ANALYSIS AND COMPARISON OF MULTI - AGENT DATA MINING TOOLS

*1
Mr. Sivabalan M, *2 Dr. Arulanandam K.,
*1
Guest Lecturer, PG and Research Department of Computer Science, Government Thirumagal Mills
College, Gudiyattam, Tamilnadu, India.
*2
Assistant Professor & Head of Department, PG and Research Department of Computer Science,
Government Thirumagal Mills College, Gudiyattam, Tamilnadu, India
---------------------------------------------------------------------***---------------------------------------------------------
Abstract: INTRODUCTION

The main objectives of this research is to
conduct a comparative study of the architectural
designs of Multi-Agent development toolkits,
studying the individual methodology used in the
designing the architectures within Multi-Agent
development environments. The study aims to
research widely into the various design
methodology used in implementing various Multi-
Agent design environments, analysis different
aspects relating to the components of the
development tools.The FIPA (Foundation for
Intelligent Physical Agents) a part of the analysis
and evaluation of the specification providing the
necessary requirements for agent systems.
This analysis is to compare and contrast
each agent against and other properties. To make an
effective comparison as well as evaluate individual
agent tools using Multi-Agent development tool to
compare their performance and a high level
understanding of a particular agent–oriented
development methodologies put it use. The main
objective can allow in the analysis agent
architecture entirely. The analysis of Multi-Agent
development tools aims to give insight into the
many theoretical aspects of software engineering
and the drive for greater understanding of software
engineer methodology as well as the concept
behind agent design for an agent development
environment such as Multi-Agent systems.
Keywords: Distributed Data Mining, Agent
mining, Multi Agent System (MAS),Weka,
Knime, Tanagra.
Multi Agent Systems (MAS) often deal with
complex Applications that require distributed
problem solving. In many applications the
individual and collective behavior of the agents
depends on the observed data from distributed
sources. In a typical distributed environment
analyzing distributed data is a non-trivial problem
because of many constraints such as limited
bandwidth (e.g. wireless networks), privacy-
sensitive data, distributed compute nodes, only to
mention a few. The field of Distributed Data Mining
(DDM) deals with these challenges in analyzing
distributed data and offers many algorithmic
solutions to perform different data analysis and
mining operations in a fundamentally distributed
manner that pays careful attention to the resource
constraints. Since multi-agent systems are also
distributed systems, combining DDM with MAS
for data intensive applications is appealing.
The aim is to understand the paradigm of
agent oriented development tools, studying the
methodology used in Multi-Agent design models
and the applications developed. This investigation
will be carried out in the form of a comparative
study that will include agent development tools to
understand each agent by carrying out a detailed
analysis of each Multi-Agent System (MAS)
building tools.
The objective of this study is widely to fill in the
gap in knowledge as to agent-oriented
developments tools and the methodology each has
chosen to use individually then carrying out a
comparison on many levels of its abstraction,
making an evaluation from a comparison
collectively reviewing the agent development
toolkit thoroughly as well as effectiveness.
Existing Agent Simulation Tools
In this section we would like to discuss the popular
agent simulation tools like Aglets, JATLite, FTP
Software Agent and Voyager. Aglets are Java-based
autonomous agents developed by IBM, which
provide the basic capabilities required for mobility
and has a globally unique name. A travel itinerary is
used to specify the destinations to which the agent
must travel and what actions it must take at each
location. In order for an aglet to run on a particular
system, the target system must be running an aglet
host application which provides a platform-neutral
execution environment for the aglet. The aglet
workbench includes a configurable Java security
manager. Aglets can communicate using a
whiteboard that allows agents to collaborate and
share information asynchronously. Synchronous
and asynchronous message passing is also
supported for aglet communication. Aglets are
streamed using standard Java serialization or
externalization. A network agent class loader is
supplied which allows an aglet’s byte code stream
and state to travel across a network [13]. Java Agent
Template Lite (JATLite) is a set of light-weight
Java packages being developed at Stanford
University that can be used to build multiagent
systems. It is a layered architecture which provides
a different communication protocol at each layer.
The JATLite framework is intended for developing
typed-message, autonomous agents that
communicate using a peer-to-peer protocol. Both
synchronous and asynchronous message passing are
supported. Messages can be delivered through
polling or message queuing. The framework
provides additional security which checks the agent
name and password for a more secure connection
[14]. FTP Software Agent Technology is Java-based
