for the

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 149
Youth, No. 2006F3037; Excellent Young Core Teachers
Foundation of Fujian Normal University, No.
2008100233
REFERENCES
[1] H.L He, G.R. Yi, and D. N. “Method of global solar
radiation calculation on complex territories,” Resources
Science. China science press. Beijing, vol.25, pp. 78-85,
2003.
[2] Z.W. De, S.X. Ni, H.L. Zhou, H.L. Zhang, and Q.P. Tu.
“Calculation of possible direct solar radiation around
Qinghai lake based on DEM,” Plateau Meteorology. Cold
and arid regions environmental and engineering research
institute of Chinese academy of sciences. Lanzhou, vol.22,
pp. 92-96, 2003.
[3] C. Wloczyk, R. Richter. “Estimation of incident solar
radiation on the ground from multispectral satellite sensor
imagery,” International Journal of Remote Sensing. Taylor
& Francis. London UK, vol.27, pp. 1253, 2006.
[4] K S. Reddy, R. Manish. “Solar resource estimation using
artificial neural networks and comparison with other
correlation models,” Energy Conversion and Management.
Ergamon-Elsevier Science Ltd. Oxford Uk, vol. 44, pp.
2519-2530, 2003.
[5] J.G. Corripio. “Vectorial algebra algorithms for calculating
terrain parameters from DEMs and solar radiation
modeling in mountainous terrain,” Geographical
Information Science. Taylor & Francis. London UK, vol.
17, pp. 1-23, 2003.
[6] G.Q. Qiu, Y.J. Xia, and H.Y. Yang. “An optimized
clear-day solar radiation model,” Acta Energiae Solaris
Sinica. Chinese institute of renewable energy. Beijing,
vol.25, pp. 456-460, 2001.
[7] Z.Q. Chen. “The study and application of the regional soil
and terrain digital database(SOTER),” Ms D Thesis. Fujian
Normal University. Fuzhou, 2003.
[8] H.J. Zhu. Sustainable development and land utilization, 1 rd
ed., China hong kong yearbook press: Hong Kong, 1997.
pp. 41-46.
[9] G.A. Tang, X. Yang, Q. Zhang, and F.D. Deng. “DEM
based simulation on regional solar radiation,” Science
Journal of Northwest University Online. Northwest
university press. Xian, vol.1, pp. 208-214, 2003.
[10] J.Y. Gong. On theories and technology of contemporary
GIS, 1 rd ed., Wuhan University of Science and Technology
Press: Wuhan, 1999, pp. 197-199.
[11] Z.Q. Chen, J.F. Chen. “Assessment of the relative
sensitivity of ecosystem to acid deposition in Fujian
Province,” Journal of Fujian Normal University(Natural
Science Edition). Fujian normal university. Fuzhou, vol.23,
pp. 83-86, 2007.
[12] Z.Q. Chen, J.F. Chen, and X.Q. Xie. “Suitable sites
selection for Euphoria Longana based on Fujian SOTER
database and GIS,” Acta Pedologica Sinica. soil science
society of China. Nanjing, vol.42, pp. 700-703, 2005.
© 2010 ACADEMY PUBLISHER
[13] Y.G. ZHAO, G.L. Zhang, and Z.T. Gong. “SOTER-based
soil water erosion simulation in Hainan island,”
Pedosphere. Science press. Beijing, vol.13, pp. 139-146,
2003.
[14] Z.H. Yi, D.H. Xiong, Z. Yang, Y.R. He, and Y.Y. Zeng.
“Natural productivity evaluation of cultivated land based
on SOTER database in the typical region of upper reaches
of the Yangtse river,” Chinese Journal of Soil Science. Soil
Science Society of China. Shenyang, vol.36, pp. 145-148,
2005.
[15] C.W. Lv, S.Q. Cui, and L. Zhao. “Soil organic carbon
storage and its spatial distribution characteristics in Hainan
Island: a study based on HNSOTER,” Chinese Journal of
Applied Ecology. Science press. Shenyang, vol.17, pp.
1014-1018, 2006.
[16] Y.G. Zhao, G.L. Zhang, and H. Zhang. “Systematic
assessment and regional features of soil quality in the
Hainan Island,” Chinese Journal of Eco-Agriculture. China
science press. Shijiazhuang, vol.12, pp. 13-15, 2004.
[17] M.X. Men, J. Chen, Z.R. Yu and H. Xu. “Assessment of
soil erosion based on SOTER in Hebei Province,” Chinese
Agricultural Science Bulletin. Chinese agricultural science
bulletin press. Beijing, vol.23, pp. 587-592. 2007.
[18] X.L. Zhang, G.L. Zhang, and Z.T. “Gong. Evaluation for
some tropical crops in Hainan Province by using ales based
upon HaiSOTER,” Scientia Geographica Sinica.
Changchun publishing house. Changchun, vol.21, pp.
344-349, 2001.
[19] H.Z. Zhou. “A testing of SOTER project in China,”
Pedosphere. Science press. Beijing, vol.36, pp.
153-160,1993.
[20] J.G. Sun, J. Zhao, J.G. Zhen, and D.P. Li. “The study of
computer model and application of solar radiation value in
yellow earth will ravine region based on DTM,”
Engineering of Surveying and Mapping. Engineering
college of heilongjiang. Heilongjiang, vol.12, pp. 28-30,
2003.
[21] K.A. Joudi, N.S. Dhaidan. “Application of solar assisted
Heating and Desicddant Cooling Systems for a Domestic
building,” Energy Conversion & Management.
Ergamon-Elsevier Science Ltd. Oxford UK, vol.12, pp.
995-1022, 2001.
[22] X. Li, G.D. Cheng, X.Z. Chen, and L. Lu. “Improvement
of solar radiation model in arbitrary topography,” Chinese
Science Bulletin. China science press. Beijing, vol.44, pp.
993-998, 1999.
Zhi-qiang Chen Fujian,china, 1978, Ph.D., specialized in
GIS and land resource and land use plan. E-mail:
soiltuqiang061@163.com
Jian-fei Chen E-mail:gdcjf@21cn.net

