CHAPTER 4 LEARNING STYLES ASSESSMENT IN...
Transcript of CHAPTER 4 LEARNING STYLES ASSESSMENT IN...
61
CHAPTER 4
LEARNING STYLES ASSESSMENT IN E-LEARNING
4.1 COMMON MODELS OF LEARNING STYLE
IDENTIFICATION
E-learning system is based on the common rules and features of the
learners to be engaged in the learning process. Therefore, the success of
e-learning environments is greatly influenced by the factors like learning
objects, content delivery, relevant information retrieval, performance
evaluation and the maintenance of the psychological level through
identification of the individual learning styles of the learners. This work
depicts the different learning styles which are available in the literature and
provides a comparative analysis among them. In addition, it suggests the use
of fuzzy logic for handling uncertainty in Felder–Silverman learning style
model.
A learning style is “a particular way in which an individual learns”
(Butler 1986). Numerous measures and instruments including questionnaires,
interviews and profile information have been used in the past to identify the
learning styles of the learners efficiently. These metrics are labeled as explicit
information provided to describe the characteristics of the learners during the
assessment procedure. Therefore, the main objective of this work is to analyze
the related works for identifying the individual learning style based on the
learners’ behavior. This will be helpful to provide a suitable e-content to the
62
learners based on their learning styles in an e-learning environment. The
psychological level of the learners in an e-learning environment is greatly
attributed by the learning styles of the learners involved in learning. However,
in most of the existing e-learning frameworks, the psychological level
between the learners and e-learning contents is not well balanced. This is due
to the fact that the learning styles of the individuals vary from one person to
another and hence if the same kind of e-contents is provided to all the
learners, the success of the e-learning system degrades. Therefore, the
contents developed for the e-learning system could be modified based on the
learners learning styles in such a way that all the learners could be well
benefited so that the objective of the e-learning system could be satisfied.
Table 4.1 Learning styles Models – Metrics and Dimensions
Nature of Metrics
Inclusive Dimensions
Underlying Learning Theory
Learning Style Models Addressed
Static Personality Type Educational Specialization Professional career choice Job role Adaptive competencies Environmental factors Emotional factors Sociological needs Physical needs
Experiential theory model
Kolb Model
Behavioral theory model
Honey and Mufford Model
Cognitive theory model
Gregorc model
Psychological theory model
Felder-Silvermann model
Meta-learning theory model
Fleming VAK model
Dynamic Intelligence quotient factorsBiological factors Inherent interests
Personality model Carl and Myers Brigg indicator model
Intelligence theory model
Howard Gardner
Table 4.1 depicts the general metrics that are used for identifying
the learning style of a learner which includes the factors such as personality
types, early educational specialization, professional career choice, adaptive
competencies and gender (Boyatzis and Kolb 1997). When considering the
above metrics, the learners can be identified as belonging to any one of the
63
styles of learning (Butler 1986). However, these existing metrics are suitable
only in identifying the learning styles based on questionnaires, interviews and
profile information. However, the results obtained through these measures
may not be accurate in the current e-learning scenario. Since, the learner is
independent of a tutor in e-learning, the learning style is absolutely vague in
nature. Hence, a deep understanding of the learners is necessary in predicting
their learning style in E learning environment.
A deep understanding of one’s own way of learning can lead to a
great personal empowerment and self confidence. This kind of deep
understanding can be known by analyzing the behavior of the learners
involved in an e-learning environment. One of the techniques employed for
identifying the behavior of the learners in e-learning is their web browsing
pattern since clear and deep identification of the learning style of the learner
through the browsing pattern increases the measures for restructuring the
design of an e-learning content (Melody Siadaty, Fattaneh Taghiyareh 2007).
This work focuses on the analysis of learning styles of different models and to
enhance Sanders et al model to handle uncertainty present in their model
(Sanders and Bergasa-Suso 2010). According to Sanders et al model, learners
are classified into three styles namely active, reflective and unknown.
However, the unknown category can be further investigated so that a suitable
content can be provided to these categories of learners as well. In this work,
fuzzy logic based learning style model is proposed that extends the
classification provided by Sanders et al in which the learners are classified
efficiently into active, medium active, medium reflective, reflective. For this
purpose, Gaussian membership function is suggested in this work for
effective classification of learners in an e-learning environment.
64
The major contributions of this work are the provision of
1. Ability to predict the learning styles based on the web usage
information and also on their original profile information,
2. The proposed model makes use of fuzzy membership
functions. Therefore, precise reasoning in assessing the
learners based on their learning styles is achieved and
3. It proves a facility to assess the learners accurately based on
their learning styles. For this purpose, the learners are
classified into four categories namely active, medium active,
medium reflective and reflective using fuzzy rules.
4.2 INFERENCES OF LEARNING STYLES IN EDUCATIONAL
SYSTEMS
The past literatures in the area of learning styles identification had a
great impact on the static behavior of the learners. There were some hidden
traits present within the learners which are not usually considered for learning
styles assessments. Most of the learning styles discussed in the literature
assess the learning style of a particular individual based on the profile
information given by the learners themselves. On further analysis, it was
identified that the explicit information given by the learners are alone not
enough in identifying the learning style correctly. In the proposed work,
greater importance is given to the hidden nature of the learners which could
aid in assessing the learning style of an individual efficiently.
