FUNCTIONAL TESTING OF A MICROPROCESSOR

THROUGH LINEAR CHECKING METHOD

Prof. Dr. Pervez Akhtar

National University of Science & Technology,
Karachi Campus, Pakistan
pervez@pnec.edu.pk

Prof. Dr. M.Altaf Mukati

Hamdard Institute of Information Technology,
Hamdard University, Karachi, Pakistan
altafmukati@gmail.com

ABSTRACT
The gate-level testing also called low-level testing is generally appropriate at the
design time and for small circuits. The chip-level testing and board-level testing
also called high-level testing are preferred when the circuit complexities are too
high, making it difficult to perform low level testing in a reasonable amount of
time. The cost of low-level testing is also generally very high. Such high costs and
time are only justified when some design-changes are required. In this paper, a
high level quick checking method, known as Linear Checking Method, is
presented which can be used to qualify the functionality of a Microprocessor. This
can also be used to check hard faults in Memory chips.

Keywords: Microprocessors, ALU, Control Unit, Instructions.

1 INTRODUCTION abstraction [2]. The functional fault modeling should

Due to the advances in the integrated circuit
technology, more and more components are being
fabricated into a tiny chip. Since the number of pins
on each chip is limited by the physical size of the
chip, the problem of testing becomes more difficult
than ever. This problem is aggravated by the fact
that, in nearly all cases, integrated circuit
manufacturers do not release the detailed circuit
diagram of the chip to the users [1].
The users are generally more interested to know
about the chip, whether is it functionally working
and relied upon? if not, the whole chip is replaced
with a newer one. This is contrast to the gate-level
testing of a digital circuit, which is used to diagnose
faulty gates in the given circuit, in case of failing.
The idea of using functional testing is also
augmented by the fact that in case of any functional
failure, caused due to any fault in the chip, the user
can not repair the chip. Hence the users have only
two choices: either to continue using the chip with a
particular failing function, knowing that the failing
function will not be used in the given application or
to replace the whole chip.
The functional modeling is done at a higher level
of abstraction than a gate-level modeling. This in-
fact exists between the Gate-level modeling and the
Behavioral modeling, which is the highest level of
imitate the physical defects that cause change in the
function or behavior, for example; the function of a
synchronous binary up-counter is to advance one
stage higher in binary value when clock hits it. A
physical defect, which alters this function, can be
modeled in terms of its effect on the function. Such
defect-findings are extremely important at the design
time or if the design changes are required at a later
stage.
What, if a microprocessor does not produce the
correct results of any single or more functions? From
the user’s perspective, it is enough to know which
function is failing, but from designer’s perspective,
the cause of failing is also important to know, so that
the design changes may be carried out, if necessary.
This is certainly a time-taking process. For example,
the gate level simulation of the Intel 8085
microprocessor took 400 hours of CPU-time and
only provided 70% fault coverage [3].
High level functional verification for the complex
Systems-On-Chip (SOCs) and microprocessors has
become a key challenge. Functional verification and
Automatic Test Pattern Generator (ATPG) is one
synergetic area that has evolved significantly in
recent years due to the blossoming of a wide array of
test and verification techniques. This area will
continue to be a key focus of future Microprocessor
TEST and Verification (MTV) [4].

UbiCC Journal - Volume 3 1

 Functional failures can be caused due to single or
multiple stuck-at faults in any of its functional block.
The functional-testing, which refers to the selection
of tests that verify the functional operation of a
device, is one of an efficient method to deal with the
faults existing in a processor. Functional testing can
also be carried out at a smaller level, for example, a
functional test of a flip-flop might be to verify
whether can it be set or reset? and further, can it hold
the state or not? Similarly, the other MSI chips such
as Multiplexers, Encoders, Decoders, Counters,
Hardwired-Multipliers, Binary-Adders & Subtrac-
tors, Comparators, Parity Checkers, Registers and
other similar circuits can also be verified for their
required functionalities.
Some designers and manufacturers provide built-
in self-test (BIST) these days that generate the test on
the chip and responses are checked within the chip
itself. However, the widespread use of such
testability techniques is hampered by a lack of tools
to support the designer and by the additional cost in
chip area as well as the degradation in performance
[5]. For example, the Intel 80386 microprocessor
employs about 1.8% area overhead for BIST to test
portions of the circuit [6].
The ever increasing complexity combined with
the advanced technology used in the design of the
modern microprocessors, has lead to two major
problems in producing cost-effective, high quality
chips:
1. Verification: This is related to validate the
correctness of the complex design. Simulation
is the primary means of design validation used
today. In the case of processor design
validation, the sequences are either written
manually or generated automatically by a
random sequence generator [7].
2. Testing: This is related to check the
manufactured chips for realistic defects. A
variety of test generation and design-for-
testability (DFT) techniques is used to ensure
that the manufactured chips are defect-free.
Both design verification and testing depend,
therefore, on test sequences used to expose either the
design faults or manufacturing defects. It has also
been found that manufacturing test pattern
generation can be used for design verification [8] and
that design verification techniques can be used to
find better manufacturing tests [9]. However, to find
the effective test patterns for either of the said
purposes is not simple, due to high complexities of
microprocessors. Hence the only effective method
left is to develop the functional tests. Considerable
work has been done in the field of microprocessor
functional testing. One of such work, known as
‘Linear Checking Method’ is presented in this paper.
Before performing functional testing, functional
UbiCC Journal - Volume 3
description of the chip must be known. In case of
microprocessor, this can be obtained through its
instruction set. The two most important functional
blocks of any microprocessor are the CU (Control
Unit) and the ALU (Arithmetic Logic Unit). All the
instructions, at low-level, are composed of Op-Codes
and operands. An op-code, also called the Macro-
instruction, goes to the CU, which decodes each
macro-instruction into a unique set of micro-
instructions. The operands go to the ALU, which
processes it according the tasks defined within the
micro-instructions. In between these functional
blocks, there exists several registers for the
temporary storage of op-codes, decoded-instructions
and operands.
The fault may occur at various places in the
processors, causing it to function incorrectly. Some
of the common faults are: Register Decoding Fault,
Micro-Operation Decoding Fault (caused may be due
to internal defect to the CU), Data Storage Fault
(caused may be due to Stuck-at Fault or Pattern
Sensitive Fault in the memory inside the
Microprocessor), Data Transfer Fault (caused may be
due to Stuck-at Fault or Bridging Fault on the busses
connecting the various functional blocks of a
Microprocessor) or ALU Fault (caused due to
internal defect to the ALU). In each case, the given
microprocessor results in producing incorrect
function/ functions.
In the subsequent sections, first the functional
verification has been described in general and then
the Linear Checking Method has been presented
through several examples. Based on the results
obtained, the conclusion has been drawn and the
further work has been proposed.
2 FUNCTIONAL VERIFICATION
The micro-instructions from the CU and the
operands of an instruction are sent to the ALU
simultaneously. The ALU then carries out the
intended task or function. This can be shown with
the help of a block diagram, as in Fig. 1.
Figure 1: Functional testing
The typical instructions are ADD, SUB, MUL,
2
 SHL, SHR, ROTL, ROTR, INC, DEC, COMPL,
AND, OR, XOR and many others.
3 LINEAR CHECKING METHOD
This method can be used to test and verify, not
only the functionality of a microprocessor (more
specifically ALU), but the memories as well. Linear
checking is based on computing the value of ‘K’
using the equation 3.1:
K = f i (x, y) + f i (x, y) + f i (x, y) + f i (x, y)
Equation 1 is called the ‘basic equation’. The
variables x and y are the operands, ‘i’ is the
instruction. The value of K does not depend on the
values of x and y, but only depends on the instruction
and on the size of operands (number of bits in the
operands). It means the value of K is unique for
every instruction. The chances are very little that the
two instructions may have the same constant value of
K. An 8 and 16-bit ALUs have different values of K,
for the same instruction. Hence, in this method, K is
used as a reference value to verify the functionality
of an individual instruction.
3.1 Examples of functional verifications
Consider a 4-bit ALU. The value of ‘K’ can be
computed as follows:
Suppose the instruction is ADD(x, y) = x + y
Here, n = 4. Let x = 5 (0101) and y = 3 (0011)
Therefore x = 1010 and y = 1100
The value of K can be obtained from Equation 1,
as follows:
ADD(5,3)+ADD(5,12)+ADD(10,3)+
ADD(10,12) = K
8 + 17 + 13 + 22 = 60
Hence, for a 4-bit ALU, the ADD instruction
will always be tested with respect to its reference
value of 60, regardless what values of x and y are
taken, i.e. instead of 5 and 3 as in the above example,
now these values are taken as 9 and 10 respectively.
Still the value of K remains the same, as proved
below:
i.e. for x = 9 (1001) and y = 10 (1010)
x = 6 (0110) and y = 5 (0101)
ADD(9,10)+ADD(9,5)+ADD(10,6)+ADD(6,5)=
60
The generalized formula can also be developed to
find the value of K for the ADD instruction, for any
size of ALU, as follows:
UbiCC Journal - Volume 3
K+(n) = 4(2n – 1)
Where, the subscript with K represents the
function. Hence from the generalized form, we
obtain the same value of K i.e. if n = 4, then K+(4) =
4(15) = 60.
Similarly, the value of K for any instruction can
be obtained, provided its functional description is
known. The value of K, for the various other
frequently used instructions, can be obtained
similarly, as follows:
(1) Again assume a 4-bit ALU. Taking x = 10 and y
= 12 then x = 5 and y = 3 in all the computations.
3.1.1 Multiply instruction (f(x,y) = X * Y)
fi(x,y) = MPY(x,y) = X * Y
Hence, from equation 3.1, the value of K can be
obtained as follows:
MPY(12,10)+MPY(10,3)+MPY(5,12)
+MPY(5,3)
120 + 30 + 60 + 15 = 225
Generalized form Æ K* = (2n – 1)2
3.1.2 Transfer instruction (f(x) = x)
This is a single valued function, hence only one
variable is taken in computation of K, thus ‘y’ is
ignored in equation 1.
Thus f i (x) = x
= x + x + x + x = 10 + 10 + 5 + 5 = 30
Generalized form Æ K = 2(2n – 1)
3.1.3 Shift-Right instruction (f(x) = SHR(x))
It is also a single valued function. With x = 1010:
(a) f i (x, y) & f i (x, y) reduce to f i (x) and
(b) f i (x, y) & f i (x, y) reduce to f i (x)
Now f i (x) represents the value of x, after SHR
operation i.e. 1010 Æ 0101 and f i (x) represents the
value of x after SHR operation i.e. 0101 Æ 0010.
Hence, K = 0101 + 0101 + 0010 + 0010
or K = 5 + 5 + 2 + 2 = 14
Generalized form Æ K = 2(2n-1 – 1)
3.1.4 Shift-Left instruction (f(x) = SHL(x))
With the same explanation as in section 3.1.3,
the equation 1 becomes:
K = f i (x) + f i (x) + f i (x) + f i (x)
3
 Hence, with x = 1010 Æ f i (x) = 0100 & 3.1.10 Complement instruction (f(x) Æ x )
x = 0101 Æ f i (x) = 1010 K = f i (x) + f i (x) + f i (x) + f i (x)

Therefore, K = 4 + 4 + 10 + 10 = 28 = 0101 + 0101 + 1010 + 1010 = 30

Generalized form Æ K = 2(2n – 2) Generalized form Æ K = 2(2n - 1)

3.1.5 Logical-OR instruction (f(x) = x·OR·y) 3.1.11 2’s comp. instruction (f(x) Æ x + 1)
K = f i (x, y) + f i (x, y) + f i (x, y) + f i (x, y) K = f i (x) + f i (x) + f i (x) + f i (x)

= 0110 + 0110 + 1011 + 1011 = 34

= (x·OR·y)+(x·OR· y )+( x ·OR·y)+( x ·OR· y )
= 1110 + 1011 + 1101 + 0111
=14 + 11 + 13 + 7 = 45
Generalized form Æ K = 3(2 – 1) n
3.1.6 Logical-AND instruction (f(x) = x·AND·y)
K = f i (x, y) + f i (x, y) + f i (x, y) + f i (x, y)
= (x·AND·y)+(x·AND· y )+( x ·AND·y)+
( x ·AND· y )
= 1000 + 0010 + 0100 + 0001
= 8 + 2 + 4 + 1 = 15
Generalized form Æ K = 2n – 1
3.1.7 Logical-XOR instruction (f(x) = x·XOR·y)
K = f i (x, y) + f i (x, y) + f i (x, y) + f i (x, y)
=(x·XOR·y)+(x·XOR· y )+( x ·XOR·y)+
( x ·XOR· y )
= 0110 + 1001 + 1001 + 0110
= 6 + 9 + 9 + 6 = 30
Generalized form Æ K = 2(2n – 1)
3.1.8 Increment instruction (f(x) = INC(x))
This is also a single valued function:
K = f i (x) + f i (x) + f i (x) + f i (x)
Generalized form Æ K = 2(2n + 1)
3.2 Memory error correction
Linear checks can also be used to verify
memories. For example, let the multiplication x*y
function is stored in the memory. Let the operands
are 4-bit length, with x = 1010 and y = 0010, it
means x = 0101 and y = 1101. Hence, all the four
components of equation 1 are computed and the
results are stored in memory, as shown in Table 1.
Table 1: linear checks on memories
f(x,y) 20 00010100
f(x, y ) 130 10000010
f( x ,y) 10 00001010
f( x , y ) 65 01000001
K 225 11100001
If the sum of four components is not equal to the
value of K, then there must be some fault existing in
the memory. Similarly, any of the preceding
functions can be used to verify the memories. The
testing can be done more accurately if the contents of
f(x,y) are stored at address (x,y). In the above
example, the contents can be stored on the
corresponding addresses as shown in Table 2. If
addition of the contents does not come equal to the
value of K, then it will indicate some fault in the
memory. Here, the location of fault is also obtained.

= 1011 + 1011 + 0110 + 0110 Table 2: address versus contents in memory testing
= 11 + 11 + 6 + 6 = 34
Address
Generalized form Æ K = 2(2n + 1) x y
Contents

3.1.9 Decrement instruction (f(x) = DEC(x)) 1010 0010 00010100

K = f i (x) + f i (x) + f i (x) + f i (x) 1010 1101 10000010
= 1001 + 1001 + 0100 + 0100 0101 0010 00001010
= 9 + 9 + 4 + 4 = 26 0101 1101 11100001
Generalized form Æ K = 2(2 - 3) n

UbiCC Journal - Volume 3 4

 4 RESULTS
Add x+y 4(2n – 1) 1020
All the computations done in the previous section
are summarized in tables 3 & 4 for n = 4 & 8 Multiply x ∗ y (2n – 1)2 65025
respectively:
Subtract x–y 0 0
Table 3: Values tabulated through linear checking
method for n = 4 Logical OR x∨y 3(2n – 1) 765

Instruction i f i (x, y) K i (n) K i (4) Logical x∧y 2n – 1 255

AND
Clear 0 0 0 Logical x⊕y 2(2n – 1) 510
XOR
Transfer x 2(2n – 1) 30 Complement 2(2n - 1) 510
x

Add x+y 4(2n – 1) 60 2’s comp. 2(2n + 1) 514

x +1

Multiply x ∗ y (2n – 1)2 225 Increment x+1 2(2n + 1) 514

Subtract x–y 0 0 Decrement x–1 2(2n - 3) 506

Logical OR x∨y 3(2n – 1) 45 Shift-Left (x2,x3,...xn,0) 2(2n – 2) 508

Logical x∧y n
2 –1 15 Shift-Right (0,x2,x3,...xn-1) 2(2n-1 – 1) 254
AND
Logical x⊕y 2(2n – 1) 30 Rotate-Left (x2,x3,...xn,x1) 2(2n - 1) 510
XOR
Complement x 2(2n - 1) 30 Rotate-Right (xn,x2,...xn-1) 2(2n - 1) 510

2’s comp. x +1 2(2n + 1) 34

5 CONCLUSION
n
Increment x+1 2(2 + 1) 34
It is concluded that the value of K can be
n obtained for any given instruction. The ‘CLR’ (clear)
Decrement x–1 2(2 - 3) 26
instruction is a special one, since it does not have any
operand; all the four components of the equation 3.1
Shift-Left (x2,x3,...xn,0) 2(2n – 2) 28 are taken as 0. Note that almost all the values
obtained for K are unique, except for ‘Transfer’,
Shift-Right (0,x2,x3,...xn-1) 2(2n-1 – 1) 14 ‘Complement’, ‘Logical XOR’ and ‘Rotate-
Left/Right’ instructions, which means if the
Rotate-Left (x2,x3,...xn,x1) 2(2n - 1) 30 instructions are transformed due to any fault in the
CU (or in any associated circuit) then these particular
Rotate-Right (xn,x2,...xn-1) 2(2n - 1) 30 functional failures cannot be distinguished but the
processor as a whole can be declared to have
contained a fault. The last column in tables 3 & 4 can
be obtained directly from the generalized forms. This
Table 4: Values tabulated through linear checking column is stored in the memory along with the
method for n = 8 relevant function.

Instruction i f i (x, y) K i (n) K i (8) 6 FUTURE WORK

Clear 0 0 0 Further research is proposed on the given method,

especially in the case when the ‘Reference Value
Transfer x 2(2n – 1) 510 (K)’ of two or more functions is obtained same i.e. to
distinguish or identify an individual failing function,
in case when their reference values happen to be the
same, as mentioned under the conclusion.

UbiCC Journal - Volume 3 5

 7 REFERENCES
[1] Su Stephen Y.H., Lin Tony S., Shen Li:
Functional Testing of LSI/VLSI Digital
Systems, Defense Technical Information Center,
Final Technical Report, School of Engineering
Applied Science and Technology Binghampton,
NY, August 1984
[2] Altaf Mukati: Fault Diagnosis and Testing of
Digital Circuits with an Introduction to Error
Control Coding, Pub. Higher Education
Commission Pakistan, ISBN: 969-417-095-8,
2006
[3] Jian Shen, Abraham J.A.: Native mode
functional test generation for processors with
applications to self test and design validation,
Test Conference (1998), Proc. International
Volume, Issue, 18-23 Oct 1998 pp. 990 – 999.
[4] Magdy S.Abadir, Li-C. Wang, Jayanta Bhadra:
Microprocessor Test and Verification (MTV
2006), Common Challenges and Solutions,
Seventh International Workshop, Austin, Texas,
USA. IEEE Computer Society 4-5 December
2006, ISBN: 978-0-7695-2839-7
UbiCC Journal - Volume 3
[5] Jian Shen, Jacob A. Abraham: Native Mode
Functional Test Generation for Processors with
Applications to Self Test and Design Validation,
Proc. Computer Engineering Research Center,
The University of Texas at Austin, 1998.
[6] P. P. Gelsinger: Design and Test of the 80386,
IEEE Design and Test of Computers, Vol. 4, pp.
42-50, June 1987.
[7] C. Montemayor et al: Multiprocessor Design
Verification for the PowerPC 620
Microprocessor, Proc. Intl. Conf. on Computer
Design, pp. 188-195, 1995.
[8] M. S. Abadir, J. Ferguson, and T. Kirkland:
Logic design verification via test generation,
IEEE Trans. on Computer-Aided Design of
Integrated Circuits and Systems, Vol. 7, pp. 138-
148, 1988.
[9] D. Moundanos, J. A. Abraham, and Y. V.
Hoskote: A Unified Framework for Design
Validation and Manufacturing Test, Proc. Intl.
Test Conf., pp. 875-884, 1996.
6
 OPERATING SYSTEMS FOR
WIRELESS SENSOR NETWORKS: AN OVERVIEW

Daniele De Caneva, Pier Luca Montessoro and Davide Pierattoni

DIEGM – University of Udine, Italy
{daniele.decaneva; montessoro; pierattoni}@uniud.it

ABSTRACT

The technological trend in the recent years has led to the emergence of complete
systems on a single chip with integrated low power communication and transducer
capabilities. This has opened the way for wireless sensor networks: a paradigm of
hundreds or even thousands of tiny, smart sensors with transducer and
communication capabilities. Manage such a complex network that has to work
unattended for months or years, being aware of the limited power resouces of
battery-supplied nodes is a challenging task. Attending that task requires an
adequate software platform, in other words an operating system specifically suited
for wireless sensor networks. This paper presents a brief overview of the most
known operating systems, highlighting the key challenges that have driven their
design.

Keywords: wireless sensor networks, operating systems.

1 INTRODUCTION the most known operating systems designed for

Thanks to the well known “Moore’s Law”
integrated circuits are becoming smaller, cheaper
and less power consuming. This trend has led to the
emergence of complete systems on a chip with
integrated low power communication and
transducer capabilities. The consequence is the
opening of the ubiquitous computing era, in which
electronic systems will be all around us, providing
all kind of information services to users in a
distributed, omnipresent but nearly invisible fashion.
One of the most important applications that new
technologies are enabling is the paradigm of
Wireless Sensor Networks (WSNs), where
hundreds or even thousands of tiny sensors with
communication capabilities will organize
themselves to collect important environmental data
or monitor areas for security purposes.
The hardware for WSNs is ready and many
applications have become a reality, nevertheless the
missing of a commonly accepted system
architecture and methodology constitute a curb to
the expansion and the improvement of such
technologies. Aware of that, many research groups
in the world have proposed their own system
architecture. The key point in all these proposals is
the capability of the software to manage a
considerable number of sensors. In particular, there
is a tradeoff between the responsiveness of the
system and the extremely scarce resources of the
nodes in terms of power supply, memory and
computational capabilities.
In this article will be presented an overview of
WSNs. Without proposing direct comparisons, we
describe the key features of these architectures, the
challenges that led their development, with the aim
of helping the reader to choose among these
systems the one that best suites his/her purposes.
2 OVERVIEW
2.1 TinyOS
TinyOS [1] is virtually the state-of-the-art of
sensor operating systems. Berkeley University
researchers based their work aiming to face two
issues: the first was to manage the concurrency
intensive nature of nodes which need to keep in
movement different flows of data simultaneously,
while the second was to develop a system with
efficient modularity but believing that hardware and
software components must snap together with little
processing and storage overhead. The purpose of
the researchers was also to develop a system that
would easily scale with the current technology
trends, supporting smaller devices as well as the
crossover of software components into hardware.
Considering power as the most precious
resource and trying to achieve high levels of
concurrency, the system was designed following an
event-based approach, which avoids reserving a
stack space for each execution context. This design
guideline was drawn from a parallelism with high
performance computing, where event-based
programming is the key to achieve high
performance in concurrency intensive applications.

UbiCC Journal - Volume 3 7

In TinyOS neither blocking nor polling operation is
permitted and the CPU doesn’t waste time in
actively looking for interesting events; on the
contrary, unused CPU cycles are spent in a sleep
state.
System configuration can be summarized in a
tiny scheduler and a set of components. The
scheduler is a simple FIFO utilizing a bounded size
scheduling data structure for efficiency, nonetheless
a more sophisticated scheduling policy could be
implemented. When the task queue is empty, the
CPU is forced to the sleep state waiting for an
hardware event to trigger the scheduling of the
event-associated tasks. Tasks in the TinyOS
architecture are atomic and run to completion
although they can be preempted by events. This
semantics of tasks allows the allocation of a single
stack, which is an important feature in memory
constrained systems.
Three types of components were thought to
constitute the TinyOS architecture. The first type of
components are the “hardware abstraction” which
map physical hardware like I/O devices into
component models. The second type of components
is called “synthetic hardware” and simulates the
behavior of advanced hardware; often synthetic
hardware sits on top of the hardware abstraction
components. The last type of components are the
“high level software” components, which perform
control, routing and all data transformations such as
data aggregation and manipulation. This kind of
abstraction of hardware and software in the
component model is intended to ease the
exploitation of tradeoffs between the scale of
integration, the power requirements and the cost of
the system. Every component owns a fixed-size
frame that is statically allocated: this allows to
know exactly the memory requirements of a
component at compile time and prevents the
overhead associated with dynamic allocation.
TinyOS was originally developed in C, giving
the system the capability of targeting multiple CPU
architectures. However, the system was afterwards
re-implemented in nesC: this is a programming
language specific for networked embedded systems,
whose key focus is holistic approach in design.
It’s remarkable that for TinyOS a byte-code
interpreter has been developed that makes the
system accessible to non-expert programmers and
enables quick and efficient programming of a
whole WSN. This interpreter, called Maté, depicts
the program’s code as made of capsules. Thanks to
the beaconless, ad-hoc routing protocol
implemented in Maté, when a sensor node receives
a newer version of a capsule, it will install it.
Through a hop-by-hop code injection, Maté can
update the code of the entire network.
UbiCC Journal - Volume 3
2.2 MANTIS
The MultimodAl system for NeTworks of In-
situ wireless Sensors [3] was developed focusing on
two design key: the need for a small learning curve
for users and the need for flexibility. The first
objective led to fundamental choices in the
architecture of the system and the programming
language used for its implementation. In fact, to
lower the entry barrier the researchers decided to
adopt a largely diffuse design methodology, that is,
the classical structure of a multithreaded operating
system. For this reason MANTIS includes features
like multithreading, preemptive scheduling with
time slices, I/O synchronization via mutual
exclusion, and standard network stack and device
drivers. The second choice is associated with the
purpose of flattening the learning curve for users
and determinates the use of standard C as
developing language for the kernel and the API.
The choice of C language additionally entails the
cross-platform support and the reuse of a vast
legacy code base.
MANTIS kernel resembles UNIX-style
schedulers providing services for a subset of
POSIX threads along with priority based scheduling.
The thread table is allocated statically, so it can be
adjusted only at compile time. The scheduler
receives notes to trigger context switches from an
hardware timer interrupt. This interrupt is the only
kind of hardware interrupt handled by the kernel: in
fact all the others interrupts are sent directly to
device drivers. Context switches are triggered not
only by timer events but also by system calls or
semaphore operations. Besides drivers and user
threads, MANTIS has a special idle, low-priority
thread created by the kernel at startup. This thread
could be used to implement power-aware
scheduling: thanks to its position, it can detect
patterns of CPU utilization and adjust kernel
parameters to conserve energy.
MANTIS researchers thought that wireless
networking management was a critical matter, so
they developed the layered network stack as a set of
user level threads. In other words, they
implemented different layers in different threads
advocating that this choice promotes flexibility to
the detriment of performance. This flexible
structure is useful in particular for dynamic
reprogramming, because it enables application
developers to reprogram network functionalities
such as routing by simply starting, stopping and
deleting user-level threads.
Drawing from the experience in WSNs, the
developers of MANTIS gave their system a set of
sophisticated features like dynamic reprogramming
of sensors nodes via wireless communication,
remote debugging and multimodal prototyping.
MANTIS prototyping environment provides a
framework for testing devices and applications
8
across heterogeneous platforms. It extends beyond
simulation permitting the coexistence of both
virtual and physical nodes in the same network.
This feature is derived directly by the system code
architecture, which can run without modifications
on virtual nodes within an x86 architecture.
Dynamic reprogramming in MANTIS is
implemented as a system call library, which is built
into the kernel. There are different granularities of
reprogramming: entire kernel reflashing,
reprogramming of a single thread, changing of
variables within the thread. Along with dynamic
reprogramming, an important feature has been also
developed: the Remote Shell and Command Server
which allows the user “logging in” into a node and
taking control of it. The server is implemented as an
application thread and gives the user the ability to
alter node’s configuration, run or kill programs,
inspect and modify the inner state of the node.
2.3 Contiki
Contiki is an operating system based on a
lightweight event-driven kernel. It was developed
drawing from previous operating systems’ works
with the goal of adding features like run-time
loading and linking of libraries, programs and
device drivers, as well as support for preemptive
multithreading.
Event-based systems have shown good
performance for many kind of WSN’s applications;
however, purely event-based systems have the
penalty of being unable to respond to external
events during long-lasting computations. A partial
solution to this problem is adding multithreading
support to the system, but this would cause
additional overhead. To address these problems
Contiki researchers have done the compromise of
developing an event-driven kernel and
implementing preemptive multithreading features
as a library, which is optionally linked with
programs that explicitly require it.
Contiki operating system can be divided in three
main components: an event driven kernel that
provides basic CPU multiplexing and has no
platform-specific code, a program loader and
finally a set of libraries that provide higher level
functionalities.
From a structural point of view, a system
running Contiki can be partitioned in two parts: a
core and a set of loadable programs. The core is
compiled into a single binary image and is
unmodifiable after nodes’ deployment. The
programs are loaded into the system by the program
loader, which may obtain the binaries either from
the communication stack (and thus from the
network) or from the system’s EEPROM memory.
Shared libraries like user programs may be
replaced in deployed systems by using the dynamic
UbiCC Journal - Volume 3
linking. Dynamic linking is based on synchronous
events: a library function is invoked by issuing an
event generated by the caller program. The event
broadcasts the request to all the libraries and a
rendezvous protocol is used to find out the library
that implements the required function. When the
correct library has completed the call, the control
returns back to the calling process. Since dynamic
linking bases its functioning on synchronous events,
it is essential that context switching overhead is as
small as possible, in order to have a good system
performance. Contiki developers have granted this
by implementing processes as event handlers,
which run without separate protection domains.
The flexible mechanism of dynamic linking
allowed Contiki researchers to implement
multithreading as a library optionally linked with
programs. Another important component based on a
shared library is the communication stack.
Implementing the communication stack as a library
allows its dynamic replacement and, more precisely,
if the stack is split into different libraries it
becomes easy to replace a communication layer on
the run.
2.4 PicOS
PicOS is an operating system written in C and
specifically aimed for microcontrollers with limited
RAM on chip. In the attempt to ease the
implementation of applications with constrained
resource hardware platforms, PicOS creators leaned
towards a programming environment, which is a
collection of functions for organizing multiple
activities of “reactive” applications. This
environment is capable to provide services like a
flavor of multitasking and tools for inter-process
communication.
Each process is thought as a FSM that changes
its state according to the events. This approach is
very effective for reactive applications, whose
primary role is to respond to events rather than
processing data or crunching numbers. The CPU
multiplexing happens only at state boundaries: in
other words FSM states can be viewed as
checkpoints, at which PicOS processes can be
preempted. Owing the fact that processes are
preemptible at clearly defined points, potentially
problematic operations on counters and flags are
always atomic. On the other hand, such non-
preemptible character of PicOS processes makes
this system not well suitable for real time
applications. In PicOS active processes that need to
wait for some events may release the CPU by
issuing a “wait request”, which defines the
conditions necessary to their recovery. This way the
CPU resources could be destined to other processes.
The PicOS system is also equipped with several
advanced features, like a memory allocator capable
9
of organizing the heap area into a number of
different disjoint pools, and a set of configurable
device drivers including serial ports, LCD displays
and Ethernet interfaces.
2.5 MagnetOS
Applications often need to adapt not only to
external changes but also to the internal changes
initiated by the applications themselves. An
example may come from a battlefront application
that may modify its behavior switching from the
defensive to the offensive mode: this application
could change its communication pattern and
reorganize the deployed components. Focusing on
this point, researchers at the Cornell University
argued that network-wide energy management is
best provided by a distributed, power-aware
operating system. Those researchers developed
MagnetOS aiming to the following four goals. The
first was the adaptability to resources and network
changes. The second was to follow efficient
policies in terms of power consumption. The third
goal was giving the OS general-purpose
characteristics, allowing it to execute applications
over networks of nodes with heterogeneous
capabilities and handling different hardware and
software choices. The fourth goal was providing the
system with facilities for deploying, managing and
modifying executing applications.
The result was a system providing a SSI,
namely a Single System Image. In this abstraction,
the entire network is presented to applications as a
single unified Java virtual machine. The system,
which follows the Distributed Virtual Machine
paradigm, may be partitioned in a static and in a
dynamic component. The static component rewrites
regular Java applications into objects that can be
distributed across the network. The dynamic
component provides on each node services for
application monitoring and for object creation,
invocation and migration. In order to achieve good
performance an auxiliary interface is provided by
the MagnetOS runtime that overrides the automatic
object placement decisions and allows
programmers to explicitly direct object placement.
MagnetOS uses two online power-aware
algorithms to reduce application energy
consumption and to increase system survival by
moving application components within the entire
network. In practice, these protocols try to move he
communication endpoints in order to conserve
energy. The first of them, called NetPull, works at
the physical layer whereas the second one, called
NetCenter, works at the network layer.
2.6 EYES
This operating system was developed within the
UbiCC Journal - Volume 3
EYES European project and tries to address the
problems of scarce resources in terms of both the
memory and power supply and the need for
distribution and reconfiguration capabilities.
The researchers found solution to these
problems developing a event-driven system. In fact,
EYES OS is structured in modules that are executed
as responses to external events, leaving the system
in a power saving mode when there is no external
event to serve. Every module can ask for several
tasks to be performed; each task in turn defines a
certain block of code that runs to completion. In
this paradigm, no blocking operation is permitted
and no polling operation should be instantiated: the
programmer instead will use interrupts to wake up
the system when the needed input becomes
available.
The system provides a scheduler which can be
implemented as a simple FIFO or a more
sophisticated algorithm. The interrupts are also seen
as tasks scheduled and ready to be executed.
In the EYES architecture there are two system
layers of abstraction. The first layer is the “Sensor
and Networking Layer”, which provides an API for
the sensor nodes and the network protocols. The
second layer is the “Distributed Services Layer”,
which exposes an API for mobile sensor
applications support. In particular, two services
belong to this layer: the “Lookup Service” and the
“Information Service”. The first supports mobility,
instantiation and reconfiguration, while the latter
deals with aspects of collecting data. On top of
cited layers stand the user applications.
The EYES OS provides a four-step procedure
for code distribution, designed to update the code
into nodes, including the operating system. This
procedure is resilient to packet losses during the
update, using as few communication and local
resources as possible and halting the node
operations only for a short period.
3 CONCLUSIONS
The operating systems here described present
different approaches to the common problems of
WSNs. It is not in the aim of this article to express
opinions about the presented systems; nevertheless,
some general guidelines could be drawn from the
work experience made by all the esteemed
researchers.
We present now some guidelines for the
development of the next generation of WSN
operating systems, that should help both researchers
and users.
The constrained nature of resources in
embedded systems is definitely evident, so a
small, efficient code is a primary goal, as well
as power-aware policies are an obligatory
10
 Table 1: summary of WSN OS features.

TinyOS Mantis
Objectives Manage concurrent data flows Small learning curve
Scale easily with technology
Modularity
Structure Event-based approach Multithreaded OS,
Tiny scheduler and a set of components UNIX-style scheduler
No blocking or polling Statically-allocated thread table
Developed in nesC Developed in C
Special A byte code interpreter for non-expert Specific idle task that adjusts kernel
features programmers parameters to conserve energy
Remote debugging and reprogramming

Contiki PicOS
Objectives Preemptive multithreading support Aimed for microcontrollers with tiny RAM
Runtime loading and linking of libraries
Structure Lightweight event-driven kernel Each process thought as a FSM
Multithreading features as an optionally Multiplexing at state boundaries
linked library Written in C
Special Capable of changing communication layer on Memory allocator
features the run A set of configurable device drivers

MagnetOS EYES
Objectives Adaptability to resource and network changes Address problems of scarce memory and
Manage nodes with heterogeneous power supply
capabilities
Structure Single System Image, the entire network is a Event driven OS
unified Java virtual machine Structured in modules executed as responses
to external events
Each task runs to completion
Special Two on-line special algorithms to reduce Two layers of abstraction with specific API
features energy consumption for applications and physical support
Four-step procedure to update the code

Table 2: the seven expected features of the next generation WSN operating systems.

Power-aware policies

Self organization

Easy interface to expose data

Simple way to program, update and debug
network applications
Power-aware communication protocols

Portability
Easy programming language for non-tech
users

UbiCC Journal - Volume 3 11

 condition to exploit the efficiency in WSN
applications.
To ensure a proper functioning of a network,
which is constituted by unattended nodes that
could have been deployed in a harsh
environment, the operating system must provide
a mechanism for self-organization and
reorganization in case of node failures.
A WSN, especially if composed of a huge
number of nodes, must behave as a distributed
system, exposing an interface where data and
processes are accessible and manageable like it
happens with databases.
A large number of nodes carries also the need
for an easy, yet power efficient way to program
the network, which should be also usable after
the deployment and without affecting normal
functioning. Such a programming (and re-
programming) procedure must be robust to
interference and to all other causes of
transmission failures during the dissemination
of code chunks. While the entire re-
programming of the core of the system may not
be necessary, the applications must be patched,
updated or even totally changed if the main
purpose of the WSN is changed. This leads to
the preference, if possible, of different levels of
re-programming granularity.
The operating system must treat wireless
communication interfaces as special resources,
providing a set of different power aware
communication protocols. The system has to
choose the proper protocol, according to the
current environment state and application needs.
The operating system should be portable to
different platforms: this is necessary both for the
possible presence of nodes with different tasks
and for the opportunity of a post-deployment of
newer sensors, which could be placed in order
to reintegrate the network node set.
The operating system should provide a
platform for fast prototyping, testing and
debugging application programs. In this context
it is remarkable to note that, if the WSN
paradigm will spread in a kaleidoscopic set of
applications, touching many aspects of our life,
UbiCC Journal - Volume 3
then program developers will not be just
communication and computer engineers. It
appears clear that, in order to support non-
technical developers, a really simple API or
even an application-typology programming
language must be provided, alongside with the
“normal” and more efficient API. Making WSN
easy to use will make them more attractive and
step up their diffusion.
4 REFERENCES
[1] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D.
Culler, K. Pister: System Architecture
Directions for Networked Sensors, ASPLOS.
(2000).
[2] K. Sohraby, D. Minoli, T. Znati: Wireless
Sensor Networks: technology, Protocols and
Applications, John Wiley & Sons Inc. (2007).
[3] H. Abrach, S. Bhatti, J. Carlson, H. Dai J. Rose,
A. sheth, B. Sheth, B. Shucker, J. Deng, R. Han:
MANTIS: System Support for MultimodAl
NeTworks of In-situ Sensors, Proceedings of the
2nd ACM International Conference on Wireless
Sensor Networks and Applications. (2003).
[4] A. Dunkels, B. Grönvall, T. Voigt, J. Alonso:
The Design for a Lightweight Portable
Operating System for Tiny Networked Sensor
Devices, SICS Technical Report (2004).
[5] E. Akhmetshina, P. Gburzynski, F. Vizecoumar:
PicOS: A Tiny Operation System for Extremely
Small Embedded Platforms, Proceedings of the
Conference on Embedded System and
Applications ESA’02 (2002).
[6] R. Barr, J. Bicket, D. S. Dantas, B. Du, T.W.D.
Kim, B. Zhou, E. Sirer: On the need for system-
level support for ad hoc and sensor networks,
SIGOPS Oper. Syst. Rev. (2002).
[7] S. Dulman, P. Havinga: Operating System
Fundamentals for the EYES Distributed Sensor
Network, Proceedings of Progress’02 (2002).
12
 Performance Evaluation of Deadline Monotonic Policy
over 802.11 protocol

Ines El Korbi and Leila Azouz Saidane

National School of Computer Science
University of Manouba, 2010 Tunisia
Emails: ines.korbi@gmail.com Leila.saidane@ensi.rnu.tn

ABSTRACT
Real time applications are characterized by their delay bounds. To satisfy the
Quality of Service (QoS) requirements of such flows over wireless
communications, we enhance the 802.11 protocol to support the Deadline
Monotonic (DM) scheduling policy. Then, we propose to evaluate the performance
of DM in terms of throughput, average medium access delay and medium access
delay distrbution. To evaluate the performance of the DM policy, we develop a
Markov chain based analytical model and derive expressions of the throughput, the
average MAC layer service time and the service time distribution. Therefore, we
validate the mathematical model and extend analytial results to a multi-hop
network by simulation using the ns-2 network simulator.

Keywords: Deadline Monotonic, 802.11, Performance evaluation, Average

medium access delay, Throughput, Probabilistic medium access delay bounds.

1 INTRODUCTION use a distributed scheduling and introduce a new

Supporting applications with QoS requirements
has become an important challenge for all
communications networks. In wireless LANs, the
IEEE 802.11 protocol [5] has been enhanced and the
IEEE 802.11e protocol [6] was proposed to support
quality of service over wireless communications.
In the absence of a coordination point, the IEEE
802.11 defines the Distributed Coordination
Function (DCF) based on the Carrier Sense Multiple
Access with Collision Avoidance (CSMA/CA)
protocol. The IEEE 802.11e proposes the Enhanced
Distributed Channel Access (EDCA) as an extension
for DCF. With EDCA, each station maintains four
priorities called Access Categories (ACs). The
quality of service offered to each flow depends on
the AC to which it belongs.
Nevertheless, the granularity of service offered
by 802.11e (4 priorities at most) can not satisfy the
real time flows requirements (where each flow is
characterized by its own delay bound).
Therefore, we propose in this paper a new
medium access mechanism based on the Deadline
Monotonic (DM) policy [9] to schedule real time
flows over 802.11. Indeed DM is a real time
scheduling policy that assigns static priorities to flow
packets according to their deadlines; the packet with
the shortest deadline being assigned the highest
priority. To support the DM policy over 802.11, we
medium access backoff policy. Therefore, we focus
on performance evaluation of the DM policy in terms
of achievable throughput, average MAC layer
service time and MAC layer service time
distribution. Hence, we follow these steps:
 First, we propose a Markov Chain
framework modeling the backoff process of
n contending stations within the same
broadcast region [1].
Due to the complexity of the mathematical
model, we restrict the analysis to n
contending stations belonging to two traffic
categories (each traffic category is
characterized by its own delay bound).
 From the analytical model, we derive the
throughput achieved by each traffic
category.
 Then, we use the generalized Z-transforms
[3] to derive expressions of the average
MAC layer service time and the service
time distribution.
 As the analytical model was restricted to
two traffic categories, analytical results are
extended by simulation to different traffic
categories.
 Finally, we consider a simple multi-hop
scenario to deduce the behavior of the DM
policy in a multi hop environment.