software designed to manage heterogeneous
networks across the Internet using agent
technology. The agents are autonomous and mobile,
and can move to any system in the network which
has an Agent Responder installed. As the agent
moves from system to system, its tasks may change,
depending on the environment of the system it is
visiting. The agents can interact with other agents
or with the user, as needed [15]. But FTP agents do
not require any user interaction—based on push
technology, they can move from system to system,
respond to events, and perform tasks according to
criteria predefined by the user. An Agent Manager
is responsible for launching the agent. Voyager,
from Object Space, Inc., is an agent-enhanced
Object Request Broker (ORB) coded in Java. An
ORB provides the capability to create objects on a
remote system and invoke methods on those objects
[16]. Voyager augments the traditional ORB with
agent capabilities. Voyager agents have mobility
and autonomy which is provided in the base class,
Agent. An Agent can move itself from one location
to another and can leave behind a forwarding
address with a secretary so that future message scan
be forwarded to its new location. Specialized
agents, called Messengers, are used to deliver
messages.
Need of Agents:
In Data mining perspective, an agent can be defined
as an object whose behavior is described by a
"script", with its own calculation ways, and can
move from place to place to communicate with
other agents. With its "script", the agent is able to
follow a life behavior that will be instilled at the
time of implementation and that will allow him to
have as main feature to be fully autonomous. One
of the discriminate characteristics of the agents is
the representation and reasoning on the
environment (the external world and other agents),
based on this feature, we find two different classes,
which are Cognitive agents, Reactive agents.
A cognitive agent is an agent that has an explicit
representation of its purpose and its environment.
The actions it performs to achieve its goal are the
result from a reasoning on the state of the
environment. Usually a cognitive system includes a
small number of agents; each is similar to a more or
less complex expert system. In this case we speak
of high granularity agent.
Challenges in Distributed Data mining and 3.1.1Weka
Mining Agents:
Weka is the tool most commonly used due to its
Agents can enhance data mining through involving vast functionality and supported features. This java
agent intelligence in data mining systems, while an based data mining tool provides user with both GUI
agent system can benefit from data mining via and simple CLI for performing and managing tasks
extending agents’ knowledge discovery capability to be performed. It supports all data mining tasks
[12]. Nevertheless, the agent mining interaction from preprocessing, classification, and clustering to
symbiosis cannot be established if mutual issues are visualization and feature selection.[9]
not solved [13]. These mutual issues involve
fundamental challenges hidden on both sides and
particularly within the interaction and integration.
Issues in agent-mining interaction highlighting the
existence of mutual issues. Mutual issues
constraining agent-mining interaction and
integration consist of many aspects such as
architecture and infrastructure, constraint and
environment, domain intelligence, human
intelligence, knowledge engineering and
management, and nonfunctional requirements.
Figure 1: Weka GUI
Architecture and infrastructure Data mining always
3.1.2 Knime
faces a problem in how to implement a system that
can support those brilliant functions and algorithms KNIME (Konstanz Information Miner) is an open
studied in academia. API workflow based data mining tool that provides
easy accessibility to new nodes to be added into the
Nonfunctional requirements Nonfunctional requests
workflow. It provides its user with the GUI which
are essential in real-world mining and agent
aid with the simplification of workflow generation
systems. The agent-mining simians may more or
by the user. It also provides with features to modify
less address nonfunctional requirements such as
a particular node accordingly a nd execution of
efficiency, effectiveness, action ability, and user and
partial data flow[10].
business friendliness.
3.1.3 Tanagra
Constraint and environment both agent and mining
systems need to interact with the environment, and This extension of SIPINA provides the users an
tackle the constraints surrounding a system [15]. In easy to use interface for the analysis of either real
agent communities, environment could present or artificial data. It allows the researchers to easily
characters such as openness, accessibility, add their own data mining research methodology or
uncertainty, diversity, temporality, spatiality, and/or any newly identified data mining processing
evolutionary and dynamic processes. These factors technique and also supports byproviding them with
form varying constraints on agents and agent architecture and a means to compare their
systems. methodology performances. It provides the
beginners or naives with a platform where they can
Human intelligence both agent and mining need to
carry out their experimental procedures[6].
consider the roles and components of human
intelligence. Many roles may be better played by
humans in agent-mining interaction.

