An Agent for Selecting Learning Strategy

cobblerbeggarΤεχνίτη Νοημοσύνη και Ρομποτική

15 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

57 εμφανίσεις

An Agent
for Selecting Learning

Strategy



Sassine C. Abou
-
Jaoude
and


Claude Frasson



Computer Science Department

University of Montreal

C.P. 6128, Succ. Centre
-
Ville

Montreal, Quebec Canada


{jaoude,

frasson
}@iro.umontreal.ca




Abstract

Traditional ITS presents very little flexibility regarding the pedagogical strategy they use.
Although research proves that the proper choice of the pedagogical strategy could highly
effect the learning process, ITS are usually developed followi
ng a fixed strategy that would
basically apply to all learners. This paper not only represents the possibility of a new
generation of ITS with multiple pedagogical strategies, but also introduces the concept of a
pedagogical strategy selector agent that w
ould evaluate the student’s model and his/her
performance in order to pick the best tutorial strategy that would suit him/her.


Keyword

Intelligent tutoring systems, Strategy Selector Agent, Learning Strategy, Learning Profile,
Learning by Disturbing.













Introduction

Why do we need to select pedagogical strategies and why not stick to the same strategy
through the whole learning session? What do we need to know in order to make a good
switch from one pedagogical strategy to the other?

How often should this switch take place?
Does the amelioration in performance justify the complexity introduced with the introduction
of a strategy selector ? We will discuss all these questions through this paper and provide
answers as much as possible.
We do not claim to have answered all these questions, and we
confess that, as a result of this research we also manage to raise further questions of our own
that needs to be answered. We feel that further research should be made, as this subject is
really

important in ITS and we believe that this is the impression that the reader will be
finishing this paper with.

Traditional ITS have been designed on the basis of a
tutor
-
tutee

strategy

whereby the system
is the tutor (ie the expert) and the learner is t
he tutee, who is usually supposed to assimilate
the information presented to him in a learning session. Those systems were composed of four
main components : the domain expertise, the pedagogical expertise, the student model and
the interface [Aïmeur, Alex
e & Frasson, 1995]. The development of new learning and
teaching strategies and the introduction of software agents started influencing many computer
based applications among them the ITS. And eventually, the idea of a tutor
-
tutee started
being replaced w
ith innovative new challenging strategies. Moreover the possibility to switch
from one strategy to the other started being felt.

Why is that ?

The extreme flexibility of human teachers to switch teaching methods have proven effective
in the teaching par
adigm. Till now, human acquire any given information from a human
more efficiently than any computer, that is supposedly the human teacher and the machine
have the same knowledge level. The whole reason lays in the idea that human can change
their teaching

methods and strategies on the spot. They can switch to simpler ideas if they felt
that the subject is being too complicated, they can provide explanations, introduce examples,
challenge the learner with some questions, jump to simpler preliminary notions,

and many
other teaching direct strategies that human teachers use daily. The only problem is that
teachers can have this behaviour while teaching with few learners as it is then possible to
know the status of their knowledge. The idea of the strategy sele
ctor is nothing but a modest
essay to represent this inner human flexibility in the traditional ITS. In that case, it can be
adapted to each learner.

The other analogy that exists between the human teacher and the ITS is based on the fact that
human teache
rs constantly evaluates the learners in front. In the world of ITS this would be
nothing but a well informative student model. The student model we used in our work, take
into consideration the knowledge level and the learning profile. We have considered t
wo
families of pedagogical strategies : cooperative strategies and direct strategies. Finally we
managed to run our “mini” ITS with two operating modes. The first being operating with a
strategy selector agent that would choose the proper strategy upon eva
luation of the student
model. The second, with no strategy selector agent, whereby the user picks the strategy.

Our ultimate goal was to check if the agent selector would improve the efficiency of the ITS
and if so by how much. To do that we compared the
performance of the students under both
operation.

Did we get any results ? Yes. The results we got will be presented at the conclusion of this
paper. At this point, we would like to mention that the results were highly encouraging.

The first thing to star
t with at this level would be the notion of pedagogical strategies in
general, and what does the term holds when considering strategy selection .


Pedagogical Strategies

Tutorial strategies are the set of teaching events (actions and decisions) that

motivate and
interest the learner while improving his performance. They provide tools to present the
information and pedagogical means to favour learning [Frasson, Aïmeur & Serroud, 1995].
Traditionally, tutorial strategies in ITS meant cooperative stra
tegies. The strategy was the
whole context of operation of the system. Changing the strategy would mean changing the
whole learning situation, from, for example, a tutor
-
tutee context to a learning by disturbing
context. These strategies were developed by

researchers who worked on ITS and on
improving the learning context. Many of them were computer scientist with some
pedagogical background. Among cooperative strategies we note :

The
tutor
-
tutee

traditional method, whereby the computer is the teacher and

the user is the
learner.