UbiCC Journal - Volume 3 13

 The rest of this paper is organized as follows. In
section 2, we review the state of the art of the IEEE
802.11 DCF, QoS support over 802.11 mainly the
IEEE 80.211e EDCA and real time scheduling over
802.11. In section 3, we present the distributed
scheduling and introduce the new medium access
backoff policy to support DM over 802.11. In section
4, we present our mathematical model based on
Markov chain analysis. Section 5 and 6 present
respectively throughput and the service time
analysis. Analytical results are validated by
simulation using the ns-2 network simulator [16]. In
section 7, we extend our study by simulation, first to
take into consideration different traffic categories,
second, to study the behavior of the DM algorithm in
a multi-hop environment where factors like
interferences or routing protocols exist. Finally, we
conclude the paper in section 8.
2 LITTERATURE REVIEWS
2.1 The 802.11 protocol
2.1.1 Description of the IEEE 802.11 DCF
Using DCF, a station shall ensure that the
channel is idle when it attempts to transmit. Then it
selects a random backoff in the contention
window 0 , CW  1 , where CW is the current
window size and varies between the minimum and
the maximum contention window sizes. If the
channel is sensed busy, the station suspends its
backoff until the channel becomes idle for a
Distributed Inter Frame Space (DIFS) after a
successful transmission or an Extended Inter Frame
Space (EIFS) after a collision. The packet is
transmitted when the backoff reaches zero. A packet
is dropped if it collides after maximum
retransmission attempts.
The above described two way handshaking
packet transmission procedure is called basic access
mechanism. DCF defines a four way handshaking
technique called Request To Send/Clear To Send
(RTS/CTS) to prevent the hidden station problem. A
station S j is said to be hidden from S i if S j is
within the transmission range of the receiver of S i
and out of the transmission range of S i .
2.1.2 Performance evaluation of the 802.11
DCF
Different works have been proposed to evaluate
the performance of the 802.11 protocol based on
Bianchi’s work [1]. Indeed, Bianchi proposed a
Markov chain based analytical model to evaluate the
saturation throughput of the 802.11 protocol. By
saturation conditions, it’s meant that contending
stations have always packets to transmit.
Several works extended the Bianchi model either
to suit more realistic scenarios or to evaluate other
performance parameters. Indeed, the authors of [2]
incorporate the frame retry limits in the Bianchi’s
model and show that Bianchi overestimates the
UbiCC Journal - Volume 3
maximum achievable throughput. The native model
is also extended in [10] to a non saturated
environment. In [12], the authors derive the average
packet service time at a 802.11 node. A new
generalized Z-transform based framework has been
proposed in [3] to derive probabilistic bounds on
MAC layer service time. Therefore, it would be
possible to provide probabilistic end to end delay
bounds in a wireless network.
2.2 Supporting QoS over 802.11
2.2.1 Differentiation mechanisms over 802.11
Emerging applications like audio and video
applications require quality of service guarantees in
terms of throughput delay, jitter, loss rate, etc.
Transmitting such flows over wireless
communications require supporting service
differentiation mechanisms over such networks.
Many medium access schemes have been
proposed to provide some QoS enhancements over
the IEEE 802.11 WLAN. Indeed, [4] assigns
different priorities to the incoming flows. Priority
classes are differentiated according to one of three
802.11 parameters: the backoff increase function, the
Inter Frame Spacing (IFS) and the maximum frame
length. Experiments show that all the three
differentiation schemes offer better guarantees for
the highest priority flow. But the backoff increase
function mechanism doesn’t perform well with TCP
flows because ACKs affect the differentiation
mechanism.
In [7], an algorithm is proposed to provide
service differentiation using two parameters of IEEE
802.11, the backoff interval and the IFS. With this
scheme high priority stations are more likely to
access the medium than low priority ones. The above
described researches led to the standardization of a
new protocol that supports QoS over 802.11, the
IEEE 802.11e protocol [6].
2.2.2 The IEEE 802.11e EDCA
The IEEE 802.11e proposes a new medium
access mechanism called the Enhanced Distributed
Channel Access (EDCA), that enhances the IEEE
802.11 DCF. With EDCA, each station maintains
four priorities called Access Categories (ACs). Each
access category is characterized by a minimum and a
maximum contention window sizes and an
Arbitration Inter Frame Spacing (AIFS).
Different analytical models have been proposed
to evaluate the performance of 802.11e EDCA. In
[17], Xiao extends Bianchi’s model to the prioritized
schemes provided by 802.11e by introducing
multiple ACs with distinct minimum and maximum
contention window sizes. But the AIFS
differentiation parameter is lacking in Xiao’s model.
Recently Osterbo and Al. have proposed
- 14
 different works to evaluate the performance of the
IEEE 802.11e EDCA [13], [14], [15]. They proposed
a model that takes into consideration all the
differentiation parameters of the EDCA especially
the AIFS one. Moreover different parameters of QoS
have been evaluated such as throughput, average
service time, service time distribution and
probabilistic response time bounds for both saturated
and non saturated cases.
Although the IEEE 802.11e EDCA classifies the
traffic into four prioritized ACs, there is still no
guarantee of real time transmission service. This is
due to the lack of a satisfactory scheduling method
for various delay-sensitive flows. Hence, we need a
scheduling policy dedicated to such delay sensitive
flows.
2.3 Real time scheduling over 802.11
A distributed solution for the support of real-
time sources over IEEE 802.11, called Blackburst, is
discussed in [8]. This scheme modifies the MAC
protocol to send short transmissions in order to gain
priority for real-time service. It is shown that this
approach is able to support bounded delays. The
main drawback of this scheme is that it requires
constant intervals for high priority traffic; otherwise
the performance degrades very much.
In [18], the authors introduced a distributed
priority scheduling over 802.11 to support a class of
dynamic priority schedulers such as Earliest
Deadline First (EDF) or Virtual Clock (VC). Indeed,
the EDF policy is used to schedule real time flows
according to their absolute deadlines, where the
absolute deadline is the node arrival time plus the
delay bound.
To realize a distributed scheduling over 802.11,
the authors of [18] used a priority broadcast
mechanism where each station maintains an entry for
the highest priority packet of all other stations. Thus,
stations can adjust their backoff according to other
stations priorities.
The overhead introduced by the broadcast
priority mechanism is negligible. This is due to the
fact that priorities are exchanged using native DATA
and ACK packets. Nevertheless, authors of [18]
proposed a generic backoff policy that can be used
by a class of dynamic priority schedulers no matter if
this scheduler targets delay sensitive flows or rate
sensitive flows.
In this paper, we focus on delay sensitive flows
and propose to support the fixed priority Deadline
Monotonic (DM) policy over 802.11 to schedule
delay sensitive flows. For instance, we use a priority
broadcast mechanism similar to [18] and introduce a
new medium access backoff policy where the
UbiCC Journal - Volume 3
backoff value is inferred from the deadline
information.
3 SUPPORTING DEADLINE MONOTONIC
(DM) POLICY OVER 802.11
With DCF all the stations share the same
transmission medium. Then, the HOL (Head of Line)
packets of all the stations (highest priority packets)
will contend for the channel with the same priority
even if they have different deadlines.
Introducing DM over 802.11 allows stations
having packets with short deadlines to access the
channel with higher priority than those having
packets with long deadlines. Providing such a QoS
requires distributed scheduling and a new medium
access policy.
3.1 Distributed Scheduling over 802.11
To realize a distributed scheduling over 802.11,
we introduce a priority broadcast mechanism similar
to [18]. Indeed each station maintains a local
scheduling table with entries for HOL packets of all
other stations. Each entry in the scheduling table of
 
node S i comprises two fields S j , D j where S j is
the source node MAC address and D j is the
deadline of the HOL packet of node S j . To
broadcast HOL packets deadlines, we propose to use
the two way handshake DATA/ACK access mode.
When a node S i transmits a DATA packet, it
piggybacks the deadline of its HOL packet. Nodes
hearing the DATA packet add an entry for S i in
their local scheduling tables by filling the
corresponding fields. The receiver of the DATA
packet copies the priority of the HOL packet in ACK
before sending the ACK frame. All the stations that
did not hear the DATA packet add an entry for S i
using the information in the ACK packet.
3.2 DM medium access backoff policy
Let’s consider two stations S 1 S2and
transmitting two flows with the same deadline D1
( D1 is expressed as a number of 802.11 slots). The
two stations having the same delay bound can access
the channel with the same priority using the native
802.11 DCF.
Now, we suppose that S 1 and S 2 transmit flows
with different delay bounds D1 and D2 such as
D1  D2 , and generate two packets at time instants
t 1 and t 2 . If S 2 had the same delay bound as S 1 ,
its packet would have been generated at time t '2 such
as t '2  t 2  D 21 , where D 21  D 2  D1  .
At that time, S 1 and S 2 would have the same
priority and transmit their packets according to the
- 15
 802.11 protocol.
Thus, to support DM over 802.11, each station
uses a new backoff policy where the backoff is given
by:
 The random backoff selected in 0 , CW  1
according to 802.11 DCF, referred as BAsic
Backoff (BAB).
 The DM Shifting Backoff (DMSB):
corresponds to the additional backoff slots that
a station with low priority (the HOL packet
having a large deadline) adds to its BAB to
have the same priority as the station with the
highest priority (the HOL packet having the
shortest deadline).
Whenever a station S i sends an ACK or hears
an ACK on the channel its DMSB is revaluated as
follows:
DMSBS i   DeadlineHOLS i   DTmin S i  (1)
Where DTmin S i  is the minimum of the HOL
packet deadlines present in S i scheduling table and
DeadlineHOLS i  is the HOL packet deadline of
node S i .
Hence, when S i has to transmit its HOL packet
with a delay bound Di , it selects a BAB in the
contention window 0 , CW min  1 and computes the
WHole Backoff (WHB) value as follows:
WHBS i   DMSBS i   BAB S i  (2)
The station S i  decrements its BAB when it
senses an idle slot. Now, we suppose that S i senses
the channel busy. If a successful transmission is
heard, then S i  revaluates its DMSB when a correct
ACK is heard. Then the station S i adds the new
DMSB value to its current BAB as in equation (2).
Whereas, if a collision is heard, S i reinitializes its
DMSB and adds it to its current BAB to allow
colliding stations contending with the same priority
as for their first transmission attempt. S i transmits
when its WHB reaches 0. If the transmission fails, S i
doubles its contention window size and repeats the
above procedure until the packet is successfully
transmitted or dropped after maximum
retransmission attempts.
4 MATHEMATICAL MODEL OF THE DM
POLICY OVER 802.11
In this section, we propose a mathematical
model to evaluate the performance of the DM policy
using Markov chain analysis [1]. We consider the
UbiCC Journal - Volume 3
following assumptions:
Assumption 1:
The system under study comprises n contending
stations hearing each other transmissions.
Assumption 2:
Each station S i transmits a flow Fi with a delay
bound Di . The n stations are divided into two
traffic categories C1 and C 2 such as:
 C1 represents n1 nodes transmitting flows
with delay bound D1 .
 C 2 represents n 2 nodes transmitting flows
with delay bound D2 , such as D1  D2 ,
D 21  D 2  D1  and n1  n 2   n .
Assumption 3:
We operate in saturation conditions: each station has
immediately a packet available for transmission after
the service completion of the previous packet [1].
Assumption 4:
A station selects a BAB in a constant contention
window 0 ,W  1 independently of the transmission
attempt. This is a simplifying assumption to limit the
complexity of the mathematical model.
Assumption 5:
We are in stationary conditions, i.e. the n stations
have already sent one packet at least.
Depending on the traffic category to which it
belongs, each station S i will be modeled by a
Markov Chain representing its whole backoff (WHB)
process.
4.1 Markov chain modeling a station of category
C1
Figure 1 illustrates the Markov chain modeling a
station S 1 of category C1 . The states of this Markov
chain are described by the following quadruplet
R , i , i  j , D21  where:
 R : takes two values denoted by C 2 and
~ C 2 . When R ~ C 2 , the n 2 stations of
category C 2 are decrementing their shifting
backoff (DMSB) during D21 slots and
wouldn’t contend for the channel. When
R  C 2 , the D21 slots had already been
elapsed and stations of category C 2 will
contend for the channel.
 i : the value of the BAB selected by S 1 in
0 ,W  1 .
- 16
 Figure 1: Markov chain modeling a category C1 Station

 i  j  : corresponds to the current backoff moves to the state ~ C 2 , i  j , i  j , D 21 ,

of the station S 1 .
 D21 : corresponds to D 2  D1  . We choose
the negative notation  D21 for stations of
C1 to express the fact that only stations of
category C 2 have a positive DMSB equal to
D21 .
Initially S 1 selects a random BAB and is in
one of the states ~ C 2 , i , i , D 21  , i  0..W  1 .
During D21  1 slots, S 1 decrements its backoff if
none of the n1  1 remaining stations of category
C1 transmits. Indeed, during these slots, the n 2
stations of category C 2 are decrementing their
DMSB and wouldn’t contend for the channel.
When S1 is in one of the states
~ C 2 , i , i  D21  1, D21  , i  D 21 ..W  1 and
senses the channel idle, it decrements its th
D21 slot.
But S 1 knows that henceforth the n 2 stations of
category C 2 can contend for the channel (the D21
slots had been elapsed). Hence, S 1 moves to one of
the states C 2 , i , i  D21 , D 21  , i  D21 ..W  1 .
However, when the station S 1 is in one of the
states ~ C 2 , i , i  j , D 21  , for i  1..W  1 ,
j  0.. minD 21  1, i  1 and at least one of the
n1  1 remaining stations of category
C1
transmits, then the stations of category C 2 will
reinitialize their DMSB and wouldn’t contend for
channel during additional D21 slots. Therefore, S 1
i  1..W  1 , j  0.. minD21  1, i  1 .
Now, If S1 is in one of the states
C 2 , i , i  D21 , D21  , i  D21  1..W  1 and at
least one of the n  1 remaining stations (either a
category C1 or a category C 2 station) transmits,
then S1 moves to one of the states
~ C 2 , i  D21 , i  D21 , D21  , i  D21  1..W  1 .
Markov chain modeling a station of
4.2
category C2
Figure 2 illustrates the Markov chain modeling
a station S 2 of category C 2 . Each state of S 2
Markov chain is represented by the quadruplet
i , k , D21  j , D21  where:
 i : refers to the BAB value selected by S 2 in
0 ,W  1 .
 k : refers to the current BAB value of S 2 .
 D 21  j : refers to the current DMSB of S 2 ,
j  0 , D21  .
 D21 : corresponds to D  D  .
2 1
When S 2 selects a BAB, its DMSB equals
D 21 and is in one of the states i , i , D 21 , D 21  ,
i  0..W  1 . During D 21 slots, only the n1
stations of category C1 contend for the channel.
If S 2 senses the channel idle during D21 slots, it
moves to one of the states i , i ,0 , D 21  , i  0..W  1 ,
where it ends its shifting backoff.

UbiCC Journal - Volume 3 - 17

 Figure 2: Markov chain modeling a category C 2 Station

When S 2 is in one of the states i , i ,0 , D 21  ,  2  i , i , D 21  j , D 21 , i  0..W  1,

i  0..W  1 , the n 2  1 other stations of category j  0..D 21  1
C 2 have also decremented their DMSB and can   2 : the set of states of S 2 , where stations
contend for the channel. Thus, S 2 decrements its of category C 2 contend for the channel
BAB and moves to the state i , i  1,0 , D 21  , (pink states in figure 2).
i  2..W  1 , only if none of the n  1 remaining  2  i , i ,0 , D21 , i  0..W  1
stations transmits.  i , i  1,0 , D 21 , i  2..W  1

If S 2 is in one of the states i , i  1,0 , D 21  , Therefore, when stations of category C1 are in

i  2..W  1 , and at least one of the n  1 one the states of  1 , stations of category C 2 are in
remaining stations transmits, the n 2 stations of one of the states of  2 . Similarly, when stations of
category C 2 will reinitialize their DMSB and S 2 category C1 are is in one of the states of  1 ,
moves to the state i  1, i  1, D21 , D21  , stations of category C 2 are in one of the states of
i  2..W  1 .
2.
4.3 Blocking probabilities in the Markov chains Hence, we derive the expressions of S1
According to the explanations given in blocking probabilities p 11 and p 12 shown in
paragraphs 4.1 and 4.2, the states of the Markov figure 1 as follows:
chains modeling stations S 1 and S 2 can be divided
into the following groups:  p 11 : the probability that S 1 is blocked given
that S 1 is in one of the states of  1 . p 11 is
  1 : the set of states of S 1 where none of the
the probability that at least a station S 1' of
n 2 stations of category C 2 contends for the
the other n1  1 stations of C1 transmits
channel (blue states in figure 1).
 1  ~ C 2 , i , i  j , D 21 , i  0..W  1, given that S 1' is in one of the states of  1 .
j  0.. minmax0 , i  1, D 21  1 p 11  1  1   11 n1 1 (3)
where  11 is the probability that a station S 1'
  1 : the set of states of S 1 where stations of
of C1 transmits given that S 1' is in one of
category C 2 can contend for the channel
the states of  1 :
(pink states in figure 1).
 1  C 2 , i , i  D 21 , D 21 , i  D 21 ..W  1 
 11  Pr S 1' transmits  1 
~ C2 ,0 ,0 , D21 
1
  2 : the set of states of S 2 where stations of  (4)
W 1  min max 0 ,i 1,D21 1 
category C 2 do not contend for the channel  

i 0 
  1~ C2 ,i ,i  j , D21  

(blue states in figure 2). j 0 

UbiCC Journal - Volume 3 - 18

  1R ,i ,i  j , D21  is defined as the probability of
 p 22 : the probability that S 2 is blocked
the state R , i , i  j , D21 , in the stationary
given that S 2 is in one of the states of  2 .
conditions and  1   1  R ,i ,i  j , D21 
 is the p 22  1  1   12 n1 1   22 n2 1 (9)
probability vector of a category C1 station.
The blocking probabilities described above
 p 12 : the probability that S 1 is blocked allow deducing the transition state probabilities and
given that S 1 is in one of the states of  1 . having the transition probability matrix Pi , for a
p 12 is the probability that at least a station station of traffic category C i .
S 1' of the other n1  1 stations of C1 Therefore, we can evaluate the state
probabilities by solving the following system [11]:
transmits given that S 1' is in one of the states
of  1 or at least a station S '2 of the n 2  i Pi   i

stations of C 2 transmits given that S '2 is in 
  ij  1
 j
(10)
one of the states of  2 . 
p 12  1  1   12 n1 1 1   22 n2 (5) 4.4 Transition probability matrices
4.4.1 Transition probability matrix of a
where  12 is the probability that a station category C1 station
S 1' of C1 transmits given that S 1' is in one Let P1 be the transition probability matrix of
the station S 1 of category C1 . P1 i , j is the
of the states of  1 .

 12  Pr S 1' transmits  1  probability to transit from state i to state j . We
have:
C2 ,D21 ,0 , D21 
1
 W 1
(6) P1 ~ C 2 , i , i  j , D 21 , ~ C 2 , i , i   j  1, D 21 

i  D21
1C2 ,i ,i  D21 , D21 
 1  p11 , i  2..W  1, j  0.. mini  2 , D 21  2 
(11)
P1 ~ C 2 , i ,1, D 21 , ~ C 2 ,0 ,0 , D 21   1  p11 ,
and  22 the probability that a station S 2' of i  1.. minW  1, D 21  1
C 2 transmits given that S 2' is in one of the (12)
states of  2 . P1 ~ C 2 , i , i  D 21  1, D 21 , C 2 , i , i  D 21 , D 21 
 1  p 11 , i  D 21 ..W  1

 12  Pr S 2' transmits  2  (13)
2 0 ,0 ,0 ,D21  P1~ C2 , i , i  j , D21 , ~ C2 , i  j , i  j , D21 
(14)
 W 1 W 1
(7)  p11 , i  2..W  1, j  1.. mini  1, D21  1
 i ,i ,0 ,D21 
2   i ,i 1,0 ,D21 
2
i 0 i2 P1~ C2 , i , i , D21 , ~ C2 , i , i , D21   p11 ,
(15)
i  1..W  1
 2i ,k ,D21  j ,D21  is defined as the probability
of the state i , k , D21  j , D 21 , in the P1C2 ,i ,i  D21 , D21 ,~ C2 ,i  D21 ,i  D21 , D21 
 p12 ,i  D21  1..W  1
stationary condition.  2   2  i ,k ,D21  j ,D21 
 (16)
is the probability vector of a category C 2
station. P1C2 ,i ,i  D21 , D21 ,C2 ,i  1,i  1  D21 , D21 
 1  p12 ,i  D21  1..W  1
In the same way, we evaluate p 21 and p 22 the
(17)
blocking probabilities of station S 2 shown in
figure 2: 1
P1~ C2 ,0 ,0 , D21 , ~ C2 , i , i , D21   ,
 p 21 : the probability that S 2 is blocked W (18)
given that S 2 is in one of the states of  2 . i  0..W  1
p 21  1  1   11  n1
(8)
If D 21  W  then:

UbiCC Journal - Volume 3 - 19

 P1C2 , D21 ,0 , D21 , ~ C2 , i , i , D21  
1
,  11  f  11 , 12 , 22 
W (19)   f  , , 
 12 11 12 22
i  0..W  1
 22  f  11 , 12 , 22 
under the constraint
By replacing p11 and p 12 by their values in 
equations (3) and (5) and by replacing P1 and  1  11  0 , 12  0 , 22  0 , 11  1, 12  1, 22  1
in (10) and solving the resulting system, we can (28)
R ,i ,i  j , D21 
express  1 as a function of  11 ,  12 and Solving the above system (28), allows deducing
 22 given respectively by equations (4), (6) and the expressions of  11 ,  12 and  22 , and deriving
(7). the state probabilities of Markov chains modeling
category C1 and category C 2 stations.
4.4.2 Transition probability matrix of a
category C2 station
Let P2 be the transition probability matrix of 5 THROUGHPUT ANALYSIS
the station S 2 belonging to the traffic category C 2 .
The transition probabilities of S 2 are: In this section, we propose to evaluate Bi , the
normalized throughput achieved by a station of
P2 i , i , D21  j , D21 , i , i , D21   j  1, D21  traffic category C i [1]. Hence, we define:
(20)
 1  p21 , i  0..W  1, j  0..D21  1
 Pi ,s : the probability that a station Si
P2 i , i , D21  j , D21 , i , i , D21 , D21   p21 , belonging to traffic category C i transmits a
(21)
i  0..W  1, j  0..D21  1 packet successfully. Let S 1 and S 2 be two
stations belonging respectively to traffic
P2 i , i ,0 , D21 , i , i  1,0 , D21   1  p22 , categories C1 and C 2 . We have:
(22)
i  2..W  1
P1,s  Pr S1 transmits successfully 1  Pr1 
P2 1,1,0 , D21 , 0 ,0 ,0 , D21   1  p22 (23)  Pr S1 transmits successfully  1  Pr 1 
  11 1  p11  Pr 1    12 1  p12  Pr  1 
P2 i , i ,0 , D21 , i , i , D21 , D21   p22 ,
(24) (29)
i  1..W  1
P2 ,s  Pr S 2 transmits successfully  2  Pr  2 
P2 i , i  1,0 , D21 , i  1, i  1, D21 , D21   p22 ,
(25)  Pr S 2 transmits successfully  2  Pr 2 
i  2..W  1
  22 1  p 22  Pr  2 
P2 i , i  1,0 , D21 , i  1, i  2 ,0 , D21   1  p22 , (30)
(26)
i  3..W  1  Pidle : the probability that the channel is idle.
1
P2 0 ,0 ,0 , D21 , i , i , D21 , D21   , i  0..W  1 (27)
W The channel is idle if the n1 stations of
category C1 don’t transmit given that these stations
By replacing p 21 and p 22 by their values in are in one of the states of  1 or if the n stations
equations (8) and (9) and by replacing P2 and  2 (both category C1 and category C 2 stations) don’t
in (10) and solving the resulting system, we can
transmit given that stations of category C1 are in
i ,k ,D21  j ,D21 
express  2 as a function of  11 ,  12 one of the states of  1 . Thus:
and  22 given respectively by equations (4), (6)
R ,i ,i  j , D21 
and (7). Moreover, by replacing  1 and Pidle  1   11 n1 Pr  1   1   12 n1 1   22 n2 Pr  1 
 2i ,k ,D21  j ,D21  by their values, in equations (4), (6) (31)

and (7), we obtain a system of non linear equations Hence, the expression of the throughput of a
as follows: category C i station is given by:

UbiCC Journal - Volume 3 - 20

 Pi ,s T p 8 and we depict the throughput achieved by the
Bi  different stations present in the network as a
 2 
PIdle Te  Ps T s   1  PIdle 


n P
i 1
i i ,s
Tc


function of the contention window size W ,
D 21  1 . We notice that the throughput achieved
(32) by category C1 stations (stations numbered from
S 11 to S 14 ) is greater than the one achieved by
Where Te denotes the duration of an empty
category C 2 stations (stations numbered from S 21
slot, Ts and Tc denote respectively the duration of
to S 24 ).
a successful transmission and a collision.
 2 
 1  PIdle 



i 1
ni Pi ,s 


corresponds to the

probability of collision. Finally T p denotes the

average time required to transmit the packet data
payload. We have:

 
Ts  T PHY  TMAC  T p  T D  SIFS 
(33)
TPHY  T ACK  T D   DIFS


Tc  TPHY  TMAC  T p  TD  EIFS  (34)

Where T PHY , TMAC and T ACK are the

durations of the PHY header, the MAC header and
the ACK packet [1], [13]. TD is the time required to
transmit the two bytes deadline information.
Stations hearing a collision wait during EIFS before
resuming their packets.
For numerical results stations transmit 512
bytes data packets using 802.11.b MAC and PHY
layers parameters (given in table 1) with a data rate
equal to 11Mbps. For simulation scenarios, the
propagation model is a two ray ground model. The
transmission range of each node is 250m. The
distance between two neighbors is 5m. The EIFS
parameter is set to ACKTimeout as in ns-2, where:
ACKTimeout  DIFS  T PHY  T ACK  T D   SIFS
(35)
Table 1: 802.11 b parameters.
Figure 3: Normalized throughput as a function of
the contention window size D 21  1, n  8 
Analytically, stations belonging to the same
traffic category have the same throughput given by
equation (32). Simulation results validate analytical
results and show that stations belonging to the same
traffic category (either category C1 or category
C 2 ) have nearly the same throughput. Thus, we
conclude the fairness of DM between stations of the
same category.
For subsequent throughput scenarios, we focus
on one representative station of each traffic
category. Figure 4, compares category C1 and
category C 2 stations throughputs to the one
obtained with 802.11.

11 Mb/s Curves are represented as a function of W and

Data Rate
Slot 20 µs for different values of D 21 . Indeed as D 21
SIFS 10 µs increases, the category C1 station throughput
DIFS 50 µs increases, whereas the category C 2 station
PHY Header 192 µs throughput decreases. Moreover as W increases,
MAC Header 272 µs the difference between stations throughputs is
ACK 112 µs reduced. This is due to the fact that the shifting
Short Retry Limit 7 backoff becomes negligible compared to the
contention window size.
For all the scenarios, we consider that we are in
n Finally, we notice that the category C1 station
presence of n contending stations with stations obtains better throughput with DM than with
2
for each traffic category. In figure 3, n is fixed to

UbiCC Journal - Volume 3 - 21

 802.11, but the opposite scenario happens to the
category C 2 station.
expression of the average service time and the
service time distribution. The service time depends
on the duration of an idle slot Te , the duration of a
successful transmission Ts and the duration of a
collision Tc [1], [3],[14]. As Te is the smallest
duration event, the duration of all events will be
T 
given by  event  .
T
 e 

6.1 Z-Transform of the MAC layer service time

6.1.1 Service time Z-transform of a category

Figure 4: Normalized throughput as a function of
the contention window size (different D 21 values)
In figure 5, we generalize the results for
different numbers of contending stations and fix the
contention window size W to 32.
C1 station:
Let TS 1 Z  be the service time Z-transform of
a station S 1 belonging to traffic category C1 . We
define:
H 1R ,i ,i  j , D21  Z  : The Z-transform of the
time already elapsed from the instant S 1 selects a
basic backoff in 0 ,W  1 (i.e. being in one of the
states ~ C 2 , i , i , D21  ) to the time it is found in the
state R , i , i  j , D 21  .
Moreover, we define:

11
 Psuc : the probability that S 1 observes a
successful transmission on the channel,
while S 1 is in one of the states of  1 .
11
Psuc  n1  1 11 1   11 n1  2 (36)

12
 Psuc : the probability that S 1 observes a
successful transmission on the channel,
while S 1 is in one of the states of  1 .
Figure 5: Normalized throughput as a function of 12
Psuc  n1  1 12 1   12 n1  2 1   22 n2
the number of contending stations (37)
 n2 22 1   22 n2 1 1   12 n1 1
All the curves show that DM performs service
differentiation over 802.11 and offers better We evaluate H 1R ,i ,i  j , D21  Z  for each state
throughput for category C1 stations independently
of S 1 Markov chain as follows:
of the number of contending stations.

  Ts 
1  11  Te 
6 SERVICE TIME ANALYSIS H 1~ C2 ,i ,i , D21 Z    Psuc Z 
W 

In this section, we evaluate the average MAC
 Tc  
layer service time of category C1 and category C 2   min i  D21 1,W 1

stations using the DM policy. The service time is 

p11  Psuc 11
 T 
Z e   H 1~ C 2 ,k ,i , D21  Z 
the time interval from the time instant that a packet  k i 1

becomes at the head of the queue and starts to   Ts   Tc  

 12  Te 
 
 
contend for transmission to the time instant that
 Ĥ 1C 2 ,i  D21 ,i , D21  Z  Psuc Z 12 T
 p11  Psuc Z  e  
either the packet is acknowledged for a successful  
transmission or dropped [3].  
(38)
We propose to evaluate the Z-Transform of the Where:
MAC layer service time [3], [14], [15] to derive an

UbiCC Journal - Volume 3 - 22

  Ĥ 1C ,i  D ,i , D  Z   H 1C ,i  D ,i , D  Z   Ts 

1  p11 H 1~ C ,0 ,0 , D

 
 TS1 Z   Z Z 
2 21 21 2 21 21
 Te 
 if i  D 21   W  1 (39) 2 21 

 m   c
T 
 Ĥ 1C2 ,i  D21 ,i , D21  Z   0 Otherwise  1  p12 H 1C2 ,D21 ,0 , D21  Z      Te 
Z  p11H 1~C ,0 ,0 , D Z 
2 21 

i 0
We also have: 
 p12 H 1C 2 ,D21 ,0 , D21  Z  i

1  p11 Z  j H 1~ C2 ,i ,i , D21  Z  m 1

H 1~ C2 ,i ,i  j , D21  Z     Tc 
 T


 Ts   Tc 

  Z  e  p11 H 1~ C2 ,0 ,0 , D21 Z   p12 H 1C 2 ,D21 ,0 , D21  Z   
 
   
11  Te  11 T
1 Psuc Z  p11  Psuc Z  e   
 
i  2..W  1, j  1..mini  1, D21  1
(40) (44)

1  p11 Z D21 H 1~ C2 ,i ,i , D21 Z  6.1.2 Service time Z-transform of a category
H 1C2 ,i ,i  D21 , D21  Z   C2 station:
 Ts   Tc 

  In the same way, let TS 2 Z  be the service

   
11
1  Psuc Z  Te   p11  11 T
Psuc Z  e 
 1  p12 ZH 1C2 ,i 1,i 1 D21 , D21  Z ,i  D21 ..W  2 time Z-transform of a station S 2 of category C 2 .
We define:
(41)
H 2i ,k ,D21  j ,D21  Z  : The Z-transform of the
H 1C2 ,W 1,W 1 D21 , D21 Z  time already elapsed from the instant S 2 selects a
1  p11 Z  D21
H 1~ C 2 ,W 1,W 1, D21  Z  basic backoff in 0 ,W  1 (i.e. being in one of the
 (42)
 Ts   Tc  states i , i , D 21 , D 21  ) to the time it is found in the
 
   
11
1  Psuc Z  Te  11
 p11  Psuc Z  Te  state i , k , D21  j , D21  .
Moreover, we define:
1  p11 ZH 1~ C2 ,1,1, D21  Z 
H 1~ C2 ,0 ,0 , D21  Z   21
 Ts   Tc   Psuc : the probability that S 2 observes a
 
   
11  Te  11 T
1 Psuc Z  p11  Psuc Z e  successful transmission on the channel,
min W 1,D21 1 while S 2 is in one of the states of  2 .
1
 1  p11 Z  H 1
i2
~ C 2 ,i ,1, D21  Z  
W
21
Psuc  n1 11 1   11 n1 1 (45)
(43)
22
 Psuc : the probability that S 2 observes a
If S 1 transmission state is ~ C 2 ,0 ,0 , D 21  , successful transmission on the channel,
the transmission will be successful only if none of while S 2 is in one of the states of  2 .
the n1  1 remaining stations of C1 transmits. 22
Psuc  n1 12 1   12 n1 1 1   22 n2 1
Whereas when the station S 1 transmission state is (46)
 n2  1 22 1   22 n2  2 1   12 n1
C 2 , D21 ,0 , D 21  , the transmission occurs
successfully only if none of n  1 remaining We evaluate H 2i ,i ,D21  j ,D21  Z  for each state
stations (either a category C1 or a category C 2
of S 2 Markov chain as follows:
station) transmits.
1
H 2i ,i ,D21 ,D21  Z   , i  0 and i  W  1 (47)
If the transmission fails, S 1 tries another W
transmission. After m retransmissions, if the   Ts 
1  22  Te 
packet is not acknowledged, it will be dropped. H 2i ,i ,D21 ,D21  Z     Psuc Z 
Thus, the Z-transform of station S 1 service time is: W 

 Tc  
 Z 
 
 Te 
 H 2i 1,i ,0 ,D21  Z , i  1..W  2
22
p 22  Psuc


(48)

UbiCC Journal - Volume 3 - 23

 To compute H 2i ,i ,D21  j ,D21  Z  , we define   Tc  
m 1
   
Z  , such as: TS 2 Z    p 22 Z H 20 ,0 ,0 ,D21  Z 
 Te 
j
Tdec 
 
 
0
Tdec Z   1 (49)  Ts    Tc  
i
  m
   
1  p 21 Z
1  p 22 Z  Te  H 2

0 ,0 ,0 ,D21  Z   p 22 Z

 Te 
H 20 ,0 ,0 ,D21  Z 

j
Tdec Z   i 0
 
  Ts   Tc  
 21  Te 
 
  (55)
1   Psuc Z 21 T j 1
 p 21  Psuc Z  e  Tdec Z 
 
 
6.2 Average Service Time
for j  1..D 21
From equations (44) (respectively equation
(50) (55)), we derive the average service time of a
category C1 station ( respectively a category C 2
So:
station). The average service time of a category C i
station is given by:
H 2i ,i ,D21  j ,D21  Z   H 2i ,i ,D21  j 1,D21  Z  j
Tdec Z ,
X i  TS i1 1 (56)
i  0..W  1, j  1..D21 , i , j   0 , D 21 
(51)
And:
Where TS i1 Z  , is the derivate of the service
H 2i ,i 1,0 ,D21  Z   1  p 22 ZH 2i 1,i ,0 ,D21  Z  time Z-transform of a category C i station [11].
1  p 22 ZH 2i ,i ,0 ,D21  Z  By considering the same configuration as in

  Ts   Tc   figure 3, we depict in figure 5, the average service
 22  Te 
 
 
1   Psuc Z 22 T D21
 p 22  Psuc Z  e  Tdec Z  time of category C1 and category C 2 stations as a
  function of W . As for the throughput analysis,
 
stations belonging to the same traffic category have
i  2..W  2 nearly the same average service value. Simulation
(52) service time values coincide with analytical values
given by equation (56). These results confirm the
H 2W 1,W  2 ,0 ,D21  Z  fairness of DM in serving stations of the same
1  p 22 ZH 2W 1,W 1,0 ,D21  Z  category.

  Ts   Tc   (53)
 22  Te 
 
 
Tdec Z 
22  Te  D21
1   Psuc Z  p 22  Psuc Z
 
 

According to figure 2 and using equation (51),

we have:

H 20 ,0 ,0 ,D21  Z   H 20 ,1,0 ,D21  Z Tdec

D21
Z 
1  p 22 ZH 21,1,0 ,D21  Z 

  Ts   Tc   (54)
 22  Te 
 
 
Tdec Z 
22  Te  D21
1   Psuc Z  p 22  Psuc Z
 
 
Figure 6: Average service time as a function of the
contention window size D 21  1, n  8 
Therefore, we can derive an expression of S 2
Z-transform service time as follows:
In figure 7, we show that category C1 stations
obtain better average service time than the one
obtained with 802.11 protocol. Whereas, the
opposite scenario happens for category C 2 stations

UbiCC Journal - Volume 3 - 24

 independently of n , the number of contending
stations in the network.
Figure 7: Average service time as a function of the
number of contending stations
6.3 Service Time Distribution
Service time distribution is obtained by
inverting the service time Z transforms given by
equations (44) and (55). But we are most interested
in probabilistic service time bounds derived by
inverting the complementary service time Z
transform given by [11]:
~ 1  TS i Z 
X i Z   (56)
1 Z
In figure 8, we depict analytical and simulation
values of the complementary service time
distribution of a category C1 and a category C 2
stations for different values of D21 and W  32  .
Figure 8: Complementary service time distribution
for different values of D21 , W  32 
All the curves drop gradually to 0 as the delay
increases. Category C1 stations curves drop to 0
faster than category C 2 curves. Indeed, when
D21  4 slots, the probability that S 1 service time
UbiCC Journal - Volume 3
exceeds 0.01s equals 0.2%. Whereas, station S 2
service time exceeds 0.01s with the probability
57,6%. Thus, DM offers better service time
guarantees for the stations with the highest priority.
In figure 9, we double the size of the contention
window size and set it to 64. We notice that
category C1 and category C 2 stations service time
curves become closer. Indeed, when W becomes
large, the BAB values increase and the DMSB
becomes negligible compared to the basic backoff.
The whole backoff values of S 1 and S 2 become
closer and their service time accordingly.
Figure 9: Complementary service time distribution
for different values of D21 ( W  64 )
In figure 10, we depict the complementary
service time distribution for both category C1 and
category C 2 stations and for different values of n ,
the number of contending nodes.
Figure 10: Complementary service time
distribution for different values of the contending
stations
Analytical and simulation results show that
complementary service time curves drop faster
when the number of contending stations is small for
both category C1 and category C 2 stations. This
means that all stations service time increases as the
- 25
 number of contending nodes increases.
7 EXTENTIONS OF THE ANAYTICAL
RESULTS BY SIMULATION
The mathematical analysis undertaken above
showed that DM performs service differentiation
over 802.11 protocol and offers better QoS
guarantees for highest priority stations
Nevertheless, the analysis was restricted to two
traffic categories. In this section, we first generalize
the results by simulation for different traffic
categories. Then, we consider a simple multi-hop
and evaluate the performance of the DM policy
when the stations belong to different broadcast
regions.
7.1 Extension of the analytical results
In this section, we consider n stations
contending for the channel in the same broadcast
region. The n stations belong to 5 traffic categories
where n  5m and m is the number of stations of
the same traffic category. A traffic category C i is
characterized by a delay bound
Dij  Di  D j is the difference between the
deadline values of category C i and category C j
stations. We have:
Dij  i  j K
Where K is the deadline multiplicity factor
and is given by:
Di 1,i  Di 1  Di  K (58)
Indeed, when K varies, the Dij the difference
between deadline values of category C i
category Cj stations also varies.
belonging to the traffic category C i are numbered
from S i 1 to S im .
Figure 11: Normalized throughput for different
traffic category stations
In figure 11, we depict the throughput achieved
UbiCC Journal - Volume 3
by different traffic categories stations as a function
of the minimum contention window size CW min
such as CW min is always smaller than CW max ,
CW max  1024 and K =1.
Analytical and simulation results show that
throughput values increase with stations priorities.
Indeed, the station with the lowest delay bound has
the maximum throughput.
Moreover, figure 12 shows that stations
belonging to the same traffic category have the
same throughput. For instance, when n is set to 15
(i.e. m  3 ), the three stations each traffic category
have almost the same throughput.
Di , and
(57) Figure 12: Normalized throughput: different
stations belonging to the same traffic category
In figure 13, we depict the average service
time of the different traffic categories stations as a
function of K , the deadline multiplicity factor. We
notice that the highest priority station average
and service time decreases as the deadline multiplicity
Stations factor increases. Whereas, the lowest priority
station average service time increases with K .
Figure 13: Average service time as a function of
the deadline multiplicity factor K
In the same way, the probabilistic service time
bounds offered to S 11 (the highest priority station)
- 26
 are better than those offered to station S 51 (the
lowest priority station). Indeed, the probability that
S 11 service time exceeds 0.01s=0.3%. But, station
S 51 service time exceeds 0.01s with the probability
of 36%.
Figure 14: Complementary service time
distribution CWmin  32 , n  8 
The above results generalize the analytical
model results and show once again that DM
performs service differentiation over 802.11 and
offer better guarantees in terms of throughput,
average service time and probabilistic service time
bounds for flows with short deadlines.
7.2 Simple Multi hop scenario
In the above study, we considered that
contending stations belong to the same broadcast
region. In reality, stations may not be within one
hop from each other. Thus a packet can go through
several hops before reaching its destination. Hence,
factors like routing protocols or interferences may
preclude the DM policy from working correctly.
In the following paragraph, we evaluate the
performance of the DM policy in a multi-hop
environment. Hence, we consider a 13 node simple
mtlti-hop scenario described in figure 15. Six flows
are transmitted over the network.
Figure 15: Simple multi hop scenario
UbiCC Journal - Volume 3
Flows packets are routed using the Ad-hoc On
Demand (AODV) protocol. Flows F1 and F2 are
respectively transmitted by stations S 1 and S 2
with delay bounds D1 and D2 and
D 21  D 2  D1 =5 slots. Flows F3 and F4 are
transmitted respectively by S 3 and S 4 and have
the same delay bound. Finally, F5 and F6 are
transmitted respectively by S 5 and S 6 with delay
bounds D5 and D6 and D65  D6  D5 = 4 slots.
Figure 16 shows that the throughput achieved
by F1 is smaller than the one achieved by F2 .
Indeed, both flows cross nodes 6 and 7, where F1
got a higher priority to access the medium than F2
when the DM policy is used. We obtain the same
results for flows F and F . Flows F3 and F4
5 6
have almost the same throughput since they have
equal deadlines.
Figure 16: Normalized throughput using DM
policy
Figure 17 show that the complementary
service time distribution curves drop to 0 faster for
flow F1 than for flow F2 .
Figure 17: End to end complementary service time
distribution
- 27
 The same behavior is obtained for flow F5 and
F6 , where F5 has the shortest delay bound.
Hence, we conclude that even in a multi-hop
environment, the DM policy performs service
differentiation over 802.11 and provides better QoS
guarantees for flows with short deadlines.
8 CONCLUSION
In this paper we proposed to support the DM
policy over 802.11 protocol. Therefore, we used a
distributed backoff scheduling algorithm and
introduced a new medium access backoff policy.
Then we proposed a Markov Chain based
mathematical model to evaluate the performance of
the DM policy in terms of throughput , average
medium access delay and medium access delay
distribution. Analytical and simulation results
showed that DM performs service differentiation
over 802.11 and offers better guarantees in terms of
throughput, average service time and probabilistic
service time bounds for flows with small deadlines.
Moreover, DM achieves fairness between stations
belonging to the same traffic category.
Then, we extended by simulation the analytical
results obtained for two traffic categories to
different traffic categories. Simulation results
showed that even if contending stations belong to
K traffic categories, K  2 , the DM policy offers
better QoS guarantees for highest priority stations.
Finally, we considered a simple multi-hop scenario
and concluded that factors like routing messages or
interferences don’t impact the behavior of the DM
policy and DM still provides better QoS guarantees
for stations with short deadlines.
9 REFERENCES
G. Bianchi: Performance Analysis of the IEEE
802.11 Distributed Coordination Function,
IEEE J-SAC Vol. 18 N. 3, (March 2000).
H. Wu1, Y. Peng, K. Long, S. Cheng, J. Ma:
Performance of Reliable Transport Protocol
over IEEE 802.11 Wireless LAN: Analysis and
Enhancement, In Proceedings of the IEEE
INFOCOM’02, June 2002.
H. Zhai, Y. Kwon, Y., Fang: Performance Analysis
of IEEE 802.11 MAC protocol in wireless
LANs, Wireless Computer and Mobile
Computing, (2004).
I. Aad and C. Castelluccia: Differentiation
mechanisms for IEEE 802.11, In Proc. of IEEE
Infocom 2001, (April 2001).
IEEE 802.11 WG: Part 11: Wireless LAN Medium
Access Control (MAC) and Physical Layer
UbiCC Journal - Volume 3
(PHY) specification, IEEE (1999).
IEEE 802.11 WG: Draft Supplement to Part 11:
Wireless Medium Access Control (MAC) and
physical layer (PHY) specifications: Medium
Access Control (MAC) Enhancements for
Quality of Service (QoS), IEEE 802.11e/D13.0,
(January 2005).
J. Deng, R. S. Chang: A priority Scheme for IEEE
802.11 DCF Access Method, IEICE
Transactions in Communications, vol. 82-B,
no. 1, (January 1999).
J.L. Sobrinho, A.S. Krishnakumar: Real-time traffic
over the IEEE 802.11 medium access control
layer, Bell Labs Technical Journal, pp. 172-
187, (1996).
J. Y. T. Leung, J. Whitehead: On the Complexity of
Fixed-Priority Scheduling of Periodic, Real-
Time Tasks, Performance Evaluation
(Netherlands), pp. 237-250, (1982).
K. Duffy, D. Malone, D. J. Leith: Modeling the
802.11 Distributed Coordination Function in
Non-saturated Conditions, IEEE/ACM
Transactions on Networking (TON),
Vol. 15 , pp. 159-172 (February 2007)
L. Kleinrock: Queuing Systems,Vol. 1: Theory,
Wiley Interscience, 1976.
P. Chatzimisios, V. Vitsas, A. C. Boucouvalas:
Throughput and delay analysis of IEEE 802.11
protocol, in Proceedings of 2002 IEEE 5th
International Workshop on Networked
Appliances, (2002).
P.E. Engelstad, O.N. Osterbo: Delay and
Throughput Analysis of IEEE 802.11e EDCA
with Starvation Prediction, In proceedings of
the The IEEE Conference on Local Computer
Networks , LCN’05 (2005).
P.E. Engelstad, O.N. Osterbo: Queueing Delay
Analysis of 802.11e EDCA, Proceedings of
The Third Annual Conference on Wireless On
demand Network Systems and Services
(WONS 2006), France, (January 2006).
P.E. Engelstad, O.N. Osterbo: The Delay
Distribution of IEEE 802.11e EDCA and
802.11 DCF, in the proceeding of 25th IEEE
International Performance Computing and
Communications Conference (IPCCC’06),
(April 2006), USA.
The network simulator ns-2,
http://www.isi.edu/nsnam/ns/.
Y. Xiao: Performance analysis of IEEE 802.11e
EDCF under saturation conditions, Proceedings
of International Conference on
Communication (ICC’04), Paris, France, (June
2004).
V. Kanodia, C. Li: Distribted Priority Scheduling
and Medium Access in Ad-hoc Networks”,
ACM Wireless Networks, Volume 8,
(November 2002).
- 28
TEMPORAL INFORMATION SYSTEMS AND THEIR APPLICATIONS
TO MOBILE AD HOC ROUTING

V. Mary Anita Rajam, V.Uma Maheswari and Arul Siromoney

Department of Computer Science and Engineering,
Anna University, Chennai -- 600 025, India

Contact email: asiro@vsnl.com

ABSTRACT
A Temporal Information System, that incorporates temporal information into the
traditional Information System of Rough Set Theory (RST) and Variable Precision
Rough Sets (VPRS), is presented. Mobile Ad hoc Networks (MANETs)
dynamically form a network without an existing infrastructure. The Dynamic
Source Routing (DSR) protocol of MANETs is modified in this paper to use recent
routes. Weighted elementary sets are introduced in temporal information systems
and used to route packets in mobile ad hoc networks. Notions from VPRS are also
brought into weighted temporal information systems and used in routing. The
performance of these proposed routing protocols is studied.