The next category of learning style prediction systems could
effectively make use of internet education systems and are very suitable to
many courses in the internet. However, these systems fail to consider the
65
learners capacity or provide any intelligent advice on potential websites.
WebCT is the first internet educational system that could provide learning
course materials through web but cannot provide any intelligent technique
which could consider the learners capacity (Sanders 1993). The next suite of
internet educational systems like INSPIRE (Papanikolaou et al 2002),
ARTHUR (Gilbert and Han 1999), AES-CS (Triantafillou et al 2002),
Tangow (Paredes and Rodriguez 2002), could predict the learning styles of
the learners effectively, but content adaptation was made based on previous
knowledge, but these systems needed to assess the learners learning styles
offline and questionnaires evaluation.
The system developed by Bergasa (Bergasa-Suso et al 2005) called
AHA which is a set of tools termed iLessons that overcame the difficulties of
the earliest systems. The model proposed by him considered Felder Silverman
learning style preferences as a base model to categorize the learners into any
one of the four dimensions namely active/reflective, sensing/intuitive,
visual/verbal, and sequential/global. Therefore, iLessons is an important
contribution for e-Learning by identifying the learners as either active or
reflective for the first dimension. This model was satisfactory when the
learners tend towards a particular dimension most of the time. However, this
scenario is not true always. This condition was rectified subsequently, by the
model developed by Sanders and Bergasa-Suso (2010) known as new
intelligent system to categorize the learners into any one of the four
dimensions described in Felder Silverman learning style model which had an
effective user interface. According to this system, the learners were classified
into three categories for the first dimension (active/reflective) as active,
reflective, or unknown. The accuracy of classifying the learners in Sanders
and Bergasa-Suso (2010) has increased to 81% compared to the earlier work
done by Bergasa-Suso et al (2005). However, this model has to be enhanced
66
further by resolving the unknown category of the learners in order to improve
the classification accuracy. This kind of uncertainty in the inference found in
the unknown category can be handled by introducing fuzzy rules for effective
classification of the learners learning through web environments.
From the analysis made, there were many internet educational
systems which could effectively predict the learning styles through offline or
questionnaires or through online activity information. In spite of good
contributions offered by various researchers in predicting the learners learning
styles through web usage information, it will be worthier if some machine
learning algorithms are used for effective classification. Queries based on the
degree of activeness or reflectiveness is not addressed in any of the existing
recent models for accurate prediction of the learners. Hence, it is necessary to
provide a rule-based approach like neural networks or fuzzy logic for
effective classification of the learners’ learning through web environments
(Shyi-Ming Chen and Yu-Chuan Chang 2011). The model proposed in this
thesis makes use of fuzzy membership functions which helps in identifying
the degree of activeness or reflectiveness and also categorizes the learners in
the unknown category of the existing models into active, medium active,
medium reflective and reflective types of learners.
4.3 FUZZY LOGIC BASED LEARNING STYLE PREDICTION
IN E-LEARNING
In this work, a new learning style identification system based on
fuzzy logic has been proposed and the subsystem is shown in Figure 4.1. The
proposed model is based on Felder Silverman learning style model that relies
on psychological theory. The model is tested for the learners who are learning
‘C’ programming language through e-learning environments. The proposed
67
system uses the learners’ web interface information and Felder Silverman
learning style preferences as inputs.
Figure 4.1 Learning Style Identification Subsystem
The major features of this proposed model are the introduction of
metrics analyzer and fuzzy inference engine for effective learning style
identification.
4.3.1 Prediction System Input Setting
The proposed model uses mediawiki e-Learning servers for the
e-Learning contents to be posted in various formats. This e-learning server
consists of C programming language course contents in textual, audio, and
video formats. The learner after proper authentication could access any type
of contents that are available in mediawiki e-Learning server. The target
learners in this proposed model tend towards textual format of the course
contents, and therefore this model was evaluated and tested for identifying the
Felder Silverman Learning Styles
Models
Learner’s Web Interface
Information
Learning Style Identification Subsystem
Fuzzy Inference Engine
Metrics Analyzer
Metrics Meta Data Student Profile
information
Rule Base
Media Wiki E-Learning Server
Content
68
first dimension of Felder Silverman learning style preferences. The learners’
during the authentication are asked to provide their original profile
information for the purpose of learning style identification. Table 4.2 depicts
the characteristics of the earlier learning style models that are used in various
application domains.
Table 4.2 Parameters of Learners Web Interface
Parameters List Parameters Included
Deadband model
Proposed Fuzzy Logic based model
Number of mouse movement in the y-axis
Yes Yes
Ratio of document length to the time spent on a page
Yes Yes
Ratio of images area to document length and scroll distance
Yes Yes
Number of visits to a document No Yes
Fuzzy Rules No Yes
Therefore, the learners are accurately classified based on their
learning styles using their own profile information and their online web usage
activity. To facilitate the experimental evaluation, a rule base has been
constructed which is fully loaded with input and output fuzzy rules. The
learners’ activities were carefully monitored and recorded for the Learning
Styles prediction. These activities were recorded for analysis with respect to
the parameters described in Table 4.2. According to the evaluation, the
learners were classified into four kinds of learning styles namely active,
medium active, medium reflective, and reflective.