Data Mining Tools


Figure 2: Tanagra explorer for result
visualization
3.2 Experimental Analysis
The performance of these tools has been analyzed
by first running them with several datasets available
on UCI repository. Several different algorithms for
classification and clustering were implemented in
this analysis and performance of these algorithms
was observed. In this section a sample of the
experiment performed during the research is
presented and conclusion of the results of different
tools is discussed
Data Set:
Data set of (1)Audiology isused with data type
multivariate, attribute type categorical, number of
attributes 69, number of instances number of
attributes, number of instances 226.(2) Zoo: data
type: multivariate, attribute type: categorical,
integer, number of attributes: 18, number of
instances: 101[13].
Preliminaries:
The classification was carried out of data set with
percentage split methodology of 60% for training
data and remaining 40% for the test data. The
obtained measure of accuracy is used as the
criterion for the performance analysis of the tools.
Experiment:
The k nearest neighbor algorithm did not run on the
dataset, the possible reason being the incompatible
data type as the data was of discrete values. The
accuracy measure of Weka with OneR is 42.85%, to
38.23%, for ZeroR it is 27.27% to 38.23%, for k-
nearest neighbor it is 58.44% to 94.11%, for C4.5 it
is 86.12% to 92.13%.The KNIME does not provide
any implementation for the ZeroR and OneR
algorithms which proved to be a shortcoming in the
accuracy of the tool. K-nearest neighbor is not
compatible with the data type for Audiology but
produced an accuracy of 41.54% for instances in
Zoo data set. C4.5 achieved the accuracy between
62.55% and 89.36%.Tanagra, similar to KNIME,
does not provide with the implementation of ZeroR
and OneR algorithm. K-nearest neighbor algorithm,
like the case with other tools, does not conclude
over the discrete values but runs well over the Zoo
data set values. K nearest neighbor algorithm
attains an accuracy percentage of 75.32% with
percentage split. C4.5 obtains result between
74.38%and 80.51%.
Table 1: Accuracy for classification over
audiology data set.
Table 2: Accuracy achieved over Zoo data set.
it has been observed that Weka has successfully run
and implemented all the algorithms and produced
appropriate results for the algorithms but with
lowest accuracy that of ZeroR. Though ZeroR and
OneR did not provide result, KNIME produced an
accuracy of 89.36% with C4.5 over the zoo data.set
which is very close then 92.13% of that confirmed
by C4.5 in Weka. Beside the non availability of
ZeroR and OneR algorithm implementation,
Tanagra’s accuracy with C4.5 is satisfactory with
result between 74.38% and 80.51% which is more
stable then compared to 62.55% to 89.36% of what
it is with KNIME.
3.3Features and Functionalities:
3.3.1Weka:
Database system support: Weka supports reading
of files from several different data bases. It also
provide feature to import the data over internet,
from web pages or from a remotely located SQL
database server by entering merely the URL of
resource. This hence allows Weka to support variety
of different data formats[9].
Graphical representation:Weka provides limited
support to the visualization of concluded data.
Although the Weka API provides with various data
mining and processing methods, it lacks in the
representation of the result of processing. The
representation of the results, graphs and plots, lacks
in detailed representation. But that really is not of
much concern as Weka provides appreciable
support for other functionalities and also the
visualization provided is sufficient to represent the
view of data on which the analyses has been
performed and the results of the data analyses and
preprocessing. Weka also allow for available add-
on functions to be included in order to be able to
interface with R statistical package with improved
statistical analysis and representation of result[4].
Analysis and Processing Capabilities: It allows
users to use R application as can directly interact
with R package. Weka also provides with separate
GUI for knowledge flow, for Experimenter, used to
compare various results and explorer to analyze
different data sets. It also provides for creating own
filters for filtering out instances[4].
Issues:However, Weka is not better suitable option
for the large data sets as theyare roughly handled.
The databases with large unstructured data are not
suitable as it hinders the pre-processing and
computing time of Weka. It can perform well and
provide more accurate results with smaller
databases. This tool has limited ability to partition
dataset to training and test sets. It does facilitate to
save parameters for future application. For cross
validation or independent validation, Weka lacks to
allow saving model in order to avoid rebuilding for
other datasets[1].
3.3.2Knime:
Database system support: KNIME provides with
a great strength being able to establish database
connections with any number of databases that
provides JDBC. This tool also provides with a
unique functionality of ports to different data