Keywords: Rough Set Theory, Mobile ad hoc networks, Temporal Information

System, Routing.

1 INTRODUCTION symptoms and tests; and the decisions (or decision

Rough Set Theory is a mathematical tool
that deals with vagueness and uncertainty. Mobile ad
hoc networks are a collection of wireless mobile
nodes that can dynamically form a network. The
state of each link in a mobile ad hoc network
changes with time.
This paper introduces temporal information
systems that are then applied to mobile ad hoc
routing.
Each mobile node maintains a route cache
of known routes. It is shown in this paper that giving
more importance to the recent routes in the route
cache is useful. This has led to the notion of
weighted elementary sets in temporal information
systems, where more recent elementary sets are
given more importance.
attribute) may be diseases.
In an information system, elements that
have the same value for each attribute are
indiscernible and are called elementary sets. Subsets
of the universe with the same value of the decision
attribute are called concepts. A positive element is an
element of the universe that belongs to the concept.
For each concept, the greatest union of
elementary sets contained in the concept is called the
lower approximation of the concept and the least
union of elementary sets containing the concept is
called the upper approximation of the concept. The
set containing the elements from the upper
approximation of the concept that are not members
of the lower approximation is called the boundary
region. The lower approximation of the concept is
also known as the positive region.

1.1 Rough Set Theory A set is said to be rough if the boundary

In Rough Set Theory (RST) [21],
introduced by Zdzislaw Pawlak, a data set is
represented as a table, where each row represents an
event or an object or an example or an entity or an
element. Each column represents an attribute that can
be measured for an element. This table is called an
information system. The set of all elements is known
as the universe. For example, if the information
system describes a hospital, the elements may be
patients; the attributes (condition attributes) may be
region is non-empty. A set is said to be crisp if the
boundary region is empty.
Variable Precision Rough Sets (VPRS)
[31], proposed by Ziarko, is a generalization of the
rough set model, aimed at modelling classification
problems involving uncertain or imprecise
information. Classification with a controlled degree
of uncertainty is possible with this model. It is
possible to generalize conclusions obtained from a
smaller set of observations to a larger population.

UbiCC Journal - Volume 3 29

 In RST, the lower approximation of a
concept is defined using an inclusion relation. Here
in VPRS, the lower approximation is defined using a
majority inclusion relation. The β-positive region is
the union of elementary sets which are either
completely contained in the concept or are almost
contained in the concept, with a maximum error of
1 – β.
The conditional probability of an element
being positive in an elementary set is the probability
that the element is positive, given that the element
belongs to that elementary set. It is the ratio of the
number of positive elements in that elementary set to
the number of elements in that elementary set. When
this conditional probability is greater than a threshold
β (0.5 < β ≤ 1) the elementary set is said to fall in the
β-positive region.
1.2 Rough Sets in Temporal Contexts
A temporal system is a time based system
which shows the temporal variation of some specific
data or attribute. Time-series data are a kind of
temporal data and are results of some observations
usually ordered in time. Time-series data often
possess content that is conflicting and redundant.
The data may be imprecise; therefore a precise
understanding of information cannot be derived from
the data. Rough set theory offers a powerful toolset
for confronting this situation.
Analysis of time-series data and
constructing suitable data from the time-series that
can be used by rough sets are investigated in [15],
[18], [11]. Reducts can be found from the original
data and rules can be generated from the acquired
reducts using rough sets [15]. For constructing
suitable data from the time-series, different methods
are tried. In the mobile window method [4], a
window is moved along the time-series; the data
points falling into the window are transferred into a
rough sets object. The window method poses
restrictions on how far back in time dependencies
can be traced. In the columnizing method [18], the
time series are organized in columns, such that each
row represents an object where each column is an
economic indicator, and each row represents a
different point in time.
In [11], time-series is represented using a
series of events or states. An event is something that
occurs, and is associated with a time. The values of
attributes are trends, rather than values measured at
points in time, so that dependencies can be traced
back as long as required.
A decision table for time-series called a
time series decision table, is proposed in [16]. A time
UbiCC Journal - Volume 3
stamp is introduced as an attribute, in addition to the
condition attributes and the decision attribute of the
traditional information system.
A temporal information system is
introduced by Bjorvand [29]. The temporal
information system has a sequence attribute in
addition to the condition attributes and the decision
attribute present in the traditional information
system. The value of the sequence attribute is an
integer, based on the time of occurrence of the
objects in the information system. Methods are
proposed in [29] and [28] to convert this temporal
information system into the traditional information
system, so that rough set techniques can be applied.
The method proposed in [29] depends on time
intervals that must be fixed and defined in advance.
Trend expressions [28] are added to transform the
traditional method of translating a temporal
information system into an information system. So, a
new attribute is added to the information system, for
which values are set based on the trend. A real time
temporal information system is proposed in [29], in
which the difference between the time of occurrence
of the current row and the previous row (if rows are
sorted according to time) is also stored in the
information table.
Having a linearly ordered universe, based
on time, is also used in bringing in the notion of time
into a rough set information system. This temporal
information system with a linearly ordered universe
[19], [8], [1], [3], [20], [2] provides information
about the behaviour of objects in time and state of an
object is described by some attributes. The elements
can be the behaviour over time of the same object,
multiple objects, or objects independent of each
other.
Temporal templates [19], which are
homogeneous patterns occurring in some periods, are
extracted from temporal information systems. The
temporal templates are then used to discover
behaviour of temporal features of objects.
Using temporal templates [20], a temporal
multiple information system [19] is introduced that
describes many objects along the time axis e.g.
several users visiting a website. For each object, a
sequence of temporal templates is found. Collections
of patterns, called episodes, appearing in sequences
frequently are found. New attributes are generated
from the found frequent episodes.
Temporal patterns that can be potentially
specific for one musical instrument or a group of
instruments are searched [8]. Such patterns are used
as new descriptors. A time window based technique
is used to measure the values of the descriptors in
30
time. Optimal temporal templates that respond to
temporal patterns are then determined. From the
temporal templates, episodes (collections of
templates that occur together) are found.
Information maps can be constructed from
data represented by temporal information systems
[1]. The temporal information system at the current
time t is also viewed as a family of decision systems
[1], where the universe of the decision system at time
t1, is a subset of the decision system at time t2, for
t1 < t2.
Some work is based on how the data varied
over a particular duration of time (say, two years,
two months etc.). Different attributes are assigned
for each time period in [30].
A dynamic information system model
based on time sequence is proposed in [13]. The
attributes in the information system can have
different values at different time points.
1.3 Routing in Mobile Ad Hoc Networks
Sending data from a source mobile node to
a destination mobile node through a route (or
sequence of intermediate mobile nodes) is routing.
Routing is one of the most difficult issues in mobile
ad hoc networks. Each node in an ad hoc network is
responsible for routing. Hence each node should
maintain the information necessary for routing, that
is, the next hop or the path through which the data
has to be routed to the destination. This information
is either available even before it is needed
(proactive) or is got only when necessary (reactive).
Proactive routing is usually done using table
driven protocols where tables are formed initially,
and are updated either periodically or when some
change is known.
Reactive routing protocols are also known
as on-demand routing protocols. In this kind of
protocols, a route is found for a destination only
when there is a need to send information to the
destination. Each node does not have knowledge of
the location of all other nodes in the network. All
nodes when they learn or use routes, store the routes
to the destinations in a routing table or a cache.
Most of the on-demand routing protocols
[5], [25], [9], discover routes when needed by
flooding route request packets in the network. A
route reply is sent back to the source node by either
an intermediate node that has a route to the
destination or by the destination.
UbiCC Journal - Volume 3
1.3 Application of Rough Sets to Computer
Networks
Very little work has been done in the
application of rough set theory to Mobile Ad Hoc
Networks and Mobile Ad Hoc Routing. A few papers
have applied rough set theory to networks. Rough set
theory is used in intrusion detection in computer
networks [6], [14] and [32]. Rough set approach is
also applied to the flow control of a UDP-based file
transfer protocol [33] to accomplish a real time data-
transferring rate adjustment task; to network fault
diagnosis [24]; and to achieve the rules for object
recognition and classification for mobile robots [12].
Rough set logics is combined with artificial neural
networks for failure domain exploration of
telecommunication network [7]. The basic concepts
of rough set theory are illustrated using an example
concerning churn modeling in telecommunications
[23].
1.4 Overview of the Paper
This paper presents a Temporal Information
System that brings temporal information into the
traditional Information System of Rough Set Theory
and Variable Precision Rough Sets. The DSR
protocol is modified to study the use of recent routes.
The paper then introduces the notion of weighted
elementary sets in temporal information systems and
uses this to route packets in mobile ad hoc networks.
This paper then uses VPRS in weighted temporal
information systems and uses this in routing. The
performance of these proposed routing protocols are
studied.
2 TEMPORAL INFORMATION SYSTEMS
2.1 Information Systems and Decision Systems
Consider a universe U of elements. An
information system I is defined as I = (U, A, V, ρ),
where
A is a non-empty, finite set of attributes;
is the set of attribute values
of all attributes,
where is the set of possible values of
attribute ;
is an information function,
such that for every element ,
is the value of attribute for
element .
The information system can also be viewed as an
information table, where each element
corresponds to a row, and each attribute
corresponds to a column.
I = (U, A, V, ρ), is known as a decision
system, when an attribute is specified as the
31
decision attribute. A decision system is used for
predicting the value of the decision attribute.
is known as the set of condition attributes.
The concept is the set of elements of that
have a particular value (say, ) of the decision
attribute . That is,
Normally, is a boolean attribute that takes one of
two possible values. When is known as
a multi-valued decision attribute.
These definitions are based on the definition
of Rough Set Information System in [21, 22, 17].
2.2 Regions of the Universe
An equivalence relation , called
indiscernibility relation, is defined on the universe
as
In the information system , the elementary
set containing the element with respect to the
indiscernibility relation , is
The lower approximation of the concept
, with respect to and equivalence relation
on , is the union of the elementary sets of with
respect to that are contained in , and is denoted
as
The upper approximation of is the union
of the elementary sets of with respect to that
have a non-zero intersection with , and is denoted
as
The lower approximation of is also
known as the Positive region of . The set
is called the Boundary region
of . The set is called the Negative region of
. sets, where the first elementary set consists of
The conditional probability that an element
in an elementary set is positive is
The conditional probability that the element
in the elementary set is negative is
When the context is clear, the conditional probability
of an elementary set is taken to be . than the node information. A threshold was used in
The -positive region is the union of the
UbiCC Journal - Volume 3
elementary sets whose conditional probability is
greater than or equal to where 0.5. The -
negative region is the union of the elementary sets
whose conditional probability is less than where
0.5. These are based on the definitions in [34].
When , we denote it as , and note that
.
The range of is (0.5,1] in the original
VPRS definition. This indicates probabilistically that
the elementary set is positive, when the decision
attribute is boolean. It appears that when the decision
attribute is multi-valued with as the number of
possible values, the range of is (1/k,1].
2.3 Temporal Extensions
A Generic Temporal Information System
(GTIS) is defined as a set of information tables,
with each information table
located at a time on the time
axis. A time interval can also be considered instead
of a time instance.
A special case is when , that is, the
same universe of elements that appears in each
information table. A particular element , for
each attribute , would have a value
at each time . For example, patient $x$
has fever at , and does not have fever at . This
special case is when each .
In this paper, this is treated as a single
Temporal Information System defined as
, where is the time attribute with
as a set of pairs with
as a sequence of time instances,
For each elementary set that is formed from
the set of attributes (ie the set of attributes without
the time information), there are now elementary
elements that occurred between time instance and
, the second elementary set of elements between
and , and the last elementary set of elements
between and . This can be pictured as vertical
blocks of elementary sets along a time axis.
2.4 Information System in a Mobile Node
The use of an information system in mobile
ad hoc routing was introduced in [26]. The
information system was modified in [27] to represent
the route better, by using the link information rather
the identification of a good next hop.
32
 Let be a set of mobile nodes. A route is a hop. The decision attribute is taken as the next hop,
path through mobile nodes in and is denoted as a and the predicted next hop is also known as the
sequence of mobile nodes predicted value of the decision attribute.
Each mobile node maintains a
route cache that stores all the routes that knows. The destination can possibly be reached
Any route in the route cache is a path starting from through several different sequences of intermediary
that mobile node , and so , the first node in the nodes. In other words, several different
route, is itself. Any route in the route cache is a combinations of attribute values make it possible for
simple path, where no node repeats, that is, the destination to be reached. That is, elements in
for in the path, . So, . several different elementary sets correspond to routes
that lead to this particular destination. Thus several
Each mobile node has an elementary sets play a role in identifying the best
information table associated next hop for a particular destination. So, it is not
with it. Each row in the information table possible to use a single elementary set to predict the
corresponds to a route in the route cache maintained value of the decision attribute. The union of these
by that mobile node . elementary sets is used. In other words, for a
particular destination, the union is taken of all the
Let be the set of all possible links elementary sets that correspond to valid routes from
between the nodes. Each condition attribute the current mobile node to the destination.
corresponds to a particular link in the set of all
possible links between the nodes. So, is the same A stringent method of predicting the next
hop is when all the elements in this union of
for any and is denoted as . Each condition
elementary sets have the same value of the decision
attribute is a boolean attribute, with , and
attribute, then this value is taken as the predicted
is set to 1 or 0 depending on whether or not that link
next hop. In other words, all known routes to this
is present in the route corresponding to that element.
destination should have this particular node as the
So, since for every , and
next hop. This can also be considered as that value of
is denoted as since it is the same in each mobile
the decision attribute for which all these elementary
node .
sets are in its lower approximation. It is to be
remembered that the decision attribute is a multi-
A mobile node knows a route either because
valued attribute, and so the lower approximation is
the route is in a packet that passes through this
with respect to a value of the decision attribute.
mobile node, or because this mobile node is in
promiscuous mode, and this route is in a packet that
Another method is to have the predicted
passes between two nodes that are within range of
next hop as that value of the decision attribute where
this mobile node. When a mobile node knows a route
the union of these elementary sets is in the -positive
it is added to the route cache only if it is not identical
region. The conditional probability is determined
to a path or a sub-path of any other route already
using the union of elementary sets, and not a single
present in the route cache. However, every time a
elementary set. The probability that a particular next
mobile node knows a route, a row corresponding to
hop occurs given that the route leads to a particular
this route is always added to its information table.
destination is taken as the conditional probability.
This conditional probability should be greater than a
Consider an element x corresponding to
threshold . In other words, a large number of
a route . When a
known routes to this destination have this particular
row is added to the information table, the values of
node as the next hop.
the condition attributes corresponding to the links
are set as 1.
2.6 Temporal Decision System in a Mobile Node
2.5 Decision System in a Mobile Node
In a Temporal Information System (TIS) for
a mobile node, each element (corresponding to a
In traditional Rough Set Theory and VPRS,
route) has a particular value of the time attribute, that
the value of the decision attribute of a new element
is, each element falls in a particular time interval.
(or unknown element or test case) is predicted, based
This is determined by the time stamp of the next hop
on which elementary set it falls into. The elementary
of the route that corresponds to this element. The
set into which it falls is determined by the values of
Temporal Decision System (TDS) in a mobile node
the attributes of that element.
is used to predict the next hop.
The decision system in a mobile node is
An appropriate method (as described in the
used to predict the next hop for a particular
previous section) is used to determine the predicted
destination. This next hop is called the predicted next

UbiCC Journal - Volume 3 33

next hop in each time interval. The predicted next
hop for the TDS is then determined based on the
number of time intervals in which it is the predicted
next hop.
The predicted value of the decision attribute
is determined from the TDS based on the probability
of a particular value of the decision attribute being
the predicted value in the different time intervals.
This probability is the number of time intervals in
which that value of the decision attribute is the
predicted value divided by the total number of time
intervals. The predicted value of the decision
attribute is the value for which this probability is
greater than a threshold . In other words, that
particular next hop has been the predicted next hop
in most of the time intervals.
In Weighted Temporal Information Systems
(WTIS) and Weighted Temporal Decision Systems
(WTDS), weights are assigned to the
elementary sets between time instances and , the
elementary sets between and , and the
elementary sets between and , respectively.
The predicted value of the decision attribute is
determined after associating weights with the time
intervals. The predicted value of the decision
attribute is that value of the decision attribute where
the probability is greater than a threshold . The
probability is the sum of the weights of time intervals
in which that value of the decision attribute is the
predicted next hop divided by the sum of the weights
of all the time intervals.
When the more recent time intervals play a
more important role, the weight of a more recent
time interval is higher than the weight of a less
recent time interval.
3 MOBILE AD HOC ROUTING USING
RECENT ROUTES
This section first describes the original
Dynamic Source Routing (DSR) protocol, an on-
demand routing protocol. The DSR protocol is then
modified so that the most recent applicable route in
the route cache is used at each intermediate node.
The performance of the modified protocol is
evaluated.
3.1 Dynamic Source Routing
In Dynamic Source Routing (DSR) [5],
each mobile node has a route cache to store the
routes that are known to that mobile node.
The source node is the node that wants to
send a data packet. It uses the shortest route present
in its route cache to the destination. If there is no
UbiCC Journal - Volume 3
route in the route cache, it initiates a route discovery,
and gets back a route reply with the route to the
destination. This source route is placed in the data
packet and the data packet is sent to the next hop in
the route.
When a data packet reaches an intermediate
node, the source route in the data packet is used to
forward the data packet to the next hop. If a node
while sending the data packet to the next hop, finds
that the link does not exist, it uses the shortest route
present in its route cache to that destination. If the
route is not found in the route cache, route discovery
is done.
When a route discovery is required, the
node broadcasts route request packets. If any
intermediate node receiving the route request has a
route to the required destination in its route cache, it
sends a route reply to the initiator of the route
discovery. Else, the node appends its own address to
the route in the route request and re-broadcasts the
route request. If the route request reaches the
destination, the destination reverses the route and
sends back a route reply to the initiator of the route
discovery.
When a path is to be added to the cache, the
following are done. If a prefix of the path to be
added is present in the cache, the rest of this path is
appended to the path present in the cache. If the
whole of the new path to be added is not present in
the cache, the path is added to the cache. If there is
no free space in the cache, then a victim is picked up
and the route to be added is put in the victim's place.
If there is any path or a subpath in the cache entry
that is a prefix or the same as that of the path that is
added, then the time stamps of those links are set to
be equal to the time stamps of the links in the path
that is added.
When a link error occurs, a route error with
information about the dead link is sent to the original
sender of the data packet. Paths, or subpaths starting
with the given dead link, are removed from the cache
in nodes that receive the route error.
3.2 DSRrecent
This section describes the proposed
DSRrecent protocol and the modifications made to
the existing DSR protocol.
In the source node, in DSR, the shortest
route in the route cache is used, whereas in
DSRrecent, the route to the destination in the route
cache, that has the most recent next hop, is used
(using algorithm findRecentRoute).
34
 In intermediate nodes, in DSR, the source
route is used to determine the next hop. Only if there
is a link error, the route cache is used. The shortest
route to the destination in the route cache is then
placed in the data packet, instead of the original
source route.
However, in DSRrecent, the best route is
determined from the route cache (using algorithm
findRecentRoute). The route in the data packet is
then modified such that the route from the route
cache replaces the subpath (from the current node) in
the route in the data packet.
If no route is found (to the destination) in
the route cache of the intermediate node, or if the
found route will result in a loop, the data packet is
forwarded according to the existing route in the data
packet.
If a link error occurs in any node, the best
route found in the route cache (using algorithm
findRecentRoute) is used. If there is no route in the
route cache, route discovery is done.
(iv) Average hop count: The average
number of hops from the source node to the
destination node.
The network simulator ns2 [10] is used for
the experiments. The following parameters are ones
that have been often used in such studies. The
random waypoint mobility model is used in a
rectangular field. Constant bit rate traffic sources are
used. The radio model in the simulator is based on
the Lucent Technologies WaveLAN 802.11.
providing a 2Mbps transmission rate. A transmission
range of 250 m is used. The link layer modeled is the
Distributed Coordination Function (DCF) of the
IEEE 802.11 wireless LAN standard. The source-
destination pairs (connections) are spread randomly
over the network. 512 byte data packets are used.
Nodes move in a field with dimensions
1500 m X 300 m with a maximum speed of 2 m/sec.
The pause time is 20 seconds. The number of nodes
is kept fixed as 50. The number of communicating
source-destination pairs is varied from 5 to 40.
Simulations are run for 1000 simulated seconds.

findRecentRoute() {
besttime = 0;
Packet Delivery Ratio (%)

bestroute = NULL; 120

foreach possible-route do 100
t = time of next hop in possible-route; 80
DSR
if t > besttime then 60
besttime = t; DSRrec
40
bestroute = possible-route; 20
end 0
end 5 10 20 30 40
return bestroute;
Number of Connections
}

3.3 Performance Evaluation Figure 1: Packet delivery ratio vs. number of

connections for DSR and DSRrecent
The performance of DSRrecent is evaluated
using the following metrics, that are normally used in
such studies:
6
Normalized control

(i) Packet delivery ratio: The ratio of the 5

overhead

data packets delivered to the application layer of the 4

DSR
destination to those sent by the application layer of 3
DSRrec
the source node. 2
1
(ii) Normalized control overhead : The ratio 0
of the number of control packets sent to the number 5 10 20 30 40
of data packets received in the application layer.
Number of connections

(iii) Average end-to-end delay: The average

delay from when a packet is sent by the source node
until it is received by the destination node. Figure 2: Normalized control overhead vs. number of
connections for DSR and DSRrecent

UbiCC Journal - Volume 3 35

 The packet delivery ratios for DSR and
DSRrecent are very similar for 5 and 10 connections.
When the number of connections is increased from
20 to 40 there is an improvement in the packet
delivery ratio from 3% to 11% (Fig. 1). The
normalized control overhead for DSRrecent is more
than that for DSR when the number of sources is 5.
With increase in the number of connections from 10
to 40, there is an improvement of 4% to 25% over
DSR (Fig. 2). In average hop length and average
end-to-end delay, DSRrecent is seen to perform
worse than DSR as the number of connections is
increased (Fig. 3, Fig. 4).
more recent time intervals are assigned higher
weights than the less recent time intervals.
The use of the WTIS to predict the value of
the decision attribute has already been described in
section 2.6. This uses the predicted value of the
decision attribute in different time intervals. The
experiment described here uses a simple approach to
determine the predicted value of the decision
attribute in a particular time interval. A predicted
value of the decision attribute in a time interval has
atleast one element, with that value of the decision
attribute, in the union of elementary sets. That is, the
union is in the upper approximation for that value of
the decision attribute.

18
Average hop count

16
14 4.1 Routing Based on WTIS
12
10 DSR
8
Here, the route cache of the mobile node is
DSRrec used as the WTIS. Routes that are learnt and used are
6
4 added to the cache of the mobile node. When routes
2
0 are added, the time stamp of each link is added along
with the routes. However, unlike DSR, even if the
5 10 20 30 40
same route is present in the cache earlier, the new
Number of connections route is added with the new stamp stamps. So, the
cache now has the same route multiple times, but
with different time stamps.
Figure 3: Average hop count vs. number of
connections for DSR and DSRrecent In the source node, initially, as in DSR, a
shortest route in the route cache, if available, is
placed as the source route in the data packet. If not
available, route discovery is done.
Average end-to-end delay

10
9
8 Then in the source node, and in any
7 intermediate forwarding node, the WTIS is used to
6 DSR
5 determine the best next hop (using algorithm
4 DSRrec
3 findWeightBasedHop). If the next hop is found, and
2 does not result in a loop, the data packet will be
1
0 forwarded to this next hop. If this next hop is
different from the one in the source route that is
5 10 20 30 40
already in the data packet, this new next hop is
Number of Connections appended to the source route in the data packet at the
current node and the route is invalidated by setting a
flag in the data packet.
Figure 4: Average end-to-end delay vs. number of
connections for DSR and DSRrecent If a next hop cannot be determined from the
WTIS, or if the next hop results in a loop, if the
source route in the data packet has not been
4 MOBILE AD HOC ROUTING USING invalidated earlier, the data packet is forwarded
WEIGHTED TEMPORAL INFORMATION according to the source route. Else, a route discovery
SYSTEMS (WTIS) is done.

In Temporal Information Systems, each
elementary set is associated with a particular time
interval. In Weighted Temporal Information
Systems, elementary sets in different time intervals
have weights. Since it is seen in the previous section
that the recent route in the route cache is useful,
The total time is divided into time intervals.
The list of next hops to the destination that are
present in the route cache is found. For each possible
next hop, from the current time interval till the initial
time interval, a weighted sum of the number of times
that the particular next hop is used is found. More

UbiCC Journal - Volume 3 36

weight is assigned if the next hop has been used in
the recent past. That is, the weights assigned
decrease for earlier time intervals.
findWeightBasedHop(){
Find all possible next hops that will lead to the
destination from this node;
foreach possible next hop nh
timeInterval = currentInterval;
The average hop length and the average
end-to-end delay for TIME_WT are more than that
for DSR when the number of sources is 5 and 10.
But when the number of connections is increased
from 20 to 40, it is seen that there is a slight
improvement of about 2% in average hop length and
of about 5\% in average end-to-end delay over DSR
(Fig.7, Fig.8).

Packet Delivery Ratio (%)

weightedSum = 0; 120
weight = maxWeight; 100
totalWeight = 0; 80
DSR
while timeInterval >= 0 do 60
if nh is used as a nexthop in TIME_WT
40
timeInterval then 20
weightedSum = weightedSum + 0
weight; 5 10 20 30 40
end
Number of Connections
timeInterval = timeInterval -1;
//previous timeInterval
totalWeight = totalWeight + weight;
weight = weight -1; Figure 5: Packet delivery ratio vs. number of
end connections for DSR and TIME_WT
ratio[nh] = weightedSum / totalWeight;
end
Find the nexthop nh for which the value of ratio
6
Normalized control

is the maximum and return

} 5
overhead

4
DSR
3
TIME_WT
The ratio of the weighted sum of the usage 2
of the node to the total weight is found. The node for 1
which the ratio is greater than a threshold $\beta'$ is 0
chosen as the next hop. 5 10 20 30 40
Number of connections
4.2 Performance Evaluation

The parameters used are the same as those

given in section 3.3. The size of the time interval is Figure 6: Normalized control overhead vs. number of
taken as 40 seconds. The value of $\beta'$ used is connections for DSR and TIME_WT
0.5.

The proposed protocol used in this section 16

Average hop count

is referred to as TIME_WT. The packet delivery 14

ratios for DSR and TIME_WT are nearly similar for 12
5 and 10 connections. When the number of 10 DSR
connections is increased from 20 to 40 there is a 8
6 TIME_WT
slight improvement in the packet delivery ratio from 4
5% to 7% (Fig. 5). 2
0
The normalized control overhead for 5 10 20 30 40
TIME_WT is more than that for DSR when the Number of connections
number of sources is 5 and 10. With increase in the
number of connections from 20 to 40, there is an
average improvement of 14% to 22% over DSR
(Fig.6).
Figure 7: Average hop count vs. number of
connections for DSR and TIME_WT

UbiCC Journal - Volume 3 37

 Average end-to-end delay The routing protocol is similar to that of the
7 previous section. The next hop is chosen using the
6 notion of threshold $\beta$ ($\beta$--positive
5 regions) as described in algorithm
4 DSR
findVPRSWeightBasedHop().
3 TIME_WT
2
1
5.2 Performance Evaluation
0
The parameters used are the same as that
5 10 20 30 40
given in section 3.3. The size of the time interval is
Number of Connections taken as 40 seconds. The value of $\beta$, $\beta'$
used are 0.6,0.5 respectively.

Figure 8: Average end-to-end delay vs. number of
connections for DSR and TIME_WT
5 MOBILE AD HOC ROUTING USING
BETA-POSITIVE REGIONS IN WTIS
5.1 Routing Based on Beta-Positive Regions
The experiment described in this section
determines the predicted next hop as that value of the
decision attribute where the union of these
elementary sets is in the $\beta$-positive region, as
described in section 2.5.
The proposed protocol used in this section
is referred to as VPRS_WT. The packet delivery
ratio for VPRS_WT is less than that for DSR for 5
and 10 connections. When the number of
connections is increased from 20 to 40 there is a
slight improvement in the packet delivery ratio from
2% to 6% (Fig. 9).
The normalized control overhead for
VPRS_WT is more than that for DSR when the
number of sources is 5, 10. With increase in the
number of connections from 20 to 40, there is an
average improvement of 14% to 19% over DSR (Fig.
10).

findVPRSWeightBasedHop(){ The average hop length and the average

Find all possible next hops that will lead to the end-to-end delay for VPRS_WT is more than that for
destination from this node; DSR when the number of sources is 5, 10. But when
Foreach possible next hop nh do the number of connections is increased from 20 to
TimeInterval = currentInterval - 1; 40, it is seen that there is a slight improvement of
weightedSum = 0; about 4% in average hop length over DSR (Fig. 11).
weight = maxWeight;
totalWeight = 0; The average end-to-end delay for
while timeInterval >= currentInterval – k VPRS_WT is similar to that of DSR when the
do number of sources is 5. But when the number of
nhopCount = the number of routes with connections is increased from 10 to 40, it is seen that
next hop nh and willlead to the destination there is an improvement of about 6% in average hop
in timeInterval ; length over DSR (Fig. 12).
totalCount = the number of routes that will
lead to the destination in timeInterval
ratio1 = nhopCount/totalCount
Packet Delivery Ratio(%)

if ratio1 > $\beta$ then 120

weightedSum = weightedSum + 100
weight; 80
DSR
end 60
timeInterval = timeInterval -1; //previous VPRS_WT
40
timeInterval 20
totalWeight = totalWeight + weight;
0
weight = weight -1;
5 10 20 30 40
end
ratio[nh] = weightedSum / totalWeight; Number of Connections
end
Find the nexthop nh for which the value of ratio is
greater than $\beta'$ Figure 9: Packet delivery ratio vs. number of
} connections for DSR and VPRS_WT

UbiCC Journal - Volume 3 38

 Using recent routes (DSRrecent) was found to
Normalized control 6 improve packet delivery ratio and normalized control
5 overhead. Temporal information was brought into
overhead

4 information systems. Recent elementary sets were

DSR
3 given more importance in the two proposed methods,
VPRS_WT TIME_WT and VPRS_WT. The VPRS_WT method
2
1 uses notions from VPRS. It was seen that the control
0
overhead is much better, while the packet delivery
ratio, average hop length and average end-to-end
5 10 20 30 40
delay are slightly better than that of DSR. It was also
Number of connections seen that the improvement in performance increases
with the number of connections.

Figure 10: Normalized control overhead vs. number 7 REFERENCES

of connections for DSR and VPRS_WT
[1] A. Skowron and P. Synak. Patterns in
Information Maps. In Rough Sets and Current
Trends in Computing, volume 2475 of Lecture Notes
16 in Artificial Intelligence, pages 453 – 460. Springer,
Average hop count

14
12
002.
10 DSR
[2] A. Skowron and P. Synak. Reasoning Based on
8 Information Changes in Information Maps. In Rough
6 VPRS_WT
Sets, Fuzzy Sets, Data Mining, and Granular
4
Computing, volume 2639 of Lecture Notes in
2
0 Artificial Intelligence, pages 229 – 236. Springer
2003.
5 10 20 30 40
[3] A. Wieczorkowska, J. Wroblewski, P. Synak and
Number of connections D. Slezak. Application of temporal descriptors to
musical sound recognition. Journal of Intelligent
Information Systems, 21(1):71 – 93, 2003.
Figure 11: Average hop count vs. number of [4] J.K. Baltzersen. An attempt to predict stock
connections for DSR and VPRS_WT market data: a rough sets approach. Master’s thesis.
1996.
[5] David B.Johnson and David A. Maltz. Dynamic
source routing in ad hoc wireless networks. In
Average end-to-end delay

7
Imielinski and Korth, editors, Mobile Computing,
6
volume 353. Kluwer Academic Publishers, 1996.
5
4 DSR [6] Zhongmin Cai, Xiaohong Guan, Ping Shaoa, and
3
Guoji Sun. A rough set theory based method for
VPRS_WT
2 anomaly intrusion detection in computer network
1 systems. In Expert Systems, volume 20, pages 251 –
0 259, 2003.
5 10 20 30 40 [7] Frank Chiang and Robin Braun. Intelligent
failure domain prediction in complex
Number of Connections
telecommunication networks with hybrid rough sets
and adaptive neural nets. In 3rd International
Information and Telecommunication Technologies
Figure 12: Average end-to-end delay vs. number of Symposium, 2004.
connections for DSR and VPRS_WT [8] Slezak D., Synak P., Wieczorkowska A., and
Wroblewski J. Kdd-based approach to musical
instrument sound recognition. In Proc. of the 13th
6 CONCLUSIONS International Symposium on Foundations of
Intelligent Systems, Vol. 2366, Lecture Notes in
This paper presents temporal extensions to Artificial Intelligence, Springer, pages 28 – 36, 2002.
Rough Set Theory and Variable Precision Rough [9] Rohit Dube, Cynthia D. Rais, Kuang-Yeh Wang,
Sets. These extensions are applied to Mobile Ad hoc and Satish K. Tripathi. Signal stability-based
routing. Illustrative experiments are described and adaptive routing (SSA) for ad hoc mobile networks.
the results are presented. IEEE Personal Communications, 4(1):36{45, 1997.
[10] K. Fall and K. Varadhan. The ns manual

UbiCC Journal - Volume 3 39

(formerly ns notes and documentation), 2002.
http://www.isi.edu/nsnam/ns/doc/index.html.
[11] Roed G. Knowledge extraction from process
data: A rough set approach to data mining on time
series. In Mater's thesis, 1999.
[12] Wang Haijun and Chen Yimin. Sensor data
fusion using rough set for mobile robots system. In
Proceedings of the 2nd IEEE/ASME International
Conference on Mechatronic and Embedded Systems
and Applications, pages 1 – 5, 2006.
[13] Xiaowei He, Liming Xu, and Wenzhong Shen.
Dynamic information system and its rough set model
based on time sequence. In Proc. of 2006 IEEE
International Conference on Granular Computing,
pages 542 – 545, 2006.
[14] Peng Hong, Dongna Zhang, and Tiefeng Wu.
An intrusion detection method based on rough set
and svm algorithm. In 2004 International Conference
on Communications, Circuits and Systems, ICCCAS
2004, pages 1127 – 1130 vol. 2.
[15] Herbert J. and Yao J. T. Time-series data
analysis with rough sets. In Proc. of 4th International
Conference on Computational Intelligence in
Economics and Finance, pages 908 – 911, 2005.
[16] Li J., Xia G., and Shi X. Association rules
mining from time series based on rough set. In Proc.
of the Sixth International Conference on Intelligent
Systems Design and Applications, pages 509 – 516,
2006.
[17] J. Komorowski, Z. Pawlak, L. Polkowski, and A.
Skowron. Rough sets: A tutorial. In S. K. Pal and A.
Skowron, editors, Rough Fuzzy Hybridization: A
New Trend in Decision-Making, pages 3-98.
Springer-Verlag, 1999.
[18] Shen L. and Loh H. T. Applying rough sets to
market timing decisions. Decision Support Systems,
Special Issue: Data Mining for Financial Decision
Making, 37(4):583 – 597, 2004.
[19] Synak P. Temporal templates and analysis of
time related data. In Rough Sets and Current Trends
in Computing, volume 2005 of Lecture Notes in
Computer Science, pages 420 – 427. Springer, 2000.
[20] Synak P. Temporal feature extraction from
temporal information systems. In Ning Zhong,
Zbigniew W. Ras, Shusaku Tsumoto, Einoshin
Suzuki, editors, Foundations of Intelligent Systems,
14th International Symposium, ISMIS 2003, Vol.
2871, Lecture Notes in Computer Science, Springer,
pages 270 – 278, 2003.
[21] Z. Pawlak. Rough sets. International Journal of
Computer and Information Sciences, 11(5):341– 356,
1982.
[22] Z. Pawlak. Rough Sets — Theoretical Aspects
of Reasoning about Data. Kluwer Academic
Publishers, Dordrecht, The Netherlands, 1991.
[23] Zdzislaw Pawlak. Rough set theory and its
applications. In Journal of Telecommunications and
Information Technology, 2002.
[24] Yuqing Peng, Gengqian Liu, Tao Lin, and
UbiCC Journal - Volume 3
Hengshan Geng. Application of rough set theory in
network fault diagnosis. In Proceedings of the Third
International Conference on Information Technology
and Applications (ICITA 05), pages 556 – 559 vol. 2.
[25] C. Perkins and E. Royer. Ad-hoc on demand
distance-vector routing for mobile computers. In
Proceedings of the Second international workshop on
Mobile Computing Systems and applications, pages
90 - 100, 1999.
[26] V. Mary Anita Rajam, V. Uma Maheswari, and
Arul Siromoney. Mobile ad hoc routing using rough
set theory. In 2006 International Conference on
Hybrid Information Technology -Vol2 (ICHIT'06),
pages 80 - 83, November 2006.
[27] V. Mary Anita Rajam, V. Uma Maheswari, and
Arul Siromoney. Extensions in mobile ad hoc
routing using variable precision rough sets. In IEEE
International Conference on Granular Computing,
pages 237 – 240, November 2007.
[28] Lin S., Chen S., and Ning Z. Several techniques
usable in translating a TIS into IS. 30(5), 2003.
[29] Bjorvand A. T. Mining time series using rough
sets -a case study. In Komorowski H. J. and Zytkow
J. M., editors, Principles of Data Mining and
Knowledge Discovery, Vol. 1263, Lecture Notes in
Computer Science, Springer, pages 351 - 358, 1997.
[30] Kowalczyk W. and Slisser F. Analyzing
customer retention with rough data models. In
Komorowski H. J. and Zytkow J. M., editors,
Principles of Data Mining and Knowledge Discovery,
Vol. 1263, Lecture Notes in Computer Science,
Springer, pages 4 – 13, 1997.
[31] W.Ziarko. Variable precision rough set model.
Journal of Computer and Systems Sciences, 46(1):39
– 59, 1993.
[32] Wang Xuren, He Famei, and Xu Rongsheng.
Modeling intrusion detection system by discovering
association rule in rough set theory framework. In
Proceedings of the International Conference on
Computational Intelligence for Modelling Control
and Automation and International Conference on
Intelligent Agents, Web Technologies and Internet
Commerce, page 24, 2006.
[33] J. Zhang and R. D McLeod. A udp-based file
transfer protocol (uftp) with flow control using a
rough set approach. 2005.
[34] W. Ziarko. Set approximation quality measures
in the variable precision rough set model. In Proc. of
2nd Intl. Conference on Hybrid Intelligent Systems,
Santiago, Chile, 2002.
40
 INTEGRATION OF FUZZY INFERENCE ENGINE WITH RADIAL
BASIS FUNCTION NEURAL NETWORK FOR SHORT TERM LOAD
FORECASTING

Ajay Shekhar Pandey , S.K.Sinha

Kamla Nehru Institute of Technology ,Sultanpur, UP, INDIA
shekhar.ajay04@rediffmail.com , sinhask98@engineer.com

D. Singh
Institute of Technology,Banaras Hindu University,Varanasi, UP, INDIA

ABSTRACT

This paper proposes a fuzzy inference based neural network for the forecasting of
short term loads. The forecasting model is the integration of fuzzy inference engine
and the neural network, known as Fuzzy Inference Neural Network (FINN). A
FINN initially creates a rule base from existing historical load data. The parameters
of the rule base are then tuned through a training process, so that the output of the
FINN adequately matches the available historical load data. Results show that the
FINN can forecast future loads with an accuracy comparable to that of neural
networks, while its training is much faster than that of neural networks. Simulation
results indicate that hybrid fuzzy neural network is one of the best candidates for
the analysis and forecasting of electricity demand. Radial Basis Function Neural
Network (RBFNN) integrated with Fuzzy Inference Engine has been used to create
a Short Term Load Forecasting model.