69
4.3.2 Overview of Felder-Silverman Learning Style Model
This learning style model is often used in technology-enhanced
learning. This learning style model is usually preferred in web-based learning
environment. The model is very useful because of the four dimensions for
learning style preferences which considers the technological aspects of the
e-learning learners. The four dimensions for categorizing the learners are as
follows:
A. Active/Reflective: Active learners are active by themselves and their user
activity across the web environments are very fast, whereas the reflective
learners slowly think about concepts and the user activity across the web
environments will be usually very slow.
B. Sensing/Intuitive: Sensing learners try to learn the underlying facts
whereas intuitive learners prefer to identify relationships existing among the
concepts.
C. Visual/Verbal: Visual learners prefer to learn through pictures, flow
charts, and cartoons whereas verbal learners tend towards learning through
textual representations
D. Sequential/Global: Sequential learners learn through incremental learning
whereas global learners learn at a single scratch of time. The proposed model
concentrates on the first dimension of active/reflective since the learners
learning through e-learning environments usually fall in either of the
categories mentioned.
70
4.3.3 Metric Analyzer
The Learning style prediction of the proposed model is analyzed
using various metrics described in Table 4.3. These metrics includes both the
learners online web activity and some of their original profile information.
Using the information from Table 4.3, the metric analyzer present in the
proposed system analyzes the learners’ behavior and aids in classifying the
learners with respect to their learning styles.
Table 4.3 Metrics for Learning Styles Prediction
Online Web Information Offline Profile InformationNumber of mouse movement in the y-axis Domain of Interest Ratio of document length to the time spent on a page
Educational Background
Ratio of images area to document length and scroll distance
Professional Career
Number of visits to a document
4.3.4 Rule Base
A rule based approach is proposed in this work for the effective
classification of learners which can handle uncertain information as well is
used. Moreover, this model aids in recommending and providing suitable e-
learning materials based on the identification of learning styles. The
knowledge editor rule base consists of approximately 30 fuzzy rules for the
prediction of learners based on their learning styles using online web activity
information and offline profile information. The set of rules used in this
proposed work are shown in Table 4.4.
71
Table 4.4 Fuzzy Rules in the Proposed System
Rule No Fuzzy Rules
1 if(mousemove is more) and (doclentime is more) and (imagedoclen is more) and (visitsdoc is more) then (student is reflective)
2 if(mousemove is less) and (doclentime is lesss) and (imagedoclen is less) and (visitsdoc is less) then (student is active)
3 if(mousemove is less) and (doclentime is less) and (imagedoclen is less) and (visitsdoc is more) then (student is mactive)
4 if(mousemove is less) and (doclentime is less) and (imagedoclen is more) and (visitsdoc is less) then (student is mactive)
5 if(mousemove is less) and (doclentime is more) and (imagedoclen is less) and (visitsdoc is less) then (student is mactive)
6 if(mousemove is more) and (doclentime is less) and (imagedoclen is less) and (visitsdoc is less) then (student is mactive)
7 if(mousemove is more) and (doclentime is more) and (imagedoclen is more) and (visitsdoc is less) then (student is mreflective)
8 if(mousemove is more) and (doclentime is more) and (imagedoclen is less) and (visitsdoc is more) then (student is mreflective)
9 if(mousemove is more) and (doclentime is less) and (imagedoclen is more) and (visitsdoc is more) then (student is mreflective)
10 if(mousemove is less) and (doclentime is more) and (imagedoclen is more) and (visitsdoc is more) then (student is mreflective)
11 if(mousemove is less) and (doclentime is less) then (student is active)
12 if (imagedoclen is more) and (visitsdoc is more) then (student is active)
13 if(mousemove is more) and (doclentime is more) then (student is reflective)
14 if(imagedoclen is more) and (visitsdoc is more) then (student is reflective)
15 if(mousemove is less) and (doclentime is more) then (student is mreflective)
72
Table 4.4 (Continued)
Rule No Fuzzy Rules
16 if(mousemove is more) and (doclentime is less) then (student is rmactive)
17 if(imagedoclen is less) and (visitsdoc is more) then (student is mactive)
18 if(imagedoclen is more) and (visitsdoc is less) then (student is mreflective)
19 if(mousemove is less) and (doclentime is less) and (imagedoclen is more) and (visitsdoc is more) then (student is mreflective)
20 if(mousemove is more) and (doclentime is less) and (imagedoclen is less) and (visitsdoc is more) then (student is mreflective)
21 if(mousemove is more) and (doclentime is more) and (imagedoclen is less) and (visitsdoc is less) then (student is mactive)
22 if(mousemove is less) and (doclentime is more) and (imagedoclen is more) and (visitsdoc is less) then (student is mactive)
23 if(mousemove is less) and (visitsdoc is less) then (student is active)
24 if(mousemove is less) and (visitsdoc is more) then (student is mreflective)
25 if(mousemove is more) and (visitsdoc is less) then (student is mactive)
26 if(mousemove is more) and (visitsdoc is more) then (student is reflective)
27 if (doclentime is less) and (imagedoclen is less) then (student is mreflective)
28 if(doclentime is more) and (imagedoclen is more) then (student is mactive)
29 if(doclentime is more) and (imagedoclen is less) then (student is active)
30 if(doclentime is less) and (imagedoclen is more) then (student is reflective)
73
4.3.5 Fuzzy Inference Engine
The learners membership degrees for active/reflective type of
dimensions were studied using Engineering students of Anna University. On
successfully providing the e-learning course materials of ‘C’ programming
language, the symmetric Gaussian membership function for a fuzzy set A,
that represents the learning styles of the learners is represented by,
) = ( ) (4.1)
The parameters (width) and c (center) alter the width of the
membership function curve of the fuzzy set A, based on the input value x.