150 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Multiplicate Particle Swarm Optimization
Algorithm
Shang Gao and Zaiyue Zhang
School of Computer Science and Engineering, Jiangsu University of Science and Technology, Zhenjiang 212003, China
Email: gao_shang@hotmail.com yzzjzzy@sina.com
Cungen Cao
Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100080, China
Email: cgcao@ict.ac.cn
Abstract—Using Particle Swarm Optimization to handle
complex functions with high-dimension it has the problems
of low convergence speed and sensitivity to local
convergence. The convergence of particle swarm algorithm
is studied, and the condition for the convergence of particle
swarm algorithm is given. Results of numerical tests show
the efficiency of the results. Base on the idea of
specialization and cooperation of particle swarm
optimization algorithm, a multiplicate particle swarm
optimization algorithm is proposed. In the new algorithm,
particles use five different hybrid flight rules in accordance
with section probability. This algorithm can draw on each
other ' s merits and raise the level together The method uses
not only local information but also global information and
combines the local search with the global search to improve
its convergence. The efficiency of the new algorithm is
verified by the simulation results of five classical test
functions and the comparison with other algorithms. The
optimal section probability can get through sufficient
experiments, which are done on the different section
probability in the algorithms.
Index Terms—particle swarm optimization algorithm,
convergence, parameter
I. INTRODUCTION
Particle swarm optimization (PSO) is one of the
evolutionary computational techniques. Since its
introduction (Eberhart R C , Kennedy J.1995; Kennedy J,
Eberhart R.1995; Shi Y H , Eberhart R C.1998)[1,2,3,4],
PSO has attracted much attention from researchers
around the world. It is a population-based search
algorithm and is initialized with a population of random
solutions, called particles. Each particle in PSO moves
over the search space at velocity dynamically adjusted
according to the historical behaviors of the particle and its
companions. Using Particle Swarm Optimization to
handle complex functions with high-dimension it has the
problems of low convergence speed and sensitivity to
local convergence. Although PSO is a fairly recent
algorithm, it is quickly gaining momentum in the
Supported by the National Natural Science Foundation of China
under Grant No.60773059.
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.150-157
engineering research community. Several researchers
have analyzed the performance of the PSO with different
settings, e.g., neighborhood settings, cluster analysis, etc.
It has been used for approaches that can be used across a
wide range of applications. For example Fourie and
Groenwold applied the algorithm to structural shape and
sizing [5] and topology optimization [6] problems. The
authors applied the algorithm to a cantilevered beam [7]
and multi-disciplinary design optimization of a transport
aircraft wing [8]. The University of Florida has applied
the algorithm to biomechanical system identification
problems e.g. [9]. Some authors have worked on parallel
PSO algorithms in an attempt to alleviate the associated
high computational cost. Most parallel implementations
of the PSO algorithm presented to date are based on a
synchronous implementation e.g. [9], where all design
points within a design iteration are evaluated, before the
next design iteration is started. Particle swarm
optimization can be and has been used across a wide
range of applications. Areas where PSOs have shown
particular promise include multimodal problems and
problems for which there is no specialized method
available or all specialized methods give unsatisfactory
results. PSO applications are so numerous and diverse
that a whole paper would be necessary just to review a
subset of the most paradigmatic ones. Here we only have
a limited space to devote to this topic. So, we will limit
ourselves to listing the main application areas where
PSOs have been successfully deployed. Base on the idea
of specialization and cooperation of particle swarm
optimization algorithm, a multiplicate particle swarm
optimization algorithm is proposed.
II. THE BASIC PSO ALGORITHM
In the PSO algorithm, the birds in a flock are
symbolically represented as particles. These particles can
be considered as simple agents “flying” through a
problem space. A particle’s location in the multidimensional
problem space represents one solution for
the problem. When a particle moves to a new location, a
different problem solution is generated. This solution is
evaluated by a fitness function that provides a
quantitative value of the solution’s utility.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 151
The velocity and direction of each particle moving
along each dimension of the problem space will be
altered with each generation of movement. In
combination, the particle’s personal experience, Pid and
its neighbors’ experience, Pgd influence the movement of
each particle through a problem space. The random
values rand1 and rand2 are used for the sake of
completeness, that is, to make sure that particles explore a
wide search space before converging around the optimal
solution. The values of c1 and c2 control the weight
balance of Pid and Pgd in deciding the particle’s next
movement velocity. At every generation, the particle’s
new location is computed by adding the particle’s current
velocity, vid, to its location, xid. Mathematically, given a
multi-dimensional problem space, the ith particle changes
its velocity and location according to the following
equations[1][2]:
vid
= c0
× vid
+ c1
× rand1
× ( pid
− xid
)
(1)
+ c2
× rand 2 × ( pgd
− xid
)
x id = xid
+ vid
(2)
where c0 denotes the inertia weight factor; pid is the
location of the particle that experiences the best fitness
value; Pgd is the location of the particles that experience a
global best fitness value; c1 and c2 are constants and are
known as acceleration coefficients; d denotes the
dimension of the problem space; rand1, rand2 are random
values in the range of (0, 1). For equation (1), the first
part represents the inertia of pervious velocity; the second
part is the “cognition” part, which represents the private
thinking by itself; the third part is the “social” part, which
represents the cooperation among the particles. If the sum
of accelerations would cause the velocity vid, on that
dimension to exceed vmax,d ,then vid, is limited to vmax,d.
vmax,d determines the resolution with which regions
between the present position and the target position are
searched. Weighted combination of three possible moves
is shown in Figure 1.
The PSO algorithm can be described as follows:
I) For each particle:
Initialize particle
II) Do:
a) For each particle:
1) Calculate fitness value
2) If the fitness value is better than the best
fitness value pbest in history
3) Set current value as the new pbest
End
b) For each particle:
1) Find in the particle neighborhood, the
particle with the best fitness gbest.
2)Calculate particle velocity according
to the velocity equation (1)
3) Apply the velocity constriction
4) Update particle position according to the
position equation (2)
5) Apply the position constriction
End
While maximum iterations or minimum error
criteria
© 2010 ACADEMY PUBLISHER
xk
is not attained
vk
pbestk
xk
+ 1
Figure 1. Weighted combination of three possible moves
III. PARAMETERS ANALYSIS OF PSO
gbestk
The basic PSO described above has a small number of
parameters that need to be fixed. One parameter is the size
of the population. This is often set empirically on the basis
of the dimensionality and perceived difficulty of a problem.
Values in the range 20–50 are quite common.
c in (1) determine the
The parameters 1
c and 2
magnitude of the random forces in the direction of personal
best pid and neighborhood best Pgd. These are often called
acceleration coefficients. The behavior of a PSO changes
radically with the value of c 1 and c 2 . Interestingly, we
can interpret the components c1 × rand1
× ( pid
− xid
)
and c2 × rand 2 × ( pgd
− xid
) in (1) as attractive forces
produced by springs of random stiffness, and we can
approximately interpret the motion of a particle as the
integration of Newton’s second law. In this interpretation,
c 1 / 2 and c 2 / 2 represent the mean stiffness of the
springs pulling a particle. It is no surprise then that by
changing c 1 and c 2 one can make the PSO more or less
“responsive” and possibly even unstable, with particle
speeds increasing without control. The value c 1 = c 2 =2,
almost ubiquitously adopted in early PSO research, did
just that. However, this is often harmful to the search and
needs to be controlled. The technique originally proposed
to do this was to bound velocities so that each component
of vid is kept within the range [−vmax,d,+ vmax,d]. The
choice of the parameter vmax,d required some care since it
appeared to influence the balance between exploration
and exploitation. The use of hard bounds on velocity,
however, presents some problems. The optimal value of
vmax,d is problem-specific, but no reasonable rule of
thumb is known. Further, when vmax,d x was implemented,
the particle’s trajectory failed to converge. Where one
would hope to shift from the large-scale steps that typify
exploratory search to the finer, focused search of
exploitation, vmax,d simply chopped off the particle’s
oscillations, so that some hopefully satisfactory
compromise was seen throughout the run.
Motivated by the desire to better control the scope of the
search, reduce the importance of vmax,d, and perhaps
eliminate it altogether, the following modification of the
PSO’s update equations was proposed (Shi and Eberhart):

152 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
The inertia weight c 0 can be interpreted as the
fluidity of the medium in which a particle moves. This
perhaps explains why researchers have found that the best
performance could be obtained by initially setting c 0 to
some relatively high value (e.g., 0.9), which corresponds
to a system where particles move in a low viscosity
medium and perform extensive exploration, and gradually
reducing c 0 to a much lower value (e.g., 0.4), where the
system would be more dissipative and exploitative and
would be better at homing into local optima. It is even
possible to start from values of c 0 > 1 , which would
make the swarm unstable, provided that the value is
reduced sufficiently to bring the swarm in a stable region
(the precise value of c 0 that guarantees stability depends
on the values of the acceleration coefficients).
Naturally, other strategies can be adopted to adjust the
inertia weight. For example, in (Eberhart and Shi 2000)
the adaptation of c 0 using a fuzzy system was reported
to significantly improve PSO performance. Another
effective strategy is to use an inertia weight with a
random component, rather than time-decreasing. For
example, (Eberhart and Shi 2001) successfully used
c 0 = U ( 0.
5,
1)
. There are also studies, e.g., (Zheng et
al. 2003)[10], in which an increasing inertia weight was
used obtaining good results.
Clerc and Kennedy (2002)[11] noted that there can be
many ways to implement the constriction coefficient. One
of the simplest methods of incorporating it is the
following:
vid
= χ(
vid
+ c1
× rand1
× ( pid
− xid
)
+ c2
× rand 2 × ( pgd
− xid
))
Where c = c1
+ c2
> 4 and
2
χ =
2
c − 2 + c − 4c
When Clerc’s constriction method is used, c is
commonly set to 4.1, c 1 = c2
, and the constant
multiplier χ is approximately 0.7298. These results in
the previous velocity being multiplied by 0.7298 and each
of the two pid − xid
terms being multiplied by a random
number limited by 0.7298×2.05≈1.49618.
The constricted particles will converge without
using any vmax,d at all. However, subsequent experiments
and applications (Eberhart and Shi 2000) [12] concluded
that a better approach to use as a prudent rule of thumb is
to limit vmax,d to xmax,d, the dynamic range of each variable
on each dimension. The result is a particle swarm
optimization algorithm with no problem-specific
parameters. And this is the canonical particle swarm
algorithm of today.
Yuhui Shi, Russell C(1998) [13]have analyzed the
impact of the inertia weight and maximum velocity
allowed on the performance of PSO. A number of
© 2010 ACADEMY PUBLISHER
experiments have been done with different inertia weights
and different values of maximum velocity allowed. It is
concluded that when vmax,d. is small, an inertia weight of
approximately 1 is a good choice, while when vmax,d.is not
small, an inertia weight c 0 = 0.
8 is a good choice.
When we lack knowledge regarding the selection of
vmax,d, it is also a good choice to set vmax,d equal to xmax,d
and an inertia weight c 0 = 0.
8 is a good starting point.
Furthermore if a time varying inertia weight is employed,
even better performance can be expected.
In (Kennedy, J. 1997) based on c 1 and c 2 , Kennedy
introduces four models of PSO, defined by omitting or
restricting components of the velocity formula.
(1) The complete formula above defines the Full
Model. That is 1 0 > c and 2 0 > c .
(2) Dropping the social component results in the
Cognition-Only Model. That is 1 0 > c and 2 0 = c .
(3) dropping the cognition component defines the
Social-Only Model. That is 1 0 = c and 2 0 > c .
(4) Selfless is the Social-Only Model, but the
neighborhood best is chosen only from the neighbors,
without considering the individual particle’s gbest vector.
That is 1 0 = c , 2 0 > c and g ≠ i.
Regarding the inertia weight, it determines how the
previous velocity of the particle influences the velocity in
the next iteration: If c 0 = 0 , the velocity of the particle
is only determined by the pbest and gbest positions. This
means that the particle may change its velocity instantly
if it is moving far from the best positions in its
knowledge. Thus, low inertia weights favor exploitation
(local search). If c 0 is high, the rate at which the particle
may change its velocity is lower (it has an "inertia" that
makes it follow its original path) even when better fitness
values are known. Thus, high inertia weights favor
exploration (global search).
In this paper, an analysis of the impact of this inertia
c and
weight c 0 , together with acceleration constants 1
c 2 on the performance of PSO is given, followed by
experiments that illustrate the analysis and provide some
insights into optimal selection of the. inertia weight c 0 ,
together with acceleration constants c 1 and c 2 .
The parameters of PSO includes: number of particles
m, inertia weight c 0 , acceleration constants c 1 and c 2 ,
maximum velocity vmax,d. This paper discuss inertia
c .
weight c 0 , acceleration constants c 1 and 2
IV. CONVERGENCE OF PSO
A. Convergence Analysis
In ( Gao S, Yang J Y,2006;Gao S et al.2006) [14,15],
the convergence of particle swarm algorithm is
studied,and the condition for the convergence of particle
swarm algorithm is given.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 153
From equation (1) and (2), We can see that although
and are multi-dimensional variables, but they are
independent each other. So we can simplify multidimensional
space into one dimension space. We suppose
that the swarm’s global best position and the best position
attained by particle self are unchanged, and they are
denoted as p b and g b .The equation(1)and(2)can
be simplified as
v(
k + 1)
= c0v(
k)
+ c1(
pb
− x(
k))
(3)
+ c2
( gb
− x(
k))
x ( k + 1)
= x(
k)
+ v(
k + 1)
(4)
From equation(3)and(4), we obtain
v(
k + 2)
= c0v(
k + 1)
+ c1(
pb
− x(
k + 1))
(5)
+ c2
( gb
− x(
k + 1))
x ( k + 2)
= x(
k + 1)
+ v(
k + 2)
(6)
The formula (4) and (5) are substituted into formula
(6), we can get
x(
k + 2)
+ ( −c0
+ c1
+ c2
−1)
x(
k + 1)
(7)
+ c0
x(
k)
= c1
pb
+ c2
g b
Equation (7) is a second order inhomogeneous
difference equation. We can use characteristic equation
method to solve this equation.
The characteristic equation of Equation (7) is
2
λ + ( −c0
+ c1
+ c2
−1)
λ + c0
= 0 (8)
2
(1) If ∆ = ( −c
0 + c1
+ c2
−1)
− 4c0
= 0 ,
then λ = λ1
= λ2
= −(
−c0
+ c1
+ c2
−1)
/ 2 .
So
k
x( k)
= ( A0
+ A1k
) λ
Where A 0 , A 1 are undetermined coefficients, and
they are determined by v( 0)
and x ( 0)
. Therefore,
⎧A0
= x(
0)
,
⎪
⎨ ( 1−
c)
x(
0)
+ c0v(
0)
+ c1
pb
+ c2
gb
⎪A1
=
− x(
0)
⎩
λ
2
(2) If ∆ = −c
+ c + c −1)
− 4c
> 0 , then
( 0 1 2
0
c0
− c1
− c2
+ 1±
λ 1,
2 =
2
∆
. So
( k)
= A
k k
+ A λ + A λ ,
x 0 1 1 2 2
Where A 0 , 1 A , 2
A are undetermined coefficients.
It is assumed that b1 = x(
0)
− A0
and
b ( 1−
c)
x(
0)
+ v(
0)
+ c p + c g − A , we
2
= c0 1 b 2 b 0
can get
© 2010 ACADEMY PUBLISHER
⎧ c1
pb
+ c2
gb
⎪A0
=
⎪ c
⎪ λ2b1
− b2
⎨A1
=
⎪ λ2
− λ1
⎪ b2
− λ1b1
⎪A2
=
⎩ λ2
− λ1
(3) If ∆ = −c
+ c + c
2
−1)
− 4c
< 0 , then
( 0 1 2
0
c0
− c1
− c2
+ 1±
i
λ 1,
2 =
2
− ∆
. So
( k)
= A
k k
+ A λ + A λ
x 0 1 1 2 2
Where A 0 , 1 A , 2
A are undetermined coefficients.
⎧ c1
pb
+ c2
gb
⎪A0
=
⎪ c
⎪ λ2b1
− b2
⎨A1
=
⎪ λ2
− λ1
⎪ b2
− λ1b1
⎪A2
=
⎩ λ2
− λ1
The convergence condition is : λ 1 and
λ 1.
2 <
1 <
If ∆ = 0 , the convergence zone is
2 2
c + c − c c − 2c
− 2c
+ 1 = 0 and 0 ≤ c 0 < 1
0
2 0 0
c = c + c .
If ∆ > 0 , the convergence zone is
2 2
c + c − c c − 2c
− 2c
+ 1 > 0 , c > 0 and
,where 1 2
0
2 0 0
2c 0 − c + 2 > 0 .
If ∆ < 0 , the convergence zone is
2 2
c + c − c c − 2c
− 2c
+ 1 < 0 and c 1 .
0
2 0 0
0 <
So the convergence zone is c 0 < 1 , c > 0 and
c − c + 2 > 0 (shown in Figure 2).
2 0
Figure 2 The convergence zone of particle swarm algorithm

