Agent-Assisted Collaborative Learning
Using Agent Teamwork as a Collaborative Method to Facilitate e-Learning
Mario Mallia-Milanes and Matthew Montebello
Faculty of Information and Communications Technology, Department of Artificial Intelligence,
University of Malta, Msida, Malta
Keywords: Artificial Intelligence, Agent, Teamwork, Reasoning.
Abstract: e-Learning was a major shift in the learning medium to reach out to vast amounts of people and enable the
possibility for them to catch up on lost time or acquire new skills from the comfort of their home and at the
time most suitable to them. However numerous issues have been attributed to e-learning over the years
amongst which is the low retention rate that sheds a shadow on its validity and effectiveness. In this paper
we discuss how we propose to employ artificially intelligent agents that collaborate together and with
human counterparts to optimise the medium and extract academic benefits.
1 INTRODUCTION
Learning has been a part of man ever since he has
been. With time, learning has become more
sophisticated and necessary. This necessity has
become entrenched in society to the point that
education is now considered a basic right as much as
work and food. In the latter part of the last century
we find forward thinkers like Paolo Freire who
wedded the idea of education to politics (Freire,
2005). Fusing the idea of education into life. The
way education should be delivered to learners should
not take the form of dominance, or rather
oppression. In his ideology the teacher’s dominant,
oppressive, position over the learner is removed.
Freire insisted that the teacher should be a part of the
learning cycle. Teacher and learner should form a
synergy of continuous exchange. In this way
education becomes a gift of the teacher to the
disadvantaged student.
Moving into the twenty first century one is not
really concerned any more about the availability of
education. But now it is the delivery that has started
to become an issue. In the 1960s Seymour Papert
started publishing his visionary idea that computers
help students understand better and achieve better
(Papert, 1993). He was then heavily criticised, and
cited as an elitist, by wanting to focus attention to
privileged children over the rest who could not
afford computers (Papert, n.d.). In his seminal book
“Mindstorms: Children, Computers and Powerful
Ideas” Papert explains the development of a new
computer language called Logo which can be used
to help students better understand abstract
mathematical concepts (Papert, 1993)
Now that most governments worldwide insist
that education should be a life-long process which
helps people throughout all their life, the delivery of
education has become an issue. More people at
varying levels and ages need to be reached
effectively. So now we attempt to deliver education
on demand to broader masses, and it’s through
technology that this promise can be realised. But by
delivering on-line lessons, en-masse, we have also
unwittingly altered the student-teacher-class
relationship. The class now is not a physical class
anymore. The student is burdened with greater
responsibilities, namely that of self-motivation
(Rees, 2013). Consequently it has been noted that
learner engagement and retention suffers (Rees,
2013). The way technology has been employed to
date has done little to levitate the situation.
Computers should not be used to program learners,
but to assist them (Papert, 1993).
The allure of making automated teaching
systems quickly caught ground, hoping to make up
for sterile computer programs which are inflexible
and without emotions. Researchers are reverting to
Artificial Intelligent (AI) techniques to solve
problems presented in the educational domain. AI
has been proven to offer solutions that adapt to
Mallia-Milanes, M. and Montebello, M.
Agent-Assisted Collaborative Learning - Using Agent Teamwork as a Collaborative Method to Facilitate e-Learning.
In Doctoral Consortium (DCAART 2017), pages 3-8
3

circumstances, even to situations that have not been
encountered yet (REFERENCE). So an intelligent
technology is a natural choice where one attempts to
develop an approach that fits every learner according
to his or her needs.
Although we shall not be concerned with the
political or pedagogical aspects of education in this
work, we shall attempt to offer a solution to improve
learner engagement and retention. This cannot be
done without understanding the forces that shape
education and its delivery. So prior to delving into a
computer solution, one has to understand the way
humans learn. Then adapt a feasible technical
solution that best facilitates learning. The rest of this
paper is organised as follows. In the next section we
will delve into a number of related past research
avenues to identify the main issues related to the
area and justify the use of agents in our proposal.