The
learning companion
whereby the system simulates another learner that would
accompany the user. This strategy was first hinted by Self [Gilmore, Self, 1988], and fully
introduced by Chan [Chan, Baskin, 1990].

Learning by dis
turbing
or learning with a trouble maker whereby the user is accompanied
with a trouble maker that occasionally misleads him. This strategy was evaluated by [Frasson
& Aïmeur, 1996].

Learning by teaching.
Derived originally from the learning companion. The

human learner
is encouraged to teach the companion. This strategy was further developed by [Palthepu,
Greever & McCalla, 1991] and [Van Lehn, Ohlsson & Nason, 1994].

n, 1995].

Learning with a co
-
teacher.
This learning form was suggested also by [Aïmeur,
Alexe &
Frasson, 1995]. It consists of having a teacher and a co
-
teacher in front of the learner.


Selecting a strategy or switching from one to the other means changing the cooperative
strategy context. That is fine. But what if in the same context (in th
e tutor
-
tutee strategy, for
example) the tutor decided to give and elaborate on more examples before furthering the
information given, or decide to present an analogy with an, already, well known material, or
start his teaching with a session of questions

asking, in order to further the int
erest of the
learners etc... .
Isn’t that also changing the teaching strategy ?

Thus another family of teaching strategies came up. It is known by the direct strategies.
These strategies does not alter the session
context, but they deal more with how the
information is presented. These strategies are numerous and some might argue that any shift
in the teaching pattern or any change in the way the information is presented would make a
new strategy. Well it is true to

a certain extent.

Defining and grouping these strategies is not really important, what is, is making sure that
the agent is equipped with the right evaluative techniques to be able to make a switch
between them.

These direct strategies include among ot
hers : Learning by examples, learning by story
-
telling, learning by doing, learning by games
, learning by analogy, learning

by induction, by
abduction, deduction [Aïmeur, Alexe & Frasson, 1995].

In our model we included both families. Thus, the selector
agent can switch cooperative
strategies between teaching sessions (usually a cooperative strategy does not vary in one
teaching session) and direct strategies whenever it is needed. In order to choose the proper
strategy, the selector agent should have a s
ufficient enough idea on the learner. Which
naturally brings us to the next paragraph: the student model.


Student Model

In order to achieve an efficient strategy selection technique the learner should be well
represented by

a student model. Traditionally
, the student model was
sort of an intelligent
data base

that would reflect the knowledge

status of the learner [Frasson & Kaltenbach,
1993
]. With time, believability have been added to the student’s model, and now it
presumably reflects the emotiona
l status.

Beaumont defined the knowledge and beliefs, the goals and plans, the attitude and the
potential capabilities as the entities that would form a student model [Beaumont, 1994].

Frasson stated
more recently
that the student model includes cognitive, believable
and
inferential entities. And the model is consulted periodically by the system in order to choose
how to present

the next information [Frasson & Aimeur, 1996
].

On the implementation side, a lot of work have been presented on the student model by the
ITS community.

The existing models are not thorough, due to the fact that a real complete
representation of the learner is too complicated and almost impossible to implement, on one
hand. And emphasis on certain aspects in a student model are not always the same, they
depend on the real application, on the other hand. Relatively simple student models have
been doing a great job to instruct us on the student status [Frasson, Aïmeur & Serroud, 1995].

In our student model, we defined three modules that suffice for the impl
ementation of a
strategy selector agent. The three modules are : the knowledge level, the learning profile and
the believable aspect. The third module was practically not used in this particular application.
And this is due to the fact that, in tutoring th
e emotional factor (which usually includes
instantaneous emotional states and not stable personality types ) could be represented in the
learning profile which includes the pedagogical personality and preferences of the learner
(note that this is not proba
bly the case in common believable agents like the Synthetic actors
for example [Rousseau, Hayes
-
Roth, 1997]).

The knowledge Level

Gagné in his theory of education has roughly defined seven levels of knowledge [Gagné,
1985]. In the same line [Frasson, Aï
meur & Serroud, 1995] have defined a student model
with four different knowledge level : 1) Novice, 2) Beginner, 3) Intermediate and 4) the
expert. We have adopted this approach and table 1 explains roughly what each level implies.
To determine the knowled
ge level the system has multiple traditional techniques, either
explicitly by asking the user to answer a set of questions, specifically prepared to give a
somehow accurate answer of his knowledge level, or by referring to its memory and
retrieving the las
t performance of the user (in this case the date of this last pe
rformance
should be considered
), or by simply asking the user what level he thinks he is in (our simple
classification makes it easy for a learner to self cla
ssify, usually a person knows if he is an
expert or a beginner in a subject).