Keywords: STLF, RBFNN ,Fuzzy Inference, Fuzzy Inference Neural Networks.

1 INTRODUCTION conditions and the short time required for their

Short term forecasts in particular have become
increasingly important since the rise of the
competitive market. Forecasting the power demand
is an important task in power utility companies
because accurate load forecasting results in an
economic, reliable and secure power system
operation and planning. Short Term Load
Forecasting (STLF) is important for optimum
operation planning of power generation facilities, as
it affects both system reliability and fuel
consumption. The complex dependence of load on
human behaviour, social and special events &
various environmental factors make load forecasting
a tedious job. It is an important function performed
by utilities for planning operation and control and is
primarily used for economic load dispatch, daily
operation and control, system security and assurance
of reliable power supply. The impacts of
globalization and deregulation demands improved
quality at competitive prices, which is the reason
why development of advanced tools and methods for
planning, analysis, operation and control are needed.
Important decisions depend on load forecast with
lead times of minutes to months. The ability of ANN
to outperform the traditional STLF methods,
especially during rapidly changing weather
development, have made ANN based STLF models a
very attractive alternative for on line implementation
in energy control centers. In this era of competitive
power market, it is of main concern that how to
improve accuracy of STLF.
In recent years use of intelligent techniques have
increased noticeably. ANN and fuzzy systems are
two powerful tools that can be used in prediction and
modeling. Load forecasting techniques such as ANN
[4], [5], [6], [7], [11], [15] , [18], Expert systems [14],
fuzzy logic, fuzzy inference [2], [3], [10], [12], [13],
[16] have been developed, showing more accurate
and acceptable results as compared to conventional
methods. A wide variety of conventional models for
STLF have also been reported in the literature. They
are based on various statistical methods such as
regression [1], Box Jenkins models [9] and
exponential smoothing [19]. Conventional ANN
model based STLF have several drawbacks, such as
long training time and slow convergent speed. The
RBF model is a very simple and yet intrinsically
powerfully network, which is widely used in many
fields because of its extensive learning and highly
computing speed [6],[7]. A neuro-fuzzy approach
has been applied successfully in a price sensitive
environment [2]. Soft Computing (SC) introduced by
Lotfi Zadeh [20] is an innovative approach to

UbiCC Journal - Volume 3 41

construct computationally intelligent hybrid systems
consisting of Artificial Neural Network (ANN),
Fuzzy Logic (FL), approximate reasoning and
optimization methods.
Fuzzy system is another research area which is
receiving increased attention. The pioneering work
of Zadeh in fuzzy set theory has inspired work in
many research areas with excellent results. A fuzzy
expert system for STLF is developed in [15]. It uses
fuzzy set theory to model imprecision in the load
temperature model and temperature forecasts as well
as operator’s heuristic rules. Fuzzy set theory
proposed by Zadeh [20] provides a general way to
deal with uncertainty, and express the subjective
knowledge about a process in the form of linguistic
IF-THEN rules.
Fuzzy Systems exhibit complementary
characteristics, offering a very powerful framework
for approximate reasoning as it attempts to model the
human reasoning process at a cognitive level. It
acquires knowledge from domain experts and this is
encoded within the algorithm in terms of the set of
If-Then rules. Fuzzy systems employ this rule based
approach and interpolative reasoning to respond to
new inputs. Fuzzy systems are suitable for dealing
with problems caused by uncertainty, inexactitude
and noise, so the uniting of fuzzy system and neural
networks can exert respective advantages.
In this paper, a fuzzy inference neural network is
presented to improve the performance of STLF in
electric power systems. A Fuzzy Inference Neural
Network initially creates a fuzzy rule base from
existing historical load data. The parameters of the
rule base are then tuned through a training process so
that the output of the network adequately matches
the available historical load data. The fuzzy system
combines the fuzzy inference principles with neural
network structure and the learning ability into an
integrated neural network based fuzzy decision
system. Combining the specific characteristic that the
variety of power systems load is non-linear, we set
up a new short-term load forecasting model based on
fuzzy neural networks and fuzzy getting smaller
inference algorithm. The flexibility of the fuzzy logic
approach, offering a logical set of IF-THEN rules,
which could be easily understood by an operator,
might be a good solution for easy practical
implementation and usage of STLF models. The
hybrid FNN approach is finally used to forecast
loads with greater accuracy than the conventional
approaches when used on a stand- alone mode.
2 RADIAL BASIS FUNCTION NEURAL
NETWORK
Radial Basis Function (RBF) Network consists
of two layers, a hidden layer with nonlinear neurons
and an output layer with linear neurons. Thus, the
transformation from the input space to the hidden
UbiCC Journal - Volume 3
unit space is non-linear whereas the transformation
from the hidden unit space to the output space is
linear. The basis functions in the hidden layer
produce a localized response to the input i.e.
each hidden unit has a localized receptive field.
RBFNNs exhibit a good approximation and learning
ability and are easier to train and generally converge
very fast. It uses a linear transfer function for the
output units and Gaussian function (radial basis
function) for the hidden units. The transform
function of hidden layer is a non-negative and
nonlinear function. In RBF neural network, three
parameters are needed to study: the center and the
variance of the basis function and the weight
connecting hidden layer to the output layer. The RBF
network has many study methods according to the
different methods of selecting the center. In this
paper, a method of the self-organizing study
selecting RBF center is adopted. The method
consists of two-step procedure: the first one is self-
organizing study, which is to study the basis function
center and variance; then the next step is supervisory
study, which is the weight connecting hidden layer to
the output layer. A RBF neural network embodies
both the features of an unsupervised learning based
classification and a supervised learning layer. The
network is mainly a feed forward neural network.
The hidden unit consists of a function called the
radial basis function, which is similar to the Gaussian
Density function whose output is given by
⎛
⎜ r (x − W ) 2 ⎞⎟
jp ij (1)
o = exp − ⎜ ∑ ⎟
i ⎜ j = 1 σ ⎟
⎝ ⎠
where,
Wij = Center of the i th RBF unit for input variable j
σ = Spread of the RBF unit
x = j th variable of the input pattern
The RBF neural network generalizes on the
basis of pattern matching. The different patterns are
stored in a network in form of cluster centers of the
neurons of the hidden units. The number of neuron,
determines the number of cluster centers that are
stored in the network. The response of particular
hidden layer node is maximum (i.e. 1) when the
incoming pattern matches the cluster center of the
neuron perfectly and the response decays
monotonically as the input patterns mismatches the
cluster center; the rate of decay can be small or large
depending on the value of the spread. Neurons with
large spread will generalize more, as it will be giving
same responses (closer to 1) even for the wide
variation in the input pattern and the cluster centers
whereas a small spread will reduce the generalization
property and work as a memory. Therefore, spread is
42
an important parameter and depends on the nature of
input pattern space.
The output linear layer simply acts as an optimal
combiner of the hidden layer neuron responses. The
weights ‘w’ for this layer are found by multiple
linear regression technique. The output of the linear
layer is given by
N n
y mp = ∑i =1
w mi oi + bi
where,
N = number of hidden layer nodes (RBF units)
y mp = output val ue of the m th node in the output layer
for the i th incoming pattern
w mi = weight betweeni th RBF unitandmth outputnode
bl = biasing strength of the m th output node
oi = i th input to the linear layer.
The values of the different parameters of the
RBF networks are determined during training. These
parameters are spread, cluster centers, and weights
and biases of the linear layer. The number of neurons
for the network and spread is determined through
experimentation with a large number of
combinations of spread and number of neuron. The
best combination is one which produces minimum
Sum Squared Error (SSE) on the testing data.
3 FUZZY INFERENCE
Fuzzy inference is the process of formulating the
mapping from a given input to the output using fuzzy
logic. This process numerically evaluates the
information embedded in the fuzzy rule base. The
fuzzy rule base consists of “IF-THEN” type rules.
For a set of input variables, there will be fuzzy
membership in several fuzzy input variables. By
using the fuzzy inference mechanism, the
information is processed to evaluate the actual value
from the fuzzy rule base. A good precision can be
achieved by applying appropriate membership
definitions along with well-defined membership
functions. This is an information processing system
that draws conclusions based on given conditions or
evidences. A fuzzy inference engine is an inference
engine using fuzzy variables. Fuzzy inference refers
to a fuzzy IF-THEN structure. The fact that fuzzy
inference engines evaluates all the rules
simultaneously and do not search for matching
antecedents on a decision tree makes them perfect
candidates for parallel processing computers.
A fuzzy set is a set without a crisp, clearly
defined boundary, and can contain fuzzy variables
with a partial degree of membership, which is
presented by the membership functions within the
UbiCC Journal - Volume 3
range. There are two types of fuzzy models. The first
kind is known as Mamdani model [8]. In this model,
both fuzzy premise part and consequence part are
represented in linguistic terms. The other kind is
Takagi-Sugeno model [17] that uses linguistic term
only for the fuzzy premise part. In this paper the
Takagi-Sugeno reasoning method is used.
The fuzzification interface is a mapping from the
observed non-fuzzy input space U ⊆ R to the
(2) fuzzy sets defined in U. Hence, the fuzzification
interface provides a link between the non-fuzzy
outside world and the fuzzy system framework. The
fuzzy rule base is a set of linguistic rules or
conditional statements in the form of: "IF a set of
conditions is satisfied, THEN a set of consequences
are inferred". The fuzzy inference engine is a
decision making logic performing the inference
operations of the fuzzy rules. Based on the fuzzy IF-
THEN rules in the fuzzy rule base and the
compositional rule of inference [14], the appropriate
fuzzy sets are inferred in the output space.
Supposing the mapping µ A from discussed
region U to the range [0, 1]: U → [0,1] ,
x → µ A (x) confirms a fuzzy subset of U, named A,
the mapping µ A (x) is known as membership
function of A. The size of the mapping
µ A (x) shows the membership degree of x to fuzzy
set A, which is called membership degree for short.
In practice, membership function can be selected
according to the characteristic of the object.
Fuzzy inference based on fuzzy estimation is a
method by which a new and approximate fuzzy
estimation conclusion is inferred using fuzzy
language rule. This paper adopts composite fuzzy
inference method, which is inference method based
on fuzzy relation composing principle. A fuzzy
inference engine can process mixed data. Input data
received from the external world is analyzed for its
validity before it is propagated into a fuzzy inference
engine. The capability of processing mixed data is
based on the membership function concept by which
all the input data are eventually transformed into the
same unit before the inference computations. A
fuzzy inference engine normally includes several
antecedent fuzzy variables. If the number of
antecedent variables is k then there will be k
information collected from the external world.
Fuzzification and normalization are the two typical
transformations. Another important property is that
when an input data set is partially ambiguous or
unacceptable, a fuzzy inference engine may still
produce reasonable answers.
4 FUZZY INFERENCE NEURAL NETWORK
A fuzzy Inference neural network approach,
43
which combines the important features of ANN and
fuzzy using inference mechanism is proposed .This
architecture is suggested for realizing cascaded fuzzy
inference system and neural network modules, which
are used as building blocks for constructing a load
forecasting system. The fuzzy membership values of
load and temperature are the inputs to the ANN, and
the output comprises the membership value of the
predicted load. To deal with the linguistic values
such as high, low, and medium, architecture of ANN
that can handle fuzzy input vectors is propounded.
Each input variable is converted into a fuzzy
membership function in the range [0-1] that
corresponds to the degree to which the input belongs
to a linguistic class. RBFNN has been integrated
with fuzzy inference to form a FINN for Short Term
Load Forecasting. The RBFNN is used to extract the
features of input and output variables. It is
noteworthy that the input variables are extended to
include a output variable and extract the relationship
between inputs.
4.1 Input Variable Selection and Data Processing
The most important work in building our Short
Term Load Forecasting (STLF) models is the
selection of the input variables. It mainly depends on
experience and is carried out almost entirely by trial
and error. However, some statistical analysis can be
very helpful in determining the variables, which have
significant influence on the system load. Normally
more input neurons make the performance of the
neural network worse in many circumstances.
Optimal input parameters would result in a compact
ANN with higher accuracy and also at the same time
with good convergence speed. Parameters with effect
on hourly load can be categorized into day type,
historical load data and weather information.
Input Variables
Fuzzy Radial Basis
Inference Function Neural
Engine Network
Learning
Algorithm
Figure 1: Forecasting Model
UbiCC Journal - Volume 3
Temperature is the most effective weather
information on hourly load. Data has been taken for
Trans Alta Canada System.
In order to make minimum inference case, the
input load is sorted into 5 categories and labeled as
low (L), low medium (LM), medium (M), medium
high (MH) and high (H). The input temperature is
also sorted into 5 categories same as above. Design
data consists of hourly data, integrated load data and
temperature of two places. Keeping in view the large
geographical spread of the data , for which the utility
supply, the hourly temperature of two places have
been taken in the historical data. Firstly data are
normalized. The n rows thus give for each group the
value of m feature denoting the characteristics of
these groups. In the present work features correspond
to characterization of data model i.e. hrs., two hours
before load, one hour before load, temp.1, temp.2 In
this paper, fuzzy IF-THEN rules of the form
suggested by Takagi- Sugeno [19] are employed,
where fuzzy sets are involved only in the premise
part of the rules while the consequent part is
described by a non-fuzzy function of the input
variables. The historical data is used to design data
which are further fuzzified using IF-THEN rule.
The data model involves the range of data low
(L), low medium (LM), medium (M), medium high
(MH) and high (H), five linguistic variables for each
crisp data type. These five linguistic value are
defined as L(3800 MW-4200 MW), LM
(4280.001MW- 4760 MW), M(4760.001 MW-
5240 MW), MH (5240.001 MW -5720 MW) and
H(5720.001 MW-6200 MW)and the linguistic values
for temperature are as L (-370°C to -230 °C), LM
(-229.999°Cto -90°C), M(-89.999°C to +50°C),
MH (+50.001°C to +190 °C) and H (+190.001°C to
Neural Network
Output
44
 +330°C), using IF-THEN rule. These data are Data Set
normalized and fuzzified using inference engine as
shown in demand table (Table-1). The five linguistic
variables using IF-THEN rule for load as well as
temperature are as follows. Input Data Set
If P1 is low (L) and P2 is low (L) then α=LL (Load1, Load2, Temp1, Temp2, Hr-Load)
If P1 is low (L) and P2 is low medium (LM) then
α=LLM
If P1 is low (L) and P2 is medium (LM) then α=LM Making of Rule Base
If P1 is low (L) and P2 is medium high (MH) then
α=LMH
If P1 is low (L) and P2 is high (H) then α=LH
If P1 is low medium (LM) and P2 is low (L) then
α=LML Categorization and Distribution of
If P1 is low medium (LM) and P2 is low medium Data Set
(LM) then α=LMLM and so on.

4.2 Forecasting Model

In FINN the RBFNN plays an important role to Data Conversion
classify input data into some clusters while the fuzzy Normalization of Data
inference engine handles the extraction of rules. Fig.
1 shows the structure of FINN that has two layers;
input/output and rule layers. The input/output layer
has input and output node. The input nodes of the Crisp set of data
input/output layer are connected to neurons on the
topological map of the rule layer. The fuzzy
membership neural networks are assigned to the Fuzzification (Fuzzified input data)
weight between the input nodes and rule layer. Also,
the consequent constant is assigned between the
output node and rule layer. The parameter selection
method can be considered as a rule base initialization Training and Testing through RBFNN
process. Essentially, it performs a fuzzification of the
selected input points within the premise space. The
mean values of the memberships are centered Forecasting
directly at these points, while the membership
deviations reflect the degree of fuzzification and are
selected in such a way that a prescribed degree of Actual Data
overlapping exists between successive memberships.
The fact that the initial parameters of the FINN are
not randomly chosen as in neural networks but are
assigned reasonable values with physical meaning Mean Absolute Percentage Error
gives the training of an FNN a drastic speed
advantage over neural networks.
With fusing the strongpoint of fuzzy logic and Figure 2: Flow chart of Forecasting Process
neural networks, a fuzzy inference neural networks
model, which effectively makes use of their Table 1: Demand table
advantages, has been developed. The training
patterns for the ANN models are obtained from the
historical loads by classifying the load patterns
according to the day-types of the special days and
linearly scaling the load values. The block diagram
of the proposed system and the flow chart of the
forecasting process are shown in the Fig.1. and Fig.2.

5 SIMULATION RESULTS

The most widely used index for testing the

performance of forecasters is the MAPE. The

UbiCC Journal - Volume 3 45

 Table 2: Forecast errors in MAPE on seasonal transition weeks

Winter Spring Summer Average

January 25-31 May 17-23 July 19-25
Day Day Week Day Week Day Week Day Week
Ahead Ahead Ahead Ahead Ahead Ahead Ahead Ahead
Monday 2.5711 2.5711 1.9990 1.9990 2.2050 2.2050 2.2584 2.2584
Tuesday 1.6763 1.5041 1.8121 1.8797 2.0467 1.9221 1.8450 1.7686
Wednesday 2.0342 2.0527 2.0369 1.9750 2.4277 1.9505 2.1663 1.9927
Thursday 2.4767 2.6438 2.2687 2.0208 1.5584 1.5206 2.1013 2.0617
Friday 2.9492 1.9225 1.8399 1.8356 1.5065 1.5079 2.0985 1.7553
Saturday 2.4953 2.3185 2.4913 2.3826 1.9120 1.9915 2.2995 2.2309
Sunday 2.7416 2.8998 2.6638 2.6110 1.6234 1.5122 2.3429 2.3410
Average 2.4206 2.2732 2.1588 2.1005 1.8971 1.8014 2.1588 2.0584

Table 3: Comparison with MLR and simple RBFNN

Winter Spring Summer

Day January 25-31 May 17-23 July 19-25
MLR RBFNN FINN MLR RBFNN FINN MLR RBFNN FINN
Monday 2.3863 1.0776 2.5711 2.7664 1.0856 1.9990 2.8015 1.2466 2.2050
Tuesday 1.6070 1.0727 1.5041 2.8966 0.7082 1.8797 2.2284 2.2017 1.9221
Wednesday 2.2656 1.1105 2.0527 3.3757 0.9606 1.9750 2.6688 0.8057 1.9505
Thursday 1.8675 0.7494 2.6438 2.3315 2.2876 2.0208 3.0628 1.2365 1.5206
Friday 1.6801 1.1171 1.9225 2.9397 1.1114 1.8356 2.6345 0.9062 1.5079
Saturday 2.8921 1.6459 2.3185 1.0263 0.7726 2.3826 2.4133 1.0312 1.9915
Sunday 2.3560 1.5838 2.8998 2.2336 1.7412 2.6110 2.1984 1.1475 1.5122
Average 2.3228 1.1939 2.2732 2.5100 1.2310 2.1005 2.5725 1.2246 1.8014

6000
A c tual
5800 Forec as ted

5600

5400
(MW
)

5200
oad

5000
L

4800

4600

4400

4200
0 20 40 60 80 100 120 140 160
Hour

Figure 3: Forecast for Winter (January 25-31)

5600

5400 Actual
Forecasted
5200

5000
(M
d W
)

4800
o
La

4600

4400

4200

4000
0 20 40 60 80 100 120 140 160
Hour

Figure 4: Forecast for Summer (July 19-25)

UbiCC Journal - Volume 3 46

 5600
5400
5200
5000
(MW
)
4800
oad
4600
L
4400
4200
4000
3800
0 20 40 60
Figure 5: Forecast for Spring (May 17-23)
designed network is used to forecast the day ahead
and week ahead forecast on an hourly basis.
Forecasting has been done on the one year load data
of Trans Alta Electric Utility for Alberta, Canada.
Load varies from 3900 MW to 6200MW. The
FINN is trained using last four weeks hourly load
data and then they are used to forecast the load for
the next 168 hours i.e. one week. The results are
reported for three weeks, one each for winter,
spring and summer seasons. This reflects the
behaviour of the network during seasonal changes
and corresponding results are shown in Table 2. It
is observed that the performance of the day ahead
and week ahead forecast are equally good. Load
shape curves for three weeks are shown in Fig. 3,
Fig. 4 and Fig. 5.The errors are tabulated in Table 2.
It is observed from the figures that the forecaster
captures the load shape quite accurately and the
forecasting error on most of the week days are low
with slightly higher error on weekend days.
For having a comparative study the proposed
FINN method is compared with other two methods,
conventional Multi Layer Regression and RBF
neural networks for the same period of time. The
result (Table 3) shows that the average MAPE for
FINN is better than MLR in all seasons and the
average MAPE for RBFNN is even better than
FINN. But at the same time it is also noticeable that
the training time required in the forecasting through
RBFNN integrated with Fuzzy Inference is
approximately ten times less than the training time
required for simple RBFNN.
6 CONCLUSION
The benefit of the proposed structure is to
utilize the advantages of both, i.e. the
generalization capability of ANN and the ability of
fuzzy inference for handling uncertain problems
and formalizing the experience and knowledge of
the forecasters. Load forecasting method proposed
above is feasible and effective. A comparative
study shows that FINN and RBFNN are more
UbiCC Journal - Volume 3
Actual
Forecasted
80 100 120 140 160
Hour
accurate as compared to MLR. The error depends
on many factors such as homogeneity in data,
network parameters, choice of model and the type
of solution. The flexibility of the fuzzy logic
offering a logical set of IF-THEN rules, which
could be easily understood by an operator, will be a
good solution for practical implementation. FINN
training time was much faster and also effectively
incorporated linguistic IF-THEN rules. Load
forecasting results show that FINN is equally good
for week ahead and day ahead forecasting and
requires lesser training time as compared to other
forecasting techniques, conventional regression
MLR and simple RBF neural network.
ACKNOWLEDGEMENT
The authors would like to thank TransAlta,
Alberta, Canada for providing the load data used in
the studies.
7 REFERENCES
[1.] A.D.Papalexopoulos, T.Hasterberg: A
Regression based Approach to Short Term
System Load Forecast , IEEE Trans. On
Power Systems. Vol.5, No.4, pp 1535-1544,
(1990).
[2.] A. Khotanzad, E. Zhou and H.Elragal: A
Neuro-Fuzzy approach to Short-term load
forecasting in a price sensitive environment,
IEEE Trans. Power Syst., vol. 17 no. 4, pp.
1273–1282, (2002).
[3.] A. G. Bakirtzis, J. B. Theocharis, S. J.
Kiartzis, and K. J. Satsios: Short-term load
forecasting using fuzzy neural networks,
IEEE Trans. Power Syst., vol. 10, pp.
1518–1524,(1995).
47
[4.] C.N. Lu, H.T. Wu and S. Vemuri: Neural
Network Based Short Term Load
Forecasting , IEEE Transactions on Power
Systems, Vol. 8, No 1, pp. 336-342, (1993).
PAS-101, pp. 71-78. (1982)
[5.] D.C.Park M.A.,El-Sharkawi, R.J.Marks,
L.E.Atlas and M.J.Damborg: Electric Load
Forecasting using an Artificial Neural
Networks , IEEE Trans. on Power Systems,
vol.6,No.2, pp. 442-449,(1991).
[6.] D.K.Ranaweera, .F.Hubele and
A.D.Papalexopoulos: Application of Radial
Basis Function Neural Network Model for
Short Term Load Forecasting , IEE Proc.
Gener. Trans. Distrib., vol. 142, No.1,
(1995).
[7.] D. Singh and S.P. Singh: Self selecting
neural network for short-term load
forecasting , Jour. Of Electric Power.
Component and Systems, vol. 29, pp 117-
130, (2001).
[8.] E. H. Mamdani and S. Assilian: An
experiment in linguistic synthesis with a
fuzzy logic controller, Int. J. Man–Mach.
Stud., vol. 7, no. 1, pp. 1–12, (1975).
[9.] F. Meslier: New advances in short term
load forecasting using Box and Jenkins
approach , Paper A78 051-5, IEEUES
Winter Meeting,( 1978).
[10.] Hiroyuki Mori and Hidenori Kobayashi:
Optimal fuzzy inference for short term load
forecasting, IEEE Trans. on Power Systems,
vol.11, No.2, pp. 390-396, (1996).
[11.] I. Mogram and S. Rahman : Analysis and
evaluation of five short term load forecast
techniques, IEEE Trans. On Power Systems.
Vol.4, No.4, pp 1484-1491, (1989).
[12.] Kwang-Ho Kim, Hyoung-Sun Youn, Yong-
Cheol Kang: Short-tem Load Forecasting
for Special Days in anomalous Load
Conditions Using Neural Network and
Fuzzy Inference Method, IEEE Trans. on
Power Systems, Vol. 15, pp. 559-569,
(2000).
[13.] K.H. Kim, J.K. Park, K.J. Hwang, and S.H.
Kim: Implementation of Hybrid Short-term
UbiCC Journal - Volume 3
Load Forecasting System Using Artificial
Neural Networks and Fuzzy Expert
Systems, IEEE Trans. on Power Systems,
vol. 10, no. 3, pp. 1534–1539, ( 1995).
[14.] K.L.Ho, Y.Y.Hsu, C.F.Chen, T.E.Lee,
C.C.Liang, T.S.Lai and K.K.Chen : Short
Term Load Forecasting of Taiwan Power
System using a Knowledge Based Expert
System, IEEE Trans.on Power Systems,
vol.5, pp. 1214-1221, (1990).
[15.] K.Y. Lee, Y.T. Cha, and J.H. Park: Short-
Term Load Forecasting Using An Artificial
Neural Network,” IEEE Trans. on Power
Systems, vol. 7, no. 1, pp. 124–132, (1992).
[16.] Ranaweera D.K., Hubele N.F. and Karady
G.G: Fuzzy logic for short-term load
forecasting, Electrical Power and Energy
Systems,” Vol. 18, No. 4, pp. 215-222,
(1996).
[17.] T. Takagi and M. Sugeno: Fuzzy
identification of systems and its
applications to modeling and control, IEEE
Trans. Syst., Man, Cybern., vol. 15, pp.
116–132, (1985).
[18.] T. S. Dillon, S. Sestito, and S. Leung: Short
term load forecasting using an adaptive
neural network, Elect. Power Energy Syst.,
vol. 13, (1991).
[19.] W.R.Christiaanse: Short Term Load
Forecasting using general exponential
smoothing, IEEE Trans. On Power Appar.
Syst.,PAS-3,pp 900-911 (1988)
[20.] Zadeh L.A: Roles of Soft Computing and
Fuzzy Logic in the Conception, Design and
Deployment of Information /Intelligent
Systems, Computational Intelligence, Soft
Computing and Fuzzy-Neuro Integration
with Applications, O Kaynak, LA Zadeh, B
Turksen, IJ Rudas (Eds.), pp 1-9. (1998).
48
 EXPLORING PERFORMANCE LANDSCAPE OF UNSTRUCTURED
SEARCH SCHEMES

Hong Huang and Rajagopal Reddy Manda

Klipsch School of Electrical and Computer Engineering, New Mexico State University, USA
{hhuang, rgopal}@nmsu.edu

ABSTRACT
Search plays an important role in ubiquitous computing. In this paper, we
investigate the expected cost and latency of three unstructured search schemes:
broadcast, TTL-based, and random-walk-based search schemes. We build a unified
analytical model for unstructured search schemes. We demonstrate, through
simulation, that the different search schemes exhibit very different cost and latency
tradeoffs, leaving large gaps in the performance landscape. We propose
randomized mixing schemes to bridge such performance gaps and bring out new
Pareto frontier that offers more diverse performance choice for applications.

Keywords: performance modeling, randomized mixing, search methods.

1 INTRODUCTION target without high cost broadcast. However, a TTL-

Search for information in a network has many
important applications in ubiquitous computing, such
as route discovery of on-demand routing protocols in
wireless ad hoc networks [1], event query in sensor
nets [2][3], information lookup in peer-to-peer
networks [4], etc. Unstructured search refers to a
type of search where no a priori knowledge about
the search target is available. Unstructured search is
applicable in cases where the network is highly
dynamic and maintaining an infrastructure for data
lookup is too costly [4]. Unstructured search can be
and sometimes is implemented by a broadcast search.
However, a broadcast search is very costly and not
scalable. A broadcast search is particularly
unjustifiable when the target is replicated for
robustness and latency reduction as is common with
today’s distributed applications. As mentioned in [4],
the fundamental problem of broadcast search lies at
the lack of granularity of the search action, i.e., either
no action or a very costly one--broadcast.
To reduce cost and provide for finer granular
search actions than broadcast search, a variety of
other unstructured search schemes were proposed.
Here we focus on two types of such schemes
appearing in recent literature: iterative broadcast
(TTL-based) schemes, and random-walk-based (RW-
based) schemes. In a TTL-based scheme [5][6], a
series of broadcasts with increasing scopes are
carried out. The scope of a broadcast is determined
by the TTL (Time to Live) value carried by a query
packet, which limits the number of hops the packet
can travel. A TTL-based scheme offers finer granular
search actions than a simple broadcast by varying its
scope; and promises to reduce cost by trying search
actions with smaller cost first in the hope to find the
based scheme can be wasteful in overlapped
coverage of successive broadcasts and generally
causes larger search latency than simple broadcast.
In a RW-based search, a query packet carries out
a random walk in the network, which continues until
the target is found [2][3]. A RW-based scheme has
the finest granular search action possible, i.e.,
visiting a single node; but it can cause large latency.
Variations on the basic random walk scheme are
possible to reduce latency, e.g., using multiple
walkers.
There are two main performance metrics for a
search scheme [6]: expected cost and expected
latency. Expected cost is defined as the expected
total number of hops traveled by the query packets
generated a particular search scheme. Expected
latency is defined as the expected time duration, in
the unit of hops (i.e., one hop takes one unit of time),
between initiation of the search and the discovery of
the target. We do not include the time for the result
to travel from the target to the originator, because it
has nothing to do with the merits of a search scheme.
Previous work most closely related to ours
includes the following. TTL-based search on a line is
first treated in [5] where an optimization problem is
formulated and the competitive ratio to the optimal
offline algorithm is obtained. In [6], a dynamical
programming formulism for TTL-based scheme is
developed, and a randomized strategy for selecting
TTL values is shown to achieve the minimum worst-
case cost if the target distribution is unknown. Such
randomized strategy is pursued further in [7] to
develop schemes that achieves lowest cost
competitive ratio under constraint of varying delay
competitive ratio. An analytical model for TTL-
based search using generating function of degree

UbiCC Journal - Volume 3 49

distribution is described in [15]. Random walk is of Fi-1. The search continues until success (i=l). So
shown to be an effective search method in sensor the expected cost of a search scheme can be written
nets in [2], and its behavior is examined in [3][12]. as
Unstructured search in peer-to-peer networks is
treated in [4] focusing on target replication strategy. l
Search in graphs with a power-law degree
distribution is treated in [14]. Hybrid search schemes
E[C ] = ∑F
i =1
i −1Ci (1)

combining broadcast and random walk is discussed

in [12]. It is easy to show the above can be rewritten as
The contributions of this paper are as follows. l ⎛ i ⎞
Although there are much previous works on
unstructured search schemes, there is no systematic
E[C ] = ∑∑ ⎜
⎜
i =1 ⎝ j =1
C j ⎟ ( Fi −1 − Fi )
⎟
(2)
⎠
exploration of the performance landscape of all
unstructured search schemes in one place,
The above expression can recognized as a
delineating the feasible regions of performance standard formula for computing expected value. For
tradeoffs. This paper hopes to make progress in this
each term in the summation, (Fi - Fi+1) is the
direction: We build a unified analytical model for
probability that search sequence does not succeed in
unstructured search schemes, which is parameterized the (i-1)th action but succeeds in the ith action, and
by the granularities of the search sequences. We
the summation in the parenthesis is cost of a search
demonstrate through simulation that different sequence terminating at ith action.
unstructured search schemes exhibit very different We can use a similar approach to write the
cost and latency tradeoffs, leaving large gaps in the average latency of a search scheme but for a
performance landscape. We propose randomized cautionary note. The latency incurred by Ai generally
mixing schemes to bridge such performance gaps depends on whether the search is successful or not.
and bring out new Pareto frontier that offers more Consider, for example, a search action with TTL set
diverse performance choice for applications. to be 10. If the target is within the TTL scope, say 7
The paper is organized as follows. In Section 2, hops away, the latency is 7. But if the target is more
we build a unified model for the three unstructured
than 10 hops away, the search action fails and incurs
search schemes. In Sections 3-5, we deal with a latency of 10 regardless where the target is located.
broadcast, TTL-based, RW-based search schemes, We write D”i as the latency if the target is within the
respectively. We introduce mixed schemes in Section
scope of ith action, i.e., if the action is successful;
6, and conclude in Section 7. and D’i as the latency if the target is out of the scope,
i.e., if the search action is unsuccessful, which is
2 A UNIFIED PERFORMANCE MODEL fixed for a particular action. So, with probability (Fi
FOR UNSTRUCTURED SEARCH SCHEMES - Fi+1), i.e., the search sequence does not succeed in
the (i-1)th action but succeeds in the ith action, the
We consider the problem of searching for a
search latency is
target in a network of N nodes. A target is replicated
in m copies, locating any of the replicas caused the
D '1 + D '2 + ... + D 'i −1 + D "i
search successful. We consider a search scheme
consisting of a sequence of actions A = [A1, A2,…, Al],
where Al is the first search action, A2 the second Thus the expected latency can be written as
search action, and so forth, and Al is the terminating
action in which the target is found. For an l ⎛ i −1 ⎞
unstructured search scheme, there is no outside clue E[ D ] = ∑ ⎜⎜ ∑ D ' + D " ⎟⎟ ( F j i i −1 − Fi ) (3)
about the next search action except the history of i =1 ⎝ j =1 ⎠
previous search actions. In a broadcast search, there
is a single action: broadcast. In a TTL-based scheme, Rearranging terms, we have
each action is a limited broadcast with a particular
TTL value. In a RW-based scheme, each action is a E[D] = E[ DBCast ] + E[ D '] (4)
step in the walk. As we can see, the actions of
different search schemes have different granularities, where
which have performance implications. l
We write Ci as the cost of performing search
action Ai, Di as the average latency caused by Ai,Fi as
E[ DBCast ] ≡ ∑ D" (F
i =1
i i −1 − Fi ) (5)
the probability that Ai fails, with the convention F0 =
1, Fl = 0. The cost of the first search action is always l⎛ i −1 ⎞ l −1

paid outright, but that of a later action, Ci, is paid out E[ D '] ≡ ∑∑ ⎜
⎜
i =1 ⎝
D ' j ⎟ ( Fi −1 − Fi ) =
⎟ ∑F D' i i (6)
only if the previous action failed with the probability j =1 ⎠ i =1

UbiCC Journal - Volume 3 50


In the above, E[DBcast] expresses the expected latency
for a search that expands in scope until successful
without incurring the latency of failed actions. This
part of latency is the same as the latency of a
broadcast search, which is independent of the
particular search scheme used and represents the
minimum latency of any search scheme. E[D’]
collects the latency incurred due to failed actions,
and is dependant of the particular search scheme in
question.
The above formulism is general without regard
to any particular search scheme in question. Before
proceed further, we list our assumptions.
A1 The cost of a search action (a single step) in a
nodes in the network if the topology is a complete
graph; whereas it incurs n/2-hop latency if the
topology is a ring of n nodes. Q.E.D.
In the following, we provide bounds on
minimum expected cost and latency, and identify
search schemes that achieve these bounds. First, we
define an idealized random walk.
Definition 2.1 An idealized random walk is one
that visits a distinctive node in every step.
Proposition 2.2 The minimum expected cost
among all unstructured search schemes is
N
Cmin = ∑ F ( j − 1)
j =1
(8)

random walk is 1.
A2 The cost of a search action that requires where F(j-1) is defined in the sense of (7). This
broadcasting to n nodes is n, i.e., Ci = Ni in minimum expected cost can be achieved by an
equations (1) and (2). idealized random walk.
A3 The m target replicas are independently, Proof: First, we show that for an arbitrary search
identically distributed among the nodes. scheme, its expected cost is at least Cmin. Suppose the
The above assumptions are admittedly idealistic, scheme’s search action sequence is [A1, A2,…, Al],
but they are used here to focus on the issue intrinsic and the corresponding node coverage sequence being
to the merits of a particular search scheme and [N1, N2,…, Nl = N]. It is clear that Ci >= ΔNi =Ni - Ni-
exclude external factors such as network conditions, 1, since it takes at least a cost of ΔNi to cover ΔNi
implementation efficiency, etc. Assumption A1 holds nodes. So we have
only if the links are lossless, which is not true in
practical situations. However, introducing loss l l l
requires specification of loss probability, which is
some extraneous detail outside the scope of our
E[C ] = ∑i =1
Fi −1Ci ≥ ∑
i =1
Fi −1ΔNi = ∑ F (N
i =1
i −1 ) ΔN i

discussion. Similarly, A2 holds only if the

implementation of broadcast is perfect, again more The right-hand side of the inequality cab be
idealism than realism. Broadcast is known to cause recognized as the discrete integration of Fi with a
redundancy and inefficiency, to mitigate which step size of ΔNi on the N-axis, with support intervals
methods have been devised [8]. Again, such details as N1 + ΔN2 + ΔN2 +… + ΔNl = [1, 2, … N]. Since Fi
fall out of the scope of the present discussion, and is a non-increasing function of Ni, so we have
similar approximation is used in [4]. Assumption A3
implies the probability of success depends only on l N
the number of nodes visited, and has nothing to do
with the identities of visited nodes. In other words,
∑
i =1
Fi −1ΔN i ≥ ∑ F ( j − 1) =
j =1
the failure probability Fi is solely a function of nodes l
visited so far, Ni, i.e., ∑( F (N
i =1
i −1 ) + F ( Ni −1 + 1) + ... + F ( Ni −1 + ΔNi − 1) )

Fi = F ( Ni ) (7)
The particular form of F depends on the distribution
of the replicas. Now we are ready for the following
proposition.
Proposition 2.1 A: The expected cost is
parameterized by and thus depends only on the
search sequence [N1, N2,…, Nl], regardless the
network topology. B: The expected latency depends
on both the search sequence and network topology.
Proof: The validity of the statement A is
apparent by combing equations (1) and (7). The
validity of the statement A derives from the fact that
different latencies are incurred in different network
topologies to cover the same number of nodes. For,
example, it incurs just one-hop latency to cover all
Second, an idealized random walk is a sequential
visit of each node. Without loss of generality, let the
sequence be j=1, 2,…, N, and thus incurs the expect
the minimum cost of
N
∑ F ( j − 1) Q.E.D.
j =1
The next proposition is obvious, and we state
without proof.
Proposition 2.2 The minimum expected latency
is the minimum expected distance in hops between
the originator and any one of target replicas, which
can be achieved by a broadcast search.