154 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
B. Numerical Simulation
Numerical simulations are implemented for six
different types. It is assumed that x ( 0)
= 2 , v ( 0)
= 1,
and c p c g = 0 .
1
b + 2 b
(1) If c 0 =0.81 and c =3.61, then ∆ = 0 and
k
x ( k)
= ( 2 + 2.
9k)(
−0.
9)
. x (k)
is convergent. It is
illustrated in Figure 3.
(2) If c 0 =4 and c =1, then ∆ = 0 and
+ 1
( ) = 2
k
x k . (k)
Figure 4.
x is divergent. It is illustrated in
Figure 3 Relationship between particle location and iterative times with
c 0 =0.81 and c =3.61
Figure 4 Relationship between particle location and iterative times with
c 0 =4 and c =1
(3) If c 0 =0 and c =0.5, then ∆ > 0 and
−k
x k = 1
( ) 2
Figure 5.
. x (k)
is convergent. It is illustrated in
(4) If c 0 =0 and c =3, then ∆ > 0 and
x ( k)
Figure 6.
− )
k
= 2 ⋅ ( 2 . (k)
© 2010 ACADEMY PUBLISHER
x is divergent. It is illustrated in
Figure 5 Relationship between particle location and iterative times with
c 0 =0 and c =0.5
Figure 6 Relationship between particle location and iterative times with
c 0 =0 and c =3
(5) If c 0 =0.5 and c =0.5, then ∆ < 0 and
k
kπ
kπ
2 x(
k)
= 2 ( 2cos
+ sin )
4 4
−
. x (k)
is
convergent. It is illustrated in Figure 7.
(6) If c 0 =2 and c =5, then ∆ < 0 and
k
1
2
x(
k)
= 2
+
3 3
(cos kπ
− 2sin
kπ
)
4 4
divergent. It is illustrated in Figure 8.
. x (k)
is

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 155
Figure 7 Relationship between particle location and iterative times with
c 0 =0.5 and c =0.5
Figure 8 Relationship between particle location and iterative times with
c 0 =2 and c =5
V. MULTIPLICATE PARTICLE SWARM
OPTIMIZATION (MPSO) ALGORITHM
The first particle swarms (Kennedy and Eberhart
1995)[2] evolved out of bird-flocking simulations of a
type described by (Reynolds 1987)[14] and (Heppner and
Grenander 1990)[15]. In these models, the trajectory of
each bird’s flight is modified by application of several
rules, including some that take into account the birds that
are nearby in physical space. So, early PSO topologies
were based on proximity in the search space. However,
besides being computationally intensive, this kind of
communication structure had undesirable convergence
properties. Therefore, this Euclidean neighborhood was
soon abandoned.
In fact, the real lives work not only specially but also
cooperatively. In this paper, different particles are
assigned specific tasks. The particles use five different
hybrid flight rules in accordance with section probability.
© 2010 ACADEMY PUBLISHER
This algorithm can draw on each other ' s merits and raise
the level together The method uses not only local
information but also global information and combines the
local search with the global search to improve its
convergence.
The first flight rule: c 0 , c 1 and c 2 are not equal to
0, and they are restricted in the convergence zone (Figure
2).In this paper, we set c 0 = 1,
1 2 = c , 2 2 = c .
The second flight rule : c 0 = 0 , 1 c and c 2 are
restricted in the convergence zone (fig 2). In this paper,
we set c 0 = 0 , 1 0.
5 = c , 5 . 0 2 = c .
The third flight rule : 1 0 = c , c0 and c 2 are
restricted in the convergence zone (fig 2). In this paper,
we set c 0 = 1,
1 0 = c , 2 2 = c .
The fourth flight rule: 2 0 = c , c 0 and c 1 are
restricted in the convergence zone (fig 2). In this paper,
we set c 0 = 1,
1 2 = c , 2 0 = c .
The fifth flight rule: c 0 , c 1 and c 2 are not
restricted in the convergence zone (fig 2). The alternate
will emanate, but the algorithm can escape the local
minimum. In this paper, we set c 0 = 2 , 1 2 = c , 2 2 = c .
The five flight rules are incomplete, but they are
feasible in somewhere. The particles use five different
hybrid flight rules in accordance with section probability.
The simplest selection scheme is roulette-wheel selection,
also called stochastic sampling with replacement. As an
example, consider 5 flight rules select values (0.6, 0.1,
0.1, 0.1,0.1). The spin of the wheel results in a random
number r ,
If 0 < r ≤ 0.
6 ,the first flight rule is chosen,
else if 0. 6 < r ≤ 0.
7 , the second flight rule is
chosen,
else if 0. 7 < r ≤ 0.
8 , the third flight rule is chosen,
else if 0. 8 < r ≤ 0.
9 , the fourth flight rule is
chosen,
else if 0 . 9 < r < 1,
the fifth flight rule is chosen.
For comparison, five benchmark functions that are
commonly used in the evolutionary computation
literature are used. All functions have same minimum
value, which are equal to zero.
F 1
2
= x1
+ x , −1 ≤ x i ≤ 1
F = 100( x
2
2
− x ) + ( 1−
x ) , 2 ≤ x ≤ 2
2
2
2
2
1
2
1
− i
2 2 2
sin x1
+ x2
− 0.
5
F 3 =
+ 0.
5
2 2 2
[ 1+
0.
001(
x1
+ x2
)]
,
1 ≤ x ≤ 1
F4 2 2
= x1
+ 2x2 − 0.
3cos3πx
1 − 0.
4cosπx2
+ 0.
7
, −1 ≤ x i ≤ 1
F
2 2
= x + x , 1 ≤ x ≤ 1
− i
5
1
2
− i