This fuzzy set A represents the learning style, and the parameter c denotes the
mean of the membership function curve in this function
(Shyi-Ming Chen and Yu-Chuan Chang 2011). The symmetric Gaussian
fuzzy membership function used in the proposed model for the four categories
of the learners namely active, medium active, medium reflective and
reflective is shown in Figure 4.2.
Figure 4.2 Symmetric Gaussian Fuzzy Function
74
Based on the rule base developed in the proposed model, and on the
application of the symmetric Gaussian fuzzy membership function described
in equation (4.1), the learners could be predicted on their suitable learning
styles.
4.4 PERFORMANCE EVALUATION
The applicability of this model is tested in various online
e-learning environments. Based on the authors’ domain of interest, the
proposed model is tested for learners interested in learning ‘C’ programming
language course with varying learning styles. The main objective of the model
is to handle the uncertainty in the learners’ behavior and to classify them
accurately into four categories namely active, medium active, medium
reflective, and reflective as present in the first dimension of Felder Silverman
Learning Style model preferences. The described objective was tested and
compared with the other existing algorithms and the results are shown in this
section.
4.4.1 Experimental Set-Up
The learners during authentication procedure are solicited to
provide their original profile information including their age, gender,
educational background, domain of interest, professional career and hobbies.
Subsequently, the learners are provided with various ‘C’ programming
language course contents from any of the e-learning servers like mediawiki,
moodle, and joomla. However, these experiments were based on the
e-contents posted in mediawiki e-learning servers. The experiments were
carried out using 120 numbers of students of Anna University from various
branches namely Civil Engineering, Mechanical Engineering, Computer
75
Science and Engineering, and Electronics and Communication Engineering.
The rule base included a set of 30 fuzzy rules for Learning Styles prediction.
In this work, the inference for Learning Styles was done using Matlab R2009a
7.8.0.347. The experiments were repeated, analyzed and compared with the
other existing models provided by Bergasa-Suso et al (2005), Sanders and
Bergasa-Suso (2010). The proposed model is also compared with the
traditional Bayesian Classification algorithm to test the significant difference
in the accuracy of the proposed model. In all the repeated experiments, the
Learning Styles of the learners were predicted based on their e-learning of ‘C’
programming language course contents. However, this work is an ongoing
work. The proposed work actually had its base from Felder Silverman
learning style model preferences for Learning Style prediction and this work
is limited to predicting the first dimension of learning styles only
(Active/Reflective).
4.4.2 Results and Discussion
The experimental evaluation results shown in Figure 4.3 are
obtained from the experiments conducted in this work on predicting the
learning styles of the learners through offline profile information and online
web activity information. The learners provide their complete information
through a web interface. Subsequent, they e-learn the ‘C’ programming
language course contents in the textual format available in mediawiki
e-learning servers. On successfully collecting and recording this information,
the metric analyzer analyzes for the available metrics to be used for learning
styles prediction.
76
Figure 4.3 Three Dimensional Views of Learners X = Number of mouse movement in the y-axis
Y = Ratio of document length to the time spent on a page Z = Ratio of images area to document length and scroll distance
In this experiment, the metrics include both the profile information
and the online web activity information. With the help of these metrics, the
fuzzy inference engine classifies the learners into four categories namely
active, medium active, medium reflective and reflective. This fuzzy inference
engine makes use of a rule base fully loaded with 30 input and output fuzzy
rules. The evaluation of the proposed model was estimated in terms of
percentage of accuracy. Figure 4.3 shows the three dimensional view of the
different kinds of the categorized learners namely active, medium active,
medium reflective, and reflective. The X, Y, Z axis represents the first three
parameters described in Table 4.3.
77
4.4.3 Inference
In this research work, Felder-Silverman learning style model is
chosen for learning style prediction especially in web learning environments.
Moreover, fuzzy rules derived from Gaussian membership functions are used
in this work to handle the uncertainty in Learning Styles prediction based on
the learners’ membership degree for active/reflective type of dimensions in
Felder–Silverman dimensions of learning style model. To facilitate this, the
learners’ web usage activities are recorded and some of the parameters in such
activities of the learners are considered for learning styles identification.
Table 4.5 shows the results obtained by applying various classification
algorithms used in learning styles identification. This experiment considered
120 learners for classification. From the table 4.5, it can be observed that the
proposed fuzzy logic based classifier provides accurate results when it is
compared with the existing techniques namely Bayesian, Bergasa and
Deadband. The graphical representation of the classified and unclassified
learners is shown in Figure 4.4.
Table 4.5 Classified and Unclassified Learners
Algorithms Number of Users (120)
Classified UnclassifiedBayesian 73 47
Bergasa 88 32
Deadband 94 26
Fuzzy logic 120 0
78
Figure 4.4 Analysis of Classified Vs Unclassified Learners
The psychology of the learners may change periodically. Moreover,
the learners tend towards a particular dimension depending on circumstances,
mood and need. Therefore, tight coupling of learners to any of the two
dimensions namely active or reflective may not be complete in e-contents
recommendation to the learners. In such a scenario, the learners can be
loosely categorized as more active and less reflective known as medium
active and more reflective and less active known as medium reflective.