UbiCC Journal - Volume 3 51

 In next three sections, we derive cost and latency
models for three classes of unstructured search
schemes: broadcast, TTL-based, and RW-based. To
compute the expected cost and latency, one needs to
have a specific probability distribution for replicas in
the network. In the following, we assume replicas are
distributed uniformly in the network, which is a
common case, e.g., in DHT-based data lookup
scheme in peer-to-peer networks [4]. Thus, given m
replicas, the failure probability of a search action that
covers Ni nodes of N nodes in the network is
m E[ DBCast ] = m
⎛ N ⎞
Fi = ⎜1 − i ⎟
⎝ N⎠
number of nodes per square hop, depends on the
nodal degree and the network topology in question.
Expressions of α can be obtained only for some
special cases. For example, it is straightforward to
show that α = 1 for a square grid with nodal degree
of four, α = 2/√3 for hexagonal grid with nodal
degree of six, etc. Since the density only introduces a
scalar factor, we will, without loss of generality, use
the value of one in the following. Thus, in a 2-D
space, we have R = (N/π)1/2, and,
N Γ(m)
(11)
3
(9) 2Γ ( m + )
2

3 BRAODCAST SEARCH In passing, we note that for large m [9],

A broadcast search consists of a single action, 1

m−
2 e− m
⎛ ⎛ 1 ⎞⎞
i.e., broadcast. Its cost is fixed with E[CBCast] = N, Γ(m) ≅ 2π m ⎜1 + O ⎜ ⎟ ⎟
however the expected latency is not so trivial. To ⎝ ⎝ m ⎠⎠
make tractable analysis, we assume that a network
lives in a d-dimensional space, where the originator We have, for arbitrary dimension and large m,
resides in the center of the space. According to A3,
the probability that a target resides in a particular ⎛1 1⎞
−⎜ + ⎟
space is proportional to the volume of the space. The E[ DBCast ] ≅ Rm ⎝ 2 d ⎠ Γ( 1 + 1) (12)
probability that a target (at random distance x) is less d
than r hops away can be computed as

d
⎛r⎞ 4 TTL-BASED SEARCH SCHEMES
Pr ≡ P ( x ≤ r ) = ⎜ ⎟
⎝R⎠
A TTL-based search consists of a sequence of
broadcasts with increasing TTL values, with the last
where R is the radius of the network in hops. The action being a network-wide broadcast. To motivate
probability that minimum distance between the our analysis, we consider the first two actions in a
originator and the m target replicas is no larger than r TTL-based search, i.e., A1 and ,A2 with increasing
hops is TTL values, covering N1 and N2 nodes, incurring
costs of C1 and C2, respectively. Note that the node
m
⎡ ⎛ r ⎞d ⎤ coverage of A2 includes that of A1. We consider two
m m
Pr ≡ P ( x ≤ r ) = 1 − ⎢1 − ⎜ ⎟ ⎥ options: a) a sequence of search actions [A1, A2], or
⎢⎣ ⎝ R ⎠ ⎥⎦ b) a single action A2. Clearly, these two options have
the same probability of success at the conclusion of
And according to Proposition 2.2, the expected the search. However, their costs are different, i.e., Ca
latency of a broadcast search can be calculated as the = C1 + F1 C2 for option a, and Cb = C2, for option b.
expected minimum distance between originator and Option a is preferable to b only if C1 + F1C2 < C2.
the targets. Rearranging terms, we have the following lemma,
which was similarly stated in [6] but is derived here
m using simpler argument.
R L ⎡ ⎛ r ⎞d ⎤ Lemma 4.1 The cost of a search action A2 can
∫ ∫
m
E[ DBCast ] = rdPr = − rd ⎢1 − ⎜ ⎟ ⎥
0 0 ⎢⎣ ⎝ R ⎠ ⎥⎦ be reduced by using a sequence of search actions [A1,
A2], if the inequality holds:
1
Γ(m)Γ( + 1)
= mR d (10) C1
1 F1 < 1 − (13)
Γ(m + + 1) C2
d
Specializing to our model, and under the
where Γ( ) is the gamma function. For d = 2, which is assumptions A1-A3, we have C1 = N1, C2 = N2, and
our main focus here, we have N = απR2, with α being the inequality becomes
the node density. The node density, defined as the

UbiCC Journal - Volume 3 52

 the condition below
m
⎛ N ⎞ N
F1 = ⎜1 − 1 ⎟ < 1 − 1 ∂E[CTTL ]
⎝ N⎠ N2 = − fi −1m + mpi +1 fi m −1 = 0 (15)
∂fi
It holds only if m > 1, since N is the total number of
nodes in the network and can not be less than N2. So Equation (15) provides a recursive relationship that
we have: determines fi+2 given fi+1 and fi (note that pi = 1 - fi).
Corollary 4.1 If m = 1, the cost of a search The optimal solution is given by testing the values of
action A2 is no more than that of a sequence of search the expected cost at fi’s given by the recursive
actions [A1, A2]. equation and at the boundary fi = 0 (since fi <= 1, and
We generalize Corollary 4.1 in the following fi’s are strictly decreasing after each search action).
proposition. Exact solution of the above set of nonlinear
Proposition 4.1 Under assumptions A1 and A3, equations is difficult. However, we can try different
and if the number of replicas is one, the cost of a f1’s (note that f0 = 0) and determine a family of
broadcast search is no more than that of a TTL-based sequences. The family of sequences provides
search. different cost and latency tradeoffs and includes one
Proof: We use Corollary 4.1 recursively. that has the minimum cost.
Consider a TTL-based search scheme consists of a We calculate the expected latency using (4) and
search sequence [A1, A2,…, Al]. Apply Corollary 4.1 (6).
to the subsequence [A1, A2], which can be replaced
l −1
with a single action A2, resulting in a new search
sequence [A2, A3,…, Al], which has a cost no more E[ DTTL ] = E[DBCast ] + ∑F D
i =1
i i
than that of the original sequence. Applying
l −1 l −1 m l −1 1
Corollary 4.1 again to the new sequence, and so forth, ⎛ N ⎞
and eventually we can reduce the original sequence ∑
i =1
Fi Di = ∑
i =1
hi ⎜ 1 − i ⎟ = β
⎝ N ⎠ ∑
i =1
pi d fi m
to a single action Al, i.e., a broadcast, with a cost no
more than that of the original sequence. Q.E.D.
The above proposition implies that a TTL-based where hi is the TTL value (hop count) for the ith
search is competitive to a broadcast search only if m action, which can be expressed as
>1.
1
We now proceed to calculate the expected cost
and latency of a TTL-based scheme. In our model, a hi = β pi d
TTL-based search scheme consists of a sequence of
actions, A1, A2, …Al, with Ci = Ni, and Cl = N, where where β being a constant determined by network
Ni is the number of nodes covered in the ith round. topology.
Using (1), We calculate the expected cost as For 2-D space, which is our interest here, we
have, noting that is node density is 1, π hi 2 = pi N ,
l l m
⎛ N ⎞
E[CTTL ] = ∑i =1
Fi −1Ci = ∑
i =1
Ni ⎜ 1 − i −1 ⎟
⎝ N ⎠
and thus β = π , and therefore
(14)
l l −1
1
=N ∑ pi fi −1m E[ DTTL ] = E[ DBCast ] +
π
∑ i =1
N i fi m (16)
i =1
Ni
where pi ≡ and fi ≡ 1 − pi . Simulation results are shown in Figure 1 for a
N
network with N = 1000, and m = 2, 4,.., 10. The
network topology is generated randomly with
Clearly, the expected cost depends on the vectors
average degree of five. The expected cost and
{fi} or its complement {pi}, where fi indicates the
latency as a function of p1 (the proportion of nodes
proportion of nodes covered in the ith search action,
covered in the first action) are plotted in the figure.
and pi being its complement. We choose {pi} or {fi}
From the figure one can see that the expected cost is
to minimize the expected cost. Obviously, fi and pi,
small in the region of small p1, deceasing in a very
being probabilities, are constrained within the
gradual way to a minimum, and then rising quickly
interval of [0, 1]. Thus, we have a constrained
to become large at large p1, see Figure 1 (a). The
optimization problem. According to the Kuhn-
implication is that the expected cost is insensitive to
Tucker conditions for constrained optimization, the
the selection of p1 as long as it is small. This turns
fi’s that minimize the expected cost fall into two
out to be quite useful in practical implementation
cases: a) located at the constrain boundary, i.e., fi = 0
because there is no need to painstakingly search for
or 1, or b) located inside the interval and subject to

UbiCC Journal - Volume 3 53

the optimal p1, any choice would be fine as long as it fundamental problem for MRW is that there is no
is a small number. communication between walkers once they are
released. So even if one walker finds the target, other
1000
walkers would continue until they individually find
m=2
the target, incurring large cost. One way to rectify
900 m=4
m=6 this problem is to make random walkers terminate
m=8
800 m=10 probabilistically at each additional step. However,
700 this leaves open the possibility that all the walkers
600
might terminate before the target is found. Our
solution is to make one persistent walker that never
E[C]

terminates, assisted by k non-persistent walkers,

500

400
which survive with probability of q at each step. The
300 persistent walker guarantees the search is always
200
successful, while the non-persistent walkers
speculatively wander away trying to reduce latency.
100
In the following, we discuss expected cost and
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 latency of SRW and MRW sequentially.
p1
We model a SWR search scheme as a sequence
(a) of actions, A1, A2, …Al, with Ci = 1, Di= 1, and l >= N.
20
Let si indicate the number of distinct nodes visited in
the first i steps, then probability of failure to find any
18
m=2
m=4
of m replicas of the target is given by (since search
16 m=6
m=8 for each replica has to fail),
m=10
m
14
⎛ s ⎞
Fi = ⎜ 1 − i ⎟ = fi m (17)
12
⎝ N ⎠
E[D]

10

8
si
where fi = 1 − (18)
6
N
4

2 We calculate the expected cost using (1)

0
l −1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
l

∑ ∑f m
p1

E[CSRW ] = Fi −1Ci = i (19)

(b)
i =1 i =0

Figure 1: Expected cost E[C] (a) and latency E[D] (b)

versus p1 for TTL-based schemes. We compute expected latency as

l l −1

The expected latency shows an opposite trend: it E[ DSRW ] = ∑F

i =1
i −1 Di = ∑f
i =0
i
m
(20)
increases quickly to a maximum in the small p1
regime, then decreases monotonically and eventually
saturates at large p1, see Figure 1 (b). Particularly, The expected latency can be computed the same
the maximum latency occurs roughly around the way as the cost since each step incurs a fixed latency
point where minimum cost occurs, indicating that (one unit), which is not true in other search schemes.
reduced expected cost comes at the expense of A distinguishing fact for SRW is that the expected
increased latency. cost is the same as the expected latency. For other
search schemes, expected cost is generally larger
5 RW-BASED SEARCH SCHEMES than the expected latency. This is because, for an
action covering Ni nodes, the cost is paid outright as
In a RW-based search, a query packet visits Ni, but the latency can be smaller than Ni if the target
nodes sequentially and terminates when the target is is found before every node is covered. Such
found. There are a variety of ways to carry out a inequality does not apply to SRW because each step
RW-based search. One way is to employ a single is elementary, incurring unit cost and latency.
random walker (SRW). Such a search, if it is ideal, There remain to be determined the values of si.
incurs minimal cost according to Proposition 2.2, but There is no exact expression for si, though
it also causes large latency. To lower latency, asymptotic expressions do exist [10]. One such
multiple random walkers (MRW) can be released expression says
simultaneously, but with large increase in cost. A

UbiCC Journal - Volume 3 54

si → (1 − PR )i as i → ∞
where RR is probability that the walker will return to
the originator. But such asymptotic expression is not
useful. For the use of a RW-based search makes little
sense if i gets as large as N, let alone infinity,
because one might as well use broadcast with the
same cost but much less latency. We think that, in a
and are avoided unless all neighbors are flags, in
which the walker marches on a flagged node
regardless. The expected cost and latency as a
function of q, the survival probability, are plotted in
Figure 2, which is the case for m = 2. Plots for other
m values show similar feature, thus are omitted.
750

practical implementation, one will not use pure k=1

700
random walk simply for the sake of its theoretical k=2
k=3

purity, because it is patently inefficient, i.e., visiting

k=4
650
k=5

previously visited nodes repeatedly. Instead, some 600

optimization will be used in practical implementation,

such as using a flag to indicate previous visits.
550

E[C]
Therefore, in a practical implementation, the 500

behavior of a random walker approaches that of a 450

idealized walker, i.e., si Æi, to the extent that

optimization is done.
400

We now turn to MRW. In MRW, in addition to 350

one persistent random walker that terminates only if 300

the target is found, k independent, identical, non-

0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

persistent random walkers are also released, each

(a)
having a probability of 1-q to terminate at each step.
Since the random walkers are independent, we 340

calculate the expected cost of MRW as

330

⎡ l −1 ⎤ l −1

∑ ∑f
k=1

E[CMRW ] = k ⎢ Pi fi m ⎥ + m 320

i (21) k=2
k=3

⎣⎢ i =0 ⎦⎥
k=4
i =0 310
k=5
E[D]

where Fi is given by (17), and Pi is the probability 300

that a walker is still alive after ith step, and is given

by
290

280

q i − q l +1
Pi = (22)
1 − q l +1
270
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
q

(b)
The expected latency is determined by the
minimum of those among multiple walkers, whose Figure 2: Expected cost E[C] (a) and latency E[D] (b)
population at ith step is versus q for RW-based scheme

Ki = 1 + kPi (23)
The search continues only if all Ki walkers fail to
find any of m replicas, so the expected latency is
l −1
E[ DMRW ] = ∑f i =0
i
mKi
(24)
By varying k and q, we can construct a family of
RW-based schemes. Simulation results are shown in
Figure 2 for a network with N = 1000, random
topology and average degree being five, and with
one persistent walker plus one to five non-persistent
walkers. To approximate practical, optimized
implementation, previously visited nodes are flagged,
From the figure, one can see that large
performance variation occurs only where q is close
to 1. Roughly after q is smaller than 0.9, the
expected cost and latency fast converges to those
achieved by a single persistent walker. Figure 2 also
shows the tradeoff between expected cost and
latency, i.e., latency can be reduced, at the expense
of increased cost, by either increasing the number of
non-persistent walkers (k) or reducing the survive
probability (q). An important point, which will be
elaborated in the next section, is that the cost-latency
tradeoff here exhibits a very different feature from
that of TTL-based schemes, comparing Figures 1 and
2.

UbiCC Journal - Volume 3 55

6 MIXED SEARCH SCHEMES 1000

BCast
In this section, we examine more closely the 900

performance landscape of different unstructured 800

search schemes, and introduce mixed schemes to 700

TTL

provide more diverse cost and latency tradeoffs. In 600

RW

the following, we will simplify notion by using CA,

E[C]
DA to indicate E[CA], E[DA] of a particular search 500

scheme A, respectively. We start by laying down 400

some preparations. 300

Definition 6.1 The feasible region of a search 200

scheme in the performance space consists of the

union of expected cost and latency pairs, each
100

realizable using an instance of the search scheme. 0

0 50 100 150 200 250 300 350

Definition 6.2 A search scheme A dominates a E[D]

search scheme B if C A ≤ CB ,and DA ≤ DB . (a)

Definition 6.3 A search scheme is Pareto- 1000

optimal [11] if it is dominated by no other search 900

scheme, and the set of all search schemes that are BCast

Pareto-optimal forms the Pareto-frontier. 800

Definition 6.4 A mixed scheme between two 700

search schemes A and B is a randomized scheme that 600 TTL

selects A with a certain probability p and B with the

E[C]

500

probability 1-p.
Proposition 6.1 The feasible region of a mixed
400
RW

scheme is convex. 300

Proof: Consider two schemes A and B in the 200

feasible region, with expected cost and latency of CA, 100

CB, DA, DB, respectively. For any p in [0, 1], a mixed 0

scheme with expected cost and latency of pCA+(1- 0 20 40 60 80 100 120 140 160 180

E[D]
p)CB, pDA+(1-p)DB, is realizable. Therefore the
feasible region is convex. Q.E.D. (b)
Equipped with the above background, let us 1000

examine the cost-latency tradeoffs of three types 900 BCast

unstructured search schemes discussed previously.
Simulation results for a network of 1000 nodes,
800

random topology and average degree being five, are 700

shown in Figure 3. Broadcast consists of a point in 600

TTL

the performance region, incurring minimum latency

E[C]

500

but largest cost. TTL-based schemes consist of a 400

family of schemes parameterized by p1, the
proportion of nodes covered by the first action. TTL- 300

RW
based schemes achieve a slightly larger latency than 200

broadcast but can incur significantly lower cost, 100

especially for large m (number of replicas) and 0

optimized p1, see Figure 3(c). RW-based schemes 0 10 20 30 40 50

E[D]
60 70 80 90 100

consist of a family of schemes parameterized by k (=

1-5), the initial number of walkers, and q, the
survival probability. In the figure, individual curves
correspond to a particular value of k, with large k
curves lying higher. RW-based schemes can achieve
lowest cost, especially for small m, see Figure 3(a);
but it incurs largest latency.
A remarkable fact about Figure 3 is that different
search schemes exhibit very different cost-latency
tradeoffs, leaving conspicuous gaps between them.
(c)
Figure 3: Expected cost E[C] versus latency E[D] for
broadcast (BCast), TTL-based, and RW-based search
schemes, shown with m=2 (a), m=5 (b) and m=10 (c).
Dotted lines indicate Pareto frontiers of the mixed
scheme.
In the following, we focus on TTL-based and
RW-based schemes, omitting broadcast because of
its deficiency. We call a TTL-based scheme as
latency-preferred since its expected latency is low
and insensitive to the parameter (p1), which is largely

UbiCC Journal - Volume 3 56

limited by the E[DBCast] term in (16). However, its
expected cost can vary widely with the selection of
the parameter.
On the other hand, we call RW-based scheme as
cost-preferred since it exhibits the opposite behavior:
its expected cost is low and relatively insensitive to
choice of the parameters (k, q), but its expected
latency vary widely with selection of the parameters.
Further explanation, based on the analytical model,
of difference in cost-latency characteristic appears in
the appendix.
TTL-based and RW-based schemes are modeled
after broadcast and idealized random walk, which
incurs minimum latency and cost, respectively.
Therefore, the fact that TTL-based schemes are
latency-preferred and RW-based schemes are cost-
preferred is not surprising in retrospect. Due to the
peculiar performance characteristic of each type of
scheme, any particular type may not satisfy the
requirement of an application. Applications with
varying cost-latency requirements can be best served
if a more diverse Pareto frontier is accessible. This
can be accomplished by mixing TTL-based and RW-
based schemes. The dotted lines in Figure 3 indicate
the Pareto frontiers of the mixed scheme, and is
7 CONCLUSION
Unstructured search have high potential for
large-scale, highly dynamic networks, and it
provides a rich field for research endeavors.
Broadcast, TTL-based, RW-based schemes represent
some typical instances of unstructured search.
However, there exists large gap in performance
landscape between these pure schemes. Mixed
schemes can help to bridge the performance gap.
This is important because different applications
require different performance tradeoff, which the
mixed scheme can provide where a pure scheme can
not.
E[C]
LA
P1
PA
P2
C0
LB

constructed based on the following proposition.

Proposition 6.2 The Pareto frontier of mixing PB
two families of search schemes A and B consists of
parts of the Pareto frontiers of A and B, namely LA D0
and LB, and the Pareto frontier of the linear E[D]
combination of points on LA and LB.
Figure 4: Pareto frontier of a mixed scheme can satisfy
Proof: We prove that any point, P, in the
application requirement (shaded region) otherwise not
feasible region of the mixed scheme is dominated by satisfied by a pure scheme.
the proposed Pareto frontier of in the proposition. P
can come from only three sources: feasible region of
a pure A scheme, that of a pure B scheme, and linear 8 APPENDIX
combinations (mixing) of scheme A and B. In case
that P is from a pure A region, P is dominated by We provide more detailed explanation, based on
some point on LA. In case that P is from a pure B the analytical models developed earlier, the cost-
region, P is dominated by some point on LB. In case latency tradeoffs of TTL-based and RW-based
that P is from a linear combination of points P1 of A schemes, which can be quantified by the cost-latency
scheme and P2 of B scheme, P is dominated by a slope, defined in an obvious way as dC/dD.
linear combination of points PA and PB, where PA and For TTL-based scheme, using (14) and (16), we
PB are on LA and LB and dominates P1 and P2, have
respectively, refer to Figure 4. Q.E.D.
We provide a simple example to demonstrate the dCTTL
value of a mixed scheme. Suppose an application has − ( N + C '2 + )
dCTTL dp1
the requirement: E[C] < C0 and E[D] < D0. Such = =
requirement is satisfied only by a mixed scheme and dDTTL dDTTL f1m ⎛ p1 1 ⎞
dp1
m N ⎜ − ⎟ + D '2 +
is not satisfied by either of the pure schemes, refer to π p1 ⎝ f1 2m ⎠
Figure 4.
A cautionary note before we conclude: a mixed
where in the last expression we separate out the
scheme achieves the cost-latency point only in the
leading term and the rest of the terms in dCTTL/dp1
expected sense, but the variance can be large.
and dDTTL/dp1, respectively. Exact evaluation of
However, often times application quality of service
dCTTL/dDTTL is difficult, but we can gain some insight
requirements are expressed in terms of expected
by just examining the ratio of the leading terms,
performance metrics, and mixed schemes are useful
since the individual terms decays rapidly as O(fim),
to achieve performance tradeoffs otherwise
and with significant jumps in TTL value (thus fi) for
inaccessible to pure schemes.
successive search actions. The ratio of the leading

UbiCC Journal - Volume 3 57

terms (LT) is Randomization, in Proc. ACM MobiCom (2004).
[7] N. Chang and M. Liu: Controlled Flooding
⎛ dCTTL ⎞ ⎛ N ⎞ Search with Delay Constraints, in Proc. IEEE
⎜ ⎟ = O ⎜⎜ m ⎟⎟ INFOCOM (2006).
⎝ dDTTL ⎠ LT ⎝ mf1 ⎠ [8] Ni, S.Y., Tseng, Y.C., Chen, Y.S., Sheu, J.P.:
The Broadcast Storm Problem in a Mobile Ad
which explains the large cost-latency slope of TTL- Hoc Network, in Proc. ACM MobiCom (1999).
based schemes in Figure 3 (because N is large). [9] D. Zwillinger: CRC Standard Mathematical
We proceed in a similar manner with the RW- Tables and Formulae, Chapman & Hall/CRC
based schemes, the difference being that now we (2003).
have two parameters (k, q), and that fi decreases [10] B. H. Hughes: Random Walks and Random
slowly without jump (decrement being at most 1/N). Environments, Volume 1: Random Walks,
Thus a leading term approximation does not provide Clarendon Press, Oxford (1995).
much insight here. We write the full expressions [11] M. J. Osborne, and A. Rubenstein: A Course in
below, using (21) and (24). Game Theory, MIT Press (1994).
[12] M. M. Christos Gkantsidis and A. Saberi:
l −1 Random walks in peer-to-peer networks, in Proc.

∂CRW
∂CTTL
∑P f i i
m
of INFOCOM (2004).
= ∂k = i =0 [13] Gkantsidis, C., Mihail, M., Saberi, A.: Hybrid
∂DRW ∂DTTL l −1 search schemes for unstructured peer-to-peer
q
∂k q
m ∑P f
i =0
i i
mKi
ln fi networks, in Proc. of INFOCOM (2005)
[14] L. A. Adamic, R. M. Lukose, B. Huberman, and
l −1 A. R. Puniyani: Search in power-law networks,
dPi
∂CRW
∂CTTL
∂q
∑ dq fi m Physical Review E., vol. 64, no. 046135 (2001).
= = i =0 [15] R. Gaeta, G. Balbo, S. Bruell, M. Gribaudo, M.
∂DRW ∂DTTL l −1
dPi Sereno: A simple analytical framework to
k
∂q k
m ∑ dq
i =0
fi mKi ln fi analyze search strategies in large-scale peer-to-
peer networks, Performance Evaluation, Vol. 62,
Issue 1-4, (2005)
From the above, we can explain the smaller cost-
latency slope by the absence of the √N term as with
TTL-based schemes. Further, we can attribute the m
factor in the denominator to the fact that the slope
becomes smaller with larger m, as shown in Figure 3.

9 REFERENCES

[1] E. M. Royer and C.-K. Toh: A Review of

Current Routing Protocols for Ad Hoc Mobile
Wireless Networks, IEEE Personal
Communications, Vol 6, no. 2, pp. 46-55 (1999).
[2] D. Braginsky and D.Estrin: Rumor Routing
Algorithm for Sensor Networks, in Proc.
International Conference on Distributed
Computing Systems (2002).
[3] S. Shakkottai: Asymptotics of Query Strategies
over a Sensor Network, in Proc. IEEE
INFOCOM (2004).
[4] E. Cohen and S. Shenker: Replication Strategies
in Unstructured Peer-to-Peer Networks, in Proc.
ACM SIGCOMM (2002).
[5] Y. Baryshinikov, E. Coffman, P. Jelenkovic, P.
Momcilovic, and D. Rubenstein: Flood Search
under the California Split Rule, Operations
Research Letters, vol. 32, no. 3, pp. 199-206,
(2004).
[6] N. Chang and M. Liu: Revisiting the TTL-based
Controlled Flooding Search: Optimality and

UbiCC Journal - Volume 3 58

 Impact of Query Correlation on Web Searching
Ash Mohammad Abbas
Department of Computer Engineering
Zakir Husain College of Engineering and Technology
Aligarh Muslim University, Aligarh - 202002, India.
Abstract— Correlation among queries is an important factor to
analyze as it may affect the results delivered by a search engine.
In this paper, we analyze correlation among queries and how
it affects the information retrieved from the Web. We analyze
two types of queries: (i) queries with embedded semantics, and
(ii) queries without any semantics. In our analysis, we consider
parameters such as search latencies and search relevance. We
focus on two major search portals that are mainly used by
end users. Further, we discuss a unified criteria for comparison
among the performance of the search engines.
Index Terms— Query correlation, search portals, Web infor-
mation retrieval, unified criteria for comparison, earned points.
I. I NTRODUCTION
The Internet that was aimed to communicate research ac-
tivities among a few universities in United States has now
become a basic need of life for all people who can read and
write throughout the world. It has become possible only due
to the proliferation of the World Wide Web (WWW) which is
now simply called as the Web. The Web has become the largest
source of information in all parts of life. Users from different
domains often extract information that fits to their needs.
The term Web information retrieval1 is used for extracting
information from the Web.
Although, Web information retrieval finds its roots to tra-
ditional database systems [4], [5]. However, the retrieval of
information from the Web is more complex as compared to the
information retrieval from a traditional database. This is due
to subtle differences in their respective underlying databases2 .
In a traditional database, the data is often organized, limited,
and static. As opposed to that the Webbase is unorganized,
unlimited, and is often dynamic. Every second a large number
of updates are carried out in the Webbase. Moreover, as
opposed to a traditional database which is controlled by a
specific operating system and the data is located either at
a central location or at least at a few known locations, the
Webbase is not controlled by any specific operating system
and its data may not reside either at a central site or at few
known locations. Further, the Webbase can be thought as a
collection of a large number of traditional databases of various
organization. The expectations of a user searching information
1 The terms Web surfing, Web searching, Web information retrieval, Web
mining are often used in the same context. However, they differ depending
upon the methodologies involved, intensity of seeking information, and
intentions of users who extract information from the Web.
2 Let us use the term Webbase for the collection of data in case of the Web,
in order to differentiate it from the traditional database.
UbiCC Journal - Volume 3
on the Web are much higher than the user which is simply
retrieving some information from a traditional database. This
makes the task of extracting information from the Web a bit
challenging [1].
Since the Web searching is an important activity and the
results obtained so may affect decisions and directions for
individuals as well as for organizations, therefore, it is of
utmost importance to analyze the parameters or constituents
involved in it. Many researchers have analyzed many different
issues pertaining to Web searching that include index quality
[2], user-effort measures [3], Web page reputation [6], and user
perceived quality [7].
In this paper, we try to answer the following question: What
happens when a user fires queries to a search engine one by
one that are correlated? Specifically, we wish to evaluate the
effect of correlation among the queries submitted to a search
engine (or a search portal).
Rest of this paper is organized as follows. In section II, we
briefly review methodologies used in popular search engines.
In section III, we describe query correlation. Section IV
contains results and discussion. In section V, we describe a
criteria for comparison of search engines. Finally, section VI
is for conclusion and future work.
II. A R EVIEW OF M ETHODOLOGIES U SED IN S EARCH
E NGINES
First we discuss a general strategy employed for retrieving
information from the Web and then we shall review some of
the search portals.
A. A General Strategy for Searching
A general strategy for searching information on the Web is
shown in Fig. 1. Broadly a search engine consists of the fol-
lowing components: User Interface, Query Dispatcher, Cache 3 ,
Server Farm, and Web Base. The way these components
interact with one another depends upon the strategy employed
in a particular search engine. We describe here a broad view.
An end user fires a query using an interface, say User Interface.
The User Interface provides a form to the user. The user fills
the form with a set of keywords to be searched. The query
goes to the Query Dispatcher which, after performing some
refinements, sends it to the Cache. If the query obtained after
3 We use the word Cache to mean Search Engine Cache i.e. storage space
where results matching to previously fired queries or words are kept for future
use.
59
 2 4
1 U 3
S Query 5
query E Dispatcher
R
I 8
N 7
T
WEB BASE
E
R
response F Server Farm
Cache
A
C
9 E 6

Fig. 1. A general strategy for information retrieval from the Web.

refinement4 is matched to a query in the Cache, the results are
immediately sent by the Query Dispatcher to the User Interface
and hence to the user. Otherwise, the Query Dispatcher sends
the query to one of the Server in the Server Farm which are
busy in building a Web Base for the search engine. The server
so contacted, after due consideration from the Web Base sends
it to the Cache so that the Cache may store those results
for future reference, if any. Cache sends them to the Query
Dispatcher. Finally, through the User Interface, response is
returned to the end user.
In what follows, we briefly review the strategies employed
by different search portals.
B. Review of Strategies of Search Portals
The major search portals or search engines5 which end users
generally use for searching are GoogleTM and YahooTM . Let
us briefly review the methodologies behind their respective
search engines6 of these search portals.
Google is based on the PageRank scheme described in [8].
It is somewhat similar to the scheme proposed by Kleinberg in
[9] which is based on hub and authority weights and focuses
on the citations of a given page. To understand the Google’s
strategy, one has to first understand the HITS (Hyperlink-
Induced Topic Search) algorithm proposed by Klienberg. For
that the readers are directed to [9] for HITS and to [8] for
PageRank.
On the other hand, Yahoo employs an ontology based
 1  j  m be a set of words excluding stop words.
search engine. An ontology is a formal term used to mean a
hierarchical structure of terms (or keywords) that are related.
The relationships among the keywords are governed by a set
search engine may not search words that are not part of its
ontology. It can modify its ontology with time. One step more,
an ontology based search engine may also shorten the set of
results searched before presenting it to the end users that are
not part of the ontology of the given term.
We now describe an important aspect pertaining to informa-
tion retrieval from the Web. The results delivered by a search
engine may depend how the queries are formulated and what
relation a given query has with previously fired queries, if any.
We wish to study the effect of correlation among the queries
submitted to a search engine.
III. Q UERY C ORRELATION
The searched results may differ depending upon whether
a search engine treats a set of words as an ordered set or an
unordered set. In what follows, we consider each one of them.
A. Permutations
Searched results delivered by a search engine may depend
upon the order of words appearing in a given query7. If we
take into account order of words, the same set of words may
form different queries for different orderings. The different
orderings of the set of words of the given query are called
permutations. The formal definition of permutations of a given
query is as follows.

Definition 1: Let the query Q wi  1  i  m, Q  φ, be
a set of words excluding stop words of a natural language. Let
P  x j
If P is such that wi x j for some j not necessarily equal to
i, and  wi  Q 
x j  P such that wi x j where j may not be
of rules. As a result, an ontology based search engine such
equal to i, then P is called a permutation of Q.
as Yahoo may search other related terms that are part of
In the above definition, stop words are language dependent.
the ontology of the given term. Further, an ontology based
For example in the English language, the set of stop words,
4
By refinement of a query, we mean that the given query is transformed S, is often taken as
in such a way so that the words and forms that are not so important are

eliminated so that they do not affect the results. S a an  the  is  am  are  will  shall  of  in  for 
5 A search engine is a part of search portal. A search portal provides many
other facilities or services such as Advanced Search, News etc. 7 The term ’query’ means a set of words that is given to a search engine to
6 The respective products are trademarks of their organizations. search for the information available on the Web.
UbiCC Journal - Volume 3 60
 1
Note that if there are m words (excluding the stop words) in Google
Yahoo

the given query, the number of permutations is m!.

The permutations are concerned with a single query. Sub- 0.8

mitting different permutations of the given query to a search

engine, one may evaluate how the search engine behaves for 0.6

Latency
different orderings of the same set of words. However, one
would like to know how the given search engine behaves when 0.4

an end user fires different queries that may or may not be

related. Specifically, one would be interested in the behavior 0.2

of a given search engine when the queries are related. In what

follows, we discuss what is meant by the correlation among 0
1 2 3 4 5 6 7 8 9 10
different queries. Page Number

Fig. 2. Latency versus page number for permutation P1.

B. Correlation
An important aspect that may affect the results of Web 1
Google
searching is how different queries are related. Two queries Yahoo

are said to be correlated if there are common words between 0.8

them. A formal definition of correlation among queries is as
follows.
0.6
Definition 2: Let Q1 and Q2 be queries given to a search

Latency
engine such that Q1 and Q2 are sets of words of a natural
language and Q1  Q2  φ. Q1 and Q2 are said to be correlated 0.4

if and only if there exists a set C Q1 Q2 , C  φ.

One may use the above definition to define k-correlation 0.2

between any two queries. Formally, it can be stated as a

corollary of Definition 2. 0
1 2 3 4 5 6 7 8 9 10

Corrollary 1: Two queries are said to be k-correlated if and Page Number

only if
C
k, where

denotes the cardinality. Fig. 3. Latency versus page number for permutation P2.
For two queries that are correlated, we define a parameter
called Correlation Factor8 as follows. 1
Google
Yahoo

Q1 Q2

Correlation Factor  (1)

Q1  Q2
0.8

This is based on the fact that
Q1  Q2

Q1

Q2

Q1 Q2
. 0.6

Note that 0  Correlation Factor  1. For two uncorrelated

Latency

queries the Correlation Factor is 0. Further, one can see from 0.4

Definition 1 that for the permutations of the same query,

Correlation Factor is 1. 0.2

Similarly, one may define the Correlation Factor for a

cluster of queries. Let the number of queries be O. The 0
1 2 3 4 5 6 7 8 9 10
cardinality of the union of the given cluster of queries is given Page Number

by the following equation. Fig. 4. Latency versus page number for permutation P3.

O

Qo
∑ Qi ∑



Qi  Q j
 ∑
Qi  Qj  Qk
1
Google
o 1 i i j i j  k Yahoo

O 1
 1
Q1  Q2    QO
(2) 0.8

Using (2), one may define the Correlation Factor of a cluster

of queries as follows. 0.6
Latency

O


o  1 Qo

Correlation Factor O
(3) 0.4

 o  1 Qo

A high correlation factor means that the queries in the cluster 0.2

are highly correlated, and vice versa.

In what follows, we discuss results pertaining to query
0
correlation. 1 2 3 4 5 6
Page Number
7 8 9 10

8 This correlation factor is nothing but Jaccard’s Coefficient, which is often Fig. 5. Latency versus page number for permutation P4.
used as a measure of similarity.
UbiCC Journal - Volume 3 61
 TABLE I
S EARCH LATENCIES , QUERY SPACE , AND THE NUMBER OF RELEVANT RESULTS FOR DIFFERENT PERMUTATIONS OF THE QUERY: Ash Mohammad Abbas
FOR G OOGLE .
Permutation p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
1 0.22 0.15 0.04 0.08 0.33 0.29 0.15 0.13 0.16 0.17
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
8 5 1 0 2 0 0 3 0 0
2 0.51 0.15 0.22 0.19 0.13 0.12 0.10 0.27 0.16 0.15
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
3 2 2 1 1 0 2 0 0 0
3 0.30 0.08 0.18 0.20 0.14 0.25 0.13 0.21 0.14 0.21
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
6 4 1 3 2 1 1 0 0 0
4 0.60 0.07 0.35 0.11 0.13 0.15 0.23 0.13 0.28 0.26
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
3 0 2 1 0 0 2 0 1 1
5 0.38 0.09 0.39 0.14 0.17 0.15 0.14 0.16 0.15 0.13
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
3 2 1 2 1 0 0 1 1 1
6 0.36 0.15 0.10 0.12 0.18 0.17 0.15 0.13 0.20 0.15
300000 300000 300000 300000 300000 300000 300000 300000 300000 300000
5 4 1 3 0 2 1 2 2 0

TABLE II
S EARCH LATENCIES , QUERY SPACE , AND THE NUMBER OF RELEVANT RESULTS FOR DIFFERENT PERMUTATIONS OF THE QUERY: Ash Mohammad Abbas
FOR YAHOO .
Permutation p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
1 0.15 0.15 0.27 0.24 0.25 0.23 0.34 0.21 0.27 0.30
26100 26400 26400 27000 27000 26900 26900 26900 25900 25900
10 4 1 0 0 1 0 0 0 0
2 0.18 0.13 0.20 0.15 0.19 0.10 0.15 0.09 0.12 0.13
26900 27000 27000 26900 25800 26900 26900 26800 26800 26800
4 6 1 1 0 1 1 1 0 0
3 0.12 0.11 0.15 0.14 0.11 0.10 0.11 0.12 0.09 0.13
26900 27100 26900 26900 26500 26800 26800 26500 26500 26700
10 3 1 2 0 0 0 0 0 0
4 0.03 0.10 0.14 0.13 0.12 0.20 0.10 0.19 0.12 0.17
27000 26400 26400 26700 27000 26700 26400 26900 26800 26800
7 4 0 2 1 0 0 1 0 1
5 0.12 0.12 0.20 0.08 0.13 0.10 0.12 0.09 0.13 0.20
26400 26800 26800 26800 26800 26700 26700 26800 26700 26200
8 5 1 1 0 0 0 0 0 1
6 0.16 0.10 0.16 0.12 0.13 0.11 0.10 0.11 0.12 0.15
27100 26700 27100 26700 26600 27000 26600 26900 26500 26500
10 5 0 0 0 1 0 0 0 0

1 1
Google Google
Yahoo Yahoo

0.8 0.8

0.6 0.6
Latency
Latency

0.4 0.4

0.2 0.2

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Page Number Page Number

Fig. 6. Latency versus page number for permutation P5. Fig. 7. Latency versus page number for permutation P6.

IV. R ESULTS AND D ISCUSSION

The search portals that we have evaluated are Google and
Yahoo. We have chosen them because they represent the search
portals that majority of end users in today’s world use in
their day-to-day searching. One more reason behind choosing
them for performance evaluation is that they represent different
UbiCC Journal - Volume 3
classes of search engines. As mentioned earlier, Yahoo is based
on ontology while Google is based on page ranks. Therefore,
if one selects them, one may evaluate two distinct classes of
search engines.
The search environment is as follows. The client from where
queries were fired was a Pentium III machine. The machine
62
 0.65
Google:Q1 was part of a 512Kbps local area network. The operating
Google:Q2
0.6
Yahoo:Q1
Yahoo:Q2 system was Windows XP.
0.55 In what follows, we discuss behavior of search engines for
0.5
different permutations of a query.
0.45
Latency

0.4
A. Query Permutations
0.35
To see how a search engine behaves for different permuta-
0.3
tions of a query, we consider the following query.
0.25

Ash Mohammad Abbas

0.2
1 1.5 2 2.5 3 3.5 4
Correlation
The different permutations of this query are
Fig. 8. Latency versus correlation for queries with embedded semantics.
1 Ash Mohammad Abbas
1.1
2 Ash Abbas Mohammad
Google:Q1
Google:Q2
Yahoo:Q1
3 Abbas Ash Mohammad
1 Yahoo:Q2
4 Abbas Mohammad Ash
0.9
5 Mohammad Ash Abbas
0.8 6 Mohammad Abbas Ash
0.7
We have assigned a number to each permutation to differ-
Latency

0.6
entiate from one another. We wish to analyze search results on
0.5 the basis of search time, number of relevant results and query
0.4
space. The query space is nothing but the cardinality of all
0.3
results returned by a given search engine in response to a given
query. Note that search time is defined as the actual time taken
0.2
1 1.5 2 2.5
Correlation
3 3.5 4 by the search engine to deliver the results searched. Ideally, it
Fig. 9. Latency versus correlation for random queries.
does not depend upon the speeds of hardware, software, and
network components from where queries are fired because it is
the time taken by the search engine server. Relevant results are
0.65
Google:Q1
Google:Q2
those which the user intends to search. For example, the user
Yahoo:Q1
0.6 Yahoo:Q2 intends to search information about Ash Mohammad Abbas9.
0.55 Therefore, all those results that contain Ash Mohammad Abbas
0.5 are relevant for the given query.
0.45
In what follows, we discuss the results obtained for different
Latency

0.4
permutation of a given query. Let the given query be Ash
Mohammad Abbas. For all permutations, all those results that
0.35
contain Ash Mohammad Abbas are counted as relevant results.
0.3
Since both Google and Yahoo deliver the results page wise,
0.25 therefore, we list all parameters mentioned in the previous
0.2 paragraph page wise. We go up to 10 pages for both the search
1 1.5 2 2.5 3 3.5 4
Correlation engines as the results beyond that are rarely significant.
Fig. 10. Query Space versus correlation for queries with embedded semantics. Table I shows search latencies, query space, and the number
of relevant results for different permutations of the given query.
1.1
Google:Q1
The search portal is Google. Our observations are as follows.
Google:Q2
1
Yahoo:Q1
Yahoo:Q2 For all permutations, the query space remains the same
0.9 and it does not vary along the pages of results.
0.8
The time to search the first page of the results in response
to a the given query is the largest for all permutations.
0.7
Latency

The first page of results contain the most relevant results.

0.6

0.5 9 We have intentionally taken the query: Ash Mohammad Abbas. We wish

0.4
to search for different permutations of a query and the effect of those
permutations on query space and on the number of relevant results. The
0.3 relevance is partly related to the intentions of an end-user. Since we already
know what are the relevant results for the chosen query, therefore, this is
0.2
1 1.5 2 2.5 3 3.5 4 easier to decide what relevant results out of them have been returned by a
Correlation
search engine. The reader may take any other query, if he/she wishes so. In
Fig. 11. Query Space versus correlation for random queries. that case, he has to decide what are the results that are relevant to his/her
query and this will partly depend upon what he/she intended to search.
UbiCC Journal - Volume 3 63
 TABLE III
Q UERIES WITH EMBEDDED SEMANTICS .
S. No. Query No. Query Correlation
E1 Q1 node disjoint multipath 1
Q2 edge disjoint multicast
E2 Q1 node disjoint multipath routing 2
Q2 edge disjoint multicast routing
E3 Q1 node disjoint multipath routing 3
Q2 edge disjoint multipath routing
E4 Q1 node disjoint multipath routing ad hoc 4
Q2 wireless node disjoint multipath routing

TABLE IV
Q UERIES WITHOUT EMBEDDED SEMANTICS ( RANDOM QUERIES ).
S. No. Query No. Query Correlation
R1 Q1 adhoc node ergonomics 1
Q2 quadratic power node
R2 Q1 computer node constellations parity 2
Q2 hiring parity node biased
R3 Q1 wireless node parity common mitigate 3
Q2 mitigate node shallow rough parity
R4 Q1 few node parity mitigate common correlation 4
Q2 shallow mitigate node parity common stanza

TABLE V
S EARCH TIME AND Q UERY S PACE FO QUERIES WITH EMBEDDED SEMANTICS .
S. No. Query No. Google Yahoo
Time Query Space Time Query Space
E1 Q1 0.27 43100 0.37 925
Q2 0.23 63800 0.28 1920
E2 Q1 0.48 37700 0.40 794
Q2 0.32 53600 0.32 1660
E3 Q1 0.48 37700 0.40 794
Q2 0.24 21100 0.34 245
E4 Q1 0.31 23500 0.64 79
Q2 0.33 25600 0.44 518

TABLE VI
S EARCH TIME AND QUERY SPACE FOR RANDOM QUERIES .
S. No. Query No. Google Yahoo
Time Query Space Time Query Space
R1 Q1 0.44 28500 0.57 25
Q2 0.46 476000 0.28 58200
R2 Q1 0.46 34300 0.55 164
Q2 0.42 25000 0.35 90
R3 Q1 0.47 25000 0.40 233
Q2 0.33 754 0.68 31
R4 Q1 0.34 20000 0.58 71
Q2 1.02 374 0.64 23

Table II shows the same set of parameters for different
permutations of the given query for search portal Yahoo. From
the table, we observe that
As opposed to Google, the query space does not remain
same, rather it varies with the pages of searched results.
The query space in this case is less than Google.
The time to search the first page of results is not neces-
sarily the largest of the pages considered. More precisely,
it is larger for the pages where there is no relevant result.
Further, the time taken by Yahoo is less than that of
Google.
In most of the cases, the first page contains the largest
UbiCC Journal - Volume 3
number of relevant results. For permutation 2 (i.e. Ash
Abbas Mohammad), the second page contains the largest
number of relevant results.
Let us discuss reasons for the above mentioned observations.
Consider the question why query space in case of Google is
larger than that of Yahoo. We have pointed out that Google
is based on the page ranks. For a given query (or a set of
words), it ranks the pages. It delivers all the ranked pages
that contain the words contained in the given query. On the
other hand, Yahoo is an ontology based search engine. As
mentioned earlier, it will search only that part of its Webbase
that constitutes the ontology of the given query. This is the
64
 TABLE VII TABLE IX
L ATENCY MATRIX , L, FOR DIFFERENT PERMUTATIONS . R ELEVANCE MATRIX FOR DIFFERENT PERMUTATIONS FOR G OOGLE .
P p1 p2 p3 p4 p5 p6 p7 p8 p9 p10 P p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
1 1 1
1 1 0 0 1 1 1 1 1 8 5 1 0 2 0 0 3 0 0
2
2 0 0 0 0 1 0 1 0 0 0 2 3 2 2 1 1 0 2 0 0 0
3 0 1 0 0 0 0 0 0 0 0 3 6 4 1 3 2 1 1 0 0 0
4 0 1 0 1 0 1 0 1 0 0 4 3 0 2 1 0 0 2 0 1 1
5 0 1 0 0 0 0 0 0 0 1 5 3 2 1 2 1 0 0 1 1 1
6 0 0 1 1
0 0 0 0 0 1 6 5 4 1 3 0 2 1 2 2 0
2 2

TABLE VIII TABLE X

Query Space MATRIX , S, FOR DIFFERENT PERMUTATIONS . R ELEVANCE MATRIX FOR DIFFERENT PERMUTATIONS FOR YAHOO .
P p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
P p1 p2 p3 p4 p5 p6 p7 p8 p9 p10
1 10 4 1 0 0 1 0 0 0 0
1 1 1 1 1 1 1 1 1 1 1
2 4 6 1 1 0 1 1 1 0 0
2 1 1 1 1 1 1 1 1 1 1
3 10 3 1 2 0 0 0 0 0 0
3 1 1 1 1 1 1 1 1 1 1
4 7 4 0 2 1 0 0 1 0 1
4 1 1 1 1 1 1 1 1 1 1
5 8 5 1 1 0 0 0 0 0 1
5 1 1 1 1 1 1 1 1 1 1
6 10 5 0 0 0 1 0 0 0 0
6 1 1 1 1 1 1 1 1 1 1

reason why query space in case of Google is larger than that shown in Table IV. The words contained in these queries are
of Yahoo. random and are not related semantically.
Let us answer the question why query space changes in We wish to evaluate the performance of a search engine
case of Yahoo and why it remains constant in case of Google. for k-correlated queries. For that we evaluate search time and
Note that ontology may change with time and with order of query space of a search engine for the first page of results.
words in the given query. For every page of results, Yahoo Since both Google and Yahoo deliver 10 results per page,
estimates the ontology of the given permutation of the query therefore, looking for the first page of results means that we
before delivering the results to the end user. Therefore, the are evaluating 10 top most results of these search engines. Note
query space for different permutations of the given query is that we do not consider number of relevant results because
different and it changes with pages of the searched results10 . relevancy in this case would be query dependent. Since there
However, page ranks do not change either with pages or with is no single query, therefore, evaluation of relevancy would
order of words. The page ranks will only change when new not be so useful.
links or documents are added to the Web that are relevant to Table V shows search time and query space for k-correlated
the given query. Since neither a new link nor a new document queries with embedded semantics (see Table III). The second
is added to the Web during the evaluation of permutations of query, Q2 , is fired after the first query Q1 . On the other hand,
the query, therefore, the query space does not change in case Table VI shows search time and query space for k-correlated
of Google. queries whose words may not be related (see Table IV).
In order to compare the performance of Google and Yahoo,
the latencies versus page numbers for different permutations TABLE XI
of the query have been shown in Figures 2 through 7. Let us R ELEVANCE FOR DIFFERENT PERMUTATIONS .
consider the question why search time in case of Google is P Google Yahoo
larger than that of Yahoo. Note that Google ranks the results 1 19 16
before delivering them to end users while Yahoo does not. The 2 11 15
3 18 16
ranking of pages takes time. This is the reason why search time 4 10 16
taken by Google is larger than that of Yahoo. 5 12 16
In what follows, we discuss how a search engine behaves 6 20 16
Total 90 95
for correlated queries.