Section 3 will in fact detail the formulation of the
problem at hand and the section that follows will
describe our proposed approach. Finally we dedicate
a section on results and close the paper with our
final conclusions.
2 PREVIOUS RESEARCH
2.1 Collaborative Learning
An intelligent personal environment is one that can
embrace collective knowledge of many online users
while adapting the aggregated content to a particular
user’s needs. Network and software technology
available today permit the collection of information
from various sources and then facilitate the
presentation of material to learners of mixed
abilities. Learners would be able to approach the
subject with relative ease, as there would be
assimilation between the users’ needs and the way a
course is structured. The richness of this model
provides a possibility to learners of mixed abilities
to come together and possibly even assist each other
in learning.
The advent of e-learning has made “mass education”
possible. Computers are available and powerful
enough to handle the ever-growing demands we put
on them. It is the techniques, which facilitate
learning that normally, and understandably, lag
behind. The first thing that strikes home when
thinking of such a situation is the enhancement of
the learning process which is able to assist students
in their studies.
The model that shall be proposed in this short essay
is based on the seminal idea propositioned by
Professor Matthew Montebello (Montebello, 2014).
Professor Montebello argues that crowd sourcing is
a valuable tool to assist a learner in his journey of
instruction.
Essentially a learner needs to interact with his
environment, whether it is another person, a
software agent or a machine. Learning would be
greatly enhanced by interaction. So the proposed
system should be able to:
• Support interaction;
• Assist in scaffolded learning;
• Transform data into relevant knowledge.
Interaction with an artificial environment requires
the solution of multiple problems. The most suited
approach to the situation will be the utilisation of
software agents. Multiple software agents can be
employed to interact with each other and human
operators.
2.2 Why Use Agents?
When applying computer systems to assist real life
situations the interaction between various variables
can be complex and at times unfeasible to model in
traditional ways. This becomes especially true when
human interaction is involved, and computer
systems assist, in a very ubiquitous way, humans
through the task.
Agents are generally designed to be small,
disjoint programs that work in tandem to solve a
complex situation. Their simplicity and
collaborative features makes them more adept to
such studies. In situations where distributed
computation or communication between components
are required agents fit the bill perfectly. Moreover
agents are capable of reasoning about their
environment (AgentBuilder, n.d.).
2.3 Standards
Standards help developers build products which are
interoperable. In the case of agents interoperability
is a mandatory feature as communication is
necessary between each agent.
Currently there are two popular standards, FIPA
and OMG-MASIF. FIPA (Foundation for Intelligent
Physical Agents) was set up in 1996 specifically to
produce standards for agent systems. It seems to be
that FIPA is the leading standard. FIPA focuses on
agent architecture and interoperability (Cao & Das,
2012).
DCAART 2017 - Doctoral Consortium on Agents and Artificial Intelligence
4

OMG-MASIF was formed a year after FIPA.
The Object Management Group (OMG) released a
document in 1997 called Mobile Agent System
Interoperability Facilities (MASIF). This document
proposes a specification for communication between
agents (Cao & Das, 2012).
In this work the FIPA standard shall be followed.
This decision was taken on the basis that many tools
are FIPA compliant.
2.4 Agent Building Tools
When it comes to building agents there are a lot of
tool kits that can be of assistance. There are many
toolkits that can be used to develop multi-agent
systems. The use of tool kits will allow focus on the
domain of application rather than the building of the
agent itself. It is rather difficult to choose between
the different tool kits as can be attested through
Wikipedia.com (Wikipedia, 2015). The choice was
then narrowed down by choosing FIPA compliant
tools. The Foundation for Intelligent Physical
Agents (FIPA), an IEEE organisation was formed in
1996 to produce software standards for
heterogeneous and interacting agents and agent-
based systems (FIPA, n.d.).