sources and databases. With this there is no need to
modify the SQL query. Users can use these ports in
order to integrate from several databases and
modify the dimension of the database accordingly
[4].
Graphical representation:Workbench in KNIME
provides user with an easy way to handle different
functions and data flow of the process by merely
dragging and dropping new nodes which can then
be customized as required. It also provides with its
up gradation to improve the workbench
functionality.
KNIME can also be implemented deploying further
improvement with individual efforts over a
particular function and over the data so as to be able
to represent variety of data types and their
properties [4].Its applicability varies. from
execution of basic processes to the integration with
other software for enhanced visualization and
performance. However, the manual effort required
with KNIME makes it a less suitable option for
large complex workflows.
Analysis and Processing Capabilities: KNIME
provides all the supported operation for data mining
with a graphical interface where the naïve user can
import, filter and generate workflow without any
need to formulate SQL queries and code [10].On
the other hand KNIME allows experts to write their
own programming script for any new node or to
download extensions from other users to attain
more specific functionality.
Issues: KNIME also has limited partitioning ability
like Weka but does provide the functionality to save
parameters and validation model for cross
validation and independent validation[1].
3.3.3Tanagra:
Database system support: Importing of different
data sources is not supported in Tanagra because, as
it provides only tree representation, so it would
disturb the tree nature of the stream diagram. Even
it can also not read from any other data source
format or integrated database directly[12].
Graphical representation: The visualization of
various models is although not like other tools but it
includes several statistical measures. Tanagra is
based on stream diagram paradigm where a user
defines data sources, data operation and a linking
path through the sources and operations. This path
describes the flow similar to workflow. But Tanagra
allows the graph to be represented only as a tree so
as to maintain the simplicity. However, this creates
a bottleneck with the performance of Tanagra as
now there can be only one data source to any
operation [4].Tanagra provides with different ways
of visualization of data as in scatter plots, graphs
and tables. However, it is difficult to develop
graphs by specifying several parameter values as
some may not be applied to a given data set but
should be applied to the sub-node of the data
indirectly. This leads to the missing values for some
views. Instead the visualization of data result is
more useful as high dimensional data can be
represented in 2D.The representation of clustering
and classification result is in text[12].
Analysis capabilities: Tanagra can import text files
with whitespace delimited fields. It also provides
with a conversion tool for files it cannot read from
arff. This tool provides robust statistical analysis
functionality by complementing with a range of
uni-and multivariate parametric and nonparametric
tests. It also includes correspondence analysis,
principal component analysis, and the partial least
squares methods.[11]
Issues: Tanagra cannot implement association rule
mining for large datasets. It cannot integrate
databases. It is required to integrate different
datasets outside Tanagra and then import before
performing any operations. Out of the available
operator, not all are applicable to every node. Some
of these often fail when applied to inappropriate
location and that too without proper display of any
error message. Tanagra has limited partitioning
ability and does not provide to save parameter
values. It also does not support for saving
independent validation model [12].
4.EVALUATION OF COMPARATIVE STUDY
There are conclusions drawn from the study of
these tools. The analysis of these tools has provided
us with idea for the betterment of the whole data
mining procedure. Even though these tools have
proved to be appropriate for the specific data
domains and specific data mining tasks such as
classification, clustering, etc, the shortcomings,
flaws or specificity of these tools have acted as a
pullback from the implementation of a general
framework for data mining process. The major
common drawback with theses tools is that their
processing of data, classification, clustering,
prediction and inferring of rules all is based on the
selection of the algorithm for data mining over a
particular type of data set.If the selection of
algorithm is not appropriate regarding the domain
of data then the produced patterns or predictions
cannot be completely relied. For example, the
classification over Audiology data set results with
84.41% correctly classified instances with Simple
Logistic while with ZeroR it results in only 27.27%.
If an inappropriate algorithm is used for future data
value prediction then it would produce incorrect
results. Another issue with these tools is that the