TABLE XII
B. Query Correlation Earned Points (EP) FOR DIFFERENT PERMUTATIONS .
We have formulated k-correlated queries as shown in Ta- P Google Yahoo
ble III. Since all words contained in a query are related11, Latency Query EP Latency Query EP
Space Space
therefore, we call them queries with embedded semantics. On
1 14 5 19 33 5 3 0 3
the other hand, we have another set of k-correlated queries as 2 3 11 14 14 0 14
3 4 18 22 13 0 13
10 This observed behavior may also be due to the use of a randomized
4 1 10 11 9 0 9
algorithm. To understand the behavior of randomized algorithms, readers are 5 3 12 15 10 0 10
referred to any text on randomized algorithms such as [10]. 6 25 20 22 5 16 0 16
11 More precisely, all words in these queries are from ad hoc wireless Total 118 65
networks, an area that authors of this paper like to work.
UbiCC Journal - Volume 3 65
 TABLE XIII
In order to compare the performance of Yahoo and Google,
CEP FOR DIFFERENT PERMUTATIONS FOR G OOGLE .
the latencies versus correlation for queries with embedded
P Latency Query Space
semantics is shown in Figure 8 and that for randomized queries Contribution Contribution
is shown in Figure 9. Similarly, the query space for queries 1 123 834 5700000
with embedded semantics is shown in Figure 10 and that for 2 61 262 3300000
randomized queries is shown in Figure 11. 3 116 534 5400000
4 35 918 3000000
The query space of Yahoo is much less than that of Google 5 73 458 3600000
for the reasons discussed in the previous subsection. Other 6 530 376 6000000
Total 530 376 27000000
important observations are as follows.
In case of k-correlated queries with embedded semantics,
TABLE XIV
generally the time to search for Q2 is less than that of
CEP FOR DIFFERENT PERMUTATIONS FOR YAHOO .
Q1 .
P Latency Query Space
This is due to the fact that since the queries are correlated, Contribution Contribution
some of the words of Q2 have already been searched 1 101 385 419900
while searching for Q1 . 2 107 821 404100
The query space is increased when the given query has 3 131 558 431000
4 308 197 428700
a word that is more frequently found in Web pages (e.g. 5 130 833 425000
in R1: Q2 , the word quadratic that is frequently used 6 121 591 431500
in Engineering, Science, Maths, Arts, etc.). The query Total 901 385 2540200
space is decreased when there is a word included in
the query which is rarely used (e.g. mitigate included
in R3,R4:Q1  Q2 and shallow included in R3,R4:Q2). as follows.
The search time is larger in case of randomized queries
as compared to queries with embedded semantics.
 1
1
if latency1i j  latency0i j
The reason for the this observation is as follows. In case li j 2 if latency1i j latency0i j (4)
of queries with embedded semantics, the words of a given 0 otherwise.


query are related and are found in Web pages that are not Similarly, let S si j  be a matrix where si j is defined as
too far from one another either from the point of view follows.
of page rank as in Google or from the point of view of
ontology as in Yahoo.
 1
1
if space1i j  space0i j
si j if space1i j space0i j (5)
One cannot infer anything about the search time of 2
Google and Yahoo as it depends upon the query. More 0 otherwise.
precisely, it depends upon the fact which strategy takes In matrices defined above, where there is a ’1’, it means at
more time whether page rank in Google or estimation of that place Google is the winner and a ’ 12 ’ represents that there
ontology in Yahoo. has been a tie between Google and Yahoo. We now define a
However, from Table V and Table VI, one can infer the parameter that we call Earned Points (EP) which is as follows.
following. Google is better in the sense that its query space
 ∑  pages
  L


is much larger than that of Yahoo. However, Yahoo takes EPk relevantki k
i  Sik (6)
less time as compared to Google for different permutations i 1


of the same query. For k-correlated queries with embedded
semantics, Google takes less time to search for the first query
as compared to Yahoo. It also applies to randomized queries
with some exceptions. In exceptional cases, Google takes
much more time as compared to Yahoo. We have mentioned
it previously that it depends upon the given query as well as
the strategy employed in the search engine.
In what follows, we describe a unified criteria for comparing
the search engines considered in this paper.
V. A U NIFIED C RITERIA FOR C OMPARISON
Let us denote Google by
a superscript ’1’ and Yahoo by a
superscript ’0’12 . Let L li j  be a matrix where li j is defined
12 This is simply a representation. One may consider a representation which
is reverse of it, then also, there will not be any effect on the criteria.
UbiCC Journal - Volume 3
where, superscript k  0  1  denotes the search engine.
Table VII shows a latency matrix, L, for different permu-
tations of the query as that for Table I and Table II, and has
been constructed using both of them. In the latency matrix,
there are 40 ’0’s, 17 ’1’s, and 3 ’ 12 ’. We observe from the
latency matrix that Yahoo is the winner (as far as latencies
are concerned), as there are 40 ’0’s out of 60 entries in total.
On the other hand, Table VIII shows the query space
matrix, S, for different permutations of the same query and
is constructed using the tables mentioned in the preceding
paragraph. One can see that as far as query space is concerned,
Google is the sole winner. Infact, query space of Google is
much larger than that of Yahoo.
The relevance matrix for Google is shown in Table IX and
that for Yahoo is shown in Table X. The total relevance for the
first ten pages is shown in Table XI for both Google as well
as Yahoo. It is seen from Table XI that the total relevance for
Google is 90 and that for Yahoo is 95. Average relevance per
66
 TABLE XV
Table XIII shows contributions of latency and query space
C ONTRIBUTION DUE TO QUERY SPACE IN CEP FOR DIFFERENT SETS OF
in CEP for Google. Similarly, Table XIV shows the same for
WEIGHTS .
Yahoo. We observe that contribution of latency for Google is
Weights
wl 1, wq 10


6
Google
27 00
Yahoo
2 54
530  376 and that for Yahoo is 901  385. However, contribution
wl
wl
1,
1,
wq
wq
10
10


5
4
270 00
2700 00
25 40
254 02
of query space for Google is 27000000 and that for Yahoo is
2540200. In other words, the contribution of query space for
wl
wl
1,
1,
wq
wq
10
10


3
2
27000 00
270000 00
2540 20
25402 00
Google is approximately 11 times of that for Yahoo. Adding
these contributions shall result in a larger CEP for Google as
wl 1, wq 10 1 2700000 00 254020 00
wl 1, wq 1 27000000 2540200 compared to Yahoo. The CEP defined using (7) has a problem
that we call dominating constituent problem (DCP). The larger
parameter suppresses the smaller parameter. Note that the
TABLE XVI
definition of CEP in (7) assumes equal weights for latency
CCEP FOR DIFFERENT SETS OF comparable weights.
and query space. On the other hand, one may be interested in
Weights Google Yahoo assigning different weights to constituents of CEP depending
wl 0 9, wq 01 486 3384 811 2465 upon the importance of constituents. Let us rewrite (7) to
wl 0 8, wq 02 442 3008 721 1080
wl 0 7, wq 03 398 2632 630 9695 incorporate weights. Let wl and wq be the weights assigned to
wl 0 6, wq 04 354 2256 540 8310 latency and query space, respectively. The (7) can be written
wl 0 5, wq 05 310 1880 450 6925

as follows.
wl 0 4, wq 06 266 1504 360 5540
wl 0 3,
wl 0 2,
wq
wq
07
08
222 1128
178 0752
270 4155
180 2770 CEP k ∑
pages
relevantki   1
wl  qki wq  (8)
wl 0 1, wq 09 134 0376
permutation and per page for Google is 1  5 and that for Yahoo
is 1  583. Therefore, as far as average relevance is concerned,
Yahoo is the winner.
Table XII shows the number of earned points for both
Google as well as Yahoo for different permutations of the
query mentioned earlier. We observe that the number of earned
points for Google is 118 and that for Yahoo is 65. The number
of earned points of Google is far greater than Yahoo. The
reason behind this is that query space of Yahoo is always less
than that of Google and it does not contribute to the number
of earned points.
A closer look on the definition of EP reveals that while
∑
defining the parameter EP in (6) together with (4) and (5),
we have assumed that a search engine either has a constituent
parameter (latency or query space) or it does not have that
90 1385 i 1 dik
The weights should be chosen carefully. For example, the
weights wl 1, wq 10 6 will add 27 to the contribution in
CEP due to query space for Google and 2  54 to Yahoo. On the
other hand, a set of weights wl 1, wq 10 5 shall add 270
for Google and 25  4 for Yahoo. Table XV shows contribution
of query space in CEP for different sets of weights. It is to note
that wl is fixed to 1 for all sets, and only wq is varied. As wq is
increased beyond 10 5, the contribution of query space starts
dominating over the contribution of latency. The set of weight
wl 1  wq 10 5 indicates that one can ignore contribution
of query space in comparison to the contribution of latencies
provided that one is more interested in comparing search
engines with respect to latency. In that case, an approximate
expression for CEP can be written as follows.


pages
CEP k
i 1
relevantki
1
dik  (9)

parameter at all. The contribution of some of the parameter Alternatively, one may consider an approach that is combi-
is lost due the fact that the effective contribution of other nation of the definition of EP defined in (6) (together with (4)
parameter by which the given parameter is multiplied is zero. and (5) and that of CEP defined in (7). In that we may use
Note that our goal behind introduction of (6) was to rank the definition of matrix S which converts the contribution of
the given set of search engines. We call this type of ranking query space in the form of binaries13 . The modified definition
of search engines as Lossy Constituent Ranking (LCR). We,

is as follows.
therefore, feel that there should be a method of comparison
between a given set of search engines that is lossless in nature.
CCEPk ∑pages
relevantki  1
 Sik  (10)
For that purpose, we define another parameter that we call i 1 dik
Contributed Earned Points (CEP). The definition of CEP is as
follows. where Sik is in accordance with the definition of S given by
 (5). The acronym CCEP stands for Combined Contributory

CEP k  ∑ pages
relevantki  1
qki  (7)
Earned Points. If one wishes to incorporate weights, then the


 definition of CCEP becomes as follows.
dik
 ∑ 
i 1


 
pages
where, superscript k  0  1  denotes the search engine, d
denotes the actual latency, and q denotes the actual query
CCEPk
i 1
relevantki
1
dik
wl  Sik wq  (11)

space. The reason behind having an inverse of actual latency 13 We mean that the matrix S says either there is a contribution of query
in (7) is that the better search engine would be that which space of a search engine provided that its query space is larger than that of
takes less time. the other one or there is no contribution of query space at all, if otherwise.
UbiCC Journal - Volume 3 67
In the definition of CCEP given by (11) the weights can be weights to different constituents of the criteria—latency and
comparable and the dominant constituent problem mentioned query space. Our observations are as follows.
earlier can be mitigated for comparable weights. We define We observed that performance of Yahoo is better in terms
comparable weights as follows.


of the latencies, however, Google performs better in terms
Definition 3: A set of weights W wi    wi 0, is said of query space.
to have comparable weights if and only if ∑i wi 1 and the We discussed the dominant constituent problem. We
condition 19  wwij  9 is satisfied  wi  w j  W .
Table XVI shows the values of CCEP for different sets of
comparable weights. We observe that the rate of decrease of
CCEP for Yahoo is larger than that of Google. For example, for
wl 0  9  wq 0  1, CCEP for Google is 486  3384 and that for
Yahoo is 811  2465. For wl 0  8  wq 0  2, CCEP for Google
is 442.3008 and that for Yahoo is 721  1080. In other words,
the rate of decrease in CCEP for Google is 9  05% and that for
Yahoo is 11  11%. The reason being that in the query space
matrix, S, (see Table VIII) all entries are ’1’. It means that
query space of Google is always larger than that of Yahoo.
Therefore, in case of Yahoo, the contribution due to query
space is always 0 irrespective of the weight assigned to it.
However, in case of Google the contribution due to query space
is nonzero and increases with an increase in weight assigned

due to query space. Moreover, for a set of
to the contribution
weights, W wl 0  5  wq 0  5  , the values of CCEP are
310  1880 and 450  6925 for Google and Yahoo, respectively.
It means that if one wishes to assign equal weights to latency
and query space then Yahoo is the winner in terms of the
parameter CCEP.
In case of CCEP, the effect of the dominating constituent
problem is less as compared to that in case of CEP. In other
words, the effect of large values of query space is fairly smaller
in case of CCEP as compared to that in case of CEP. This is
with reference to our remark that with the use of CCEP the
dominating constituent problem is mitigated.
VI. C ONCLUSIONS
In this paper, we analyzed the impact of correlation among
queries on search results for two representative search portals
namely Google and Yahoo. The major accomplishments of the
paper are as follows:
We analyzed the search time, the query space and the
number of relevant results per page for different per-
mutations of the same query. We observed that these
parameters vary with pages of searched results and are
different for different permutations of the given query.
We analyzed the impact of k-correlation among two
subsequent queries given to a search engine. In that
we analyzed the search time and the query space. We
observed that
– The search time is less in case of queries with
embedded semantics as compared to randomized
queries without any semantic consideration.
– In case of randomized query, the query space is
increased in case the given query includes a word
that is frequently found on the Web and vice versa.
Further, we considered a unified criteria for comparison be-
tween the search engines. Our criteria is based upon the
concept of earned points. An end-user may assign different
UbiCC Journal - Volume 3
discussed that this problem can be mitigated using the
concept of contributory earned points if weights assigned
to constituents are comparable. If both the constituent
are assigned equal weights, we found that Yahoo is the
winner.
However, the performance of a search engine may depend
upon the criteria itself and only one criteria may not be
sufficient for an exact analysis of the performance. Further
investigations and improvements in this direction forms our
future work.
R EFERENCES
[1] S. Malhotra, ”Beyond Google”, CyberMedia Magazine on Data Quest,
vol. 23, no. 24, pp.12, December 2005.
[2] M.R. Henzinger, A. Haydon, M. Mitzenmacher, M. Nozark, ”Measuring
Index Quality Using Random Walks on the Web”, Proceedings of 8th
International World Wide Web Conference, pp. 213-225, May 1999.
[3] M.C. Tang, Y. Sun, ”Evaluation of Web-Based Search En-
gines Using User-Effort Measures”, Library an Information Sci-
ence Research Science Electronic Journal, vol. 13, issue 2, 2003,
http://libres.curtin.edu.au/libres13n2 /tang.htm.
[4] C.W. Cleverdon, J. Mills, E.M. Keen, An Inquiry in Testing of Infor-
mation Retrieval Systems, Granfield, U.K., 1966.
[5] J. Gwidzka, M. Chignell, ”Towards Information Retrieval Mea-
sures for Evaluation of Web Search Engines”, http://www.imedia.
mie.utoronto.ca/people/jacek/pubs/webIR eval1 99.pdf, 1999.
[6] D. Rafiei, A.O. Mendelzon, ”What is This Page Known For: Computing
Web Page Reputations”, Elsevier Journal on Computer Networks, vol
33, pp. 823-835, 2000.
[7] N. Bhatti, A. Bouch, A. Kuchinsky, ”Integrating User-Perceived Quality
into Web Server Design”, Elsevier Journal on Computer Networks, vol
33, pp. 1-16, 2000.
[8] S. Brin, L. Page, ”The Anatomy of a Large-Scale Hypertextual Web
Search Engine”, http://www-db.stanford.edu/pub/papers/google.pdf,
2000.
[9] J. Kleinberg, ”Authoritative Sources in a Hyperlinked Environment”,
Proceedings of 9th ACM/SIAM Symposium on Discrete Algorithms,
1998.
[10] R. Motwani, P. Raghavan, Randomized Algorithms, Cambridge Univer-
sity Press, August 1995.
68
 ARTIFICIAL IMMUNE SYSTEMS FOR ILLNESSES DIAGNOSTIC

Hiba Khelil, Abdelkader Benyettou

SIMPA Laboratory – University of Sciences and Technology of Oran,
PB 1505 M’naouer, 31000 Oran, Algeria
hibakhelil@yahoo.fr, aek.benyettou@univ-usto.dz

ABSTRACT
Lately, a lot of new illnesses are frequently observed in our societies, that it can be
avoid by daily visits to the doctor. Cancer is one of these illnesses where patients
discover it only when it is too late. In this work we propose an artificial Cancer
diagnostic which can classify patients if they are affected by Cancer or no, for this
goal we have developed the artificial immune system for Cancer diagnostic. The
artificial immune system is one of the newest approaches used in several domains
as pattern recognition, robotic, intrusion detection, illnesses diagnostic…a lot of
methods are exposed as negative selection, clone selection and artificial immune
network (AINet). In this paper we’ll present the natural immune system, we’ll
develop also four versions of Artificial Immune Recognition System (AIRS) and
after we’ll present results for Cancer diagnostic with some critics and remarks of
these methods.

Keywords: Antigen, Antibody, B memory cells, Artificial Recognition Ball

(ARB), Artificial Immune Recognition System (AIRS), Cancer diagnostic.

1 INTRODUCTION of the natural immune system and immune response

Pattern recognition is very vast domain in the
artificial intelligence, where we can find faces, prints,
speech and hand writing recognition and others
patterns that are not less important as ones
mentioned; for this goal several approaches are
developed as neuronal networks, evolutionary
algorithms, genetic algorithms and others under
exploitation. The artificial immune system is a new
approach used in different domains as pattern
recognition [1] [2] [3] [4] [5], intrusions detection in
Internet networks [6], robotics [7], machine learning
[8] and other various applications in different
domains. The artificial immune functions are
inspired from natural immune system where the
responsible cells of the immune response are
simulated to give an artificial approach adapted
according the application domain and the main
problem.
The present work is an application of the
artificial immune recognition system (AIRS) for
Cancer diagnostic. AIRS is a method inspired from
the biologic immune system for pattern recognition
(classification) proposed by A. Watkins in 2001 [9]
in his Master thesis to Mississipi university, the
improvement was been in 2004 by A. Watkins, J.,
Timmis and L. Boggess [10] where authors optimize
the number of B cells generated. This method is
characterized by the distributed training proven by A.
Watkins in his PHD to the university of Kent in 2005
[11]. In this paper we will begin by a short definition
types. The second part is a representation of an
artificial simulation of the immune systems giving a
description of training algorithms. As prototype of
training we’ll present a preview of Cancer data bases
and results of application of artificial immune system
for Cancer diagnostic. Finally some critics are given
to show limits and differences between methods and
some perspectives also.
2 NATURAL IMMUNE SYSTEM
The biologic immune system constitutes a
weapon against intruders in a given body, for this
goal several cells contribute to eliminate this intruder
named antigen, these cells participate for a 'biologic
immune response'.
We distinguish two types of natural immune
response one is innate and other is acquired
explained in the following points:
2.1 Innate Immunity
It is an elementary immune that very reduce
number of antigens are used, we can find this type of
immunity in newborn not yet vaccinated. A non
adaptive immune for long time can drive infections
and death because the body is not very armed against
antigens of the environment [12].
2.2 Innate Acquired
It is an immune endowed with a memory, named
also secondary response, triggered after the

UbiCC Journal - Volume 3 69

apparition of the same antigen in the same immune
system for the second time or more, where it
generates the development of B cells memory for
this type of antigen already met (memorized) in the
system. This answer is faster than innate one [12]
and caused the increase of the temperature of the
body, which can be explained by fighting of B cells
against antigens.
The primary immune response is only slower but
it keeps information about passage of antigens in the
system; this memorization phenomenon will interest
us to use it for the artificial pattern recognition;
according this principle the artificial immune
recognition system is developed and it is the main
subject of this paper.
3 THE ARTIFICIAL IMMUNE SYSTEM
The natural immune system is very complicated
to be artificially simulated; but A. B. Watkins
succeeded to simulate the most important functions
of the natural immune system for pattern recognition.
The main factors entering in the artificial immune
system are antigens, antibodies, B memory cells.
We’ll present in the next session training algorithms
that puts in work the noted factors (antibodies, B
memory cells and antigens).
4 THE AIRS ALGORITHM
The present algorithm is inspired from A. B.
Watkins thesis [9] [13] [14] which present the
artificial immune recognition system intended for
pattern recognition named AIRS. First, antigens
represent the training data used in the training
program in order to generate antibodies (B cells) to
be used in the test step (classification). We can note
that there are four training steps in the artificial
immune training algorithm as following:
4.1 Initialization Step
In this step, all characteristic vectors of antigens
are normalized, and affinity threshold is calculated
by (1)
n−1 n
∑ ∑ affinity(agi , ag j )
i=1 j =i+1
affinity _ threshold = (1)
n(n − 1)
2 After the end of training phase, test is executed
Noting here that affinity is the Euclidian distance
between two antigens, and n is the number of
antigens (cardinality of training data)
To begin, we must initialize the B cells memory
set (MC) and ARBs population by choosing arbitrary
examples from training data.
4.2 B Cells Identification and ARBs Generation
Ones initialization was finished; this step is
executed for each antigen from training data. First,
UbiCC Journal - Volume 3
was chosen from MC which has the least value of
affinity (maximize the stimulation) with this antigen,
noting that:
stimulation ( ag , mc ) = 1- affinity ( ag , mc ) (2)
This cell will be conserved for a long time by
cloning and generate new ARBs, these ARBs will be
add to the old ARBs set; the clone number is
calculated by formula (3)
clone _ number = hyper _ clonal _ rate*clonal _ rate*stimulation(mcmatch, ag)
(3)
Every clone is mutating according a small
algorithm described in [9], which consist to alter the
characteristic cells vectors.
4.3 Competition for Resources and Development
of a Candidate Memory Cell
In this step, ARBs’s information was completed
by calculating resources allocations in function of its
stimulation as following:
resource = stimulation ( ag , ARB ( antibody )) * clonal _ rate
(4)
and calculate the average stimulation for each ARB
also. This step cans death some ARBs which are low
stimulated.
After we clone and mutate the subset of ARBs
according their stimulation level.
While the average stimulation value of each
ARB class (si) is less then a given stimulation
thresholds then we repeat the third step.
| A Bi |
∑ a b j .s tim
j =1 (5)
si = , abj ∈ A Bi
| ABi |
s i ≥ a ffin ity _ th r e s h o ld (6)
4.4 Memory Cell Introduction
Select ARBs of the same class as the antigen
with highest affinity. If the affinity of this ARB with
m c m a tch
the antigenic pattern is better than that of
mc
then add the candidate cell ca n d id a te to the
memory cells set. Additionally, if the affinity of
m c m a tch m c ca n d id a te
and is below the affinity
threshold, then remove
m c m a tch
from memory set.
So we repeat the second step until all antigens
was treated.
using the memory cells generated from training step,
in order to classify the new antigenic patterns. The
critter of classification is to attribute the new antigen
to the most appropriate class using KMeans or KNN
(K Nearest Neighbor); in this paper we’ll present
classification results using KMeans algorithm.
70
5 THE AIRS2 ALGORITHM
The changes made to the AIRS algorithm are
small, but it offers simplicity of implementation, data
reduction and minimizes the processing time.
The AIRS2 training steps are the same as AIRS
one, just some changes which are presented as
following:
1- It’s not necessary to initialize the ARB set.
2- It’s not necessary to mutate the ARBs class
feature, because in AIRS2 we are interesting only
about cells of the same class of antigen.
3- Resources are only allocated to ARBs of the same
class as antigen and are allocated in proportion to the
ARB’s stimulation level in reaction to the antigen.
4- The training stopping criterion no longer takes
into account the stimulation value of ARBs in all
classes, but only accounts for the stimulation value
of the ARBs of the same class as the antigen.
6 AIRS AND AIRS2 ALGORITHMS USING
MERGING FACTOR
In this session we’ll present other modification
of the AIRS and AIRS2. This modification carries on
the last training step (Memory cell introduction),
mainly in the cell introduction criterion; the
condition was as following:
C a n d S tim ← S tim u la tio n ( a g , m c ca n d id a te )
M a tch S tim ← S tim u la tio n ( a g , m c m a tch )
C ellA ff ← a ffin ity ( m c ca n d id a te , m c m a tch )
if ( C a n d S tim > M a tch S tim )
if ( C ellA ff < A T * ATS )
M C ← M C − m c m a tch
M C ← M C ∪ m c ca n d id a te
This source explains conditions
m c ca n d id a te mc
and delete m a tch to the memory
cells set; the modification is carried in the following
condition:
if (CellAff < AT * ATS + factor ) (8)
Noting that factor is calculated by:
factor = AT * ATS * dampener
(9)
With ATS and dampener are two parameters
between 0 and 1, and np is the number of training
programs executed in parallel (number of classes).
This change to the merging scheme relaxes the
criterion for memory cell removal in the affinity
based merging scheme by a small fraction in
logarithmic.
This modification is used in the two algorithms
(AIRS and AIRS2), and all algorithms are applied
for Cancer diagnostics and all results we’ll be
presented in the next sessions.
UbiCC Journal - Volume 3
7 RESULTS
To determine the relative performance of AIRSs
algorithms, it was necessary to test it on data base; so
we have chosen three Cancer data bases from
hospitable academic center of Wisconsin: Brest
Cancer Wisconsin (BCW), Wisconsin Prognostic
Breast Cancer (WPBC) and Wisconsin Diagnostic
Breast Cancer (WDBC). The description of this data
bases are given as following:
- Brest Cancer Wisconsin (BCW): This data base
was obtained from hospitable academic center of
Wisconsin in 1991, which describe the cancerous
symptoms and classify them into two classes:
‘Malignant’ or ‘Benin’. The distribution of patients
is given as following: (Malignant, 214) (Benin, 458).
- Wisconsin Prognostic Breast Cancer (WPBC): This
data base is conceived by the same hospitable
academic in 1995 but it gives more details then BCW
giving nucleus of cell observations. Basing of its
characteristics, patients are classify into two classes:
‘Recur’ and ‘NonRecur’, where its distribution as
following: (Recur, 47) (NonRecur, 151).
- Wisconsin Diagnostic Breast Cancer (WDBC):
This data base is conceived also by the same
hospitable academic in 1995, it has the same attribute
then WPBC, but it classify its patients into two
classes: ‘Malignant’ and ‘Benin’, where its
distribution as following: (Malignant, 212) (Benin,
357).
(7) All training data are antigens, represented by
characteristic vectors; also for antibodies have the
same characteristic vector size as antigens. The ARB
is represented as structure having antibody
characteristic vector, his stimulation with antigen and
the resources that allowed.
7.1 Software and Hardware Resources
to add In order to apply algorithms we have used the
C++ language in Linux Mandriva 2006 environment,
every machine is endowed of 512Mo memory space
and 3.0 Ghz processor frequency.
All training programs have the same number of
antigens, the same number of initial memory cells
and ARBs also. In the same way the training
* log( np)
program is an iterative process; that we have fix 50
iterations for each training program of every class.
7.2 Results and Classification Accuracy
To run programs, we must fix the most
important training parameters as following:
hyper _ clonal _ rate clonal _ rate
, and
mutation _ rate . These parameters are used in
training steps, as criterion to limit the clone number,
to calculate the ARB’s resources and in the mutation
procedure also. The parameters values are given in
table 1:
71
 Table 1: Training parameters.
After 50 iterations, the B memory cells
generated from each training program of classes are
Parameters Type Values used in the classification step (test). In the
classification we take the shortest distance between
the new antigen and the gravity centers of all
memory cells sets, and we affect this antigen in the
Hyper_clonal_rate Integer value 30 same class of the nearest center (KMeans). Using
this principle the classification accuracies are given
Clonal_rate Integer value 20 in table 2:

Mutation_rate Real value 0.1

[0,1]

Table 2: Classification accuracies

Table 3: Average classification accuracies

UbiCC Journal - Volume 3 72

 We can observe that AIRS gives in general the
best results, which can give more B cells then to
increase the recognition chance. The AIRS2 and
AIRS2 using factor are amelioration of AIRS, but B
cells generated are less than the original algorithm
(AIRS), that’s why these methods (AIRS2, AIRS2
with factor) don’t give better results in Cancer
diagnostic in exception of some cases; the evolution
of B cells is given in the next session.
The execution of AIRS2 and AIRS2 using factor
are faster than AIRS and AIRS using factor, this can
be explained by using just B cells of the same class
of antigen, and this can reduce treatments then time
processing also.
7.3 B Cells Evolution
Noting that we have given the same chance to
each training programs (quantity of antigens
introduced, initial memory cells and initial ARBs),
the B cells generated are not necessary the same in
each method; the previews table give us the B cells
generated in each one, we can observe that the cells
generated in AIRS2 and AIRS2 using factor are less
than AIRS and AIRS using factor, as mentioned
before, although we initialized all B cells sets to the
same size. The next figures represent evolution of B
cells in function of iterations for each data base of
the best rate from four methods:
Figure 1: Evolution of B cells in BCW (AIRS)
Figure 2: Evolution of B cells in WPBC (AIRS2
using factor)
UbiCC Journal - Volume 3
Figure 3: Evolution of B cells in WDBC (AIRS2)
From figures we observe that evolution of B
cells in AIRS are faster than AIRS2 and AIRS2
using factor because there are more cells deleted on
changing the condition given in equation (8) (using
factor).
8 DISCUSSION OF RESULTS
The results of experiences can be found in table
2 and 3; comparing the used methods we can observe
that AIRS in general gives best results and we can
observe also that this method generates B cells more
than others.
In all experiences we have used the Euclidian
distance, it is possible to use hamming distance or
other.
The AIRS and AIRS using factor converge
slowly to the most B cells adapted for Cancer
diagnostic, on the contrary of AIRS2 and AIRS2
using factor, which are executed quickly than others
but they don’t give us the best memory cells.
9 CONCLUSIONS
In this paper we have presented Cancer
diagnostic results for AIRS immuno-computing
algorithm and provided directions for interpretation
of these results. We are interested in immuno-
computing because is one of the newest directions in
bio-inspired machine learning and focused on AIRS,
AIRS2, AIRS using factor and AIRS2 using factor
(2001-2005) and it can be also used for classification
(illnesses diagnostic).
We suggest that AIRS is a mature classifier that
delivers reasonable results and that is can safety be
used for real world classifications tasks. The
presented results are good but it must be improved
using optimization algorithms. In our future work we
want to make more importance to the parameters
values and propose a new method to search the best
values of these ones in order to across the
performance of these algorithms.
73
10 REFERENCES
[1] Secker A., Freitas A., Timmis J.: AISEC: An artificial
immune system for e-mail classification. In proceedings
of the Congress on Evolutionary Computation, pp. 131-
-139, Canberra. Australia (2003)
[2] Lingjun M., Peter V. D. P., Haiyang W., A:
Comprehensive benchmark of the artificial immune
recognition system (AIRS). In proceeding of advanced
data mining and applications ADMA,
vol. 3584, pp. 575--582, China (2005)
[3] Deneche A., Meshoul S., Batouche M. : Une approche
hybride pour la reconnaissance des formes en utilisant
un systeme immunitaire artificiel. In proceedings of
graphic computer science, Biskra, Algeria (2005)
[4] Deneche A.: Approches bios inspirees pour la
reconnaissance de formes, Master thesis in Mentouri
University, Constantine, Algeria (2006)
[5] Goodman D., Boggess L., Watkins A.: Artificial
immune system classification of multiple class
problems, Intelligent Engineering Systems Through
Artificial Neural press (2002)
[6] Kim J., Bently P.: Towards an artificial immune system
for network intrusion detection: an investigation of
clonal selection with a negative selection operator. In
Proceeding of Congress on Evolutionary Computation,
vol. 2, pp. 1244--1252, South Korea (2001)
[7] Jun J. H., Lee D. W., Sim K. B.: Realization of
cooperative and swarm behavior in distributed
autonomous robotic systems using artificial immune
system. In proceeding IEEE international conference of
UbiCC Journal - Volume 3
Man and Cybernetics, vol. 6, pp. 614--619. IEEE Press,
New York (1999)
[8] Timmis J.: Artificial immune systems: a novel data
analysis technique inspired by the immune network
theory, PhD thesis Wales UK University (2000)
[9] Watkins A.: AIRS: A resource limited artificial
immune classifier, Master thesis, Mississippi University
(2001)
[10] Watkins A., Timmis J., Boggess L.: Artificial immune
recognition system (airs): an immune inspired
supervised learning algorithm, vol. 5, pp. 291--317,
Genetic Programming and Evolvable Machines press
(2004)
[11] Watkins A.: Exploiting immunological metaphors in
the development of serial, parallel, and distributed
learning algorithms, PhD thesis, Kent University (2005)
[12] Emilie P.: Organisation du system immunitaire felin,
PhD thesis, National school Lyon, France (2006)
[13] Watkins A., Timmis J.: Artificial immune recognition
system (airs): revisions and refinements. In proceedings
of first international conference on artificial immune
system ICARIS, pp. 173--181, Kent University (2005)
[14] Watkins A., Boggess L.: A new classifier based
on resources limited artificial immune systems. In
proceedings of congress de Evolutionary
Computation, IEEE World Congress on
Computational Intelligence held in Honolulu, HI,
USA, pp. 1546--1551, Kent University (2005)
74
 HOT METHOD PREDICTION USING SUPPORT VECTOR MACHINES

Sandra Johnson ,Dr S Valli

Department of Computer Science and Engineering, Anna University, Chennai – 600 025, India.
sandra_johnk@yahoo.com , valli@annauniv.edu

ABSTRACT
Runtime hot method detection being an important dynamic compiler optimization
parameter, has challenged researchers to explore and refine techniques to address
the problem of expensive profiling overhead incurred during the process. Although
the recent trend has been toward the application of machine learning heuristics in
compiler optimization, its role in identification and prediction of hot methods has
been ignored. The aim of this work is to develop a model using the machine
learning algorithm, the Support Vector Machine (SVM) to identify and predict hot
methods in a given program, to which the best set of optimizations could be
applied. When trained with ten static program features, the derived model predicts
hot methods with an appreciable 62.57% accuracy.

Keywords: Machine Learning, Support Vector Machines, Hot Methods, Virtual

Machines.

1 INTRODUCTION results of the evaluation. Section 7 proposes future

Optimizers depend on profile information to
identify hot methods of program segments. The
major inadequacy associated with the dynamic
optimization technique is the high cost of accurate
data profiling via program instrumentation. The
major challenge is how to minimize the overhead
that includes profile collection, optimization strategy
selection and re-optimization.
While there is a significant amount of work
relating to cost effective and performance efficient
machine learning (ML) techniques to tune individual
optimization heuristics, relatively little work has
been done on the identification and prediction of
frequently executed program hot spots using
machine learning algorithms so as to target the best
set of optimizations. In this study it is proposed to
develop a machine learning based predictive model
using the Support Vector Machine (SVM) classifier.
Ten features have been derived from the chosen
domain knowledge, for training and testing the
classifiers. The training data set are collected from
the SPEC CPU2000 INT and UTDSP benchmark
programs. The SVM classifier is trained offline with
the training data set and it is used in predicting the
hot methods of a program which are not trained. This
system is evaluated for the program’s hot method
prediction accuracy.
This paper is structured as follows. Section 2
discusses related work. Section 3 gives a brief
overview of Support Vector Machines. In Section 4
this approach and in section 5 the evaluation
methodology is described. Section 6 presents the
work and concludes the paper.
2 RELATED WORK
Machine learning techniques are currently used
to automate the construction of well-designed
individual optimization heuristics. In addition, the
search is on for automatic detection of a program
segment for targeted optimization. While no previous
work to the best of our knowledge has used ML for
predicting program hot spots, this section reviews the
research papers which use ML for compiler
optimization heuristics.
In a recent review of research on the challenges
confronting dynamic compiler optimizers, Arnold et
al. [1] give a detailed review of adaptive
optimizations used in the virtual machine
environment. They conclude that feedback-directed
optimization techniques are not well used in
production systems.
Shun Long et al. [3] have used the Instance-
based learning algorithm to identify the best
transformations for each program. For each
optimized program, a database stores the
transformations selected, the program features and
the resulting speedup. The aim is to apply
appropriate transformations when a new program is
encountered.
Cavazos et al. [4] have applied an offline ML
technique to decide whether to inline a method or not.
The adaptive system uses online profile data to
identify “hot methods” and method calls in the hot
methods are in-lined using the ML heuristics.