In order to reduce interfacing complexity the
JADE toolkit shall be chosen as the preferred toolkit.
The Java Agent Development Framework (JADE)
was purposefully developed in Java to ensure cross-
platform compatibility of the package. Agent
development occurs through middleware and a
graphical user interface. Moreover implementation
can be distributed across different machines running
different operating systems. JADE is free to use
under the Lesser General Public License version 2.
It boasts of a large community of developers
backing it up. Telecom Italia are the copyright
holders of the software (JADE, 2015).
2.5 Reasoning Mechanisms
Computational agents require rational behaviour to
be of some use as autonomous agents in a system.
The approach to simulate rationality is naturally a
complex task (Rao & Georgeff, 1995). Much of
what we have today bases itself on the study of
human organisation. One of the most popular
models in use today is the belief-desire-intention
model. The belief-desire-intention is a very popular
model of reasoning. And many of the models in
place today are either faithful implementations or
base themselves in it.
The model was developed by Michael Bratman
as a way of explaining the future-directed intention
by humans (Bratman, 1999). It has its roots in
philosophy where one tries to understand practical
reasoning in humans. Practical reasoning is
directed towards actions. This is a process where
one has to figure out what to do. Practical reasoning
comprises the weighing of considerations,
sometimes antagonistic, against the beliefs, desires
and values one has (Wooldridge, 2000). Cognitive
science forms the basis of the approach to reasoning
and as a result human awareness can be analysed
and translated successfully into a BDI framework
that can be used by software agents (Dunin-Keplicz
& Verbrugge, 2013).
The BDI model was first developed as a model
for understanding human reasoning. But it found its
way into computer science and is actively used in
programming software agents (Georgeff, et al., n.d.).
It focusses on beliefs, desires and intentions as a
way of solving problems that face an agent. Each
action performed by an agent, human or otherwise,
can be separated into two parts. A planning part, and
a doing part. In BDI the planning part of the action
is separated from the doing aspect of the same
action. Agents programmed using this framework
are able to balance the time spent in planning against
the time spent doing (Bratman, 1999).
It is worth remaking that this model, developed
in the 80’s is considered dated. Moreover Michael
Georgeff argues that it cannot reach the rigour of
modern day demands (Georgeff, et al., n.d.). But
despite this, the BDI model is still extensively used
in frameworks. Unless the outcome from our
research dictates otherwise, this work shall be based
on this model. Primarily because of its extensive
use and the availability of the number of frameworks
that support it.
3 PROBLEM FORMULATION
3.1 The Current Situation
When one follows through the evolvement of
learning it cannot be said that nothing has been done
through the ages. But until 30 years ago learning
has not been exclusively limited to obligatory school
in many countries. Many a government, both locally
and abroad, have realistically emphasised that
learning is a life-long process.
It would be fitting at this point to start off with
defining an important point, that of the meaning of
Agent-Assisted Collaborative Learning - Using Agent Teamwork as a Collaborative Method to Facilitate e-Learning
5

learning. Chen and Wei state, “learning is an active,
interactive and constructive social process” (Chen &
Wei, 2004) It entails a synergy between a number of
activities that facilitate the acquisition of new skills
or knowledge. Ultimately learning can also be
understood as that activity a person does to acquire
new skills mostly through interaction with others.
Thus making collaboration with teachers and peers
an essential part of the process. This has been seen
to be concomitant with the actual pedagogy of the
process of knowledge transfer itself. Interaction
actually places the learner within a context of
knowledge application. Technology seems to ably
remove the interaction concept and consequently
reduces student retention (Montebello, 2014).