The choice of the best strategy depends on the knowledge level of the learner (as an example
the companion strategy usually works better with a novice or a beginner etc…). But,
this is
not enough. Another module should be considered when making a choice of a strategy.

The next section will present to the reader what exactly we mean by the learning profile, how
does it affect the choice of a strategy and what means do we use to m
easure it.


Knowledge Level

Explication

% of Expert
Knowledge

Novice

No prior knowledge of the subject at all, never introduced to the
subject before


0 %

Beginner

Familiar with the subject. Knows some of the rules but lacks in
practice, expected to an
swer basic questions correctly.


10


30 B

fntermediate

iearner knows most of the rules and is expected to answerI correctly
half of the questionI while trying to perform in the other half.



-
S0 B

bxpert

Completely knows rules. eave ability to answer

most of the questions
correctly. jainly uses the system to make his knowledge perfect.



-
100B

Table 1

: Explication and % of expert’s knowledge of different levels



The Learning Profile

Not all human have the same
learning preferences
. Some learn m
ore by actively participating
in the learning process, others by simply looking at the information source, others by doing a
lot of examples etc… Even though these subjects might have the same knowledge level.
From here we define the learning profile as
an aspect of human personality that represents the
“learning preferences” of a person and instruct which teaching strategy would appeal more to
him. The profile rarely change with time. Traditionally, this aspect was neglected due to the
fact that teaching

used to be done in a collective way with a group of students and not on a
one on one basis. Recently, this profile have been used by human teachers and institutions to
improve their teaching performance [Golay, 1997] and [
inter
CONNECTIONS, 1997].

Due to
the nature of the ITS (one user in most cases), it is very helpful, especially when
strategies could be replaced, to consider this profile.

In our model we have considered a
personality pattern test in order to define the learning profile of the user.


Tab
le 2 : Personality pattern and result

Personality Pattern Test.

The test consists of four panels that would define: 1) the personality, 2) the outlook, 3) the
temperament and 4) the lifestyle. The user have to pick one set of characteristics that applies
m
ost to him or her

(
Table 2
)
. Following that, there will be 16 patterns. Each pattern would
define a personality type and provides the learning techniques that would best suit it.


The Strategy Selector Agent.

The proposed model is a mini ITS specially des
igned to simulate the operation of an agent
selector of strategies. This mini ITS contains three main parts: 1) the learner model, 2) the
session, and 3) the strategy basis.
(
Figure 1
)
. At first the user is asked to enter the student
model. Then, is asked to

choose the operation mode, with or without the agent. If the user
picks the manual operation than he will have to enter the strategy himself regardless of the
student model he entered, and this strategy will not change through the whole session. But if
he
/she picks to operate with an agent selector of strategy, then this agent will evaluate the
student model entered and wi
ll consult his strategy basis (F
igure 2) in order to in order to
pick the best strategy. Through the whole session the agent will be imp
licitly testing the
performance of the user to see whether the strategy he picked had a positive impact or not.
And a possibility to switch direct strategies in the middle of the session would also take
place.

Figure 1. Interface of the mini
-
ITS

The rules

that we chose to select the strategy were very simple, while neglecting the learning
profile, for our first essays, we defined the following: for the novice choose the tutor
-
tutee

with learning by examples. For the beginner we chose the Companion with pro
blem session.
For the intermediate and the expert we chose the trouble making with the problem solving
session
.


Figure 2. Block diagram of the system



At any time when the correct

answers are less then or equal to one third of the total questions
the s
ystem will shift to the “learning by examples” direct strategy, present two examples and
get back.


Tests and results

The programming language was Java 1.1.5, and the development took place on a Sun
platform. Audio was introduced in order to add some live
liness to the interface.

The material (curriculum) was how to calculate the derivative of a polynomial function, the
derivative of the sum, the product and the division of two polynomials. We presented the
whole material at the beginning of the session. We

conducted tests of performance in a
package of five questions at a time and we evaluated the user according to how many
questions he/she answered correctly. A good performance would be 4 or 5 over 5.

The test were conducted at our laboratory where we en
gaged a number of people, ranging
from the novice to the expert level, and fixing the learning profile to ESTJ.

We noticed the following :


The lower the knowledge level, the more efficient it is to have a strategy selector. The
intermediates and the exp
erts did well most of the time regardless of the strategy, while the
novices had a bad performance when the strategy was picked by the user and a better one
when the system selects the strategy.