UbiCC Journal - VolumeUbiquitous

3 Computing and Communication Journal 1 75
 Cavazos et al. [5, 12] have also used supervised
learning to decide on which optimization algorithm
to use: either graph coloring or Linear scan for
register allocation. They have used three categories
of method level features for ML heuristics (i.e.)
features of edges of a control flow graph, features
related to live intervals and finally, statistical
features about the size of a method.
Cavazos et al. [11] report that the best of
compiler optimizations is method dependent rather
than program dependent. Their paper describes how,
logistic regression-based machine learning technique
trained using only static features of a method, is used
to automatically derive a simple predictive model
that selects the best set of optimizations for
individual methods within a dynamic compiler. They
take into consideration the structures of a particular
method within a program to develop a sequence of
optimization phases. The automatically constructed
regression model is shown to out-perform hand-
tuned models.
To identify basic blocks for instruction
scheduling Cavazos et al. [20] have used supervised
learning. Monsifrot et al. [2] have used a decision
tree learning algorithm to identify loops for unrolling.
Most of the work [4, 5, 11, 12, 20] is implemented
and evaluated using Jikes RVM.
The authors [8, 19] have used genetic
programming to choose an effective priority function
which prioritizes the various compiler options
available. They have chosen hyper-block formation,
register allocation and data pre-fetching for
evaluating their optimizations.
Agakov et al. [9] have applied machine learning
to speed up search-based iterative optimization. The
statistical technique of the Principal component
analysis (PCA) is used in their work for appropriate
program feature selection. The program features
collected off-line from a set of training programs are
used for learning by the nearest neighbor algorithm.
Features are then extracted for a new program and
are processed by the PCA before they are classified,
using the nearest neighbor algorithm. This reduces
the search space to a few good transformations for
the new program from the various available source-
level transformations. However, this model can be
applied only to whole programs.
The authors [10] present a machine learning-
based model to predict the performance of a
modified program using static source code features
and features like execution frequencies of basic
blocks which are extracted from the profile data
collected. As proposed in their approach [9], the
authors have used the PCA to reduce the feature set.
A linear regression model and an artificial neural
network model are used for building the prediction
model which is shown to work better than the non-
feature-based predictors.
In their work Fursin et al. [14] have used
UbiCC Journal - VolumeUbiquitous
3 Computing and Communication Journal
machine learning to identify the best procedure clone
for the current run of the program. M. Stephenson et
al. [18] have used two machine learning algorithms,
the nearest neighbor (NN) and Support Vector
Machines (SVMs), to predict the loop unroll factor.
None of these approaches aims at prediction at the
method level. However, machine learning has been
widely used in work on branch prediction [21, 22, 23,
24].
3 SUPPORT VECTOR MACHINES
The SVM [15, 16] classification maps a training
data (xi,yi), i = 1,…,n where each instance is a set of
feature values xi ∈ Rn and a class label y ∈ {+1,-1},
into a higher-dimensional feature space φ(x) and
defines a separating hyperplane. Only two types of
data can be separated by the SVM which is a binary
classifier. Fig. 1 shows a linear SVM hyperplane
separating two classes.
The linear separation in the feature space is done
using the dot product φ(x).φ(y). Positive definite
kernel functions k(x, y) correspond to feature space
dot products and are therefore used in the training
algorithm instead of the dot product as in Eq. (1):
k ( x, y ) = (Φ ( x) • Φ ( y )) (1)
The decision function given by the SVM is given in
Eq. (2):
n
f ( x ) = ∑ vi .k ( x, xi ) + b (2)
i =1
where b is a bias parameter, x is the training example
and vi is the solution to a quadratic optimization
problem. The margin of separation extending from
the hyperplane gives the solution of the quadratic
optimization problem.
Optimal Hyperplane
Margin of Separation
Feature Space
Figure 3: Optimal hyperplane and margin of
separation
4 HOT METHOD PREDICTION
This section briefly describes how machine
2 76
 learning could be used in developing a model to
predict hot methods within a program. A discussion
of the approach is followed by the scheme of the
SVM-based strategy adopted in this study.
Figure 2: System architecture of the SVM-based hot
method predictive model
4.1 The approach
Static features of each method in a program are
collected by offline program analysis. Each of these
method level features forms a feature vector which is
labeled either hot or cold based on classification by a
prior execution of the program. The training data set
thus generated is used to train the SVM-based
predictive model. Next, the test data set is created by
offline program analysis of a newly encountered
program. The trained model is used to predict
whether a method is hot or cold for the new program.
An offline analysis on the Low Level Virtual
UbiCC Journal - VolumeUbiquitous
3 Computing and Communication Journal
Machine’s (LLVM) [6] bytecode representation of
the programs provides the training as well as the test
data set. The system architecture for the SVM-based
hot method predictive model is shown in Fig.2 and it
closely resembles the architecture proposed by the
authors C. L. Huang et. al. [26]. Fig. 3 outlines the
strategies for building a predictive model.
1. Create training data set.
a. Collect method level features
i. Calculate the specified feature for every
method in a LLVM bytecode.
ii. Store the feature set in a vector.
b. Label each method
i. Instrument each method in the program
with a counter variable [25].
ii. Execute the program and collect the
frequency of the execution of each
method.
iii. Using the profile information, each
method is labeled as either hot or cold.
iv. Write the label and its corresponding
feature vector for every method in a file.
c. Steps (a) & (b) are repeated for as many
programs as are required for training.
2. Train the predictive model.
a. The feature data set is used to train the
SVM-based model.
b. The predictive model is generated as output.
3. Create test data set.
a. Collect method level features.
i. Calculate the specified features for every
method in a new program.
ii. Store the feature set in a vector.
iii. Assign the label ‘0’ for each feature
vector in a file.
4. Predict the label as either hot or cold for the test
data generated in step 3 using the predictive
model derived in step 2.
Figure 3: System outline
4.2 Extracting program features
The ‘C’ programs used for training are converted
into LLVM bytecodes using the LLVM frontend.
Every bytecode file is organized into a single module.
Each module contains methods which are either user-
defined or pre-defined. Only static features of the
user-defined methods are extracted from the
bytecode module, for the simple reason that they can
be easily collected by an offline program analysis.
Table 1 lists the 10 static features that are used to
train the classifier. Each feature value of a method is
calculated in relation to an identical feature value
extracted from the entire bytecode module. The
collection of all the feature values for a method
constitutes the feature vector xi. This feature vector
xi is stored for subsequent labeling. Each feature
vector xi is then labeled yi and classified as either hot
3 77
 (+1) or cold (-1) based on an arbitrary threshold feature 1 indicates the percent of loops found in the
scheme described in the next section. method. The “hot method threshold” used being 50%,
4 out of the 8 most frequently executed methods in a
Table 1: static features for identifying hot methods. program are designated as hot methods. The first
element in each vector is the label yi (+1 or -1). Each
1. Number of loops in a method. element of the feature vector indicates the feature
Average loop depth of all the loops in the number followed by the feature values.
2.
method.
3. Number of top level loops in a method. 4.4 Creating test data set
Number of bytecode level instructions in When a new program is encountered, the test
4. data set is collected in a way similar to the training
the method.
5. Number of Call instructions in a method. data set, except that the label is specified as zero.
6. Number of Load instructions in a method.
7. Number of Store instructions in a method. 0 1:1 2:1 3:1 4:1.13098 5:2.91262 6:2.05479
Number of Branch instructions in the 7:1.09091 8:1.34875 9:1.55172 10:34
8. 0 1:0 2:0 3:0 4:0.552341 5:0.970874 6:1.14155
method.
9. Number of Basic Blocks in the method. 7:0.363636 8:0.385356 9:0.862069 10:4
0 1:1 2:1 3:1 4:1.26249 5:0 6:2.51142 7:2.90909
10. Number of call sites for each method.
8:1.15607 9:1.2069 10:40
4.3 Extracting method execution frequencies
Figure 5: Sample test data set
Hot methods are frequently executing program
segments. To identify hot and cold methods within a
4.5 Training and prediction using SVM
training program, profile information is gathered
Using the training data set file as input , the
during execution. The training bytecode modules are
machine learning algorithm SVM is trained with
instrumented with a counter variable in each user-
default parameters (C-SVM, C=1, radial base
defined method. This instrumented bytecode module
function). Once trained the predictive model is
is then executed and the execution frequency of each
generated as output. The derived model is used to
method is collected. Using this profile information,
predict the label for each feature vector in the test
the top ‘N’ most frequently executed methods are
data set file. The training and prediction are done
classified as hot. This system keeps the value ‘N’ as
offline. Subsequently, the new program used for
the “hot method threshold”. In this scheme of
creating test data set is instrumented. Executing this
classification, each feature vector (xi) is now labeled
instrumented program provides the most frequently
yi (+1) and yi (-1) for hot and cold methods
executed methods. The prediction accuracy of the
respectively. The feature vector (xi) along with its
system is evaluated by comparing the predicted
label (yi) is then written into a training dataset file.
output with the actual profile values.
Similarly, the training data set of the different
training programs is accumulated in the file. This file
5 EVALUATION
is used as an input to train the predictive model.
5.1 Method
+1 1:1 2:1 3:1 4:0.880046 5:2.51046 6:0.875912
Prediction accuracy is defined as the ratio of
7:0.634249 8:1.23119 9:1.59314 10:29
events correctly predicted to all the events
-1 1:0 2:0 3:0 4:1.16702 5:1.25523 6:1.0219
encountered. This prediction accuracy is of two
7:3.38266 8:1.50479 9:1.83824 10:2
types: hot method prediction accuracy and total
+1 1:2 2:2 3:2 4:1.47312 5:0.83682 6:1.89781
prediction accuracy. Hot method prediction accuracy
7:1.47992 8:2.59918 9:2.81863 10:3
is the ratio of correct hot method predictions to the
actual number of hot methods in a program, whereas
Figure 4: Sample training data set
total prediction accuracy is the ratio of correct
predictions (either hot or cold) to the total number of
The general format of a feature vector is
methods in a program. Hot method prediction
yi 1:xi1, 2:xi2, 3:xi3, …….j:xij
accuracy is evaluated at three hot method threshold
where the labels 1, 2, 3,.... , j are the feature numbers
levels: 50%, 40% and 30%.
and xi1, xi2, ...., xij are their corresponding feature
The leave-one-out cross-validation method is
values. Fig. 4 shows a sample of three feature vectors
used in evaluating this system. This is a standard
from the training dataset collected for the user-
machine learning technique where ‘n’ benchmark
defined methods found in the SPEC benchmark
programs are used iteratively for evaluation. One out
program. The first feature vector in Fig. 4 is a hot
of the ‘n’ programs is used for testing and the rest ‘n-
method and is labeled +1. The values of the ten
1’ programs are used for training the model. This is
features are serially listed for example ‘1’ is the
repeated for all the ‘n’ programs in the benchmark
value of feature 1 and ‘29’ of 10. The value ‘1’ of

UbiCC Journal - VolumeUbiquitous

3 Computing and Communication Journal 4 78
 suite.
Total Method Prediction Accuracy
5.2 Benchmarks Hot Method Thresholds 50% 40% 30%
Two benchmark suites, SPEC CPU2000 INT 120
[17] and UTDSP [13] have been used for training

Prediction Accuracy %
100
and prediction. UTDSP is a C benchmark and SPEC
80
CPU2000 INT has C and C++ benchmarks.
Evaluation of the system is based on only the C 60

programs of either benchmark. The model trained 40

from the ‘n-1’ benchmark programs in the suite is 20

used to predict the hot methods in the missed out 0
benchmark program.

2
p

cf

x
er
c

e
f
ol
ip
gc

r te
zi

ag
m

rs

tw
bz
g

6.

1.

vo
pa

er
4.

0.
6.
17

18

Av
16

5.
7.

5.3 Tools and platform
The system is implemented in the Low Level
Virtual Machine (LLVM) version 1.6 [6]. LLVM is
an open source compiler infrastructure that supports
compile time, link time, run time and idle time
optimizations. The results are evaluated on an Intel
(R) Pentium (R) D with 2.80 GHz and 480MB of
RAM running Fedora Core 4. This system uses the
libSVM tool [7]. It is a simple library for Support
Vector Machines written in C.
Hot Method Prediction Accuracy
Hot Method Thresholds
120
Prediction Accuracy %
30
25
25
19
SPEC CPU2000 INT
Figure 7: Total prediction accuracy on the SPEC
CPU2000 INT benchmark
The total method prediction accuracy on the
SPEC CPU2000 INT and UTDSP benchmark suites
is shown in Fig. 7 and 9. The total method prediction
accuracy for all C programs on the SPEC CPU2000
INT varies from 36 % to 100 % with an average of
68.43%, 71.14% and 71.14% for the three hot
method thresholds respectively. This averages to
70.24%. The average prediction accuracies obtained
50% 40% 30% on the UTDSP benchmark suite are 69%, 71% and
58% respectively for 50%, 40% and 30% hot method
thresholds. This averages to 66%. Overall the system

100

80
predicts both hot and cold methods in a program with
68.15% accuracy.
60

40 7 CONCLUSION AND FUTURE WORK

20

0
Optimizers depend on profile information to
identify the hot methods of program segments. The
2
p

cf

x
er
c

e
f
ol
ip
gc

rte
zi

ag
m

major inadequacy associated with the dynamic

rs

tw
bz
g

6.

1.

vo
pa

er
4.

0.
6.
17

18

Av
16

5.
7.

30
25

optimization technique is the high cost of accurate

25
19

SPEC CPU2000 INT

Figure 6: Hot method prediction accuracy on the
SPEC CPU2000 INT benchmark
6 RESULTS
Fig. 6 shows the prediction accuracy of the
trained model on the SPEC CPU2000 INT
benchmark program at three different hot method
thresholds: 50%, 40% and 30%. The hot method
prediction accuracy for all C programs on the
benchmark is found to vary from 0 % to 100 % with
an average of 57.86 %, 51.43% and 39.14% for the
three hot method thresholds respectively. This
averages to 49.48% on the SPEC CPU2000 INT
benchmark suite. Similarly, on the UTDSP
benchmark suite, in a 0% to 100% range, the hot
method prediction accuracy averages for the three
thresholds are 84%, 81% and 62% respectively.
This averages to 76% on the UTDSP benchmark
suite. Overall, this new system can obtain 62.57%
hot method prediction accuracy.
data profiling via program instrumentation. In this
work, a method is worked out to identify hot
methods in a program using the machine learning
algorithm, the SVM. According to our study, with a
set of ten static features used in training the system,
the derived model predicts total methods within a
program with 68.15% accuracy and hot methods
with 62.57% accuracy. However, hot method
prediction is of greater value because optimizations
will be more effective in these methods.
Future work in this area is aimed at improving
the prediction accuracy of the system by identifying
more effective static and dynamic features of a
program. Further research in this system can be
extended to enhance it to a dynamic hot method
prediction system which can be used by dynamic
optimizers. Applying this approach, the prediction
accuracy of the other machine learning algorithms
can be evaluated to build additional models.

UbiCC Journal - VolumeUbiquitous

3 Computing and Communication Journal 5 79
 Hot Method Prediction Accuracy
Hot Method Thresholds 50% 40% 30%
100
Prediction Accuracy %

80

60

40

20

t
ir
c
2

ff t
m

em

m
s

fi r
ss

eg

ge
m

iir
ee

ng
ct

ul
tra
lp

l li

sf
72

nr
pc

ra
te
e

ra
tr e
jp
sL

lm
Fu

ec

od
pr

t
de

og
ad

la

e
cu

dy
m

m
sp

Av
st
_
co

2.
ar
ge

en

hi

V3
M
ed

W
1.

1.
72

72

UTDSP benchmark
G

Figure 8: Hot method prediction accuracy on the UTDSP benchmark.

Total Prediction Accuracy

Hot Method Thresholds
50% 40% 30%
100
Prediction Accuracy %

80

60

40

20

e
ff t

fir

i ir
ng

eg
ct

t
c

ir
m

rm
em
ee

llis
m

ra

ul
ss

ag
lp
72

sf
te
pc

ra

m
jp
Fu

tn
L
e

tre

lm

er
ec

od
e

G
s
pr
ad

og

la
_d

cu

dy

Av
sp

m
m

st
ge

en
ar

2.
co

hi
M

V3
ed

W
1.

1.
72

72

UTDSP benchmark
G

Figure 9: Total Prediction Accuracy on the UTDSP benchmark.

International Conference on Compiler

8 REFERENCES Construction (CC 2006), 2006.
[6] C. Lattner and V. Adve: LLVM: A compilation
[1] Matthew Arnold, Stephen Fink, David Grove, framework for lifelong program analysis &
Michael Hind, and Peter F. Sweeney: A Survey transformation, In Proceedings of the 2004
of Adaptive Optimization in Virtual Machines, International Symposium on Code Generation
Proceedings of the IEEE, pp. 449-466, and Optimization (CGO’04), March 2004.
February 2005. [7] Chih-Chung Chang and Chih-Jen Lin:
[2] A. Monsifrot, F. Bodin, and R. Quiniou: A LIBSVM : a library for support vector
machine learning approach to automatic machines, 2001. Software available at
production of compiler heuristics, In http://www.csie.ntu.edu.tw/~cjlin/libsvm.
Proceedings of the International Conference on [8] M. Stephenson, S. Amarasinghe, M. Martin,
Artificial Intelligence: Methodology, Systems, and U. M. O’Reilly: Meta optimization:
Applications, LNCS 2443, pp. 41-50, 2002. Improving compiler heuristics with machine
[3] S. Long and M. O'Boyle: Adaptive java learning, In Proceedings of the ACM
optimization using instance-based learning, In SIGPLAN Conference on Programming
ACM International Conference on Language Design and Implementation
Supercomputing (ICS'04), pp. 237-246, June (PLDI’03), pp. 77–90, June 2003.
2004. [9] F Agakov, E Bonilla, J Cavazos, G Fursin, B
[4] John Cavazos and Michael F.P. O'Boyle: Franke, M.F.P. O'Boyle, M Toussant, J
Automatic Tuning of Inlining Heuristics, 11th Thomson, C Williams: Using machine learning
International Workshop on Compilers for to focus iterative optimization, In Proceedings
Parallel Computers (CPC 2006), January 2006. of the International Symposium on Code
[5] John Cavazos, J. Eliot B. Moss, and Michael Generation and Optimization (CGO), pp. 295-
F.P. O'Boyle: Hybrid Optimizations: Which 305, 2006.
Optimization Algorithm to Use?, 15th [10] Christophe Dubach, John Cavazos, Björn

UbiCC Journal - VolumeUbiquitous

3 Computing and Communication Journal 6 80
 Franke, Grigori Fursin, Michael O'Boyle and
Oliver Temam: Fast compiler optimization
evaluation via code-features based performance
predictor, In Proceedings of the ACM
International Conference on Computing
Frontiers, May 2007.
[11] John Cavazos, Michael O'Boyle: Method-
Specific Dynamic Compilation using Logistic
Regression, ACM Conference on Object-
Oriented Programming, Systems, Languages,
and Applications (OOPSLA), Portland, Oregon,
October 22-26, 2006.
[12] John Cavazos: Automatically Constructing
Compiler Optimization Heuristics using
Supervised Learning, Ph.D thesis, Dept. of
Computer Science, University of Massachusetts,
2004.
[13] C. Lee: UTDSP benchmark suite. In
http://www.eecg.toronto.edu/~corinna/DSP/infr
astructure/UTDSP.html, 1998.
[14] G. Fursin, C. Miranda, S. Pop, A. Cohen, and
O. Temam: Practical run-time adaptation with
procedure cloning to enable continuous
collective compilation, In Proceedings of the
5th GCC Developer’s Summit, Ottawa, Canada,
July 2007.
[15] Vapnik, V.N.: The support vector method of
function estimation, In Generalization in
Neural Network and Machine Learning,
Springer-Verlag, pp.239-268, 1999.
[16] S. Kotsiantis: Supervised Machine Learning: A
Review of Classification Techniques,
Informatica Journal 31, pp. 249-268, 2007.
[17] The Standard Performance Evaluation
Corporation. http://www.specbench.org.
[18] M. Stephenson and S.P. Amarasinghe:
Predicting unroll factors using supervised
classification, In Proceedings of International
Symposium on Code Generation and
Optimization (CGO), pp. 123-134, 2005.
[19] www.cag.csail.mit.edu
UbiCC Journal - VolumeUbiquitous
3 Computing and Communication Journal
/~mstephen/stephenson_phdthesis.pdf , M. W.
Stephenson, Automating the Construction of
Compiler Heuristics Using Machine Learning,
PhD thesis, MIT, USA, 2006.
[20] J. Cavazos and J. Moss: Inducing heuristics to
decide whether to schedule, In Proceedings of
the ACM SIGPLAN Conference on
Programming Language Design and
Implementation (PLDI), 2004.
[21] B.Calder, D.Grunwald, Michael Jones,
D.Lindsay, J.Martin, M.Mozer, and B.Zorn:
Evidence-Based Static Branch Prediction Using
Machine Learning, In ACM Transactions on
Programming Languages and Systems
(ToPLaS-19), Vol. 19, 1997.
[22] Daniel A. Jiménez , Calvin Lin: Neural
methods for dynamic branch prediction, ACM
Transactions on Computer Systems (TOCS),
Vol. 20 n.4, pp.369-397, November 2002.
[23] Jeremy Singer, Gavin Brown and Ian Watson:
Branch Prediction with Bayesian Networks, In
Proceedings of the First Workshop on
Statistical and Machine learning approaches
applied to Architectures and compilation, pp.
96-112, Jan 2007.
[24] Culpepper B., Gondre M.: SVMs for Improved
Branch Prediction, University of California,
UCDavis, USA, ECS201A Technical Report,
2005.
[25] Youfeng Wu, Larus. J. R. : Static branch
frequency and program profile analysis, 1994.
MICRO-27, Proceedings of the 27th Annual
International Symposium on Microarchitecture,
pp: 1 – 11, 1994.
[26] C.-L. Huang and C.-J. Wang: A GA-based
feature selection and parameters optimization
for support vector machines, Expert Systems
with Applications, Vol. 31, Issue 2, pp: 231-
240, 2006.
7 81
 PROPAGATION MODEL FOR HIGHWAY IN MOBILE
COMMUNICATION SYSTEM

K.Ayyappan *P. Dananjayan

Department of Electronics and Communication Department of Electronics and Communication
Engineering, Engineering
Rajiv Gandhi College of Engineering and Pondicherry Engineering College
Technology, Pondicherry - 605014, India.
Pondicherry, India. * pdananjayan@hotmail.com
*corresponding author

ABSTRACT

Radio propagation is essential for emerging technologies with appropriate design,

deployment and management strategies for any wireless network. It is heavily site
specific and can vary significantly depending on terrain, frequency of operation,
velocity of mobile terminal, interface sources and other dynamic factor. Accurate
characterization of radio channel through key parameters and a mathematical
model is important for predicting signal coverage, achievable data rates, specific
performance attributes of alternative signaling and reception schemes. Path loss
models for macro cells such as Hata Okumura, COST 231 and ECC 33 models are
analyzed and compared their parameters. The received signal strength was
calculated with respect to distance and model that can be adopted to minimize the
number of handoffs and avoid ping pong effect are determined. This paper
proposes a propagation model for highway environment between Pondicherry -
Villupuram which is 40 kilometers spaced out .Comparative study with real time
measurement obtained from Bharat Sanchar Nigam Limited (BSNL) a GSM based
wireless network for Pondicherry, India has been implemented.

Keywords: Handoff, Path loss, Received signal strength, ping pong, cellular mobile.

1 INTRODUCTION to model a path loss equation. To conceive these

Propagation models have traditionally focused on
predicting the received signal strength at a given
distance from the transmitter, as well as the variability
of the signal strength in a close spatial proximity to a
particular location. Propagation models that predict
the signal strength for an arbitrary transmitter-
receiver (T-R) separation distance are useful in
estimating the radio coverage area of a transmitter.
Conversely, propagation models that characterize the
rapid fluctuations of the received signal strength over
very short travel distances are called small-scale or
fading models. Propagation models are useful for
predicting signal attenuation or path loss. This path
loss information may be used as a controlling factor
for system performance or coverage so as to achieve
perfect reception [1].
The common approaches to propagation
modeling include physical models and empirical
models. In this paper, only empirical models are
considered. Empirical models use measurement data
models, a correlation was found between the received
signal strength and other parameters such as antenna
heights, terrain profiles, etc through the use of
extensive measurement and statistical analysis.
Radio transmission in a mobile communication
system often takes place over irregular terrain. The
terrain profile of a particular area needs to be taken
into account for estimating the path loss. The terrain
profile may vary from a simple curved earth profile
to a highly curved mountainous profile. A number of
propagation models are available to predict path loss
over irregular terrain. While all these models aim to
predict signal strength at a particular receiving point
or in a specific location are called sector, the methods
vary widely in their approach, complexity and
accuracy. Most of these models are based on a
systematic interpretation of measurement data
obtained in the service area. In cellular mobile
communication systems, handoff takes place due to
movement of mobile unit and unfavorable conditions
inside an individual cell or between a numbers of

UbiCC Journal - Volume 3 1 82

 adjacent cells [2, 3]. It is a seamless service to active
users while data transfer is in progress, so
unnecessary handoffs should be avoided.
Hard handoff suffers from ‘ping pong’ effect
when the mobile users are near the boundaries of
adjacent cells and is a result of frequent handoffs. The
parameters measured to determine handoff are usually
the received signal strength, the signal to noise ratio
and the bit error rate. However, a path loss model can
increase the connection reliability. Hence, the choice
of path loss model plays an important role in the
performance of handoffs. In this paper different path
loss models for macro cells such as Hata Okumura
model, Cost 231 model and ECC 33 model are
analyzed and compared their parameters. A
propagation model for highway is proposed by
modifying Cost 231 and Hata Okumura suburban
model and it is implemented in Pondicherry –
Villupuram highway and compared its parameters
with experimental values.
The work is organized as follows. Section 2
describes the path loss models. Section 3 deals with
the received signal strength for different path loss
models. Section 4 discusses the models and the
results are evaluated. Section 5 concludes with the
performance of various path loss models.
2 PATH LOSS MODELS
Path loss is the reduction in power of an
electromagnetic wave as it propagates through space.
It is a major component in analysis and design of link
budget of a communication system. It depends on
frequency, antenna height, receive terminal location
relative to obstacles and reflectors, and link distance,
among many other factors. Macro cells are generally
large, providing a coverage range in kilometers and
used for outdoor communication. Several empirical
path loss models have been determined for macro
cells. Among numerous propagation models, the
following are the most significant ones, providing the
foundation of mobile communication services. The
empirical models are
i. Hata Okumura model
ii. COST 231 model
iii. ECC 33 model
These prediction models are based on extensive
experimental data and statistical analysis, which
enable us to compute the received signal level in a
given propagation medium [5, 6]. The usage and
accuracy of these prediction models depends on the
propagation environment.
2.1 Free Space Propagation Model
In radio wave propagation models, the free space
model predicts that received power decays as a
function of T-R separation distance. This implies that
UbiCC Journal - Volume 3
received power decays with distance at a rate of 20
dB/decade.
The path loss for free space model when antenna
gains are included is given by
PL( dB) = -Gt - Gr +32.44 +20log ( d) + 20 log ( f ) (1)
where
Gt is the transmitted antenna gain in dB,
Gr is the received antenna gain in dB,
d is the T-R separation distance in kilometers and
f is the frequency in MHz.
2.2 The Hata Okumura model
The Hata-Okumura model is an empirical
formula for graphical path loss data provided by
Yoshihisa Okumura, and is valid from 150 MHz to
1500 MHz. The Hata model is a set of equations
based on measurements and extrapolations from
curves derived by Okumura. Hata presented the
urban area propagation loss as a standard formula,
along with additional correction factors for
application in other situations such as suburban, rural
among others. The computation time is short and
only four parameters are required in Hata model [7].
However, the model neglects terrain profile between
transmitter and receiver, i.e. hills or other obstacles
between transmitter and receiver are not considered.
This is because both Hata and Okumura made the
assumption that transmitter would normally be
located on hills. The path loss in dB for the urban
environment is given by
PL ( dB ) = A + B log ( d ) (2)
where
d is distance in kilometer,
A represents a fixed loss that depends on frequency
of the signal.
These parameters are given by the empirical formula
( )
A = 69.55 + 26.16 log ( f ) -13.82 log h b - a ( h m )
B = 44.9 - 6.55 log h b ( )
where,
f is frequency measured in MHz,
hb is height of the base station antenna in meters,
hm is mobile antenna height in meters and
a (hm) is correction factor in dB
For effective mobile antenna height a (hm) is given by
a [ hm ] = 1.1 log ( f ) - 0.7 hm - 1.56log ( f ) - 0.8
The path loss model for highway is given by
For without noise factor
2
f
PL ( dB) = PL ( dB)urban - 2 log - 5.4 (3)
28
For with noise factor
2 83
 2 d is distance between base station and mobile (km)
f hb is BS antenna height in meters and
PL ( dB ) = PL ( dB )urban - 2 log (4)
28 hm is mobile antenna height in meters.

3 RECEIVED SIGNAL STRENGTH

2.3 COST-231 Hata model
In mobile communication, received signal
To extend Hata-Okumura- model for personal
strength is a measurement of power present in a
communication system (PCS) applications operating
received radio signal. Signal strength between base
at 1800 to 2000 MHz, the European Co-operative for
station and mobile must be greater than threshold
Scientific and Technical Research (COST) came up
value to maintain signal quality at receiver [9].
with COST-231 model. This model is derived from
Simultaneously signal strength must not be too strong
Hata model and depends upon four parameters for
to create more co-channel interference with channels
prediction of propagation loss: frequency, height of
in another cell using same frequency band. Handoff
received antenna, height of base station and distance
decision is based on received signal strength from
between base station and received antenna [8].
current base station to neighbouring base stations.
The path loss in urban area is given by
The signal gets weaker as mobile moves far away
from base station and gets stronger as it gets closer.
( )
PL ( dB) = 46.33+ 33.9 log ( f ) -13.82 log hb The received signal strength for various path loss
models like Hata Okumura model, Cost 231 model,
(5)
- a ( hm ) + 44.9 - 6.55 log ( hb ) log ( d) and ECC-33 model are calculated as

Pr= Pt +Gt+Gr-PL - A (7)

where a ( hm ) = 1.1 log ( f ) -0.7 hm - 1.56 log ( f ) -0.8 where,
Pr is Received signal strength in dBm,
The path loss calculation for highway is similar to Pt is transmitted power in dBm,
Hata-Okumura model. Gt is transmitted antenna gain in dB,
Gr is received antenna gain in dB,
2.4 ECC-33 model PL is total path loss in dB and
A is connector and cable loss in dB.
The ECC 33 path loss model, which is developed
by Electronic Communication Committee (ECC), is 4 PERFORMANCE ANALYSIS
extrapolated from original measurements by Okumura
and modified its assumptions so that it more closely Table 1: Simulation parameter
represents a fixed wireless access (FWA) system. The
path loss model is defined as, Parameters Values
PL ( dB ) = A fs + A bm - G t - G r (6) Base station transmitter power 43dBm
Mobile transmitter power 30dBm
where,
Afs is free space attenuation, Base station antenna height 35m
Abm is basic median path loss, Mobile antenna height 1.5m
Gt is BS height gain factor and
Gr is received antenna height gain factor. Transmitter antenna gain 17.5dB
They are individually defined as, Threshold level for mobile -102dBm
Afs = 92.4 + 20 log ( d) + log ( f ) and Threshold level for base station -110dBm
2 Frequency 900 MHz
Abm = 20.41+9.83log( d) +7.894log( f ) +9.56 log( f )
Connector loss 2 dB
h 2
Gt = log b 13.958+ 5.8 log( d)
{ } Cable loss 1.5dB
200
Duplexer loss 1.5dB
for medium city environments,
Maximum uplink loss 161.5dB
G r = 42.57 +13.7 log ( f ) log ( h m )-0.585
where, Maximum downlink loss 161.8dB
f is frequency in GHz,
The performance analysis is based on the network specification obtained from Bharat Sanchar
calculation of received signal strength, path loss Nigam Limited (BSNL), India shown in Table 1 is
between the base station and mobile from the used for evaluating the performance of path loss
propagation model. The GSM based cellular models.

UbiCC Journal - Volume 3 3 84

 4.1 Path losses for various models The number of handoffs per call is relative to cell
size, smaller the cell size maximum the number of
handoff and the path loss model which cover
maximum distance will minimize the number of
handoff. The path loss calculation using Hata-
Okumura and COST 231 model are less than the
threshold value up to 19 km and ECC 33 model
exceed the threshold value at 5km.

Fig.1 Pondicherry- Villupuram highway map Path loss (dB)

Table 2 Path loss for various models

HATA
Distance COST ECC33
OKUMURA Distance (km)
(km) 231(dB) (dB)
(dB)
0.5 104.7 105.1 130.7 Fig.2 Comparison of Path loss

4.2 Received signal strength for various models

1 115.2 115.5 139.3
1.5 121.3 121.7 144.7 Table 3: BS to MS received power
2 125.6 126.0 148.6 COST HATA ECC
Distance
2.5 129.0 129.4 151.8 231 OKUMURA 33
(km)
(dBm) (dBm) (dBm)
3 131.7 132.1 154.5
0.5 -49.2 -49.6 -75.2
3.5 134.1 134.5 156.8
1 -59.7 -60.0 -83.8
4 136.1 136.5 158.8
1.5 -65.9 -66.1 -89.2
The path losses for various models are
calculated using eq (2),(5) and (6) between 2 -70.1 -70.5 -93.2
Pondicherry-Villupuram highway as shown in Fig.1
which is connected via Villianur, Kandamangalam, 2.5 -73.5 -73.9 -96.3
Madagadipet and Valavanur. The circle shown in
this figure are the base station transceiver (BTS) 3 -76.2 -76.6 -99.0
and identified by BTS number.The path loss values
are calculated and given in Table 2 and the 3.5 -78.6 -79.0 -101.3
comparison for various models are given in
Fig.2.The maximum allowable uplink loss for 4 -80.6 -81.0 -103.3
Nortel S8K base station transceiver is 161.5 dB and
maximum allowable downlink loss is 161.8 dB.

UbiCC Journal - Volume 3 4 85

 Received signal strength (dBm)
Received signal strength (dBm)

Distance (km)
Distance (km)
Fig.3 Received signal strength for suburban models
Fig. 4 Received signal strength for highway models
The received signal strength for COST
The general area around highway could be
231, Hata-Okumura and ECC 33 models are
suburban because of location. The path loss
calculated using eq (7) shown in Table.3 for
calculation is a major concern in highways with and
suburban area and the comparison shown in Fig.3.
without noise level. In this paper the suburban model
The received signal strength using ECC 33 model is
is modified to highway with small correction factor
-103.3 dBm at four kilometers which is greater than
with respect to the location. RSS value for highway
threshold level of mobile receiver -102 dBm. The
without noise factor was calculated using suburban
received signal strength using COST 231 and Hata-
model but highway with noise factor was calculated
Okumura model values are less than the sensitivity
using additional correction factor of 5.4 with suburban
threshold of mobile. So these two models are
model. Here up to 2.5 km the received signal strength
preferred for maximum coverage area and reduce
was calculated using suburban model and beyond 2.5
the number of handoff.
km it was calculated with additional noise
factor of 5.4.
4.3 Highway propagation model

Table 4: BS to MS received power

Distance COST HATA Experiment

(km) 231 OKUMURA Value
(dBm) (dBm) (dBm)
0.5 -49.2 -49.6 -50

1 -59.7 -60.0 -57

1.5 -65.9 -66.1 -65

2 -70.1 -70.5 -68

2.5 -73.5 -73.9 -74

3 -81.6 -82.0 -82

3.5 -84.0 -84.4 -84

4 -86.0 -86.4 -86

Fig.5 Adjacent cell RSSI

UbiCC Journal - Volume 3 5 86

 The Agilent Technologies provides Drive Test
solution for GSM networks. The Agilent Drive Test
E6474A model is used to calculate the amount of
signal received by a mobile in Pondicherry -
Villupuram highway as shown in Table 4.The
received signal strength for COST 231, Hata-
Okumura models are calculated and compared with
experimental value are shown in Fig 4. The
modified COST 231, Hata-Okumura suburban
models for highway are matches with the
experimental values.
High RSS handover margins can result in poor
reception and dropped calls, while very low values
of handover margin can produce ping pong effects
as mobile switches frequently between cells. The
existing cell broadcast control channel (BCCH-74)
and adjacent cells (BCCH-76, 66, 68,69and 78)
received signal strengths are shown in Fig.5. The
optimum handover decision taken from drive test
adjacent cell received signal strength indicator
(RSSI) as shown in Fig.5.For BCCH-74 the
corresponding base station identification code
(BSIC) is 41 and the received signal
strength is -72 dBm. Consecutively, mobile user
receives adjacent inter and intra cell received signal
strengths from BSIC 44,46and 41. The optimum
handover decision taken from the drive test adjacent
cell received signal strength indicator (RSSI) to
improve the signal reception and reduce the number
of dropped calls and also the ping-pong effect
5 CONCLUSION
In this paper, different path loss models for
macro cells were used. The calculated path loss is
compared with existing model like Hata-Okumura
model, COST 231model and ECC 33 model. The
received signal strength from base stations was
calculated and the calculated values are compared
with observed value between Pondicherry and
Villupuram highway. The result shows that
modified suburban model for highway using Hata-
Okumura and COST 231 model is closer to the
observed received signal strength and predicted to
be suitable model for highway received signal
strength calculation.
REFERENCES
[1] Armoogum.V, Soyjaudah.K.M.S, Fogarty.T
and Mohamudally.N, “Comparative Study of
Path Loss using Existing Models for Digital
Television Broadcasting for Summer Season in
the North of Mauritius”, Proceedings of Third
Advanced IEEE International Conference on
Telecommunication, Mauritius Vol. 4, pp 34-
38, May 2007.
UbiCC Journal - Volume 3
[2] Tomar G.S and Verma. S, “Analysis of handoff
initiation using different path loss models in
mobile communication system”, Proceedings of
IEEE International Conference on Wireless and
Optical Communications Networks, Bangalore,
India, Vol. 4, May 2006.
[3] Ken-Ichi, Itoh, Soichi Watanche, Jen-Shew Shih
and Takuso safo, “Performance of handoff
Algorithm Based on Distance and RSS
measurements”, IEEE Transactions on vehicular
Technology, Vol. 57, No.6, pp 1460-1468,
November 2002.
[4] Abhayawardhana V.S, Wassell.I.J, Crosby
D, Sellars. M.P. and Brown. M.G, “Comparison
of empirical propagation path loss models for
fixed wireless access systems”, Proceedings of
IEEE Conference on Vehicular Technology ,
Stockholm, Sweden, Vol. 1, pp 73-77, June 2005.
[5] Maitham Al-Safwani and Asrar U.H. Sheikh,
“Signal Strength Measurement at VHF in the
Eastern Region of Saudi Arabia”, The Arabian
Journal for Science and Engineering, Vol. 28,
No.2C, pp.3 -18, December 2003.
[6] S. Hemani and M. Oussalah, “Mobile Location
System Using Netmonitor and MapPoint server”,
Proceedings of Sixth annual Post graduate
Symposium on the Convergence of
Telecommunication, Networking and
Broadcasting, PGNet ,pp.17-22,2006.
[7] T.S. Rappaport, “Wireless Communications”,
Pearson Education, 2003.
[8] William C.Y. Lee, “Mobile Cellular
Telecommunications”, McGraw Hill International
Editions, 1995.
[9] Ahmed H.Zahram, Ben Liang and Aladdin Dalch,
“Signal threshold adaptation for vertical handoff
on heterogeneous wireless networks”, Mobile
Networks and application, Vol.11, No.4, pp 625-
640, August 2006.
6 87
 A Hardware implementation of Winograd Fourier Transform
algorithm for Cryptography

1
G.A.Sathishkumar and 2Dr.K.Boopathy bagan
1
Assistant Professor, Department of Electronics and Communication Engineering, Sri Venkateswara
College of Engineering, Sriperumbudur -602108.
2
Professor, Department of Electronics, Madras Institute of Technology, Anna University Chrompet Campus,
Chennai-600044
Tamil Nadu.
INDIA
sathish@svce.ac.in , kbb@mail.yahoo.com

ABSTRACT
This paper presents a hardware implementation of efficient algorithms that uses the
mathematical framework. The framework based on the Winograd’s Fourier Transform
Algorithm, obtaining a set of formulations that simplify cyclic convolution (CC) computations
and CRT. In particularly, this work focuses on the arithmetic complexity of a multiplication
and when there is multiplication then the product represents a CC computational operation.
The proposed algorithms is compared against existing algorithms developed making use of the
FFT and it is shown that the proposed algorithms exhibit an advantage in computational
efficiency .This design is most useful when dealing with large integers, and is required by
many modern cryptographic systems.
The Winograd Fourier Transform Algorithm (WFTA) is a technique that combines
the Rader’s index mapping and Winograd’s short convolution modules for prime-factors into
a composite-N Fourier Transform structure with fewer multipliers (O(N) ). While theoretically
interesting, WFTA’s are very complicated and different for every length. It can be
implemented on modern processors with few hardware multipliers and hence, is very useful in
practice today.

Keywords: Discrete Fourier Transform, Fast Fourier Transform, Winograd’s Theorem,

Chinese Remainder Theorem.

INTRODUCTION

Many popular crypto-systems like the RSA
encryption scheme [12], the Diffie-Hellman
(DH) key agreement scheme [13], or the
Digital Signature Algorithm (DSA) [14] are
based on long integer modular exponentiation.
A major difference between the RSA scheme
and cryptosystems based on the discrete
logarithm problem is the fact that the modulus
used in the RSA encryption scheme is the
product of two prime numbers. This allows
utilizing the Chinese Remainder Theorem
(CRT) in order to speed up the private key
operations. From a mathematical point of
view, the usage of the CRT for RSA
decryption is well known. However, for a
hardware implementation, special multiplier
architecture is necessary to meet the
requirements for efficient CRT-based
decryption. This paper presents the basic
algorithmic and architectural concepts of the
WFTA crypto chip, and describes how they
were combined to provide optimum
performance. The major design goal with the
WFTA was the maximization of performance
on several levels, including the implemented
hardware algorithms.
In digital signal processing, the design of fast
and computationally efficient algorithms has
been a major focus of research activity. The
objective, in most cases, is the design of
algorithms and their respective implementation
in a manner that perform the required
computations in the least amount of time. In
order to achieve this goal, parallel processing
has also received a lot attention in the research
community [1]. This document is organized as
follows. First, mathematical foundations
needed for the study of algorithms to compute
the DFT and FFT algorithm are summarized.
Second, identification is established between
Winograd Fourier Transform and the Rader’s
algorithm. Third, the algorithm development
for the basic problem of the multiplication,
using the conceptual framework developed in
the two previous sections, is explained. The

UbiCC Journal - Volume 3 88

 section also presents several signal flow
diagrams that may be implemented in diverse
architectures by means of very large scale
integration (VLSI) or very high-speed
integrated circuits hardware description
language (VHDL). Conclusions, contributions,
and future development of the present work are
then summarized.
1. DISCRETE FOURIER TRANSFORM
(DFT)
1.1 DEFINITION
The discrete Fourier transform (DFT) is a
powerful reversible mapping transform for
discrete data sequences with mathematical
properties analogous to those of the Fourier
transform and it transforms a function from
time domain to frequency domain. For length n
input vector x, the DFT is a length n vector X,
with n elements:
(1)
A simple description of these
equations is that the complex numbers Fj
represent the amplitude and phase of the
different sinusoidal components of the input
"signal" xk. The DFT computes the Fj from the
xk, while the IDFT shows how to compute the
xn as a sum of sinusoidal components
Fjexp(2πikn/N) / N with frequency (k/N)
cycles per sample. By writing the equations in
this form, we are making extensive use of
Euler's formula to express sinusoids in terms
of complex exponentials, which are much
easier to manipulate. The number of
multiplication and addition operations required
by the Discrete Fourier Transform (DFT) is of
order N2 as there are N data points to calculate,
each of which requires N arithmetic
operations. To be exact, the input, if complex,
would contain 2 terms and every exponential
term would contain 2 terms. So, this would
quadruple the computational complexity, thus
number of multiplications is 4N2. Hence, for
real inputs the required number of
multiplications will be 2N2.
A fast Fourier transform (FFT) [2] is
an efficient algorithm to compute the discrete
Fourier transform (DFT) and it’s inverse.
FFT’s are of great importance to a wide variety
of applications, from digital signal processing
and solving partial differential equations to
algorithms for quick multiplication of large
integers. By far the most common FFT is the
Cooley-Tukey algorithm. This is a divide and
conquer algorithm that recursively breaks
down a DFT of any composite size N = N1N2
UbiCC Journal - Volume 3
into many smaller DFT’s of sizes N1 and N2,
along with O(N) multiplications by complex
roots of unity traditionally called twiddle
factors. The most well-known use of the
Cooley-Tukey algorithm is to divide the
transform into two pieces of size N / 2 at each
step, and is therefore limited to power-of-two
sizes, but any factorization can be used in
general (as was known to both Gauss and
Cooley/Tukey). These are called the radix-2
and mixed-radix cases, respectively (and other
variants such as the split-radix FFT have their
own names as well). Although the basic idea is
recursive, most traditional implementations
rearrange the algorithm to avoid explicit
recursion. In addition, because the Cooley-
Tukey algorithm breaks the DFT into smaller
DFTs, it can be combined arbitrarily with any
other algorithm for the DFT.
1.2 Multiplication of large integers
The fastest known algorithms [1, 8,
and 10] for the multiplication of very large
integers use the polynomial multiplication
method. Integers can be treated as the value of
a polynomial evaluated specifically at the
number base, with the coefficients of the
polynomial corresponding to the digits in that
base. After polynomial multiplication, a
relatively low-complexity carry-propagation
step completes the multiplication.
2. WINOGRAD FOURIER TRANSFORM
ALGORITHM (WFTA)
The Winograd Fourier Transform
Algorithm (WFTA)[7] is a technique that
combines the Rader’s index mapping and
Winograd’s short convolution algorithm for
prime-factors into a composite-N Fourier
Transform structure with fewer multipliers.
The Winograd algorithm, factorizes zN − 1 into
cyclotomic polynomials—these often have
coefficients of 1, 0, or −1, and therefore
require few (if any) multiplications, so
Winograd can be used to obtain minimal-
multiplication FFTs and is often used to find
efficient algorithms for small factors. Indeed,
Winograd showed that the DFT can be
computed with only O(N) irrational
multiplications, leading to a proven achievable
lower bound on the number of multiplications
for power-of-two sizes; unfortunately, this
comes at the cost of many more additions, a
tradeoff no longer favorable on modern
processors with hardware multipliers. In
particular, Winograd also makes use of the
89
 PFA as well as an algorithm by Rader for
FFTs of prime sizes.
2.1 RADER’S ALGORITHM
Rader's algorithm (1968)[7,10] is a
fast Fourier transform (FFT) algorithm that
computes the discrete Fourier transform (DFT)
of prime sizes by re-expressing the DFT as a
cyclic convolution.
Since Rader's algorithm only depends
upon the periodicity of the DFT kernel, it is
directly applicable to any other transform (of
prime order) with a similar property, such as a
number-theoretic transform or the discrete
Hartley transform.
The algorithm can be modified to
gain a factor of two savings for the case of
DFTs of real data, using a slightly modified re-
indexing/permutation to obtain two half-size
cyclic convolutions of real data; an alternative
adaptation for DFTs of real data, using the
discrete Hartley transform, was described by
Johnson.
Winograd extended Rader's algorithm
to include prime-power DFT sizes pm, and
today Rader's algorithm is sometimes
described as a special case of Winograd's FFT
algorithm, also called the multiplicative
Fourier transform algorithm, which applies to
an even larger class of sizes.
2.2 Algorithm
The Rader algorithm to compute the DFT,
N−1 nk
X(K) = ∑x(n)W k, n∈ZN;ord(W) = N (2)
n=0 N N
is defined for prime length N. We first
compute the DC component with
N −1
x (0 ) = ∑
n = 0
x (n ) (3 )
Because N=p is a prime, it is known that there
is primitive element, a generator ‘g’, that
generates all the elements of n and k in the
field ZN
We substitute n with gk mod N and get the
following index transform:
UbiCC Journal - Volume 3
for k=0,…,N-2. We notice that right side of the
above equation is a cyclic convolution i.e.,
N−2 n+k mod(N−1)
X[gk mod N]− x(0) = ∑x[gn mod N] W (4)
n=0 N
x[g0 mod N], x[g1mod N],...x[gN−2 mod N]]
n−2mod( N−1)
⊕[WN,WNg ,WNg ] (5)
Thus, an N-point DFT are converted into an
N-1 point Cyclic convolution.
3.WINOGRAD’S SMALL
CONVOLUTION ALGORITHM
This algorithm performs the
convolution with minimum number of
multiplications and additions and thus the
computational complexity of the process is
greatly reduced. Cyclic convolution is also
known as circular convolution. Let h={h0, h1, …
hn-1 } be the filter coefficients and
x={x0,x1,….xn-1} be the data sequence. The
cyclic convolution can be expressed as
s(p)=hOn x=h(p)x(p)mod(pn-1). (6)
The cyclic convolution can be computed as a
linear convolution reduced by modulo pn-1.
Alternatively, the cyclic convolution can be
computed using CRT with m(p)=pn-1,which is
much simpler. Thus, Winograd's minimum-
multiply DFT's are useful only for small N.
They are very important for Prime-Factor
Algorithms, which generally use Winograd
modules to implement the short-length
DFT’s[10]. The theory and derivation of these
algorithms is quite elegant but requires
substantial background in number theory and
abstract algebra. Fortunately, for the
practitioner, the entire short logarithm one is
likely to need have already been derived and
can simply be looked up without mastering the
details of their derivation.
3.1 Algorithm
1. Choose a polynomial m(p) with degree
higher than the degree of h(p)x(p) and
factor it into k+1 relatively prime polynomials
with real coefficients, i.e.,
90
 m(p) = m(0)(p)m(1)(p).....m(k)(p). (7) Project Navigator, we can run the
implementation process in one-step, or we can
2. Let M(i)(p) = m(p) / m(i)(p) and use the run each of the implementation processes
separately.
Chinese Remainder Theorem (CRT)
(i)
algorithm to get N (p).