156 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
We write down the alternate times, as the error
between the minimum value and result is 0.00001. The 5
flight rules select values are set 4 combinations. They are
0.6:0.1:0.1:0.1:0.1, 0.4:0.2:0.2:0.1:0.1,
0.3:0.2:0.2:0.2:0.1, and 0.2:0.2:0.2:0.2:0.2. 50 rounds of
TABLE I. COMPARISON OF ALGORITHMS
computer simulation are conducted for each algorithm,
and the results are shown in Table 1.
All the MPSO algorithms are proved effective.
Especially the algorithm with selection probability (
0.3:0.2:0.2:0.2:0.1 ) is a simple and effective better
algorithm than others.
Functions Algorithms Best Average Worst
F 1
F 2
F 3
F 4
F 5
PSO 5 34.9 107
MPSO (0.6:0.1:0.1:0.1:0.1) 4 16.5 30
MPSO (0.4:0.2:0.2:0.1:0.1) 5 12.06 26
MPSO (0.3:0.2:0.2:0.2:0.1) 4 10.88 18
MPSO (0.2:0.2:0.2:0.2:0.2) 4 12.4 28
PSO 9 129.6 417
MPSO (0.6:0.1:0.1:0.1:0.1) 14 36.68 97
MPSO (0.4:0.2:0.2:0.1:0.1) 12 21.64 41
MPSO (0.3:0.2:0.2:0.2:0.1) 13 21.74 37
MPSO (0.2:0.2:0.2:0.2:0.2) 13 26.08 38
PSO 5 30.9 84
MPSO (0.6:0.1:0.1:0.1:0.1) 2 14.5 34
MPSO (0.4:0.2:0.2:0.1:0.1) 4 12.54 21
MPSO (0.3:0.2:0.2:0.2:0.1) 3 11.36 19
MPSO (0.2:0.2:0.2:0.2:0.2) 5 12.76 24
PSO 7 85.76 272
MPSO (0.6:0.1:0.1:0.1:0.1) 11 28.36 72
MPSO (0.4:0.2:0.2:0.1:0.1) 8 17.16 31
MPSO (0.3:0.2:0.2:0.2:0.1) 6 17.88 26
MPSO (0.2:0.2:0.2:0.2:0.2) 10 18.52 37
PSO 4 33.6 96
MPSO (0.6:0.1:0.1:0.1:0.1) 3 15.5 27
MPSO (0.4:0.2:0.2:0.1:0.1) 4 11.52 24
MPSO (0.3:0.2:0.2:0.2:0.1) 3 10.23 16
MPSO (0.2:0.2:0.2:0.2:0.2) 3 11.4 26
VI CONCLUSIONS
Base on the idea of specialization and cooperation of
particle swarm optimization algorithm, a multiplicate
particle swarm optimization algorithm is proposed. The
efficiency of the new algorithm is verified by the
simulation results of five classical test functions and the
comparison with other algorithms. Future work will also
address additional issues related to the use of PSO in
high-dimensional spaces, including the selection of
swarm size, the parameters of PSO. Even though good
experimental results have been obtained in this paper,
only a small benchmark problem has been tested. The
selection probability may be problem-dependent. To fully
justify the benefits of selection probability as described in
this paper, more problems need to be tested. By doing so,
a clearer understanding of PSO performance will be
obtained.
© 2010 ACADEMY PUBLISHER
ACKNOWLEDGMENT
This work was partially supported by National Basic
Research Program of Jiangsu Province University
(08KJB520003), Qing Lan Project and the Open Project
Program of the State Key Lab of CAD&CG.
REFERENCES
[1] R. C. Eberhart and J. Kennedy, “A new optimizer using
particles swarm theory”, Proceedings Sixth International
Symposium on Micro Machine and Human Science,
Nagoya, Japan, 1995,pp.39-43.
[2] J. Kennedy and R. C. Eberhart, “Particle swarm
optimization”, Proceedings IEEE International Conference
on Neural Networks. Perth, 1995, pp.1942-1948.
[3] Y. H. Shi and R. C.Eberhart, “A modified particle swarm
optimizer”, IEEE International Conference on
Evolutionary Computation. Anchorage, Alaska, May 4-9,
1998, pp.69-73.
[4] J. Kennedy, “The Particle Swarm: Social Adaptation of
Knowledge”, Proceedings of the 1997 International
Conference on Evolutionary Computation, IEEE Press,
1997,pp. 303-308.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 157
[5] P. C. Fourie and A. A.Groenwold, “The Particle Swarm
Optimization Algorithm in Size and Shape Optimization”,
Structural and Multidisciplinary Optimization, vol.
23,no.4, pp. 259–267,2002.
[6] P. C. Fourie and A. A. Groenwold, “Particle Swarms in
Topology Optimization”, Proceedings of the Fourth World
Congress of Structural and Multidisciplinary Optimization,
Dalian, China, 2001.
[7] G. Venter and J. Sobieszczanski-Sobieski, “Particle
Swarm Optimization”, AIAA Journal, 2003, vol. 41,no.8,
pp. 1583–1589.
[8] G. Venter and J. Sobieszczanski-Sobieski,
“Multidisciplinary Optimization of a Transport Aircraft
Wing using Particle Swarm Optimization”, Structural and
Multidisciplinary Optimization, 2004, vol. 26, no.1-2, pp.
121–131.
[9] J. F. Schutte, B. J. Fregly, R. T. Haftka, and A. George, “A
Parallel Particle Swarm Algorithm”. Proceedings of the
Fifth World Congress of Structural and Multidisciplinary
Optimization, Venice, Italy, 2003.
[10] Y.L. Zheng, L.H. Ma, L.Y. Zhang, and J.X. Qian,. “On the
convergence analysis and parameter selection in particle
swarm optimization”. Proceedings of the IEEE
international conference on machinelearning and
cybernetics. Piscataway: IEEE 2003,pp. 1802–1807
[11] M. Clerc., and J. Kennedy, The particle swarm—explosion,
stability, and convergence in a multidimensionalcomplex
space. IEEE Transaction on Evolutionary Computation,
2002,Vol.6, No.1, pp.58–73.
[12] R. C. Eberhart, and Y. Shi, “Comparing inertia weights
and constriction factors in particle swarm optimization”.
Proceedings of the IEEE congress on evolutionary
computation (CEC), San Diego, CA. Piscataway: IEEE,
2000, pp. 84–88.
[13] Y. H. Shi and R. C. Eberhart, “Parameter Selection in
Particle Swarm Optimization”, Proceedings of the 7th
International Conference on Evolutionary Programming
VII,1998,pp.591-600.
[14] C. W. Reynolds, “Flocks, herds, and schools: a distributed
behavioral model”, Computer Graphics, 1987, Vol.21,
No.4, pp.25–34.
© 2010 ACADEMY PUBLISHER
[15] H. Heppner and U. Grenander, “A stochastic non-linear
model for coordinated bird flocks” In S. Krasner (Ed.), The
ubiquity of chaos. Washington: AAAS. ,1990, pp. 233–238.
[16] S. Gao and J. Y. Yang, Swarm Intelligence Algorithm and
Applications. Beijing: China Water Power Press, 2006,
pp.112-117(in Chinese).
[17] S. Gao, K. Z. Tang, X. Z. Jiang, and J. Y. Yang,
“Convergence Analysis of Particle Swarm Optimization
Algorithm”, Science Technology and Engineering
,2006,vol.6, no.12, pp.1625-1627,1631(in Chinese).
Shang Gao was born in 1972, and received his M.S. degree
in 1996 and Ph.D degree in 2006. He now works in school of
computer science and technology, Jiangsu University of Science
and Technology. He is an associate professor and He is engage
mainly in systems engineering.
Zaiyue Zhang was born in 1961, and received his M.S.
degree in mathematics in 1991 from the Department of
Mathematics, Yangzhou Teaching College, and the Ph.D.
degree in mathematics in 1995 from the Institute of Software,
the Chinese Academy of Sciences. Now he is a professor of the
School of Computer Science and Technology, Jiangsu
University of Science and Technology. His current research
areas are recursion theory, knowledge representation and
knowledge reasoning.
Cungen Cao was born in 1964, and received his M.S. degree
in 1989 and Ph.D. degree in 1993 both in mathematics from the
Institute of Mathematics, the Chinese Academy of Sciences.
Now he is a professor of the Institute of Computing Technology,
the Chinese Academy of Sciences. His research area is large
scale knowledge processing.