Technology has the ability to greatly assist
learning (Chen & Wei, 2004). But despite this,
learning through technology has not yet reached its
full potential. If one takes a look back in time, it can
be noted that the use of technology to assist learning
is not an innovative idea (Papert, n.d.). But
technology was mainly used to increase the spread
of learning further. The advent of radio, and
subsequently that of television, has inspired many to
introduce programmes that help people acquire skills
such, as the learning of new languages, at their own
pace. None are apparent today, and the success of
such initiatives is dubious (Rees, 2013). But the
seed of using technology to bring learning closer to
more people was sown. Later on in the 90’s with the
advent of the Internet, that permitted global
connectivity, the idea of distance education started
to surface again. E-learning started to become a
buzzword and has been embraced by many
educational institutions. This enabled institutions to
reach far beyond the limitations of their physical
capacity. Material took the form of videos, sound
clips and soft text. Once more results from various
studies are being to show that despite the technology
is promising, the end results are not (Rivard, 2013).
e-Learning should have given the student more
freedom, but it also burdened him with more
responsibility. Commonly teaching material would
just be converted from standard printed material to a
digital form and making them available to all.
Material is in no way customised to suit different
learning styles. Normally a one-size-fits-all
situation is delivered. This, although convenient and
very cost effective, is not ideal.
3.2 Tools
The most common tools in use today for the support
of e-learning environments are typically, e-mail,
material presentation packages, and social media and
chat rooms/blogs. Taking a closer look at these
tools one can identify a potential issue. They are
very able at delivering material but they cannot
adapt to the learner’s style or wants. And mostly
remove a crucial element that of collaboration. In
other words the student is not being engaged in a
normal, or rather, natural way within his learning
environment.
The Social Learning theory, expounded by
Albert Bandura, suggests that people learn expressly
by interacting with their surroundings. A learner
follows on by observing things that happen around
him. He picks up ideas, shares them and develops
them further. This action of collaboration then helps
to develop the identity of the learner by interacting
with the environment that projects roles and values
on the person. Finally identity construction helps
motivate social participation (Bandura, 1971), (Paul,
2012), (Orit, et al., 2015). As a corollary what has
just been said, learners are demotivated and leave.
Hence the low retention rate when students are
exposed to an isolated, one-size-fits-all environment
(Rivard, 2013).
4 PROPOSED WORK
Artificial Intelligence is not a new proposal to
education, especially e-learning. But it must be
added that the impact on e-learning has not been
significant (Corbett, et al., 1997). This may be due
to the fact that the personal adaptation of knowledge
is still in its infancy.
The proposed research shall study the
collaboration capabilities of independent multi
agents and their capacity to solve problems as a team
within an e-learning environment. Moreover the
human element in this design shall be taken into
account. Collaboration comes as a result of
commitment from each participating agent, human
or otherwise, sharing similar beliefs, desires and
intentions.
4.1 Objectives
The environment selected for the study shall be an e-
learning environment which will entail close
cooperation between a system of agents and a
human actor. Naturally we have to ensure that there
is a binding factor between human and artificial
agents which will lead to teamwork. The loop in this
study will close when team work will eventually
facilitate e-learning.
DCAART 2017 - Doctoral Consortium on Agents and Artificial Intelligence
6

4.2 Research Questions
From the research objectives the following questions
are placed:
• How can commitment be negotiated between
software agents in order to improve group
interaction and problem solving capabilities?
• Could agents adjust properly to human
commitment to the same task?
• What will happen if the values of a software
agent will start to differ from that of a group?
5 RESULTS
5.1 Expected Outcome
Most of the work shall comprise the building of a
multi-agent environment and testing it out to see
whether the above research questions can be met.
The data collected from the experiments set up will
be analysed using discrete event simulation
modelling techniques.
This study is expected to reach two goals. The
primary goal is that of studying collaboration
between agents and its outcome. In this case the
collaboration of agents is elicited through their
beliefs, desires and intentions. The formation of
teams will happen only if the BDI are close enough
to make an agent co-operate. For this study, BDI
will not be restricted, but an agent will have to form
its own data set as part of its experience. Hence it
may be harder to have agents to cooperate without
“forcing” them to do so.
The second goal is tightly coupled with the first.