Therefore, this work aims at resolving this ‘unknown’ category to be
classified into any one of the four categories namely active, medium active,
medium reflective and reflective, thereby increasing the accuracy percentage
in predicting the learners based on their learning styles. The evaluation
measure called accuracy is defined as the percentage of classified users
belonging to a particular dimension. Table 4.6 shows the accuracy percentage
of the various algorithms present in the literature.
79
Table 4.6 Accuracy of the classified Users
Algorithms Accuracy (%)
Active Reflective Medium Active
Medium Reflective
Bayesian 53 47 0 0Bergasa 61 39 0 0
Deadband 65 35 0 0Fuzzy logic 40 33 14 13
From the Table 4.6 the accuracy in classifying the learners is found
where the proposed fuzzy logic based algorithm identifies four categories of
learners namely active, medium active, medium reflective and reflective since
fuzzy membership functions are used for refined classification. However, the
other algorithms present in the literatures could identify only active and
medium type of learners. The accuracy percentage of the other algorithms in
comparison with the proposed algorithm is shown in Figure 4.5.
Figure 4.5 Learning Style Identification
0
20
40
60
80
100
Active Reflective Medium Active Medium Reflective
Learning Style Dimension
Learning Style Identification
Bayesian
Bergassa
Deadband
80
Hence, the objective of obtaining fine accuracy in the Learning Styles prediction of the uncertain learners who are learning through web
environments has been achieved. The accuracy of the proposed fuzzylogic algorithm is validated using precision performance measure.
4.5 AGENT BASED PAIR PROGRAMMING
Several factors affect the performance in learning a programming course (White and Sivitanides 2002) and research is under progress to identify
why most of the students drop out from learning the programming courses. Learning a programming course is comparatively difficult than learning any
other course in e-learning since, programming knowledge requires cognitive abilities, logical thinking and mastering the abstract concepts (Van Gorp and
Grissom 2001). Therefore, this situation is a significant problem in the field of Computer Science and Engineering. There are several studies (Han et al 2010,
Palmieri 2002, Ueno 2000, Woods and Warren 1995, De Azevedo and Scalabrin 2005, Devedzic 2004, Sklar and Richards 2006, Fleming 2001,
Boyatzis and Kolb 1997, Cohen 1988, Martinez 2001, Myller et al 2002 and Chen et al 2000) which addresses these problems by providing many methods
of educational systems. Prior research works have shown that effective factors like improving self-efficacy and good metal effort can increase the
performance of the students, especially when learning a programming course (Vennila Ramalingam et al 2004).
Several factors influence the success of novice programmers namely
Prior Computing Experiences (Bunderson and Christensen 1995, Taylor and Mounfield 1994), e-learning, Computer Entertainment during Training
(Potosky 2002), Self-Efficacy (Byrne and Lyons 2002), Collaborative learning Environments (De Azevedo and Scalabrin 2005), Learning Styles
(Thomas et al 2002, Butler 1986) and Students Mental Effort (Soloway and Ehrlich 1984, Wiedenbeck et al 1999). Among the above mentioned factors,
providing a collaborated learning environment with increased self-efficacy of
81
the students is the main focus of this work which leads to the success of the students in learning programming courses.
Pair programming is a term used in learning environments where two programmers work together at one workstation on the same design,
algorithm or code (McDowell et al 2006). Since, at least two people are involved, this is a form of collaborated learning environment. This kind of
collaboration is found beneficial to the learners especially when learning a programming course.
Self-efficacy is a key component in learning activities since
learning involves more than just acquiring skills. Bandura (1977) defines self-efficacy as people’s own judgment of their capabilities to organize and
execute course of actions required to achieve a specific goal. Educational researchers recognize that, because skills and self-beliefs are so intertwined,
one way of improving learners performance is to improve learners self-efficacy (Compeau and Higgins 1995). Attempts have also been made,
with some success, to increase self-efficacy in learning by peer modeling of tasks, verbal persuasion, or other types of social influences, such as
cooperative learning environments (Bandura 1977, Compeau and Higgins 1995, Canas et al 1994, Gist et al 1989). Persons who possess higher
self-efficacy belief show more effort and resistance for completing tasks and therefore have better and more effective task-fulfillment compared with
individuals who have weak self-efficacy (Gist et al 1989). However, increasing the self-efficacy by the above methodologies is found to be
inefficient for the learners learning through web. According to Bandura's theory, self-efficacy belief is under the influence of the following elements
namely personal experience which leads to success or failure, observing the behavior of the model, the substitute, or the example, vocal encouragement
and considering the physiological conditions
82
In such a scenario, the self-efficacy of the learners can be increased
if they are allowed to judge their own abilities, before they could pair to learn
with appropriate agents to master a programming course. Moreover, as
analyzed before the students could outperform well in learning a
programming course, if they had prior programming experiences by learning
the basic concepts (Kinnunen and Malmi 2006). This is achieved, by allowing
the learners to self learn the e-contents from e-Learning servers based on their
choice of learning materials available in different formats namely documents,
audio and video lectures. This increases the level of self-efficacy because the
target learners acquire some basic knowledge about the programming course.