158 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
Research and Design of Intelligent Electric Power
Quality Detection System Based on VI
Yu Chen
Zhengzhou Institute of Aeronautical Industry Management, Zhengzhou, China
Email: chenyu3440@gmail.com
Abstract—Electric power quality problem has become an
important problem in modern society, which effects
industrial production and product quality, etc. So, we
urgently need monitor and analyze electric power
parameter to solve electric power quality problem. In this
paper, we design a kind of electric power parameter
detection system based on visual instrument (VI) technology,
etc. System uses voltage/currency transformer module and
signal processor module to realize electric power network
analog signal conversion. System adopts multiplexing
technology, which solves multi data acquisition card (DAQ)
using complexity problem as well as saving cost. In addition,
system software adopts method of wavelet resolution
analysis combining with FFT, which solving FFT
disadvantage in analyzing non-steady distortion signal.
System is applied to harmonic monitor, voltage fluctuation
and flicker monitor, etc. This paper mainly introduces
system’s structure, algorithm simulation, part of hardware
and software design in detail. Through experiment, monitor
error of harmonic wave’s THD% is 0.01%, and amplitude
wave’s frequency and amplitude of voltage flicker is
respectively less than 0.01Hz and 0.005V, which obtains
accuracy result in high monitor precision as expectation.
Index Terms—visual instrument, DAQ, FFT, harmonic
monitor, voltage fluctuation and flicker, wavelet analysis,
Matlab+LabVIEW
I. INTRODUCTION
With the development of power electric system,
especially introduced of power electronic equipment,
power system harmonic harm is serious increasing. On
production of industrial and mining enterprises, etc,
power quality is necessary. To assure equipment reliable
operation, we need detect voltage, currency, and their
state of power network, etc in real time. Traditional
detection equipment depends artificial to complete, but
there are many more problems: one is field circumstance
is poor, especially, under middle or high voltage
condition, which is not suitable to work in long term; the
other is that artificial style has larger error, and data static,
processor both require amount of time and work. In
addition, present detection equipment can not completely
detach field. So, design a kind of convenient and simple
detection style has very important significance. With the
increasing of electric component precision, speed, and
reliability, it provides base to realize high character
method and in real-time control. [1]
This system adopts technologies of data acquisition
multiplexing, VI, PC, and new algorithm of wavelet
© 2010 ACADEMY PUBLISHER
doi:10.4304/jcp.5.1.158-165
multi-resolution, etc, which an intelligent power network
detection system is carried out.
II. SYSTEM STRUCTURE AND WORK PRINCIPLE
Intelligent electric power quality detection system is
mainly composed of two parts, which is PC system
installing VI software LabVIEW8.5+Matlab and data
acquisition system. Because PC system input signal is
small voltage signal after processed by transformer and
process module transforming from three phase
voltage/currency. According to waiting signal detection
method demand, when monitor voltage phase, power
factor, etc, which needs adopt synchronization collection
style, it is to say, we need synchronization sample to two
phase voltage or voltage-currency. If we make six groups
signal amplify, filter and transmit to data acquisition
module, which needs data acquisition module at least has
six groups synchronization sampling channel. But normal
synchronization DAQ only has four groups
synchronization sampling channel, which needs two card
and increases synchronization controlling problem
between two cards, and brings cost and technology
complex problem by this kind of style. So, system adopts
multiplexing technology to solve this problem. Overall
structure of system design is shown as Fig. 1.
Signal process module
Data acquisition card
PC
Sensor monitor signal
Figure 1. Overall structure of system design
III. SYSTEM HARDWARE DESIGN
VI software
(LabVIEW8.5)
Six groups of electric power parameter of electric
power system two channel groups output through double
channel 16 groups/double 8 group analog switch AD7507,
and then it is amplified, low-passing filtered and
transmitted to DAQ. So, we only need one card to fulfill
system signal collection requirement, which not only
reduces cost, but also simplifies technology schema. To
increase monitor system generality, we design multirange
automatic control circuit, which makes system
automatically range shifting to meet monitored value
changing according DAQ A/D transfer’s factors of
voltage change limit value, monitor precision and
resolution, etc.

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 159
System completes field electric power system detection
in real time, and processes field collection data through
PC Computer adopts normal industry control computer.
System PC can realize displaying and alarming through
analyzing and processing receiving data. Because paper
coverage is limit, so we introduce some important parts.
A. Transformer Circuit
System adopts voltage transformer SPT204A and other
circuit to consist voltage signal collection module.
SPT204A is a precision current transformer of mA level
in fact. Because of size technology it usually adopts
voltage transformer of current type. And its rated
transformation ratio is 1, input rated current is 2mA, input
voltage can’t add to primary coil in use, we need to serial
into limit currency resistance, which makes voltage signal
transform into currency signal of mA level. Secondary
output mA level currency signal is transformed into
voltage signal by amplifier I/V transferring. Amplifier
uses good character chip LM118, which can easy to reach
higher precision and stability. Transformer works at
approximate zero load state, which has many advantages
such as broad dynamic range, high precision, and good
linearity, etc. Voltage signal collection circuit is shown as
Fig. 2.
Figure 2. Voltage transformer circuit
System uses currency transformer STC254AK and
other circuit to compose currency transformer module.
Circuit principle is basic same as voltage transformer
without introducing. In addition, in system designing, we
should notice transformer has homonymy and synonym
end, input signal should synonym end connects LM118
reversed phase end, which can assure amplifier signal
same with detection signal. Otherwise it is reversed phase.
[2]
B. Multi-channel Selection Switcht
System adopts America AD .co production 16
channel/double 8 channel, low leakage, CMOS analog
multiplexer AD7507, which makes sensor module output
in two channel voltage signal, and then realize subsequent
process of amplify and filter. AD7507 logic true value
table is shown as Table I.
TABLE I.
LOGIC TRUE VALUE TABLE
A2 A1 A0 EN On-off switch
× × × 0 NONE
0 0 0 1 1&9
0 0 1 1 2&10
0 1 0 1 3&11
0 1 1 1 4&12
1 0 0 1 5&13
1 0 1 1 6&14
1 1 0 1 7&15
1 1 1 1 8&16
© 2010 ACADEMY PUBLISHER
From Table I we can see, each group of chip outputs
corresponding one group of input, it can control Ua, Ub,
Uc, Ia, Ib and Ic signal in double channel acquisition by
three channel selection ports and a enable port, so, we
need four control signal of D0, D1, D2 and D3 form
DAQ. Circuit principle is shown as Fig. 3. [3]
Figure 3. Multi-channel selection switch circuit
C. DAQ Selectiont
Because this projection monitor task is simpler without
multi-module trigger problem, monitor signal is low
frequency, and has more requirements of EMI, etc. So,
thinking about hardware cost and realization complexity,
we adopt data acquisition module based on PCI bus, and
plug DAQ directly into PC PCI bus, which realizes field
data collected and processed as controlled object by PC.
This system adopts ADLINK .co DAQ-2010 4 channel
synchronization DAQ, which has 14 bit A/D resolution;
maximum sampling frequency can reach 2MHz.
According GB of electric network quality, A level
instrument frequency monitor range is 0~2500Hz, which
is 50 times of sampling maximum time harmonic signal’s
frequency. From Nyquist sampling method we can see, if
DAQ’s maximum sampling frequency is more than 5 kHz,
it can reach system requirement. This DAQ-2010 can
completely meet electric power quality Ath level monitor
requirement of three phase voltage, currency.
In addition, ADLINK .co not only provides DAQ-2010
driving program, which can assure DAQ to realize all
function well in LabVIEW platform, but also provides
part of VI function, which makes user program by calling
these functions in LabVIEW circumstance.
IV. SYSTEM SOFTWARE DESIGN
Software design of system adopts modularization
program structure. PC program control software uses
LabVIEW+Matlab, and adopts LabSQL[4] connecting
with database to design. Software system adopts modular
design, which divides into some relative independent
function module, and arranges suitable enter and exit port
parameters. We mainly introduce electric power quality
monitor method of harmonic, voltage fluctuation and
flicker, etc. In harmonic monitor, we use FFT combining
wavelet multi-resolution monitor method, which can
realize harmonic monitor when signal has mutation. In
voltage fluctuation and flicker monitor, we use
synchronization square monitor algorithm with multiresolution
monitor method, which can realize analysis
without low-passing filter. We mainly introduce part of
method and theory in use.