Can human learning really be improved if a closely
knit group of agents collaborate with a human?
5.2 Overview of the Proposed Model
As stated earlier agent interaction shall be studied
within an e-learning environment. In order to adapt
the educational content to the learner one has to be
aware of a number of situations, namely:
• The profile of the learner;
• The domain of knowledge being
experienced;
• The needs and wants of the learner.
The resultant outcome should be the intersection of
relevant material presented in such a way as to
satisfy the learner. The domain of knowledge can be
sought through a variety of sources, through
interaction with human players or computer sources.
In essence when collaborating, a learner should not
be bound to the medium delivering responses. First
we shall start by describing a system of agents that
needs to be set up. The idea is to have a set of
agents that need each other’s support to work
properly. There will be more than one agent for
each of the types listed below. Our system of agents
shall comprise the following:
• Knowledge Agent has knowledge in a
particular area.
• Knowledge Server Agent stores, retrieves,
and manages knowledge; answer queries; and
provides information by inferring or
reasoning using the stored knowledge bases.
• Interface Agent serves as an interface to
learners, monitors and learns from the user’s
actions, and then functions as an intelligent
assistant.
• Coach or Tutor Agent provides guidance to
assist in the learning process.
• Mediator Agent coordinates the activities of
other agents and resolves conflicts between
them.
• Knowledge Management Agent provides the
high-level coordination of knowledge
activities, such as creation, assembly,
manipulation, and interpretation of
knowledge, within either an individual or a
collective project.
• Information Search Agent searches for
specific information and sends the results
back to learners.
• Directory Agent points to an appropriate
agent, service, or resource.
• Mentor Agent is envisaged as acting in a
rather analogous way in the learning
environment, as a kind of coach for the
higher-level strategies of learning.
(Chen & Wei, 2004)
5.3 Comparison to Actual
Implementations
Corbett et al, in their article “Intelligent Tutoring
Systems” propose a model that may be used to help
the design of such a system (Corbett, et al., 1997).
Moreover, they singularly cite a successful project
that has been undertaken in Scotland called
SCHOLAR. The model divides a system into four
areas, each taking care of distinct parts of the
learning system. This model can be followed on, but
a more dynamic approach to the learning system will
be taken.
Contrary to what we are trying to attain in this
Agent-Assisted Collaborative Learning - Using Agent Teamwork as a Collaborative Method to Facilitate e-Learning
7

work the material developed for SCHOLAR was
manually assembled. In our case an agent will be
allowed to form its domain. The creators of this
project claimed that they observed a correlation
between SCHOLAR use and attainment of results.
Moreover they also claim that there has been an
observed improvement in autonomous learning.
Students who used the software in the evenings and
weekends have achieved better results than peers
who used the software exclusively in class with their
teacher. Curiously the report waters down its claims
as the authors defend themselves by saying that it
cannot be said that all the achievement can be
attributed to the use of SCHOLAR.
But much can be taken from this study which
really attempts to involve students by giving them a
system which helps them through their studies. In
this work we attempt to show whether artificial
intelligence can really come to the rescue of e-
learning.
6 CONCLUSIONS
In essence interaction is an integral part of learning.
People interact, and exchange ideas and grow
intellectually through this process. Removing
interaction greatly reduces interest and motivation.
So in order to improve the chances of success
technology has to be able to maintain interaction
while also being able to transform data into
knowledge. The information has always been there,
in some form or other. Digitally it is now even more
accessible. The only remaining issue is that of
transforming data into knowledge in such a way as it
engages and retains the learner (Camilleri P., 2015).
In this work we are going to seek technical solutions
that address this issue properly. (Rivard, 2013)
REFERENCES
AgentBuilder, n.d. Why, When, and Where to Use
Software Agents. [Online] Available at: www.agent
builder.com [Accessed 29 July 2015].
Bandura, A., 1971. Social Learning Theory, s.l.: General
Learning Corporation.
