BALANCING ADAPTIVE CONTENT WITH AGENTS
Modeling and Reproducing Group Behavior as Computational System
Harri Ketamo
Satakunta University of Applied Sciences, Pori, Finland
Keywords: Adaptive systems, User modelling, Software agents, Artificial intelligence, Social behaviour.
Abstract: To ensure the quality of adaptive contents, there should be continuous testing during the development phase.
One of the most important reasons to empirically test the content during the development phase is the
balance of the adaptive framework. Empirical testing is time-consuming and in many cases several iterative
cycles are needed. In 2007 we started to develop methods of testing in a computational test bench. The idea
to speed up the production process was based on software agents that could behave like real user
community. The study shows that we can construct very reliable artificial behaviour when comparing it to
human behaviour in group level. On design phase’s usability tests, we are especially interested in group
behaviour, not on single action etc., which means that the method suits for it’s purposes.
1 INTRODUCTION
Adaptation can be seen as being high end
personalization: In adaptation, the system optimizes
the output with technologies that, in general, can be
divided into two main groups: static (indirect)
adaptation and dynamic (direct) adaptation. In static
adaptation the rules are fixed beforehand by
developers. In dynamic adaptation the system tracks
the user and optimizes the navigation paths
according to the user's behaviour. Dynamic
adaptation of system requires at the least a user
model, a context model and artificial intelligence.
(Kinshuk, Patel & Russel 2002; Brusilovsky, 2001;
Manslow 2002).
Because the idea of adaptive educational systems
is to produce individual and optimized learning
experiences the high end user models as well as
methods are relatively complex (e.g. Raye, 2004;
Lucas, 2005). Furthermore, the social dimension
should not be forgotten: In very large samples, the
most successful outcomes in group level may
contain valuable guidelines for adaptation.
Before we can ensure the quality of the
adaptation, at the least we have to know if 1) the
adaptive framework is in tune, 2) is there out layered
content and 3) are the learning paths really unique?
It is possible to theoretically check the first three
questions by constructing an algorithm that
computes all the possible combinations, but there
will be no such system that can compute the problem
in a reasonable amount of time within large
domains.
Another solution to this is to model human
behaviour as a system and then run the learning
material using these artificial users. In our study the
behaviour of software agents are constructed
according to social behaviour: We have constructed
several archetypes of a user by clustering the
behaviour of human users.
2 RESEARCH TASK AND
PROCEDURE
2.1 Research Task
In this study, the adaptation is studied in terms of
Complex Adaptive Systems: self-organization,
entropy and emergence. These concepts describe and
define the high level system properties. In a
computational system that aims at simulating group
behaviour, focusing on high level phenomena might
incur the best outcome.
Earlier results, received from preliminary studies
of the application (e.g. Ketamo, 2008) are referenced
in this study. The computational core is based on the
author's previous work that had been applied, for
example, as the background of an educational game
291
Ketamo H.
BALANCING ADAPTIVE CONTENT WITH AGENTS - Modeling and Reproducing Group Behavior as Computational System.
DOI: 10.5220/0002781502910296
In Proceedings of the 6th International Conference on Web Information Systems and Technology (WEBIST 2010), page
ISBN: 978-989-674-025-2
Copyright
c
2010 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
series (e.g. Ketamo & Suominen 2008; Ketamo
2009).
2.2 Material in Testing
Mathematics Navigator is a product family
distributed by Otava Publishing Company Ltd..
Mathematics Navigator is based on an adaptive
content management system (Figure 1) and content
objects. Mathematics Navigator is a high-end
platform designed to produce dynamically adaptive
and self-organizing content. Dynamic adaptation is
done by varying the sequences of course elements
(bits of theory, examples, exercises) and supplying
individual learning paths for the students. In this
study Mathematics Navigator is used as a test and
evaluation platform.
Adaptation of content is done according to a
user's / student's learning and studying behaviour.
This kind of modelling requires at the least 1) a user
model that records the skills and learning of each
user and 2) a domain or context model on the
relations between the bits of information of learning
content. Mathematics Navigator gathers information
on a student's actions and, based on this information,
creates and modifies an understanding of his/her
mathematical competence. Mathematics Navigator
adjusts the exercises and content to support the
development of students. On the user interface
(Figure 2), a student has a graphical representation
of the development of his/her learning profile.