160 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
A. LabSQL Connecting With Database
This system selects LabSQL connects with database,
principle of LabSQL connecting with database is shown
as Fig. 4. Using LabSQL can realize data transmission
between application program and database.
LabVIEW
LabSQL VIs
DSN (Data source name)
ODBC Driver
Database
Figure 4. Frame diagram flow of LabSQL connecting with database
To avoid user error operation because no using
authority, we design login and check functions module,
etc. Login program uses LabVIEW to design. User
information includes user name, password, authority,
login times, and last time login time, etc. Database “用
户.mdb” use Microsoft Access to build. Among them, we
design a “用户” table to storage user information, which
design structure is shown as Table II.
TABLE II.
DATABASE USER TABLE
B. FFT
At present, harmonic monitor method is DFT and FFT.
DFT computational complexity is proportional to square
of transform length N. When harmonic analysis, in order
to assure calculation precision, N is a big number value,
so, computational complexity is greatly. FFT is fast speed
algorithm of DFT, which can improve 1~2 magnitude in
DFT operation efficient. So, it has broad application in
harmonic analysis. [5] Its way is shown as below: at first,
calculate basic wave, real and imaginary parts of each
harmonic, then calculate its amplitude and phase, and
each harmonic containing amount and harmonic total
amount. As voltage value for example, if we use FFT
algorithm to get real and imaginary of K times harmonic
(k=2, 3…40) is Ur( k ) and U ( k ) , then K times harmonic
i
voltage amplitude U( k ) is shown as formula (1):
2 2
U( k) = Ur ( k) + Ui ( k)
(1)
Phase of K times’ harmonic voltage is as (2):
U(k)
(2)
r
θ( k)= arctg
Ui( k)
Correspondingly, we can get harmonic containing
amount by using some times harmonic amplitude
© 2010 ACADEMY PUBLISHER
corresponding to basic amplitude percentage. K times’
harmonic containing amount is shown as (3): [6]
U( k)
HRU ( k)
= × 100%
(3)
U (1)
Total harmonic distortion (THD) is total containing
amount reflecting harmonic, which can calculate by
below formula is shown as formula (4):
40
2
∑ U k
(4)
k=2
THD= × 100%
U1
Though FFT has character of fast speed, high precision,
etc, this method has two sides of error, it is leakage error
when sample period is asynchrony with signal period, and
frequency aliasing error as sample frequency is less than
biggest signal frequency twice. Leakage error has found
some ways to solve, such as adding window technology,
and interpolation technology, etc. But, at aliasing error
aspect, normally, we adopt adding low-passing filter
behind sampling, which can filter higher of Nyquist
frequency component. But, FFT on harmonic monitor is
suggested wave is stable and period. In fact, harmonic
signal would have catastrophe point, integral action of
Fourier transform smoothing unstable signal catastrophe,
which can make Fourier algorithm effect is not ideal at
unstable harmonic monitor, such as signal catastrophe,
transient distortion, short time harmonic, etc.
C. Multi-resolution Wavelet Analysis Algorithm
When sampling frequency of signal that waiting to be
analyzed is satisfied with the sampling theory of Nyquist,
normalization frequency band is restrained between -π
and +π. So we can use parts of low frequency 0~π/2 and
high frequency 2/π~π. They are decomposed by ideal
low-pass filter H0 and high-pass filter H1, which is
corresponding to signal general picture and detail. The
two groups of output must be orthogonal because the
frequency band cannot associate with, and that the width
of frequency band is halved, so the information would not
lose if we use sample rate in half. Sample rate of bandpass
signal is decided by its width of band but its upper
limit of frequency. The analogous process can be carried
out to every part of low frequency signal after
decomposing. That is to say, output signal is decomposed
into a general approach part of low frequency and a detail
of high frequency part in each level decomposing, and
output sample rate of each level can divide into half. So
the original signal s(n) is decomposed in multi-resolution.
a ( k )
1
x( n )
2 ( ) a k d2 ( k )
3 ( ) a k d3 ( k )
d ( k )
Figure 5. Multi-resolution wavelet decomposition
Fig. 5 takes on an operation as: x(n)=a3(k)+ d3(k)+
d2(k)+d1(k). S is original analysis signal. d1(k), a1(k),
d2(k), a2(k), d3(k) and a3(k) are high and low frequency
1

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 161
of each layer’s decomposition. If we want decompose any
more, the method is same. The ultimately intention is that
construct an orthogonal wavelet base in a frequency band,
which highly approach in 2 L ( R ) dimensional. These
orthogonal wavelet bases in different frequency resources
are equal as band-pass filter of different bandwidth. The
low frequency signal is decomposed constantly in multiresolution
analysis, which makes frequency resolution
become higher and higher.
Based on the theory of multi-resolution analysis,
Mallat presents a fast algorithm of wavelet decomposition
and reconstruct—Mallat. [7] The initial value is show as
below:
j − 1
j j −1
j
c n = ∑ a i − 2 n c d i n = ∑ bi−2nci i
i .
N M
Output is c −
j
and d ( N − M ≤ j ≤ n)
.
N is signal decomposition level of analysis signal f(x),
j
f n is discrete sample value of analysis signal f(x),
c n and
j
d n are modulus of low frequency and high frequency. n a
and n b are decomposition list of low frequency and high
frequency, which is decided by scale function Φ(x) and
wavelet function Ψ(x). The detail of each signal
frequency band can be observed through focusing signal
scale based on multi-resolution Mallat arithmetic. [8]
D. FFT Combining Wavelet Transform
Ideal harmonic analysis algorithm should be not only
getting each times stable harmonic accuracy frequency,
amplitude, and phase but also monitor small change, such
as transient harmonic, and transient catastrophe. So,
single to use Fourier algorithm or wavelet transform
algorithm also make perfect harmonic analysis, so, this
paper combines the two algorithm to realize harmonic
monitor and analysis.
Before we adopt FFT to carry out harmonic analysis,
we firstly monitor if unstable interference signal has or
not, if it has, we will filter it at first. Because Fourier
transform has limited in time domain, so, it can not
analyze unstable distorted wave. But, wavelet transform
has good time domain locality, which can process
transient signal and make up Fourier transform
shortcoming. As well as provide better basic to FFT.
When monitor catastrophe point by using mode
maximum and singularity detection method. Harmonic
monitor principle is shown as Fig. 6. [9]
Basic wave, each times harmonic, THD and spectrum
Sample signal after wavelet
de-noising and filter
Multiresolution
wavelet
transform
Figure 6. Flow diagram of combining wavelet transform with FFT
From upon diagram we can see, this kind of harmonic
monitor and analysis algorithm adds multi-resolution
wavelet transform algorithm before FFT algorithm, which
© 2010 ACADEMY PUBLISHER
Stable
component
Subtract
FFT
analysis
Transient
component
is the key of this method and its function is: ①analyzing
signal transient component; ② correction-compensation
initial signal, which reduces FFT error because of
transient component, and improve spectrum analysis
precision.
E. Voltage Flicker Synchronization Square Algorithm
Voltage flicker of electric net voltage can be seen that
a carrier wave, which composed of power frequency
voltage in sinusoidal, is changed amplitude modulation
by a frequency 0.5Hz~35Hz signal. Electric net voltage
changing usually causes voltage flicker. So electric net
voltage fluctuation and flicker signal u(t) can use be
expressed as formula (5):
ut ( ) = V[1 + msin( Ω t)]sin( wt)
m
(5)
In formula (3-3), Vm is voltage amplitude of power
frequency, w0 is angle frequency of power frequency
carrier wave, m is voltage amplitude of amplitude
modulation wave/ carrier wave voltage (modulation
parameter), and m ≤ 1,
mcos(Ωt) is flicker voltage, and Ω
is angle frequency of amplitude modulation voltage.
The envelope signal of voltage flicker msin(Ωt)
includes its amplitude and frequency information. The
signal will be decomposed into a series of sinusoidal
signals’ sum by using wavelet multi-resolution signal
decomposition method, which constituted by groups of
sub-band filters, and each component compares to a kind
of corresponding frequency sinusoidal signal. Wavelet
analysis approaches original signal in different resolution,
and projection component has a certain bandwidth with
time-domain resolution. Through replacing traditional
low-pass filter of synchronization radio-detector by subband
filter, we cannot only measure envelope signal of
voltage fluctuation with non-distorted, but also precisely
monitor the start time and amplitude of voltage flicker.
This paper introduces a method with adopting IEC
commendatory of square monitor [10] and combining
wavelet analysis in allusion to the stable or unstable
voltage flicker signal. After experiment, this method is a
better way. Its measurement principle diagram is shown
as Fig. 7.
Sampling
2
u () t
Square
ut ()
Wavelet
decomposition
Envelope
signal
Figure 7. Wavelet transforms measurement
2
u () t High frequency
detail parameter
2
u () t Low frequency
approach parameter
Wavelet reconstruction
According to square measurement method, we can gain
envelope signal of u(t) from 2
u () t low frequency
approach parameter when make modulation wave voltage
to square based on wavelet reconstruction. Formula
derivation process is shown as below formula (6):
2 2 2 2 2
u ( t) = A (1+ 2msin Ω t+ m sin Ω t)sin ωt
0

162 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
2 2 2 2
2
2 2
A
=
2
m A m A
(1 + ) + A msin Ωt− (1 + )cos 2ωt− 2 2 2 4
m cos 2Ωt
2 2
A 2
A
− m sin(2 ω+Ω ) t+ msin(2 ω−Ω)
t
2 2
2 2
A 2 A 2
+ m cos 2( ω+Ω ) t+ m cos 2( ω−Ω)
t
8 8
(6)
From formula (6) we can see, modulation wave voltage
quadratic term has not only DC component, but also
2( ω ± Ω ) ,
frequency component, such as Ω , 2Ω , 2ω , 0 0
2ω 0 ±Ω. We analyze time-frequency of 2 u () t by using
Mallat wavelet decomposition arithmetic. [11] Its process
is corresponding to a decomposition process of a series of
sub-band. Considering actual modulation index m is far
from 1, multiple component of amplitude modulation
wave voltage is far from amplitude of amplitude
modulation wave, so we can neglect it. Thereby, we can
realize demodulation after wavelet transforming, and
obtain near weighting amplitude modulation voltage ' u() t,
which is shown as formula (7)
' 2
u () t ≈ mV sinΩ
t
(7)
m
V. SIMULATION ANALYSIS
A. Mutation Point Analysis and Compensation
In order to verify this algorithm accuracy, we suggest
o
signal: x( t) = sin100π t+ 0.3sin(200πt+ 15 ) + 0.15sin(400πt+
o o o
30 ) + 0.1sin(600πt+ 30 ) + 0.01sin(1200πt+ 60 ) + pt () cos(
3200πt). From signal, signal basic wave frequency is
50Hz, amplitude is 1, and containing amplitude of 0.3,
frequency of 100Hz 2 times harmonic; amplitude of 0.15,
frequency of 200Hz 4 times harmonic; amplitude of 0.1,
frequency of 400Hz 6 times harmonic; and amplitude is
0.01, frequency of 600Hz 12 times harmonic component..
In addition, structure a mutation signal p(t), signal
produce a unstable cosine shrinkage pulse signal at
0.015s, constant width is 0.000625s, p(t) signal is
structured through segment function. Suggesting we
sample signal as frequency 6.4 kHz, so pulse is between
96 and 100 sample pot in theory. Through Matlab
simulation to initial signal, Initial signal simulation wave
is shown as Fig. 8.
Figure 8. Initial signal simulation diagram.
According to signal x(t), its basic frequency f is 50Hz,
sample frequency f is 6.4kHz, from formula (3-5), we
s
can get N=5. So, after making 5 layers wavelet multiresolution
decomposition, we can get a5, which is basic
wave component range. Each times’ harmonic
respectively is distributed at high frequency band d1 and
d2. Frequency band division is shown as Fig. 9.
© 2010 ACADEMY PUBLISHER
0~3.2
kHz
1.6~3.2
kHz
d1
0~1.6
kHz
0.8~1.6
kHz
d2
0~800
Hz
400~80
0Hz
0~400
Hz
200~40
0Hz
0~200
Hz
Figure 9. Dividing of signal y frequency band.
100~20
0Hz
According to multi-resolution analysis layer, we make
signal x(t) db24 wavelet decomposition, and extract this
two layers high frequency detail d1, d2 and its neighbor
d3. According to d1 (or d2) confirming transient
disturbance continue time segment, without disturbance,
d1 value is minimal, 0.00000 or 0.00001, and we make
mode maximum value [13] monitor to judge if there is
mutation and confirm disturbance time, limited value is
mode average 10 times, and we can obtain disturbance
appear between 96 and 100 point, and its continue time
∆T we can get by ways as:
1
∆ T = × (100 − 96) = 0.000625s.
6400
From calculation, this transient signal time is shortest,
(∆T