Considerable increase in the self-efficacy can make them to outperform well
in the subsequent learning with peer-learning agents in a pair programming
strategy (Han et al 2007).
4.6 AGENT BASED PAIR PROGRAMMING MODEL
The use of Internet in education has provided a new revolution
known as e-learning or web-based learning. In spite of all the important
contributions provided by various researchers in the area
e-learning described above, it will be useful if machine learning techniques
are used to further improve the self-efficacy of the students involved in
learning. Moreover, it is proved that the exchange of roles and the meaningful
feedback between the peer-learning agent and the learners have a positive
effect on learning (Han et al 2007 and Han et al 2010). Moreover, the
peer-learning agents enhance the effect of learning by determining the
learners’ level of understanding. However, increasing the self-efficacy of the
learners can still enhance the performance of the learners (Vennila
Ramalingam et al 2004). Prior literatures has suggested several factors of
increasing the self-efficacy of the learners in terms of providing previous
computing experiences through short lectures or self learning, computer
83
entertainment during training, motivations, learning styles and assessing the
learners mental model during learning.
The first methodology of learning focuses on the recommendations
of suitable e-learning contents available in e-learning servers to self learn. This is
done in order to facilitate the learners with previous computing experiences to
understand the basic concepts in a particular programming course. Moreover,
suitable e-learning contents are provided based on identifying their learning
styles using Felder Silverman learning style model (Felder and Silverman
1988) as a base for identifying the learning styles. Subsequently, the learners
are paired with peer-learning agents for mastering the programming courses.
The second groups of learners have a different methodology of
learning where the learners learning through web environments did not have
any prior learning experiences. These groups of learners are directly paired
with peer-learning agents where the learners learn through pair programming
strategy. The learners and the agents alternate roles of the tutor and tutee in
order to learn a programming course.
The third methodology of learning is based on software agents
where the agent system is based on pedagogical learning. The agent in this
system assumes the instructor role and the learner assumes the tutee role. The
agent and the learners interact once in a while on the target e-course. After a
certain period of time, they are paired with peer-learning agents where the
learners and the agents learn using pair programming strategy. The agent and
the learner alternate roles of the tutor and tutee and learn a target course as
described earlier. The learners belonging to the different methodologies of
learning are assessed for their performances using posttest simple multiple
choice questionnaires. The proposed agent based model architecture is shown
in Figure 4.6.
84
Figure 4.6 Agent Based Pair Programming Model
4.7 COMPONENTS OF AGENT BASED MODEL
4.7.1 Learning Style Identification
The students involved in e-learning develop high interest in
learning web contents posted in any e-learning servers. A wide variety of
e-learning servers are available like moodle, mediawiki and joomla. Media
wiki server was used for testing the proposed system because the user
interface is simple in nature. The learners are identified for their individual
Self-Learned Group
Self Learning using Recommended e-contents
Pair Programming Learning with Peer-Learning Agent
Comparative performance Evaluation using multiple-choice
questionnaire
Learning Style Identification
Peer Learned Group
Conversation Agent Learned
Group
Felder Silverman Learning Style
model Recommended
E-Contents
Recommendations and the methodology of
learning
PeerLearned Group
Conversation Agent Learned
Group
85
learning styles based on Felder Silverman learning style model inventory
(Felder and Silverman 1988). According to the model, the four dimensions for
categorizing the learners are Active/Reflective, Sensing/Intuitive,
Visual/Verbal and Sequential/Global.
The target course of learning in this work is ‘C’ programming
language. The ‘C’ programming language contents are posted in various
formats like documents, audio files and video lectures. These contents are
made available in mediawiki e-learning server. The students are now analyzed
for their learning styles using the online usage activities of the learners
described in Section 4.3. Accordingly, the learners are classified into four
categories namely active, medium active, medium reflective and reflective.
In such a scenario, suitable recommendations are provided to the learners on
the choice of e-learning contents for self study based on their learning styles.
For instance, students with active learning style can be recommended to learn
the basic concepts of ‘C’ programming language available in three or more
documents, students with reflective learning style can be recommended to
learn concepts available in single document and students with either medium
active or medium reflective learning styles can be recommended to learn the
contents available in one or two documents and the corresponding hands-on
experiences.
4.7.2 Self-Learning
The learners during authentication procedure are solicited to
provide their original profile information including their age, gender,
educational background, domain of interest, professional career, and hobbies.
Subsequently, the learners are provided with various ‘C’ programming
language course contents from any of the e-learning servers like mediawiki,
moodle, and joomla. However, our experiments were based on the e-contents
86
posted in mediawiki e-learning servers. The experiments were carried out
using 120 numbers of students of Anna University from various branches
namely Civil Engineering, Mechanical Engineering, Computer Science and
Engineering, and Electronics and Communication Engineering. The rule base
included a set of 30 fuzzy rules for Learning Styles prediction. The learners
are given 4 weeks of time in order to learn the basic concepts of ‘C’
programming language like datatypes, operators, functions and looping
statements. After 4 weeks of self–learning, the students gain some knowledge
on the basic concepts of ‘C’ language. This helps in improved self-efficacy in
order to master ‘C’ programming language course. The learners are tested for
their learning experiences, but however, this test is not treated to be
mandatory.