current state of art does not provide an automated
mining technique. All the tasks such as
classification and clustering are performed
consecutively and are specific to an application[5].
A theoretical framework is required for
implementation of unified theory where these data
mining tasks can be unified and overcome with the
shortcomings of these tools.
Multi agent systems are those systems which break
a domain in components and each component
performs its part of processing to achieve a
common goal[5]. These agents are adoptive to new
environment, autonomous, flexible, and easily
modifiable and pertains some characteristics
relating to artificial intelligence. These intelligent
agents handle several independent tasks. Such multi
agent system is modular and provides robustness to
the system. A MAS based data mining tool would
provide with:
 Most suitable processing technique
 Preprocessing of new data and updating in
the database.
 A dataset of patterns obtained every time
after any new data addition and suggest with
possible knowledge that can be discovered.
 Reduced processing time for pattern
recognition and representation of analyzed
data.
Fig: Data Mining process.
In the above depicted process, the initial stages
(data gathering, data preprocessing, data cleansing,
and dataset preparation) are same as the data
mining process. However, the next stage in this
process, the mining tasks, classification, clustering
and visualization are unified and the process is
called Unified Data Mining Process UDMP.
In this technique of data mining, the only required
input is a dataset in which the proceeding tasks are
performed automatically and knowledge is mined.
This knowledge is then analyzed by the user if is
required by him according to his business rules[5].
There is no need to manually implement algorithms
for classification and clustering in the whole
process, however it is though required to specify
them at the beginning of the process.
Comparison of Data Mining Tools
Table: Comparison of Data Mining Tools
Multi-agent systems and their applications
Agents and Multi-Agent Systems (MAS) have
emerged as a powerful technology to cope with the
increasing complexity of a variety of Information
Technology scenarios. We are not going to explain
the full details of the agents because these are
covered in other chapters. We are only going to
provide a basic overview.
The most widely accepted definition for the “agent”
term is that “an agent acts on behalf of someone
else, after having been authorized”. This definition
can be applied to software agents, which are
instantiated and act instead of a user or a software
program that controls them. The difficulty in
defining an agent arises from the fact that the
various aspects of agency are weighted differently,
with respect to the application domain at hand.
Wooldridge & Jennings have succeeded in
combining general agent features into the following
generic abstract definition integrating all the
characteristics into the notion of an agent: “An
agent is an autonomous software entity that
-functioning continuously - carries out a set of goal-
oriented tasks on behalf of another entity, either
human or software system. This software entity is
able to perceive its environment through sensors
and act upon it through effectors, and in doing so,
employ some knowledge or representation of the
user's preferences” (Wooldridge, 1999).
An agent may have, depending on the domain it is
situated in, some or all of the properties listed
below (Symeonidis & Mitkas, 2005):
• Autonomy (considered a must-have feature by
many researchers in the field of agents)
• Interactivity: Reactivity or Pro-activeness
• Adaptability
• Sociability
• Cooperativity
• Competitiveness
• Mobility
• Learning
The Rough Sets algorithm implemented in
MASCE can be summarized as follows:
This algorithm takes as input a decision table S =
(U, C∪ D, V, f) and produces as output the set of
decision rules {τij} (Attia at al., 2004).
Step 1: Vertical reduction:
The vote value is calculated for all the tuples
(similar tuples are collapsed into one and their
number is added to the vote).
Then tuples, with vote values less than the noise
filter threshold, are removed from the database
table.
Step 2: Horizontal reduction: Attributes reduction
is made by calculating the best reduct RED as
follows, let all attributes be called AR and the user
preferred attributes if any be UA. Begin
2.1. Construct the modified discernibility matrix
M(C): Each entry mij contains the condition
attributes whose values are not identical on both xi
and xj where xi, xj belong to different classes of
IND(D) i.e. they represent different decision
concepts. M(C) is a diagonal symmetric matrix.
mij = 0
fx i, xj ∈ same IND(D) = {c ∈ C: f(c, xi) ≠ f(c, xj)}
if xi, xj ∈ different IND(D)
2.2. Find the CORE from discernibility matrix:
For any c ∈ C, c ∈ CORE(C) if and only if there
exists i, j, 1 ≤ j < i ≤ N such that mij = {c}. Note
that a core may be empty
2.3. Determine the attribute set UA which user
prefers to emphasize. If UA is empty that means
that the user does not have preference for any
attribute.
2.4. Let RED = CORE ∪ UA