Figure 1: The high level architecture of Navigator.
The user interface of Mathematics Navigator is
based on a menu-bar and three main areas (Figure
3). On the left side of the interface is a table of
contents and a content-related competence profile of
the user. The table of contents presents two different
views of the content: 1) a traditional book-like table
of contents and 2) an exercise-adapted table of
contents.
The exercises are presented one at a time in the
bottom-right corner of the interface. The user can't
proceed to a new exercise before the current one has
been answered by picking an answer from a total of
4 alternative answers. The exercises are selected to
support an individual user's learning needs. There
are no fixed paths for learning: everything is based
on the student's competence profile and estimated
need for practice and content. Guiding factors in
exercise selection are: 1) course structure (a
traditional table of contents), 2) the measured and
estimated learning abilities and areas of weaknesses,
and 3) critical points, derived from the learning
community's actions.
The competence profile values are indicated by
colours that vary from red (insufficient skills or
skills not yet estimated) via yellow to green (good
skills). Those skills mastered and measured within a
certain theme will be transferred with certain
estimates to other themes requiring similar (by
proximity or by hierarchy) skills.
Figure 2: The user interface of Mathematics Navigator and
a basic mathematics course (in Finnish).
The entire content - theory, examples and exercises -
is based on the idea of content objects. A content
object is a unit containing the smallest possible bit of
content that can be used independently without the
support of other content objects. Naturally, products
are based on content objects that strongly support
one another. The content objects are described by 1)
detailed rank ordered keywords (tags) that define the
structure of content and 2) single relations between
objects in a preferred order. The preferred order is
calculated as a state of semantic network, formed
according to tags. The method can be described in
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
292
general in terms of weighted proximity between
objects (Figure 3).
Figure 3: Defining proximity between the content objects.
The calculated proximities between the content
objects can be visualized as weighted graph (Figure
4) which basically visualizes the whole idea behind
the content model.
Figure 4: Calculated proximities between content objects
visualized as weighted graph.
This kind structure and content model offers
numerous possible paths through the material. A
special feature of the modelling is the capability of
self-organization: The clusters inside the material
were self-organized according to mathematical
definitions and context. This self-organization is a
key feature of Mathematics Navigator and it
remarkably reduces the costs of producing adaptive
content. In this study the aim is to reduce the costs
of empirical testing, but we can find similarities in
information modelling between Mathematics
Navigator and the test method.
In Mathematics Navigator, a learner's
performance can be estimated in relatively complex
domains with very detailed and extensible user
models, combined with a learnable system. Of
course, this kind of modelling takes some time: the
system must learn the performance level of the user.
Empirically formed decision trees were used when
the system had not yet constructed a detailed enough
profile about the learner in order to dynamically
adapt the content and exercises to his/her learning
needs.
Empirical testing of such systems takes a lot of
time and in many cases iterative cycles are needed
that incur new time costs on development.
2.3 Procedure
This study is a design study with actions like the
design of the system, implementation of the system,
empirical testing and evaluation. The design and
implementation of the system was done in the
autumn 2007-winter 2008 and the design was
revised during the summer 2008 as a consequence of
pre-tests.
Empirical testing (n=447) was done in two
phases: the data from human users was collected
during the content development projects of
Mathematics Navigator in 2005-2008. The data
collected during those studies was meant to support
the development of those specific projects.
However, the data was found to be complete enough
to serve as empirical data for this study. This
existing data makes it possible to run tests while the
work was in progress.
The empirical testing reported in this study is
based on the autumn 2008 -version of the system.
Tests were run for up to 1000 artificial workers, but
in most cases the system found its balance using less
than 100 artificial workers. The limit of the test
sample size is defined by an estimator meant to
measure the possibility of changing the visualization
structures (Figures 6-8). In fact, the testing costs
would not be higher even if we run the tests using
millions of artificial workers. Naturally, the need for
computing time would have increased.
The system is implemented with software agents
using the same application interface as the host
system, which means that they can act as human
users. The host system that they are logged in does
not recognize which connections are from artificial
users and which are from human users. This method
requires that the user interface should be described
in detail. This description focuses only on
interactions, not on visualisation. Because
Mathematic Navigator is a Web Service, the user
interface was defined with SOAP and WSDL and no
further definition for this study was needed.