4.7.3 Learning with Peer-Learning Agents
The learners have some basic knowledge about the ‘C’
programming language course. This aids in increasing their self-efficacy,
since they have been given self learning experiences of learning contents
based on the identified learning styles. The peer-learning agents and the
learners switch between the roles of tutor and tutee. Since, the learners have
obtained the basic knowledge about ‘C’ programming language, they learn
with the peer-learning agents on advanced programming tested this proposed
system for teaching and learning ‘C’ programs like Armstrong number,
Factorial number and Fibonacci series. The learners initiate the system by
providing the logic of the program and the peer-learning agent helps the
students in writing the syntactically correct program. In this case, the learner
assumes the role of tutee and the agent assumes the role of a tutor. In the next
time, the peer-learning agents explain the logic of the code to the learners and
they teach the agents in writing the syntactically correct program. In this case,
the learners and agents assume the roles of tutor and tutee respectively. In this
87
kind of mechanism, the learners and the agents alternate the roles of tutor and
tutee and master the ‘C’ programming language course. This teaching-
learning process is done for 4 weeks of time as given for self-learning. After 4
weeks of time, the learners are again assessed for their performances using a
test consisting of multiple-choice questions. Their performances are evaluated
for comparison.
4.7.4 Learning through Conversational Agents
This type of learning is the traditional pedagogical learning where
the agents assume the instructor role. This is an automated tutoring
application where the agent interacts with one learner at a time. This is a form
of collaborative learning environment where the agent teaches the learners on
the basic concepts of the target course and the learner interacts with the agents
occasionally. The learners progress is monitored once in a while by the agents
by posting simple queries and their returned feedbacks. These types of
learners learn through such conversational agents and then after a period of
time they are paired with peer-learning agents. Now, the learners learn
through Pair programming strategy as earlier. The progress of these learners is
also taken into account for comparing the group performances.
Incorporating self learning experiences in pair programming using
peer learning agents is helpful for enhancing the learning performance. In this
research work, the self-efficacy which is the belief on one’s own capabilities
is increased by providing prior-learning experiences. Moreover, the different
direct learning methodologies namely self-learning following by pair
programming, learning through peer-learning agents and pedagogical learning
assuming the instructor role are compared with the help of the post-test
questionnaires mean score values are considered in this work. The
self-learning group is assumed to have prior learning experiences. The other
two groups have either a direct interaction with peer-learning agents through
88
pair programming strategy or pedagogical agent learning in the form of
instructor. The results of the experiments conducted in this research work are
shown in Tables 4.7.
Table 4.7 Groups Performances
Number of Users
Performance (Mean Score)
Self-Learning
Peer-Learning
Pedagogical Learning
15 24 17 14
30 37 18 14
45 47 24 16
60 49 34 19
75 52 37 23
90 55 42 27
105 59 44 31
120 60 46 34
4.8 SYSTEM EVALUATION
This study sought to determine whether the impact of giving self-
learning experiences by providing suitable e-learning contents based on their
learning styles was beneficial to the learners for subsequent learning with
peer-learning agents. The self-efficacy of the learners is judged by their
learning styles (Boyatzis and Kolb 1997). This efficacy is increased by
providing suitable e-contents to the learners for self learning based on their
learning styles. The system was evaluated using 129 learners from four
Engineering Departments of Anna University and this is given by the variable
‘N’ in Tables 4.8 and 4.9. All these learners were interested in learning ‘C’
89
programming language course. Forty-three learners from the departments
were assigned to a group called prior-learned group, where these learners
were provided with self-learning e-contents recommended based on their
learning styles. The remaining 43 learners from the other departments were
assigned to a group called experimental group. The other 43 learners
belonging to different departments were assigned to a group called
pedagogical group and these learners learn through conversational agents
before they are paired with peer-learning agents.
The self-learned group had a self-learning experience through
recommended learning using e-contents posted in e-learning servers based on
their learning styles. The peer-learned group did not have self-learning
experience and had no choice of assessing their self-efficacy and learnt ‘C’
programming language directly using peer-learning agents. The performance
assessments of both the groups were tested for knowledge retention and
knowledge transfer, eight questions for each of the category. The pre test and
post test consisted of 50 questions (25 on retention and 25 on transfer) which
were reviewed for effectiveness by a professor of Anna University considered
as the domain expert in ‘C’ programming language course.
4.9 RESULTS AND DISCUSSIONS
4.9.1 Learning Styles Identification
The learners are varied in nature and same kind of e-content
delivery will not have any impact of subsequent learning. Their self-efficacy
is judged using their learning styles identification. It is a proven fact that
increasing the self-efficacy helps in creating a positive atmosphere in learning
which increases the performance of the students exponentially. The proposed
system aims at increasing the self-efficacy of the students by providing
90
suitable ‘C’ programming e-contents available in documents, audio and video
lectures in mediawiki e-learning server. The learning styles are identified
using Felder Silverman learning style model using the learners’ web usage
activity. Subsequent to the identification of learning styles, they are
recommended to learn the e-contents available in e-learning servers based on
their learning styles. This increases the self-efficacy of the students
substantially which helps in mastering the ‘C’ programming course when
learning with peer-learning agents subsequently.