2.5. AR = AR – RED
2.6. Find attribute a in AR which has the
maximum SGF(a, RED, D)
2.7 RED = RED ∪ {ai}, AR = AR – {ai} (i =1, 2,
…m)
2.8 If k(RED, D) = 1, then stop, otherwise go to
step
Step 3: Generate the reduced relation by removing
those attributes which are not in the best reduct
RED.
Step 4: Combine similar tuples in the reduced
relation.
Step 5(a): Transform tuples in the reduced relation
into decision rules for each class in D.
Step 5(b): For the same class in the reduced table,
two tuples can be combined if the values of the
condition attributes differ in only one attribute, thus
obtaining a more general set of decision rules. Or
instead of steps 5(a) and 5(b) we can use the
following alternative method for generation of
decision rules:
Step 6(a): Extract the decision rule which is called
τij as follows: τij = DesC(Xi) ⇒ DesD(Yj) such that
Xi ∩ Yj ≠ Φ for Xi ∈ C* and Yj ∈ D*.
Step 6(b): Call A rule (deterministic) if Xi ⊆ Y j i.e.
Xi ∩ Y j = Xi, otherwise a rule is nondeterministic.
The set of all decision rules {τij} for all classes Yj
∈ D * is called the decision algorithm of the
information system S.
/* End of the algorithm*/ The significance of an
individual attribute {a} added to the set A with
respect to the dependency between A and D
(Decision set) is represented by significant factor
SGF, given by:
SGF (a, A, D) = k(A+{a}, D) – k(A, D) where k(A,
D) = card(POSA(D)) / card(U) .
Decision trees mining in MASCE
Decision tree induction is a well-known discipline
in Machine Learning presented by Quinlan in 1986
(Quinlan, 1986). The basic algorithm for decision
tree induction is a greedy algorithm that constructs
decision trees in a top-down recursive divide-and-
conquer manner. In the process of constructing a
tree, the criteria of selecting test attributes
influences the classification accuracy of the tree.
Presently, there are many criteria for choosing the
test attribute in building decision tree, such as ID3
(Quinlan, 1986) and C4.5 (Quinlan, 1993) which
use an entropy-based measure known as
information gain as a heuristic for selecting the
attribute. Decision trees represent a supervised
approach to classification. A decision tree is a
simple structure where non terminal nodes
(internal) represent tests on one or more attributes
and terminal (leaf) nodes reflect decision outcomes.
The ordinary tree consists of one root, branches,
nodes (places where branches are divided) and
leaves. In the same way the decision tree consists of
nodes which stand for circles, the branches stand
for segments connecting the nodes. A decision tree
is usually drawn from left to right or beginning
from the root downwards, so it is easier to draw it.
The first node is a root. The end of the chain “root -
branch - node-...- node” is called “leaf”. From each
internal node (i.e. not a leaf) may grow out two or
more branches. Each node corresponds with a
certain characteristic and the branches correspond
with a range of values. These ranges of values must
give a partition of the set of values of the given
characteristic.
Conclusion:
The above study was conducted by using four
algorithms over a data set: Zero Rule (ZeroR), One
Rule (OneR), decision tree (C4.5), and k-nearest
neighbor (KNN). Tools were run over the data set
and results were observed for each algorithm.
Accuracy percentage served as performance
measure. Weka was indentified a better performer
with the specified algorithms, followed by KNIME
and Tanagra. This performance ranking based on
the type of data set used and how the classifier is
implemented within the tool, as task of
classification is affected by so. But Weka still
proved to be better as it provided with the
implementation of ZeroR and OneR over data types
where other tools did not. The functionality offered
by these tools, like API support and graphical
presentation along with other features aid with the
selection of tool best suitable according to the usage
by different users. The methodologies discussed for
UDMT formulation offer better approach towards
data mining but still are left with some issues due to
an incomplete theory for their correct formulation.
The MAS suffer from the problem of selection of
appropriate set of algorithms for classification,
clustering and visualization. The correct application
of an algorithm as a function is also an issues
The Inductive Databases theory addresses the need
for a language to design inductive queries and
generation of databases including both data and
patterns. All the defined patterns are then required
to satisfy the constraints imposed upon. This results
in need of better algorithms for constraint based
data mining.
Future Work:
Based on our result we concluded that there
is enough claim for Agent – oriented methodology
varies regarding the different methodology applied;
three out of the five architectures studied has
applied the BDI theory into practices. Other method
such as O-MaSE used in Agent Tool3 and the
Cognitive method used in Courage has been applied
only within limited architectures. We found
significant performance boost that proofs our agent-
oriented methodology. Enhance comparison of
agent-oriented tools.

