Abstract

Machine Learning Approach for an Advanced Agent-based Intelligent Tutoring System
Roya Aminikia
Learning Management Systems (LMSs) are digital frameworks that provide curriculum, training
materials, and corresponding assessments to guarantee an effective learning process. Although
these systems are capable of distributing the learning content, they do not support dynamic learning
processes and do not have the capability to communicate with human learners who are required to
interact in a dynamic environment during the learning process. To create this process and support
the interaction feature, LMSs are equipped with Intelligent Tutoring Systems (ITSs). The main
objective of an ITS is to facilitate students’ movement towards their learning goals through virtual
tutoring. When equipped with ITSs, LMSs operate as dynamic systems to provide students with
access to a tutor who is available anytime during the learning session. The crucial issues we address
in this thesis are how to set up a dynamic LMS, and how to design the logical structure behind an
ITS. Artificial intelligence, multi-agent technology and machine learning provide powerful theories
and foundations that we leverage to tackle these issues.
We designed and implemented the new concept of Pedagogical Agent (PA) as the main part of
our ITS. This agent uses an evaluation procedure to compare each particular student, in terms of
performance, with their peers to develop a worthwhile guidance. The agent captures global knowledge
of students’ feature measurements during students’ guiding process. Therefore, the PA retains
an updated status, called image, of each specific student at any moment. The agent uses this image
for the purpose of diagnosing students’ skills to implement a specific correct instruction. To develop
the infrastructure of the agent decision making algorithm, we laid out a protocol (decision tree) to
select the best individual direction. The significant capability of the agent is the ability to update
its functionality by looking at a student’s image at run time. We also applied two supervised machine
learning methods to improve the decision making protocol performance in order to maximize
the effect of the collaborating mechanism between students and the ITS. Through these methods,
we made the necessary modifications to the decision making structure to promote students’ performance
by offering prompts during the learning sessions. The conducted experiments showed that
the proposed system is able to efficiently classify students into learners with high versus low performance.
Deployment of such a model enabled the PA to use different decision trees while interacting
with students of different learning skills. The performance of the system has been shown by ROC
curves and details regarding combination of different attributes used in the two machine learning
algorithms are discussed, along with the correlation of key attributes that contribute to the accuracy
and performance of the decision maker components.