BALANCING ADAPTIVE CONTENT WITH AGENTS - Modeling and Reproducing Group Behavior as Computational
System
293
3 RESULTS
In this study, artificial users represent archetypes of
human users. Artificial users are based on complex
modelling which means, that there are several
archetypes of users as well as variance inside the
archetypes. The common denominator for all is that
everything is based on human user- and group
behaviour.
In the first step, data collected from human users
is used to construct a model about behaviour in
Mathematics Navigator. This model is formed from
a group behaviour point of view. The key elements
in the model were patterns of UI actions and
particularly the probabilities and uncertainties of
following action according to observed patterns. In
the model the averages and variances were not the
key point. In fact the modelling is not statistical at
all: The model aims to predict behaviour, not to test
any hypothesis. The patterns of behaviour were
clustered in order to build the archetypes of users,
which increases the similarity of artificial users to
human users.
One interesting find during the pre-tests was that,
when increasing uncertainty in decision-making
(increasing variance in decision in statistical terms)
up to specified limits, the system could point out
possible problems in a shorter time. This finding is
relatively straightforward: Because we are not
looking for statistical evidence, we can focus on
pointing out problems without proving them
statistically. This helps us in two ways. Firstly, the
computational requirements were decreased.
Secondly, the estimator, calculated to detect when
the simulation is ready, could be used in a more
reliable way. The challenges of avoiding behaviour
that never happens with human users are great. On
the other hand, we never can be sure how users
behave in digital environments. At this point, we
have to accept the fact that we might construct
agents that fail when compared to human users'
behaviour.
As a result of modelling, a finite state system
with a structure equation about action probabilities,
uncertainty and disorder was formed. In Figure 5
one part of the model behind the decision-making
system (probability network) of the artificial user is
visualized. User and context modelling was solved
technically by constructing a dynamically extensible
Semantic Networks. The user model could be
exported e.g. as an XML Topic Map and could be
manipulated by xPath.
Figure 5: Adaptation schema as a system model.
The three key elements of Complex Adaptive
Systems are self-organization, entropy and
emergence. All of these can be found in the
visualized learning paths of Navigator's users
(Figure 6). Similar results could be achieved when
the artificial user's behaviour is visualized (Figure
7). In the figures the paths are formed by connecting
content objects browsed during use of Mathematics
Navigator. In the figures the nodes are essential
pieces of content that can be defined as being 'the
backbone of the domain'. The connectors show the
paths that users had passed through. When reading
the figures, the assumption is that there were
numerous connectors between nodes, but some
directions or patterns are more frequent than others.
In terms of Complex Adaptive Systems, this
systematic organization of connectors pointed out
from disordered can be called emergence. In Figures
6 and 7 only connectors that are stronger than
average are presented. Learning paths and
progression can be read from left to right. In Figures
6-8, only the most significant paths are visualized. If
all paths were to be visualized, the figures would not
be readable at this scale.
The most important finding is that the
visualizations remain the same when comparing a
human user's visualization and visualizations
constructed according to the behaviour of artificial
users.
Naturally there are differences: artificial user's
visualization (Figure 7) is rougher than a human
user's visualization (Figure 6), because it is based on
archetypes.
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
294
Figure 6: Emergence formed by a human user's learning
paths.
Figure 7: Emergence formed by artificial user's learning
paths – current version (partial visualization in order to
show similarities to visualization in Figure 6).
However, during the development of the
algorithm, the roughness of the group behaviour of
artificial users has decreased (compare to figure 8).
When in pre-test version of artificial users (figure 8)
there were relatively few strong navigation paths, in
current version (figure 7) the visualization of
behaviour remains more versatile.
The visualizations indicate that, at a general
level, human users and artificial users cause similar
emergences and as a group they behave in a similar
way. However, this similar behaviour cannot be seen
at an individual level. We have to focus on group
behaviour at a very general level.
Figure 8: Emergence formed by artificial user's learning
paths – pre-test version (partial visualization).
From the product development point of view, the
question is about how modelling and visualizations
can be used is to check if the framework is in tune.
One common problem caused by the complexity of a
system is visualized in Figure 9: The definitions of
the content objects construct a framework that can
point out a 'local minimum' on competence for an
unexpectedly long time. In visualization this is seen
as a connector back to the object itself.