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 163
B. Harmonic Analysis
Because wavelet analysis is base on multi-resolution
signal processing theory, different scale corresponds with
different time and frequency resolution. That is to say,
different frequency component of original signal can be
separated through wavelet decomposition to realize
detection and tracking of the harmonic. In electric power
system, original signal fundamental wave frequency is
50Hz and effectively decomposition to it can obtain subband
( 0 /2 j
∼ f ) whose frequency range contains 50Hz.
s
Therefore, fundamental wave can be separated from
various harmonic. In power system, harmonic’s order,
produced actually by the voltage or current of network,
mostly may be 2K+1. In given simulation algorithm,
original current signal construct harmonic of 1, 3, 5, 7
orders and expression is: x(t)=sin(wt)+sin(3wt+3)/3+sin
(5wt+5)/5+sin(7wt+7)/7, w=100π. According to sampling
theorem, the sampling frequency of discrete signals can’t
be lower two times than the highest frequency. The
sampling frequency is fs=50×512=25600 =25.6 kHz and
carry on 7 levels decomposition to original signals by
choosing the db24 as wavelet function.
In original signals, only fundamental wave frequency
is between 0Hz and 100Hz. Therefore, we can obtain
fundamental wave component by means of the
reconstruction of a7 ( k ) and simultaneously attain the end
of fundamental wave detection. The transforming results
of scale 1 to 7 are sub-band d1~d7, and respectively
indicate the harmonic of continuous frequency. The
reconstructed images through different wavelet
coefficient decomposing are as shown in Fig. 12.
Figure 12. Reconstructed of various wavelet resolution coefficients
From simulation results of sub-band a7, we can figure
out the amplitude value of fundamental wave component
to be 0.9931V and the effective value to be 0.7023V. The
relative error of the effective values is 0.67%. Therefore,
this method can realize the separation of fundamental
wave. In addition, because wavelet db24 is un-symmetry,
therefore, the excursion of frequency’s phase appears
between the fundamental wave component, which
detected and the original fundamental component. It is
shown as Fig. 13.
© 2010 ACADEMY PUBLISHER
Figure 13. Ideal fundamental wave (broken line) and real fundamental
wave (real line)
The original signal can be decomposed into the
fundamental wave a7 and sub-harmonics d1, d2, d3, d4,
d5, d6, and d7 through wavelet transform. The frequency
of sub-band signals respectively is 445.64Hz, 276.72Hz,
150.00Hz, 50Hz, and the result is consistent with Table 1.
It can be seem that wavelet transform has good
localization characteristic in time domain and frequency
domain as well and it is suitable for harmonic analysis of
electric power system. The real-time harmonic tracking
mainly aims at obtaining the trend of harmonic. The
relative error of amplitude and phase may not demand too
much. By choosing suitable wavelet, the wavelet
transform can effectively track the changing trend of
harmonic. It is shown as Fig. 14.
Figure 14. Ideal harmonic (broken line) and real harmonic (real line)
C. Voltage Fluctuation and Flickers
Steelmaking arc furnace is main equipment causing
electric voltage fluctuation and flicker, and each
frequency sub-value of amplitude is inversely
proportional to frequency. Now we analyze arc furnace
voltage flicker signal by wavelet analysis tools of
simulation software Matlab. We suggest electric voltage
flicker signal expression causing by steelmaking arc
furnace as formula (8):
ut () = Vm[1+ msin Ω t]sin( wt)
(8)
From formula (8), Vm=1V, Ω=6πrad/s, w=314πrad/s.
when 0.1875s≤t≤1.875s, m=0.05. On Matlab simulation
experiment, at first, we make initial signal u(t) squaring,
get signal s(t), then get 8000 sample point of s(t) by
3.2kHz sample frequency, and then select Daubechies
wavelet db24, decompose signal into multi-layer wavelet
decomposition and reconstruction.
High frequency d1 and d2 is after wavelet transform
and reconstruction. Model maximum position is mutation
point position. It is shown as Fig. 15. In which, abscissa
is time, unit is s, ordinate is voltage amplitude, and unit is

164 JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010
V. The following wave diagram abscissa and ordinate is
same.
Figure 15. Reconstruction high frequency signal d1 and d2.
From Fig. 15 we can see, two position of mutation
point respectively at t1≈0.20s and t2≈1.85s, modulus
maximum of position t2 obvious larger than t1, which
express singular degree of t2 is higher than t1, simulation
experiment shows that mutation position and singular
degree are uniform with initial signal. Low frequency a6
is voltage flicker signal envelope after wavelet
transform reconstruction. It is shown as Fig. 16.
Figure 16. Reconstruction low frequency signal a6
From Fig. 16 we can see, voltage flicker signal initial
time t1 is at position 0.2s and end time t2 position is at
1.85s, which is uniform with theory value 0.1875s and
1.875s, and in Fig. 15, transient signal shows as slope
form at position t1, and position t2 shows as spike signal,
which is uniform to mutation ambiguity degree in
conclusion1. From it we can see, by selecting suitable
wavelet, orthogonal wavelet decomposition algorithm can
effectively recognize signal position of catastrophe point
and singular degree.
To verify wavelet reconstruction accuracy of sub-band
signal, we suppose a(t)=0.1sin(Ωt)+0.03sin(3Ωt)+0.02sin
(5Ωt), from Ω=6πrad/s, we can see, flicker signal
includes three amplitude wave of 3Hz, 9Hz, and 15Hz,
and theoretical amplitude respectively is 0.1, 0.03 and
0.02. From multi-resolution Mallat algorithm, each subband
frequency range is shown as Table III.
TABLE III.
SUB-BAND FREQUENCY RANGE
Scale(j) Reconstruction signal Frequency band (Hz)
1 d1 800~1600
2 d2 400~800
3 d3 200~400
4 d4 100~200
5 d5 50~100
6 d6 25~50
6 a6 0~50
7 d7 12.5~25
7 a7 0~25
8 d8 6.25~12.5
10 d10 1.5625~3.125
© 2010 ACADEMY PUBLISHER
VI. SIMULATION ANALYSIS
For easy to realize human-computer interaction
function, we adopt visual instrument technology to design
upper PC voltage flicker observe panel. According to
wavelet decomposition and reconstruction method, we
mix LabVIEW and Matlab to realize flicker monitor by
calling Matlab script node in LabVIEW [14]. Through
setting wavelet decomposition limit frequency value, we
can make signal decomposed into enough bigger scale,
which can make program automatic judge wavelet layer
number, and get all amplitude modulation wave voltage
signal of carrier wave.
A. Harmonic Monitor
Upper PC detection system receives lower PC
transmitting data, and memories into EXCEL format file.
Through read power network sample data files, where,
we introduce software program method as harmonic
detection for example. System uses wavelet transform
and FFT analysis algorism to realize harmonic analysis.
Matlab [15] wavelet analysis tool box function is perfect.
So, we use Matlab program language to realize wavelet
analysis. And adopt Matlab+LabVIEW mixed program
technology, which make harmonic analysis Matlab
program harmonic.m use at LabVIEW[16][17]program
call. Voltage steady state component after wavelet
decomposing is sent into function FFT.vi port X, FFT
result is obtained at port FFT{X} after program running.
To make wave display control X axis calibration
correspond to time, we add Bundle data packer node: first
item is wave initial position x , where valuation 0,
0
express display from point. Secondary item is step
length ∆ x , where is obtained by system sample
frequency/sample point. The third item is frequency
analysis result shaping array. The three items are sent into
display control after packing.
A phase voltage as example, its harmonic analysis
front panel measurement result is shown as Fig. 17.
Figure 17. Harmonic analysis front panel
Ideal THD% calculation formula is shown as:
2 2 2
77.77 + 40.44 + 31.11
THD% = × 100% = 29. 83%
o
312
From Fig. 17 we can see, THD% of monitor value is
29.82%, which obtains accurate measurement result in
little monitor error.
B. Voltage Fluctuation and Flicker
We also use LabVIEW combining Matlab to complete
voltage fluctuation and flicker monitor. In theory, three
kinds of amplitude wave signal locate at d7, d8 and d10