3. Compute

h(i)(p) = h(p) mod m(i)(p),

(8)

x (i)(p) = x(p) mod m(i)(p),

(9)

for i=0,1,2,...,k.

4. Compute s(i)(p) = h(i)(p)x(i)(p) mod m(i)(p), (10)

for i =0,1,...,k.

5. Compute s(p) using the equation:

s(p)= ∑(i=0 to k) s(i)(p)N(i)(p)M(i)(p) mod m(p).

(11)

The computational complexity in case of

WFTA is of the order of N, O(N). It has been
Figure 1 Simulation result for N=2
found that the number of multipliers required
by WFTA is always less than 2N, which
drastically reduces the hardware needed for
implementing a DFT block.

4. REALIZATION OF WFTA IN
VERILOG HDL

The behavioral simulation and synthesis of

WFTA for N=2 and 5 can be viewed in the
following descriptions. We focus on the
Verilog HDL [6] used for our simulation and
synthesis. It is shown in Fig.1, 2, 3, 4,5and 6.

4.1.1 SYNTHESIS RESULTS

Synthesis of WFTA is being

implemented using XILINX ISE 9.1i tool.
After design entry and optional simulation, we
run synthesis. During this step, VHDL,
Verilog, or mixed language designs become
net list files that are accepted as input to the
implementation step.
Figure 2 Simulation result for N=5
4.1.2 IMPLEMENTATION

After synthesis, we run design

implementation, which converts the logical
design into a physical file format that can be
downloaded to the selected target device. From

UbiCC Journal - Volume 3 91

 Figure 5 Schematic of DFT N=5

Figure3 Schematic of DFT N=2

Figure 4 Schematic of WDFT N=2

Figure 6 Schematic of WDFT N=5

UbiCC Journal - Volume 3 92

 Values of Multipliers Multipliers only least number of multiplier are required.
The future scope of the work is to reduce the
N required in required in minimum multiplier so that, the algorithm
DFT WFTA occupy less space and consume less power.

Thus Winograd’s Fourier transform

2 8 0 algorithm for N=2, 3 and 5 have been realized
and the WFTA has been proved accurate by
3 8 2 comparing the values obtained using WFTA
with those of DFT using MATLAB.
Behavioral simulation of the WFTA was done
5 50 5
using XILINX ISE Simulator and the synthesis
was implemented in XILINX ISE tool. A full
Table 1 Comparison of DFT and WFTA RTL schematic is obtained with logic gate
level modeling. Finally, the circuit was
4.1.3 COMPARISON OF THE OUTPUT implemented using FPGA Spartan 3 kit.
VALUES OF DFT AND WDFT FOR N=2, 3
AND 5 The WFTA for the transform lengths
equal to powers of prime numbers, i.e. N=2m,
The output of normal DFT and WDFT 3m and 5m, can be obtained by iteratively using
coefficients are obtained by using MATLAB the same skeleton that we have used. Even
software, as A0, A1, A3, A4 and A5.It is though this project drastically reduces the
shown in Fig.7. number of multipliers required to compute the
DFT, it doesn’t eliminate its use. Hence, a new
algorithm can be devised based on theories of
WFTA to implement DFT with no multipliers
and reduced adders.

5. REFERENCES

[1] H. Krishna, B. Krishna, K. Y. Lin, J. D.

Sun, Computational Number Theory and
Digital Signal Processing (CRC Boca Raton,
Florida, pp 471-485,(1994).

[2].C. S. Burrus and P.W. Eschenbacher, “An

in-place in-order prime factor FFT algorithm”,
IEEE Trans. Acoust. Speech Signal
Processing, vol. ASSP-29, no. 4, pp. 806–817,
(1981).

[3].A. M. Despain, “Very fast Fourier

transform algorithms hardware for
implementation,” IEEE Trans. Computres.,
vol. C-28, no. 5, pp. 333–341, May (1979).

[4].Keshab Parhi, “VLSI Digital Signal

Processing Systems: Design and
implementation”Wiely., pp.237-244 (1999).

Figure 7 Comparisons of normal dft and wdft [5].M. D. Macleod and N. L. Bragg, “A fast
hardware implementation of the Winograd
4.2 CONCLUSION AND FUTURE Fourier transform algorithm,” Electron. Lett.,
ENHANCEMENTS vol. 19, pp. 363–365, May (1983).

Most of the present day cryptography [6].Uwe Meyer-Baese, “Digital signal

algorithms require complex multiplications processing with Field Programmable Gate
and more multipliers are required. With the use Arrays” Springer-verlag, pp 273-276, (2007)
of Winograd’s Fourier transform algorithm

UbiCC Journal - Volume 3 93

 [7].S. Winograd, “On computing the discrete University, MIT Chrompet campus, Chennai.
Fourier transform,” Math. Comp., vol. 32, no. His areas of interest include VLSI, Image
141, pp. 175–199, Jan. (1978). processing, Signal processing and network
[8]. J. McClellan and C. Rader, Number Security.
Theory in Digital Signal Processing.
Englewood Cliffs, NJ: Prentice Hall, pp 79-85
(1979).

[9].S. Winograd, Arithmetic complexity of

computations (Society for Industrial and
Applied Mathematics, (1980).

[10]. M. Heideman, Multiplicative complexity,

convolution, and the DFT Springer Verlag,
New York ( 1988)

[11]. J. Cooley, Some applications of

computational complexity Theory to Digital
Signal Processing. 1981 Joint Automatic
Contr. Conf. University of Virginia, June 17-
19 (1981).

[12]. R. L. Rivest, A. Shamir, and L. Adleman.

A Method for Obtaining Digital Signatures and
Public Key Cryptosystems. Communications
of the Association for Computing Machinery,
21(2) pp. 120–126, February (1978).

[13]. W. Diffie and M. E. Hellman. New

Directions in Cryptography. IEEE
Transactions on Information Theory, IT-22(6)
pp. 644–654, November (1976).

[14]. National Institute of Standards and

Technology (NIST). FIPS Publication 186:
Digital Signature Standard. National Institute
for Standards and Technology, Gaithersburg,
MD, USA, May (1994).

Sathishkumar.G.A obtained his M.E from PSG

college of Technology, Coimbatore, India. He
is currently perusing PhD from Anna
University, Chennai and Faculty member in
the Electronics and Communication Dept of
Sri Venakesateswara College of Engineering,
Sriperumbudur.His research interest is VLSI
Signal processing Algorithms, Image
Processing and Network Security.

Dr.K.Boopathy Bagan completed his doctoral

degree from IIT Madras. He is presently
working as professor, ECE dept, in Anna

UbiCC Journal - Volume 3 94

 A MULTIAGENT CONCEPTULIZATION FOR SUPPLY-CHAIN
MANAGEMENT

Vivek kumar , Amit Kumar Goel , Prof. S.Srinivisan

Department of computer science & Engineering Gurgaon Institute Technology Management, India
vivek.kumar1977@rediffmail.com, goelkamit@rediffmail.com

ABSTRACT

In Global world there is a huge network consisting by different companies for their suppliers, warehouses,
distribution centers, retailers, with the help of these entities any organization acquired raw material ,
transformed , and delivered finished goods. The major challenges for Industrial organization are to reduce
product development time, improve quality, and reduce cost of production. This is done only when the
relationship among various organization/industrial houses is good, This is not only be done by the change of
Industrial process/ method but the latest electronic tools controlled by computer software is required to establish
in this competitive world. Software agents consist of one or many responsibility of supply chain, and each agent
can interact with others without human intervention in planning and execution of their responsibilities. This
paper present solution for the construction, architecture, coordination and designing of agents. This paper
integrates bilateral negotiation, Order monitoring system and Production Planning and Scheduling multiagent
system.

KeyWords: Agent, Supply chain, Multiagent, Multiagent System Architecture for supply chain
management

1. INTRODUCTION

To improve the efficiency of supply chain
management it is mandatory to take intelligent,
tactical, strategic and good operational decision at
each end of chain. Under strategic decision the agent
will take the decision about suppliers, warehouses,
production units, transportation system etc. The
tactical decision takes place to meet actual demand.
The agent on operational level is responsible to
execute whole plan. To do all things in smooth way
the coordination among agents is must otherwise if
the material do not arrive on time the production will
stop, if finished good has been ready and warehouses
are not empty then it will create a great confusion.
The ability to manage all level of supply chain
system [1], coordination, accurate and timely
dissimilation of information is the enterprise goal.
2 Agent
In Software we can define the agent that it is an
entity which consists of various attributes defines a
particular domain. Exp: An agent deals with
warehousing consist its local attribute as well as the
details which will be coordinated with other entity
(Agents). So agents emulate the mental process or
simulate the rational behavior. A multi-agent system
is a loosely coupled network of problem-solver
entities that work together to find answers to
problems that are beyond the individual capabilities
or knowledge of each entity. The first issue is how
the different activities of supply chain can be
distributed in agents. A typical example of
multiagent system is taken with the help of Coffee
maker and toast maker. Let a person wants the toast
as the coffee is ready, means the coordination
between Coffee maker and toast maker is essential.
Otherwise many situation may be raised like Coffee
is ready but toast is not prepared and it comes after
some time or the toast is ready and the Coffee is not
prepared.
1.Agents are problem solvers.
2.Agents are pro-active.
3.Agents are goal-oriented.
4.Agents are context-ware.
5.Agents are autonomous
2.1 Requirement / Logistics agent
These agents coordinate all activities of plant and
find the various demands of various sections. It holds
the data of day to day production, find how much
material has been consumed a day depending on the
working hours a machine works. Categorized each
component in different table and coordinates with
other agent like Demand agent etc. The intelligent
part of the agent is to find the efficiency of machine,
minimizing cost increasing throughput etc. It can
also consist feedback of the finished goods and
suggest appropriate changes if required.

UbiCC Journal - Volume 3 95

 2.2 Demand Agent
This agent coordinates with other agent like
requirement/logistics agent. The main objective of
this agent is to fulfill the requirement of various
section of the company/customer. The intelligent part
of this agent is to acquire orders from various
vendors, compare them on the basis of quality, price,
availability etc. In case any demand increases or
decreases automatically vendor will be
communicated.
2.3 Transport agent
This agent is responsible for the availability of the
transport, dispatching of finished goods at particular
destination. It manages all the transportation routes.
2.4 Financial agent
This agent is responsible to avail the money for
purchasing any material. It coordinates with other
agents analyze the cost and ensure that the money
has been paid to the party in definite time.
2.5 Scheduling agent
This agent is responsible for scheduling and
rescheduling activities in the factory, exploring
hypothetical “what-if” scenarios for potential new
orders
Protégé-2000
Agent System
Type
Database
Behavior
Software
Developer
Fig 2: Architecture of Multiagent
UbiCC Journal - Volume 3
and generating schedules that are sent to the
dispatching agent for execution. It assigns resources
new orders and start times to activities that are
feasible while at the same time optimizing certain
criteria such as minimizing work in progress or
tardiness. It can generate a schedule from scratch or
repair an existing schedule that has violated some
constraints. In anticipation of domain uncertainties
like machine breakdowns or material unavailability,
the agent may reduce the precision of a schedule by
increasing the degrees of freedom in the schedule for
the dispatcher to work with. For example, it may
“temporally pad” a schedule by increasing an
activity’s duration or “resource pad” an operation by
either providing a choice of more than one resource
or increasing the capacity required so that more is
available.
3 MIDDLE AGENTS
3.1 Facilitators
Agents to which other agents surrender their
autonomy in exchange for the facilitator' s services.
Facilitators can coordinate agents'activities and can
satisfy requests on behalf of their subordinated
agents.
Mediator ATS
Agent
Data Mining
Module
XML
Application
Data
96
 3.2 Mediators

Agents that exploit encoded knowledge to create • A common format for the content of
services for a higher level of applications. communication

3.3 Brokers • A shared ontology

Agents that receive requests and perform actions

using services from other agents in conjunction with 4. AGENT COMMUNICATION LANGUAGE
their own resources.
There are two main approaches to design a agent
3.4 Helpline/Yellow pages communication language [6], The first approach is
procedural and the second one is declarative .In
Agents that assist service requesters to find service procedural communication is based on executable
provider agents based on advertised capabilities. content but in declarative communication is based on
definition assumptions and declarative
3.5 Agent Interaction statement .one of the more popular declarative agent
languages (KQML)[8]
Interaction is one of the important features of an
agent [2]. In other words, agent recurrently 5. MULTIAGENT SYSTEM
interaction to share information and to perform task ARCHITECTURE FOR SUPPLY CHAIN
to achieve their goal. Researchers investigating MANAGEMENT
agent’s communication languages mention three key
elements to achieve multiagent interaction. Our framework provides a GUI application that
[3][4][5].A common agent communication language enables the design of multiagent system with protégé
and protocol -2000[7], as well as single agents or multiagent
communities using common drag and drop operation.

Retailer
Retailer Agent

Data
mining
looks

Warehouse Logistics

Warehouse
Plant
Warehouse Plant
Plant

Operation
Purchase Resource
Management

Scheduling

Supplier
Raw material

Fig-1-Architecture of Multiagent Supply Chain Management System

UbiCC Journal - Volume 3 97

 6. FORMULATION OF BEHAVIOR TYPES
Behavior depends on the generic templates and work
flow i.e. receiving and sending the message. Execute
the stored application and gives necessary deriving
decision using inference engine.
There are four types of workflow terminals.
6.1 Add-on terminals
For the addition of predefined function.
6.2 Execute terminals
Execute the particular reasoning terminal.
6.3 Agent Types
After the formulation of behavior type we get a new
agent type in order to be used later in multiagent
system development i.e.
Agent Type = Agent + Behavior
6.4 Receiving terminals
For the filtration of receiving information.
6.5 Sending terminals
For the composition and then send further.
New agent can be created by existing one which will
be a template for creating agent instances during the
design of a multiagent system architecture.
6.6 Data base for agent
This unit acts as a storage facility to ensure inters
functionality between all system components. In this
system the database stores ontologies, behavior,
types of agent and the historical data to be mined.
This unit can be designed by RMI.
6.7 Agent training system (ATS)
This system gathers information from the data
mining procedure and then takes the decision and
sends this decision into the newly created agent.
6.8 Data Mining System
This system holds the implementation of data mining
algorithm executed by data mining procedures which
gives a new decision model which are again enabled
into agent via ATS. And also responsible for
UbiCC Journal - Volume 3
embedding specific knowledge into agents. This data
mining module receives the information from the
XML document and executes the suitable data
mining functions designed by the application
developer. These models represented in Predictive
Modeling Markup Language [8] which is a data
mining standard defined by DMG (Data Mining
Group) [9] provides the agent platform with
versatility and compatibility to other. Major data
mining software are Oracle SAS SPSS and MineIT
etc
7. CONCLUSIONS
Information technology based solution frameworks
offer a way to more effectively integrate decision-
making by enabling better knowledge sharing and
facilitating more transparent economic transactions.
The multi-agent system paradigm promises to be a
valuable software engineering abstraction for the
development of computer systems. In addition, the
wide adoption of the Internet as an open environment
and the increasing popularity of machine-
independent programming languages, such as Java,
make the widespread adoption of multi-agent
technology a feasible goal
REFERENCES
1. Zhou, L.; Xu, X.; Huang, T. and Deng,
S.: Enterprise Interoperability: New Challenges and
Approaches,
2. Nwana, H. S.: Software Agents: An Overview,
The Knowledge Engineering Review,
October/November 1996, Vol. 11, No. 3, PP. 205-
244.
3. Data Mining Group, the: Predictive Model
Markup Language Specifications (PMML)[1], ver.
2.0 available at: http://www.dmg.org
4. Bradshaw, J. M.; Dutfield, S.; Benoit, P. and
Woolley, J.D. KAoS: Toward An Industrial-Strength
Open Agent Architecture, Software Agents,
Bradshaw, J.M. (Ed.), Menlo Park, Calif., AAAI
Press, 1997, PP. 375-418.
5. Russell, S. J. and Norvik, P.: Artificial
Intelligence: A Modern Approach, Prentice Hall,
Englewood Cliffs, N.J., 1995.
98
 6. Genesereth, M.: An Agent-based Framework for
Interoperability, Software Agents, Bradshaw, J. M.
(Ed.), Menlo Park, Calif., AAAI Press, 1997, PP.
317-345.

7. Noy, N. F.; Sintek, M.; Decker S.; Crubezy, M.;

Fergerson, R. W. & Musen, M. A.: Creating
Semantic Web Contents with Protégé-2000, IEEE
Intelligent Systems Vol. 16, No. 2, 2001, PP. 60-71.

8. Finin, T., Labrou, Y. and Mayfield, J.: KQML as

an Agent Communication Language, Software
Agents, Bradshaw, J.M. (Ed.), Menlo Park, Calif.,
AAAI Press, 1997, PP. 291-316.

9. Huhns, M. N. and Singh, M. P.: Agents and Multi-

agent Systems: Themes, Approaches, and Challenges,
Readings in Agents, Huhns, M. N. and Singh, M. P.
(Eds.), San Francisco, Calif., Morgan Kaufmann
Publishers, 1998, PP. 1-23.

UbiCC Journal - Volume 3 99

SCALABLE ENERGY EFFICIENT AD-HOC ON DEMAND DISTANCE
VECTOR (SEE-AODV) ROUTING PROTOCOL IN WIRELESS MESH
NETWORKS

Sikander Singh
Research Scholar, Department of Computer Science & Engineering, Punjab Engineering College (PEC),
Deemed University, Sector-12, Chandigarh-160012 (India)
er_sikander@rediffmail.com,

Dr. Trilok Chand Aseri

Sr. Lecturer, Department of Computer Science & Engineering, Punjab Engineering College (PEC), Deemed
University, Sector-12, Chandigarh-160012 (India)
a_trilok_chand@yahoo.com

ABSTRACT
A new routing protocol called Scalable Energy Efficient Ad-hoc on Demand
Distance Vector (SEE-AODV) having good scalable properties and energy efficient
than existing Ad hoc on Demand Distance (AODV) routing protocol in wireless
mesh networks has been proposed in this paper. Wireless mesh networks (WMNs)
consist of mesh routers and mesh clients, where mesh routers have minimal mobility
and form the backbone of WMNs. They provide network access for both mesh and
conventional clients. Two techniques called Clustering and Blocking-Expanding
Ring Search has been applied on existing AODV routing protocol to improve its
scalability and energy efficiency problem. Results shows that, performance of SEE-
AODV routing protocol is better than existing AODV routing protocol in wireless
mesh networks. To show the efficiency of our proposed routing protocol,
simulations have been done by using Network Simulator-2 (NS-2).

Keywords: Wireless mesh networks, Ad-hoc network, Routing, Distance vector.

1 INTRODUCTION applications have different technical requirements

Wireless mesh network (WMN) [1] technologies
have been actively researched and developed as key
solutions to improve the performance and services of
wireless personal area networks (WPANs), wireless
local area networks (WLANs) and wireless
metropolitan area networks (WMANs) for a variety
of applications, such as voice, data and video.
Compared with mobile ad hoc networks (MANETs),
wireless sensor networks (WSNs) and infrastructure-
based mobile cellular networks, WMNs are (i) quasi-
static in network topology and architecture (ii) not
resource constrained at mesh routers and (iii) easy
and flexible to deploy. These technological
advantages are especially appealing to the emerging
market requirements on future wireless networks and
services, such as flexible network architecture, easy
deployment and self-configuration, low installation
and maintenance costs and interoperable with the
existing WPAN, WLAN and WMAN networks.
Potential applications of WMNs include broadband
home networking, community and neighborhood
networking, enterprise networking, building
automation and so on. These wide ranges of
and challenges in the design and deployment of mesh
networking architectures, algorithms and protocols.
The objective of this work is to develop routing
protocols for Wireless Mesh Networks and to
analyze their performance by realizing different
environments. The analysis has been done
theoretically and through simulations using NS-2
(Network Simulator-2). Objectives of the work are:-
1. To simulate the proposed routing protocol,
Scalable Energy Efficient-Ad-hoc on-demand
Distance Vector (SEE-AODV) for wireless mesh
networks.
2. Evaluation of routing protocols based on various
parameters.
3. Comparison of proposed protocol with existing
protocol.
2 RELATED WORK
Wireless mesh networks has recently gained a
lot of popularity due to their rapid deployment and
instant communication capabilities. These networks
comprise of somewhat static multi-radio Mesh
Routers [2], which essentially provide connectivity

UbiCC Journal - Volume 3 100

between the mobile single-radio Mesh Clients.
Special routing protocols are employed which
facilitate routing between the Mesh Routers as well
as between the Mesh Routers and the Mobile Clients.
AODV is a well known routing protocol that can
discover routes on-the-fly in a mobile environment.
However, as the protocol was actually developed for
single-radio nodes, it frequently lacks the ability to
exploit the potential offered by the Mesh Routers.
There are hundreds of proposed routing protocols [3],
many of them have been standardized by IETF and
have been in use for many years. Some of those
protocols have proven themselves in the Internet and
are expected to continue to perform well for many
years to come. In the ad-hoc networking arena,
several classes of routing protocols have been
proposed and carefully analyzed. The WMN
companies [4] are using a variety of routing
protocols to satisfy their needs. Furthermore
proposed routing protocol takes advantage of the
fixed nodes in WMNs. In this paper some
enhancements are done to improve the existing
AODV protocol so that it works well in wireless
mesh networks having good scalability and energy
efficient.
3 SCALABLE ENERGY EFFICIENT AD-
HOC ON DEMAND DISTANCE VECTOR
ROUTING PROTOCOL (SEE-AODV)
In this paper to develop SEE-AODV Routing
Protocol two techniques called Clustering and
Blocking Expanding Ring Search have been applied
to improve the performance of existing AODV
routing protocol in wireless mesh networks. The
performance of wireless mesh networks is highly
dependent on routing protocol. AODV is a popular
routing protocol for wireless networks. AODV is
well suited for wireless mesh networks in that it has
low processing & memory overhead and low
network utilization. Additionally AODV provides
loop freedom for all routes through the use of
sequence numbers.
3.1 Design Goals
The design goal of SEE-AODV routing protocol
is to improve the scaling potential of AODV and to
make it energy efficient. The main feature of AODV-
Clustering includes:-
(i) Gradualness
The protocol first works on AODV method, then
gradually changes to a clustering route protocol.
There are several considerations about it: First, there
is central control node in mesh network; second,
using this method can also allow AODV nodes
coexisting with AODV-Clustering nodes; third, it
can reduce the overhead caused by frequently
changing of cluster.
(ii) Coexist with AODV
UbiCC Journal - Volume 3
AODV is a widely accepted routing protocol.
There are several implemented version of AODV
reported. One of the important principles in
designing AODV-Clustering is the coexistence with
AODV; it means nodes which implement AODV
protocol can work collaboratively with nodes that
implement proposed protocol in the same network.
To achieve this, keep all AODV route control
packets and add some new control packets as needed.
In fact AODV-Clustering let all nodes that haven’t
joined a cluster works on AODV method.
(iii) Route Discovery Mechanism
In AODV-Clustering, there are two route
discovery mechanisms for the nodes which join the
cluster. One is Blocking-ERS, it can increase the
efficiency of the route discovery; the other is
traditional RREQ flooding route discovery
mechanism which extends from AODV. Normally
Blocking-ERS is used first, if a suitable route can’t
be found before timeout, then traditional route
discovery mechanism is used.
(iv) Local Route Repair
To reduce the number of lost data packets, when
the route broke due to mobility, AODV-Clustering
let the node upstream of the break link perform a
local repair instead of issuing the RERR. If the local
repair is successful, a RREP will be returned either
by the destination or by a node with a valid route to
the destination. In the event that a RREP is not
returned, the local repair is not successful and a
RERR message is sent to the source node.
3.2 AODV-Clustering Routing Scheme
(i) Route Discovery
At the beginning the status of all nodes is
“unassigned”. Source node broadcast [5] RREQ
message to find a route, destination node or
intermediate node that has fresh route to destination
reply with RREP message to source. On the way
which RREP pass by, one or several nodes are
selected as cluster head (CH) using a defined rule.
CH node will broadcast CH message to its
neighbours, the neighbour nodes which receive this
broadcast will act differently according to its roles.
(a) For node whose status is “unassigned”, it will
issue a Join Cluster Request to the broadcasting CH
and can become an ordinary cluster member after
receiving the acknowledgment from this CH. In
AODV-Clustering, CH node can reach all his
members by one hop, so the protocol’s architecture is
one level hierarchy.
(b) For node which is an ordinary cluster member, it
will judge if the broadcast message sender is its
original CH, if yes, no action needed; else, send Join
Cluster Request to the CH and become gateway node
after receiving acknowledgement from it. Being a
gateway, it will send Gateway Message to all CH
nodes which it connect directly, let them put its
address in their Gateway Table.
101
(c) For node which is a gateway, it will check its
Gateway Node Table whether there is an entry for
this broadcasting CH node, if yes, no action needed;
else, send Join Cluster Request to it and put its
address into Gateway Node Table after receiving
acknowledgement from it. Gateway nodes have two
tables, one is Gateway Node Table which contains
cluster head address which it connect directly,
another is Joint Gateway Table, which contains
address of the CH nodes which can be reached by 2
hops and also the nodes which help to reach these
CHs which are called Joint Gateway.
(ii) Route Maintenance
AODV-Clustering extends many features from
AODV for route maintenance, for example, use
“hello” message to confirm the existence of
neighbours, use RERR message to inform the nodes
affected by the link breakage. In addition to it,
AODV-Clustering adds some cluster related
maintenance operations such as joining cluster,
leaving cluster and changing its status.
(a) Joining Cluster
The nodes whose status is “unassigned” will join
the cluster after receiving the CH Broadcasting
Message. To reduce overhead, no periodic CH
Broadcasting is used. Instead, a “on demand” method
is used for new nodes to join cluster: when the node
whose status is “unassigned” broadcast RREQ to its
neighbours, a specific mark is set in RREQ, the mark
will inform neighbouring CHs to let him join the
cluster. Using this way, unassigned node that newly
comes will have a chance to join the nearby cluster
when it has data to send.
(b) Leaving Cluster
If the ordinary cluster member finds its CH node
unreachable, it will change its status to unassigned.
This means the node leaving the cluster. CH node
should change its status to unassigned when it has no
cluster members.
(c) Changing Cluster
In AODV-Clustering, the change of cluster is not
occurred immediately. When the node leave the
cluster, its status become “unassigned”, it works on
AODV method until new CH node appear nearby
and it has a chance to join the cluster.
3.3 Route Discovery Approaches in Wireless
Mesh Networks
In Wireless mesh networks, nodes cooperating
for delivery of a successful packet form a
communication channel consisting of a source, a
destination and possibly a number of intermediate
nodes with fixed base station. In this paper some
inefficient elements have been found in well known
reactive protocols AODV and propose a new
approach for rebroadcasting in Expanding Ring
Search. This leads to the Blocking-ERS scheme, as
we call it, which demonstrates improvement in
energy efficiency at the expense of route discovery
UbiCC Journal - Volume 3
time in comparison to conventional route search
method.
3.3.1 Blocking-ERS
An alternative ERS scheme has been proposed to
support reactive protocols such as DSR (Distance
Source Routing) and AODV and it is called Blocking
Expanding Ring Search. The Blocking-ERS
integrates, instead of TTL sequences, a newly
adopted control packet, stop instruction and a hop
number (H) to reduce the energy consumption during
route discovery stage. The basic route discovery
structure of Blocking-ERS is similar to that of
conventional TTL sequence-based ERS. One of the
differences from TTL sequence-based ERS is that
the Blocking-ERS does not resume its route search
procedure from the source node every time when a
rebroadcast is required. The rebroadcast can be
initialized by any appropriate intermediate nodes. An
intermediate node that performs a rebroadcast on
behalf of the source node acts as a relay or an agent
node. Fig. 1 shows an example of Blocking ERS
approach in which the rebroadcasts are initialized by
and begins from a relay node M in rebroadcast round
2 and another relay node N in round 3 and so on. In
Fig. 1 the source node broadcasts a RREQ including
a hop number (H) with an initial value of 1. Suppose
that a neighbour M receives the RREQ with H=1 and
the first ring was made. If no route node is found,
that is, no node has the requested route information
to the destination node, the nodes in the first ring
rebroadcast the RREQ with an increased hop number,
for example RREQ with H=2 is rebroadcast in this
case The ring is expanded once again just like the
normal expanding ring search in AODV except with
an extended waiting time.
The waiting time can be defined as
Waiting Time = 2 × Hop Number
The nodes in Blocking-ERS receiving RREQs
need to wait for a period of 2H, i.e. 2×their hop-
number unit time before they decide to rebroadcast,
where the ‘unit time’ is the amount of time taken for
a packet to be delivered from one node to one-hop
neighbouring node.
Figure 1: Blocking-ERS
(A) Energy Consumption
102
 Energy consumption during the transmission of
RREQs can be saved by using the Blocking-ERS
scheme. Let the amount of energy consumption on
each node for one broadcast is the same unit energy
consumption, denoted by UnitEnergy. Assume that
each action of broadcasting a RREP, RREQ or ‘stop
instruction’ consumes the same amount of 1
UnitEnergy. This can be easily shown by the
difference of the energy consumption between the
conventional TTL sequence-based ERS and the
Blocking-ERS scheme.
(i) One route case: - First consider only the energy
consumption along the route from the source to the
route node. The energy consumption for the TTL
based ERS and for the Blocking-ERS can be
described by the following formula (Eq. (1) and Eq.
(2)) respectively, where Hr is the hop number of the
route node.
Hr
ETTL − ERS = H r + ∑ i (UnitEnergy) (1)
i =1
E Blocking − ERS = 3H r (UnitEnergy) (2)
The difference of the amount of energy
consumption is more visible from Fig. 2 where the
amounts of energy consumption by the two ERS
approaches are plotted against the number of rings.
The Blocking-ERS curve is below the TTL based
sending the ‘stop instruction’. For the conventional
TTL based ERS, the energy consumption during the
route discovery process includes that in two stages:
(a) searching for the route node and (b) return RREP.
The energy consumed for “(b) returning a RREP” is
Hr UnitEnergy for both routing schemes and Hr
UnitEnergy is consumed for the Blocking-ERS stage
‘(c) sending the stop instruction’. In the stage of ‘(a)
searching for the route node’, the energy
consumption is different for the two methods. Each
ring contains different number of nodes that
rebroadcast to form the next ring. Let ni be the
number of nodes in ring i and the hop number of the
route node be Hr.
In the Blocking-ERS, the energy consumed in
each ring is as below: -
Ring i Energy Consumed
0 1
1 n1
Hr-1 nhr-1
In the TTL based ERS, the energy consumed in
each ring is as follows: -
Ring i Energy Consumed
0 1
1 1 + n1
Hr-1 1 + n1 + n2 + · · · + nhr-1
Therefore, the total energy consumption by the
Blocking- ERS is given by Eq. (3)
Hr

ERS curve after ring 3. As it is clear from Fig. 2, the E Blocking − ERS = 2(1 + ∑ ni ) + E RREP (UnitEnergy ) (3)
i =1
difference of the amount of energy consumption
between these two mechanisms becomes larger as
the distance increases between the source and the Similarly, the total energy consumption by the
route node. conventional TTL sequence based ERS is given by
Eq. (4)

Hr i
E TTL − ERS = H r + ∑ ∑ n j + E RREP (UnitEnergy ) (4)
i −1 j =1

The difference between the E Blocking − ERS and

ETTl − ERS is given by Eq. (5)

H r −1 i
E Saved = H r − 2 + ∑ ((∑ n j ) − 2 ni )(UnitEnergy) (5)
i =1 j =1

Clearly, when ni = 1, for i = 1, · · · , Hr . The

Figure 2: Comparison of energy consumption for one
route (Energy is measured in Joule, and Distance in
number of hops)
(ii) General case: - Now consider the general case.
For the Blocking-ERS, the energy consumption
during the route discovery process can be considered
as the total energy consumption in three stages: (a)
searching for the route node (b) return RREP and (c)
above formulas represent the energy consumption for
a single route. This indicates that the energy
consumption saving achieved by the Blocking-ERS
for a single route is the minimum amount of energy
saving.
(B) Time Delay
Consider the time delay for the route discovery
period, during which from the RREQ is broadcasted
for the first time, transmitted from the source node to
the route node possibly via flooding. That is the total

UbiCC Journal - Volume 3 103

time taken from when the source node broadcasts the
first RREQ for the first time until after a Route Node
is found and the source node receive the RREP from
the Route Node. Let the UnitTime be the one-hop
transmission time, which is the time taken for a
RREQ from a broadcasting node to one of its
neighbour nodes. In case of TTL sequence based
ERS suppose H = 3, that is the route node is 3 hops
distant from the source node. The total time includes
the time when TTL = 1, 2 and 3. The final TTL
number equals to the hop number of the route node.
This gives the following formula of total time delay
for the TTL sequence-based ERS (Eq. 6): -
Hr
TTTL − ERS = 2∑ i (UnitTime) (6)
i =1
Now consider the time delay in the Blocking-
ERS. The total time includes the time for three
stages: (a) searching for the route node, (b) returning
the RREP and (c) broadcasting the ‘stop instruction’.
For stage (a), the time consists of the time to for
broadcasting and the waiting time. The broadcasting
time for 1 hop distance is 1 UnitTime. The waiting
time depends on the hop number of the node. The
route node is 3 hops distant from the source node.
Each node needs to wait for 2H before
rebroadcasting. At ring 1, the node waits for 2 × 1 =
2 UnitTime, and at ring 2, the node waits for 2 × 2 =
4 UnitTime, and at ring 3, the node waits for 2 × 3 =
6, so the total waiting time for the ‘(a) stage of
searching for the route node’ is 2 + 4 + 6 = 12, and
the total time for stage (a) is 12+Hr = 12+3 = 15. The
time for stage (b) and (c) are Hr and this gives 2Hr =
2×3 = 6. Therefore, the total time for the route
discovery and flooding control is 15 + 6 = 21
UnitTime. Mathematical formula is presented below,
where H r represents the hop number of a route node.
Figure 3: Comparison of the time delay (Time is
measured in millisec, and Distance in number of
hops)
The formula of the time delay in the Blocking-
ERS is given by Eq. (7): -
UbiCC Journal - Volume 3
Hr
TBlocking − ERS = 3H r + 2∑ i (UnitTime ) (7)
i =1
Compare this to the TTL sequence based ERS as
given in Eq. (8): -
Hr
TTTL − ERS = 2∑ i (UnitTime ) (8)
i =1
It is clear that the difference between the two is
3Hr three times of the hop serial number of the route
node, depending only on the distance between the
source node to the Route node. The time delay in
both approaches is compared in Fig. 3 As illustrated
the Blocking-ERS takes slightly more time than the
conventional TTL sequence-based ERS for the route
discovery process.
4 PERFORMANCE EVALUATION
To evaluate the performance of SEE-AODV
routing protocol (AODV-Clustering with Blocking-
ERS technique), the simulation is done by using NS-
2 [6]. The performance of SEE-AODV routing
protocol is evaluated by comparing it with existing
AODV routing protocol in the same conditions.
There is evaluation of three performance metrics: -
(i) Packet Delivery Fraction: - This is the fraction of
the data packets generated by the sources that are
delivered successfully to the destination. This
evaluates the ability of the protocol to discover
routes.
(ii) Routing Load: - This is a ratio of control packet
overhead to data packet overhead, measured by the
number of route control packets sent per data packet.
The transmission at each hop along the route was
counted as one transmission in the calculation.
(iii) Average Route Acquisition Latency: - This is
the average delay between the sending of a route
request packet by a source for discovering a route to
a destination and the receipt of the first
corresponding route reply. If there is a fresh route
already, 0 was used for calculating the latency.
Figure 4: AODV, SEE-AODV packet delivery
104
fraction
As shown in Fig. 4 the packet delivery fraction
obtained using SEE-AODV is almost identical to that
obtained using AODV when node numbers are small.
However, when there are larger numbers of nodes
(i.e., more than 200), SEE-AODV performed better.
This suggests that SEE-AODV is highly effective in
discovering and maintaining routes for delivery of
data packets.
Figure 5: AODV, SEE-AODV Routing Load
Fig. 5 shows the routing load comparison of
these two protocols. Routing load is measured by
numbers of route control packets sent per data
packets. For AODV, route control packets include
RREQ, RREP, RERR and Hello messages, for SEE-
AODV cluster related messages such as CH
Broadcast Message, Join Cluster Request, Join
Cluster ACK, and Gateway message are also
included. It is clear from Fig. 5, routing load of SEE-
AODV was significantly lower than AODV when
there are large numbers of nodes in network. So
SEE-AODV became a hierarchical protocol
gradually, hierarchical routing greatly increases the
scalability of routing in wireless networks by
increasing the robustness of routes.
Figure 6: AODV, SEE-AODV Average Route
Acquisition Latency
Fig. 6 shows the Average Route Acquisition
Latency of the two protocols. SEE- AODV
UbiCC Journal - Volume 3
performed better than AODV in more “stressful”
situations (i.e. larger number of nodes, more load),
this is greatly contributed to the Blocking-ERS
technique of AODV-Clustering and reduction of the
RREQ flooding.
5 CONCLUSION AND FUTURE WORK
In this paper Scalable Energy Efficient AODV
(SEE-AODV) routing protocol has been introduced
to solve the scalability problem of AODV by
applying Clustering and to make it energy efficient
by using Blocking-ERS techniques in wireless mesh
networks. The performance is studied by simulations
based on NS-2. The result shows, that SEE-AODV
protocol achieves better scalability than existing
AODV while keeping the merits of it. The analysis
demonstrates a substantial improvement in energy
consumption that can be achieved by the Blocking-
ERS at a margin cost of slightly longer time.
There are some limitations still in SEE-AODV
routing protocol, such as during discovery of a route
the technique used in this protocol slightly takes
more time than existing conventional one, so it need
further study to improve this drawback.
6 REFERENCES
[1] K. N. Ramachandran: On the Design and
Implementation of Infrastructure
Mesh Networks, IEEE Wksp. Wireless Mesh
Networks, Calcutta, pp. 78-95 (Sept 2005).
[2] R. Draves, J. Padhye, and B. Zill: Routing in
Multi-Radio, Multi-Hop Wireless Mesh
Networks, Mobile Communication, pp. 114–28
(Sept 2004).
[3] R. Draves, J. Padhye, and B. Zill: Comparison
of routing metrics for static multi-hop wireless
networks, Proc. of SIGCOMM’04, (Portland,
OR), pp. 12-24 (Aug 2004).
[4] I. F. Akyildiz and X. Wang: A Survey on
Wireless Mesh Networks, IEEE Commun.
Mag., vol. 43, no. 9, pp. S23–S30 (Sept 2005).
[5] V. Li, H. S. Park, and H. Oh: A Cluster-Label-
Based Mechanism for Backbone on Mobile Ad
Hoc Networks, The 4th Wired/Wireless Internet
Communications (WWIC 2006), pp.26-36 (May
2006).
[6] http://www.isi.edu/nsnam/ns/nsdocumentation.
105