References:

[1]“A comparative analysis of data mining tools in agent

based systems”, Sharon Christa, K. Laxmi Madhuri,
V.Suma, research and industry incubation centre,
Dhyanchand Sagar institutions.

[2]“A Comparisonstudy between data mining tools with

some classification methods”, Abdullah H. Wahbeh,
Qasem A. Al-Radaideh, Mohammed N. Al-Kabi, and
Emad M. Al-Shawakfa.

[3]“Evaluation of Fourteen Desktop Data Mining

Tools”, Michel A. King and John F. Elder IV, Ph.D.,
Department of Systems Engineering-summary.

[4]“Suitability analysis of data mining toolsand

methods” Samuel Kováč, Masaryk university faculty of
informatics.

[5]“Towards the Formulation of a Unified DataMining

Theory, Implemented by Means ofMultiagent Systems
(MASs)”, Dost Muhammad Khan, Nawaz Mohamudally
and D. K. R. Babajee.

[6]“A Study of Data Mining Toolsin Knowledge

Discovery Process”, Y. Ramamohan, K. Vasantharao, C.
Kalyana Chakravarti, A.S.K.Ratnam.

[7]“Towards a General Framework for Data

Mining”,Saˇso DˇzeroskiJoˇzef Stefan Institute, Jamova
39, 1000 Ljubljana, Slovenia

[8]Principles of Data Mining, by David Hand,

HeikkiMannila and Padhraic Smyth

[9]WEKA, the University of Waikato, Available

at:http://www.cs.waikato.ac.nz/ml/weka/downloading.ht
ml

[10]KNIMEAvailableat:http://www.knime.org/downloa
d-desktop[11]“Open-Source Tools for Data Mining”
Blaz Zupan, PhD, Janez Demsar, PhD.[12]“Tanagra: An
Evaluation” Jess