4.9.2 Pre-Test
The system conducted a pre-test for both self-learned group and
experimental group. At the initial stage, this test was conducted to see if the
first two groups were homogeneous in terms of the levels of achievement.
The test consisted of 50 questions (25 on knowledge retention and 25 on
knowledge transfer). The t-test results of the pre-test are shown in Table 4.8.
The results found in Table 4.8 indicate that the two groups had some
significant differences in their performances. But according to the results of
the pre-test on retention and transfer, there was not much of a difference
statistically difference between the self-learned group and the experimental
group and the difference is found to be greater than 0.05.
4.9.3 Post-Test
After 8 weeks of time for teaching and learning ‘C’ programming
language course using self-learning and peer-learning agent strategy, the
students are evaluated for their performances. The test consisted of
questionnaires as done earlier. Now, the groups are separated widely. The
self-learned group had pre-learning experience for 4 weeks of time using the
recommended e-contents available in e-Learning servers based on their
91
learning styles. They underwent pair programming learning strategy with
peer-learning agents for another 4 weeks of time. However, the experimental
group learnt with the peer-learning agents for 8 weeks of time without having
a pre-learning experience. Both the groups are evaluated using a test of
50 questions (25 on knowledge on retention and 25 on transfer). The t-test
results of the post-test are shown in Table 4.9. According to the results shown
in Table 4.9, there was a large difference which is statistically significant
between the prior-learned and experimental groups and the values are found
to be less than 0.05. The statistical difference for Knowledge Retention
between prior-learned group and experimental group is 0.0035 and the
statistical difference for Knowledge Transfer between the same groups is
0.0006. Moreover, it is also evident from the table that, the mean scores of the
prior-learned group is greater than the experimental group. Learners who had
a self-learning experience using the recommended e-contents from
e-learning servers based on their learning styles had very high performance
results compared to the learners who directly learnt with the peer-learning
agents using a pair programming strategy. The results of Table 4.9 indicate
that increased self-efficacy of the students along with pair programming
strategy using peer-learning agents had a positive effect on knowledge
retention and transfer in ‘C’ programming language.
Table 4.8 Pre-test Evaluations on the Metrics of Knowledge Retention and Transfer using t-test
Metrics Group N Mean Standard Deviation
t-testvalues
Significant difference
Knowledge Retention
Self-Learned 43 56.50 25.60 1.9215 0.416
Peer-Learned 43 55.50 25.32
Knowledge Transfer
Self-Learned 43 50.80 16.07 0.7385 0.542
Peer-Learned 43 50.40 16.67
92
Table 4.9 Post-test Evaluations on the Metrics of Knowledge Retention and Transfer using t-test
Metrics Name of the Groups N Mean Standard
Deviationt-testvalues
Significant difference
Knowledge Retention
Self-Learned 43 76.10 17.943.9489 0.0034
Peer-Learned 43 69.70 14.93
Knowledge Transfer
Self-Learned 43 88.30 5.775.1604 0.0006
Peer-Learned 43 66.40 12.95
Tables 4.8 and 4.9 shows the t-test results when the target groups
are two in number. In this case, the t-test evaluations are shown for
self-learned and peer-learned groups at a single instant of time. The final
integration testing is done using ANOVA test when all the target groups are
observed for performance levels. However, when all the three groups namely
self-learned, peer-learned and conversation agents learned groups have to be
compared ANOAV test evaluation methodology is used. Table 4.10 shows the
ANOVA test evaluation for the three groups for the post-test.
Table 4.10 Post-Test Group Evaluations using ANOVA-test Method
Metrics Name of the Groups N Mean Standard
Deviation Significant difference
Knowledge Retention
Self-Learned 43 70.256 10.045
0.0148Peer-Learned 43 68.116 9.223Conversational Agent-Learned 43 64.419 8.500
Knowledge Transfer
Self-Learned 43 71.488 11.042
0.0040Peer-Learned 43 68.116 9.223Conversational Agent-Learned 43 64.419 8.500
93
4.9.4 Discussions
In this work, new methods to enhance the learners self-efficacy are
proposed in which the self-efficacy of the learner is judged by identifying
their learning styles. The self-efficacy of the individual learner is increased by
recommending suitable e-contents available in e-learning servers namely
documents, audio and video lectures based on their learning styles. This kind
of recommendation facilitates the learners in pre-learning experiences.
Subsequently, the learners learn with peer-learning agents using a pair
programming strategy. Learning with peer-learning agents is a form of
collaborative learning where the learner and the agent can alternate the roles
of a tutor and a tutee. In this case, pair programming is used as a
teaching-learning strategy. Hence, such exchange of ideas between the learner
and the agent can have a positive effect on teaching and learning. Moreover,
the learners have increased self-efficacy by having pre-learning experience
based on their choice of e-contents.
According to Cohen (1988), the effect size of retention was 0.38,
which is a medium effect size, whereas the effect size of transfer was 0.79,
which is significant. The experimental results obtained from this work
indicate that the effect size of transfer exceeded that of retention. The
proposed learning system is found to have a positive effect on teaching and
learning strategies. Moreover, it is evident from Table 4.6 that self learned
group did extremely well compared to the other groups namely peer-learned
and conversational agent learned, since the self learned group consisted of
learners who had self learning experience on the basic concepts of ‘C’
programming language using the recommended e-contents available in
e-learning servers based on their learning styles.