Bratman, M. E., 1999. Intentions, Plans and Practical
Reason. 1st ed. s.l.:CSLI Publications.
Cao, J. & Das, S. K., 2012. Mobile Agents in Networking
and Distributed Computing. 1st ed. New Jersey: John
Wiley and Sons.
Chen, P. & Wei, D., 2004. Knowledge Management for
Agent Based Tutoring Systems. In: Designing
Distributed Learning Environments with Distributed
Software Agents. 1st ed. London: Idea Group
Pubishing, p. 311.
Chen, P. & Wei, D., 2004. Multi-Agent Based Educational
Environment. In: Designing Distributed Learning
Environments with Distributed Software Agents. 1st
ed. London: Idea Publishing Group, p. 311.
Corbett, A. T., Koedinger, K. R. & Anderson, J. R., 1997.
Intelligent Tutoring Systems. In: M. Helander, T.
Landauer & P. Prabhu, eds. Handbook of Human
Computer Interaction. Pensylvania: Elsever Science
B.V., pp. 849-874.
Dunin-Keplicz, B. & Verbrugge, R., 2013. Teamwork in
Multi Agent Systems: A Formal Approach. s.l.:J.
Wiley and Sons.
FIPA, n.d. Foundation for Intelligent Physical Agents.
[Online] Available at: www.fipa.org [Accessed 30
July 2015].
Freire, P., 2005. Pedagogy of the Oppressed. 30th
Anniversary Edition ed. s.l.:The Continuum
Internationa Publishing Group Ltd..
Georgeff, M., Pell, B. & Pollack, M., n.d. University of
Oxford, Department of Computer Science. [Online]
Available at: http://www.cs.ox.ac.uk/people/michael.
wooldridge/pubs/atal98b.pdf [Accessed 31 July 2015].
JADE, 2015. JAVA Agent DEvelopment Framework.
[Online] Available at: http://jade.tilab.com/ [Accessed
30 July 2015].
Montebello, M., 2014. Artificial Intelligence to the Rescue
of e-Learning. Barcelona, Spain, EDULearn 14.
Noe, R. A., n.d. Learning System Design: A Guide to
Creating Effecitve Initiatives. s.l.:Foundation, SHRM.
Orit, M. et al., 2015. Harnessing Technology for
Promoting Undergraduate Art Education: A Novel
Model that Streamlines Learning between Classroom,
Museum, and Home. IEEE Transactions on Learning
Technologies, January-March.8(1).
Orwell, G., 1949. Nineteen Eighty-Four. 1st ed. London:
Secker and Warburg.
Papert, S. A., 1993. Mindstorms: Children, Computers
and Powerful Ideas. 2nd ed. s.l.:Basic Books.
Papert, S., n.d. Works by Papert. [Online] Available at:
www.papert.org [Accessed 3 May 2015].
Paul, A. M., 2012. Machines are Taking Over. The New
York Times Magazine, 14 September.
Rao, A. S. & Georgeff, M. P., 1995.
BDI Agents from
Theory to Practice. Melbourne, Proceedings of the
First International Conference on Multagent Systems.
Rees, J., 2013. The MOOC Racket. [Online] Available at:
http://www.slate.com/articles/technology/future_tense/
2013/07/moocs_could_be_disastrous_for_students_an
d_professors.html [Accessed 26 April 2015].
Rivard, R., 2013. Measuring the MOOC Dropout Rate.
[Online] Available at: www.highered.com [Accessed
26 April 2015].
Wikipedia, 2015. Comparison of agent-based modeling
software. [Online] Available at: https://en.wikipe
dia.org/wiki/Comparison_of_agent-based_modeling_
software [Accessed 30 July 2015].
Wooldridge, M., 2000. Reasoning about Rational Agents.
1st ed. Cambridge Massachusetts: The MIT Press.
DCAART 2017 - Doctoral Consortium on Agents and Artificial Intelligence
8