Figure 9: Problems with content: the framework is not in
tune.
Figure 10: Unexpected, but correct re-organization of the
content.
Another example that could sound like a problem
is visualized in figure 10, where we can see a strong
parallel navigation patterns. At first it seems an
unexpected behaviour of the system. In fact, in the
visualization we can see a refresh path, organized for
those of the artificial users whose mathematical
skills were not developed enough during the course.
The refresh path was launched by 2-5 last exercises
in the course.
Artificial users cannot correct such problems but
authors of the content can check and improve the
definitions of content objects. In most cases such an
event is fixed by adding relevant keywords into the
material.
BALANCING ADAPTIVE CONTENT WITH AGENTS - Modeling and Reproducing Group Behavior as Computational
System
295
4 CONCLUSIONS
The aim of the study was to construct a method of
reducing time and workload costs when developing
new courses in the Mathematics Navigator platform.
The main questions to be studied computationally
were the following: 1) Is the adaptive framework in
tune? 2) Is there out layered content? and 3) Are the
learning paths really unique?
One of the efficient solutions was artificial users
that represents archetypes of human users. The
artificial users was implemented as software agents
using the same application interface as the host
system, which means that they can act as human
users. The host system that they are logged in does
not recognize which connections are from artificial
users and which are from human users.
Questions 2 and 3 could be studied in detail:
Learning paths are relatively unique but not
randomized. There is clear emergence as well as
remarkable entropy in paths formed by human users
and artificial users. Also, out layered content can be
noticed easily. However, Question 1 (Is the
framework in tune?) was challenging: Being in tune
is not only a computational task, it also deals with
curriculum. In the current system, the curriculum
was not in focus and therefore we could not say that
the framework is in tune according to the
curriculum. At best we can say that the framework is
very probably in tune in a solution specific
mathematical context.
In general, the studies show that we can
construct similar group behaviour between human
and artificial users and the saved resources in
development can be significant. In terms of
resources, the development project's savings are
estimated to be as high as 2-4 months of a
developers work and 2-3 months of total time in
project schedule.
The next phase of the study is to apply the
method to other relevant contexts, for example in
game development or in applications of social
media. The game development applications could
use relatively similar mechanics described in this
study. However, social media has brought an
increasing need for intelligent agents, that could
search and pick the most important pieces of content
out of the enormous information overload.
Teachable software agents could be used as
personal media readers that could pick up those
personally meaningful messages. Furthermore,
behaviour on reading messages is relatively close to
navigation behaviour in general level: In both cases
there are individual behaviour patterns and what is
different between individual behaviour and
normal/average behaviour explains something from
the goals of the behaviour.
REFERENCES
Brusilovsky, P. (2001). Adaptive Hypermedia. User
Modeling and User-Adapted Interaction. Vol 11, pp.
87-110.
Ketamo, H. (2009). Semantic networks -based teachable
agents in an educational game. WSEAS Transactions
on Computers, vol 8(4), pp. 641-650.
Ketamo, H. (2008). Cost Effective Testing with Artificial
Labour. In proceedings of 2008 Networked &
Electronic Media Summit. Saint-Malo, France, 13-
15.10.2008, pp.185-190.
Ketamo, H. & Suominen, M. (2008). Learning-by-
Teaching in Educational Games. In proceedings of Ed-
Media 2008. 30.6.–4.7.2008, Vienna, Austria., pp.
2954-2963.
Kinshuk, Patel, A. & Russell, D. (2002). Intelligent and
Adaptive Systems. In H.H. Adelsberger, B. Collins &
J.M. Pawlowski (ed.). Handbook on Information
Technologies for Education and Training. Germany:
Springler-Verlag, pp. 79-92.
Lucas, P.J.F. (2005). Bayesian network modelling through
qualitative patterns. Artificial Intelligence, Vol 163 (
2), pp. 233-263
Manslow, J. (2002). Learning and Adaptation. In Rabin,
S. (ed.) AI Game Programming Wisdom.
Massachusetts: Charles River Media, Inc., pp. 557-
566.
Reye, J. (2004). Student Modelling based on Belief
Networks. International Journal of Artificial
Intelligence in Education, Vol 14, pp. 63-96.
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
296