JOURNAL OF COMPUTERS, VOL. 5, NO. 1, JANUARY 2010 165
after wavelet transform reconstruction, we use 10 level
reconstruction. Sub-band d7, d8, and d10 simulation
result is shown as Fig. 18.
Figure 18. LabVIEW front panel running result
From Fig. 18 we can get each wavelet reconstruction
sub-band frequency and amplitude after taking single
frequency measurement function of LabVIEW function
tools. It is shown as Table IV.
TABLE IV.
RECONSTRUCTION RESULTS
Sub-band Frequency(Hz) Error(Hz) Amplitude(V) Error(V)
d10 3.01 0.01 0.097 -0.003
d8 8.99 -0.01 0.029 -0.001
d7 14.99 -0.01 0.019 -0.001
From three sub value of suggestion voltage a(t), which
theoretical amplitude are respectively 0.1V, 0.03V and
0.05V. Signal basic wave, three times, and five times
harmonic wave sub-value amplitude are respectively
0.097V, 0.029V, and 0.019V, frequency are respectively
3.01Hz, 8.99Hz, and 14.99Hz. Corresponding frequency
and amplitude monitor error are shown as Table III,
which is verified algorithm’s accuracy in smaller monitor
error.
VII. CONCLUSION
System changes traditional electric power parameter
measurement style based on hardware as core through VI
technology, which completely adopts PC’s strong
functions of calculation, display, and storage, etc.
System uses module of voltage/currency transformer,
multiplexing switch, etc, which realizes electric power
network signal detection. Software design of detection
module, data display module, alarming module, report
form drawing module, etc are done by control software of
Matlab+LabVIEW8.5 [18] System hardware has many
advantages of sample structure, higher measurement
precision, convenient extension, and easy to realize
© 2010 ACADEMY PUBLISHER
electric power parameter monitor, etc. System software
uses wavelet multi-resolution analysis method, etc in
monitor error of harmonic wave’s THD% is 0.01%, and
amplitude wave’s frequency and amplitude of voltage
flicker is respectively less than 0.01Hz and 0.005V,
which realizes high precision electric power parameters
measurement.
REFERENCES
[1] Pan Tianhong, and Sheng Zhanshi, “Desing of Power
Network Voltage Monitoring Instrument Based on SMS
Technique of GSM,” Electric Power Automation
Equipment, p.48, Sep 2005.
[2] SPT204A . “Voltage transformer SPT204A,” Beijing :
Department of Chunjing SCM Technology Development,
http://www.chunjs.com/.
[3] AD7507.pdf. “8-and 16-Channel Analog Multiplexers
AD7506/AD7507,”Analog Devices Inc, www.analog.com,
2006.
[4] Mi Xiaoyuan, Zhang Yanbin, and Xue Deqing,.”Using
LabSQL Visiting Database in LabVIEW,”Microcomputer
Information, pp.53-54, Oct 2004.
[5] Sun Xiaoming, “Reasch and Design of Electric Power
Quality Monitor System,” Shandong University, 2005.
[6] George J. Wakelih. “Electric Power System Harmonic –
Basci Principle, Analysis Method and Filter Design,”
Beijing: Mechanical Industry Publishing House, 2003.
[7] Changhua Hu, “System Analysis and Design based Matlab-
Wavelet Analysis,” Xi’an: Electronic Science and
Technology University, 1999.
[8] Lizhi Cheng, and Hanwei Guo, “Wavelet and Discrete
Transform Principal & Practice,” Beijing: Tinghua
University. 2005.
[9] Zhang Junxia, Ren Zihui, Yue Mingdao, Zhang Guoyuan.
“Design of Power Quality Monitoring System Based on Virtual
Instrument,” Instrument Technology, p. 8, Jun 2008.
[10] Xiaoli G, Jianlan L, Xiao W, Jun D, and Zhenguo S,
"Simulation on Voltage Fluctuation Signal Measurement,”
Proceedings of the Electric Power System and Automation,
pp. 41-42, Feb 2006.
[11] Lin Zhou, Xue Xia, Yunjie Wan, Hai Zhang, and Peng Lei,
“Harmonic Detection Based on Wavelet Transform,”
Transactions of China Electro technical Society, p.68, Sep
2006.
[12] Lizhi Cheng, and Hanwei Guo, “Wavelet and Discrete
Transform Principal & Practice,” Beijing:Tinghua
University, May. 2005.
[13] Zheng Chanzheng, Mao Zhe, and Xie Zhaohong, “The
grain depot temperature test system based on nRF905,”
Microcomputer Information, pp.284-285, Feb 2007.
[14] Guo Xianding, “Applying Wavelet-Transform to Inspect
the Discontinuous Dot of Signal,” Electronic Application,
pp, 90-92, Nov 2006.
[15] GeorgeJ,Wakelih,.“Electric power system Harmonic—
Basic Principle, analysis method and filter design,” Beijing:
Mechanical Industry Publisher, 2003.
[16] Dai Pengfei, Wang Shengkai, Wang Gefang, and Ma Xin,
“Monitor Engineering and LabVIEW Application,”
Electric Industry Publishing House, May 2006.
[17] Chen Xihui, and Zhang Yinhong, “LabVIEW Program
Desing,” Beijing: Tsing Hua University, July 2007.
[18] Fei Feng, and Yang Wansheng, “LabView and Matlab
Mixed Program,” Electric Technology Application, pp.4-6,
March 2004.

Aims and Scope.
Call for Papers and Special Issues
Journal of Computers (JCP, ISSN 1796-203X) is a scholarly peer-reviewed international scientific journal published monthly for researchers,
developers, technical managers, and educators in the computer field. It provide a high profile, leading edge forum for academic researchers, industrial
professionals, engineers, consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work
on all the areas of computers.
JCP invites original, previously unpublished, research, survey and tutorial papers, plus case studies and short research notes, on both applied and
theoretical aspects of computers. These areas include, but are not limited to, the following:
• Computer Organizations and Architectures
• Operating Systems, Software Systems, and Communication Protocols
• Real-time Systems, Embedded Systems, and Distributed Systems
• Digital Devices, Computer Components, and Interconnection Networks
• Specification, Design, Prototyping, and Testing Methods and Tools
• Artificial Intelligence, Algorithms, Computational Science
• Performance, Fault Tolerance, Reliability, Security, and Testability
• Case Studies and Experimental and Theoretical Evaluations
• New and Important Applications and Trends
Special Issue Guidelines
Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by
invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal.
Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the
readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length.
The following information should be included as part of the proposal:
• Proposed title for the Special Issue
• Description of the topic area to be focused upon and justification
• Review process for the selection and rejection of papers.
• Name, contact, position, affiliation, and biography of the Guest Editor(s)
• List of potential reviewers
• Potential authors to the issue
• Tentative time-table for the call for papers and reviews
If a proposal is accepted, the guest editor will be responsible for:
• Preparing the “Call for Papers” to be included on the Journal’s Web site.
• Distribution of the Call for Papers broadly to various mailing lists and sites.
• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be
informed the Instructions for Authors.
• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact
information.
• Writing a one- or two-page introductory editorial to be published in the Special Issue.
Special Issue for a Conference/Workshop
A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like
general chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is
typically made of 10 to 15 papers, with each paper 8 to 12 pages of length.
Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop:
• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”.
• Sending us a formal “Letter of Intent” for the Special Issue.
• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees.
Information about the Journal and Academy Publisher can be included in the Call for Papers.
• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus
the evaluation from the Session Chairs and the feedback from the Conference attendees.
• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors.
Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced.
• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact
information.
• Writing a one- or two-page introductory editorial to be published in the Special Issue.
More information is available on the web site at http://www.academypublisher.com/jcp/.

(Contents Continued from Back Cover)
Implementation of Low Density Parity Check Decoders using a New High Level Design
Methodology
Syed Mahfuzul Aziz and Minh Duc Pham
REGULAR PAPERS
A Formal Model for Abstracting the Interaction of Web Services
Li Bao, Weishi Zhang, and Xiong Xie
Performance Evaluation of Elliptic Curve Projective Coordinates with Parallel GF(p) Field
Operations and Side-Channel Atomicity
Turki F. Al-Somani
VPRS-Based Knowledge Discovery Approach in Incomplete Information System
Shibao Sun, Ruijuan Zheng, Qingtao Wu, Tianrui Li
NaXi Pictographs Input Method and WEFT
Hai Guo and Jing-ying Zhao
A Systematic Decision Criterion for Urban Agglomeration: Methodology for Causality Measurement
at Provincial Scale
Yaobin Liu and Li Wan
Application of Improved Fuzzy Controller for Smart Structure
Jingjun Zhang, Liya Cao, Weize Yuan, Ruizhen Gao, and Jingtao Li
Numerical Simulation of Snow Drifting Disaster on Embankment Project
Shouyun Liang, Xiangxian Ma, and Haifeng Zhang
The Simulation of Extraterrestrial Solar Radiation Based on SOTER in Zhangpu Sample Plot and
Fujian Province
Zhi-qiang Chen and Jian-fei Chen
Multiplicate Particle Swarm Optimization Algorithm
Shang Gao, Zaiyue Zhang, and Cungen Cao
Research and Design of Intelligent Electric Power Quality Detection System Based on VI
Yu Chen
81
91
99
110
117
125
131
139
144
150
158