The use of the trouble maker with three novices made two of

them abandon the system and
it took
35% more time

for the third to assimilate the knowledge and move to the beginner
level, due to the confusion that the trouble maker induced.


The use of the tutor
-
tutee model with the “teaching by examples” direct stra
tegy for the
novice, proved the best, with the least time for the novice to be able to answer three of the
five questions correctly.


With the expert, the trouble maker proved more efficient then the companion, but the tutor
-
tutee was not bad and it perfo
rmed better then we expected.


For the direct strategies, the expert did not like the “teaching by examples”, on the contrary
they preferred the problem solving session.


The knowledge level as expected did not vary within the same session.


Conclusion

By the time
, the system we constructed
is continuously evolving and improving taking into
account more factor to refine the output
. We compromised a lot in the strategies and we stick
to three in the cooperative strategies (the trouble
-
maker, the companion and the tutor
-
tutee)
and to two in the direct stra
tegies (learning by examples and learning by problem solving). In
the near future, more testing will be done and the introduction of more strategies will be
made (one of them that we are working on now, is the learning by teaching). We will
introduce the n
otion of the agent changing the direct strategy within the same teaching
session. We have already experimented on this by introducing an example presentation in the
middle of a problem solving session, when the user performs poorly.

In the far future we wo
uld like to consider three main points :

The
adaptability

of the agent and what this characteristic would add to our system.

The study of the
curriculum

and the implementation of how would a curriculum influence
the choice of the tutorial strategy.

To b
uild a more representative
student model
, if possible.

Another challenge would be to build the basis on which the agent should rely, in order to pick
the strategy upon consulting the student’s model. Since we would like to consider more
different student
models. So far we have operated in a forward way. That is we fixed the
strategies that the agent should pick for a given student model and we tested the choices we
made. We would like to consider also, the reverse way, whereby we start with no assumption
at all, and we deduce what strategy would suit best each student model by running the system
with different student models with no preferences at all.

Finally, what is encouraging, is that the result we obtained were highly informative. And
somehow we suc
ceeded in testing the agent selector of strategy without building complicated
ITS.


References


Aïmeur, E., Alexe, C., and Frasson, C. 1995.
"
Tutoring Strategies in SAFARI Project
"
,
Departmental Publication # 975, Department of Computer Science, University o
f Montreal.


Beaumont, I.H. 1994.
"
User modelling in the Interactive Anatomy Tutoring

System
ANATOM_TUTO
"
, User Modelling and user
-
Adapted Interaction, vol 4, no 1, (pp 21
-

45).


Chan, T.W. & Baskin, A.B. 1990.
"
Learning Companion Systems
"
. In C. Frasson & G.
G
authier (Eds.)
Intelligent Tutoring Systems

: At the Crossroads of Artificial Intelligence and
Education
, Chapter 1, New Jersey

: Ablex Publishng Corporation.


Frasson, C., Aïmeur, E. 1996 "A Comparison of Three Learning Strategies in Intelligent
Tutoring Systems",

Journal of Educational Computing Research,
vol 14.



Frasson, C., Aïmeur, E., and Serroud, A. 1995.
"
Le Stratège
"
, SAFARI Project, University of
Montr
eal.


Frasson, C., Kaltenbach, M.
1993

"Strengthening the Novice
-
Expert shift using the self
-
explanation effect",
Journal of Artificial Intelligence in Education
, special issue on student
m
odelling, vol 3(4), pp. 477
-
494
.


Gagné, R. 1985.
"
The Conditions of Learning and Theory of Instruction
"
.
Holt, Rinehart and
Wi
nston,
4th edition
.


Gilmore, D. and Self, J. 1988.
"
The application of machine learning to intelligent tutoring
systems Teaching
"
.
The Proceedings of International Conference on the Learning Sciences
,
AACE.


Golay,1997,
http://baynet.com/adventurous/profil.actual.html



inter
CONNECTIONS,1997,
http://www.interconnections.co.uk/Market/PCFG/learn/lrntest.htm


Palthepu, S., Greer, J., & McCalla, G. 1991.
"
Learning by Teaching
"
.
The Proceedings of
International Conference on the Learning Sciences
, AACE.


Rousseau, D., Hayes
-
Roth, B. 1997.
"
A Social
-
Psychological Model for Synthetic Actors
"
.
Report No. KSL 97
-
07. Kno
wledge Systems Laboratory. Department of Computer Science.
Stanford University.


Van Lehn, K., Ohlsson, S. & Nason, R. 1994.
"
Application of simulated students

: an
exploration
"
.
Journal of artificial intelligence in education,

vol 5, no 2, (pp135
-
175).