Agent-oriented Support Environment in Web-based Collaborative
Nagoya University, Japan
Nagoya University, Japan
Nagoya University, Japan
Abstract: Currently, the webbased learning support systems are
one of interesting and hot topics in points of the utilization of Internet
and the application of computers to education. In particular, the webbased
collaboration is very applicable means to make unfamiliar students, who
are unknown to each other, discuss together in the same virtual interaction
space. However, there are some problems derived from the gap between the
real world and virtual environment: coordination of discussions, cooperative
reactions, comprehension of learning progress, etc. These problems may
be dependent on the fact that the actions of students cannot be influenced
from the behaviors of others directly.
In this paper, we address a coordination mechanism to promote cooperative
actions/reactions for progressive discussions. Our idea is to apply
an agentoriented framework to this coordination mechanism and introduce
two different types of agents. One is a coordinator and the other is a
learner. The coordinator monitors the learning progress of groups and promotes
the discussion, if necessary, so as to reach their common goal successfully.
The learners are assigned to individual students, and act as interaction
mediators among students in place of the corresponding students. Of course,
the coordinator is a passive entity and learners are active entities in
our collaborative learning space.
Key Words: Collaborative learning environment, coordinator, learning
situation, learner, personal learning history
Categories: K.3.1, K.3.2
The fast and worldwide enlargement of Internet/Intranet has made
it possible that every person can interact instantly without depending
on their physical
1This is an extended version
of a paper presented at the ICCE/ICCAI 2000 conference in Taipei, Taiwan.
The paper has received an Outstanding Paper Award and is published in J.UCS
with the permission of ICCE/ICCAI
locations. Also, various applications, which are available on the web
environment, have been developed with respect to the contentbased
resource sharing, in addition to the traditional message exchanges. The
webbased collaborative learning is one of applications, based on such
a hot topic, and has been applied as computersupport for virtual learning
space. If their computers were connected mutually through the webbased
learning environment, students could discuss their common solving process
successively and exchange various solving methods/ideas cooperatively.
However, there are some problems with encouraging activated discussions
among students and making it possible that individual students should understand
the correct answer and solving process effectively:
- students may not participate in the discussion interactively because
of their hesitation, derived from the fact that they are unknown with each
- students cannot grasp the behaviors of others directly or indirectly
because only the direct actions and reactions are observable through the
These problems are radical drawbacks for collaborative learning.
In order to solve these drawbacks effectively, we propose an agentoriented
support environment for collaborative learning. Of course, the agentoriented
frameworks for the construction of collaborative learning mechanism /environment
have been already investigated. Florea  proposed
a multiagent collaborative learning environment in the web world.
In this environment, three kinds of agents were introduced: a personal
agent which gets the information according to the requests of each student,
tutor agent which generates advice when personal agents asked for the help,
and information agent which acquires more information from Internet. Agents
are activated by students' requests so that this system environment does
not benefit passive students. Ogata, et al.  proposed
mediator agents in the collaborative learning environment which assist
students to find suitable collaborators. The mediator agent for each student
holds the corresponding students' profile which indicates the understanding
and interesting degrees about knowledge. When a student has problems, his/her
mediator agent asks other mediator agents for the learning situations of
their corresponding students and specifies appropriate students who may
be able to help solving the problems. This research copes with the above
problem 1) indirectly because this functionality supports the creation
of appropriate learning groups, but does not manage the progress of collaborative
learning. Nakamura, et al.  and Liming, et al. 
introduced respectively pseudo students which correspond to individual
human students. These pseudo students have the same knowledge as the corresponding
students and participate into the discussion in their ways if the corresponding
students do not join in the discussion positively or cannot understand
the discussion stage. These research viewpoints focus on passive students such as problem
1), but do not solve the problem 2). So, in spite of
these various agentbased investigations, the previous drawbacks are
not always overtaken.
In this paper, we address a collaborative learning environment, organized
systematically under two different types of agents: coordinator and learner.
The coordinator takes roles to monitor the discussion situation among students,
grasps the learning progress and guides the learning process if necessary.
The learners are virtual students corresponding possibly to individual
students in our webbased collaborative learning environment. The coordinator
and learner are complementary entities in the learning environment: the
coordinator is a passive entity; and the learner is an active entity as
the autonomy for practically participating students. In our investigation,
we expect the collaborative learning of high school students who study
mathematical exercises together, especially computation for the roots of
equations. First of all, we show an overall framework of our collaborative
learning environment on the webworld in Section 2. The functionalities
of two different types of agents are stated in Sections
3 and 4, and then our prototype system is shown
in Section 5. Finally, we conclude our paper in Section
2 Collaborative Learning Environment
In the webbased collaborative learning environment, the actions/reactions
of participating students are inherently different from their behavior
in the real world. Students in the physically constrained learning space
can speak with each other by means of facetoface, feel/recognize
activities, occurring from the discussions of students, directly by various
sensitive receptors and find out some new events/facts indirectly. Although
these are not always implemented adaptively in the webbased virtual
learning space, it is necessary to organize a collaborative learning environment
in which the logical activities for support of interaction, discussion
and comprehension can be implemented successfully and effectively.
Figure 1 shows our collaborative learning environment
conceptually, which is characterized by two different types of agents:
coordinator and learner. The coordinator is placed in the center of our
virtual classroom (as a network server), monitors the interaction among
students and generates advice if necessary, according to the learning situation.
This interaction is supported by the conversation means through the public
communication line. The learner is a pseudo student in our virtual classroom
and is assigned to the corresponding student one by one. The learner takes
roles of the personal management of interaction interface for the corresponding
student, the handshaking control of public communication line, the management
of its own private learning history, and so on. In addition, the learner
can communicate with other learners directly through the private talking
line in order to exchange their personal learning histories.
Figure 1: Collaborative learning environment
Since students are studying with limited learning tools in the webbased
virtual learning space, they are sometimes not able to communicate naturally.
Furthermore, various students participate in the learning group and the
learning process is not always completed successfully: i.e. some students
are not able to solve the problem, some students are not able to understand
the derived answering process after all, and so on. The coordinator solves
such drawbacks in the webbased virtual learning space by managing
the learning situation globally: the coordinator takes the place of a teacher
in our classroom activity. For the purpose of resolving inappropriate learning
situation stepwisely and guiding the learning group effectively, how to
model and control learning situation is an important subject. If the coordinator
grasps the learning situation appropriately, the advice which were generated
may become appropriate hints in order for the learning group to proceed
to the next phase of learning process. However, it is not always necessary
to model the learning situation in detail precisely. This is, we think,
because among the learning group students are able to help each other by
discussion, so that the coordinator only has to detect the situation in
which the learning group cannot proceed the learning by itself.
The coordinator holds the right answer and the answering paths for an
exercise as knowledge to grasp the current learning situation.
Figure 2: Answer space
When the exercise has several answering paths for the goal, the answer
space of the exercise is expanded as 2dimensional network structure
as illustrated in Figure 2. In this figure, the learning progress along
xaxis means the stepwise progress of deriving answer, whereas that
along yaxis shows the extent of discussion. If the coordinator grasps
the learning situation on the basis of the answering process of network
structure as it were, it is very troublesome to manage the eventually changeable
conversation stages successively. Therefore, our coordinator manages the
learning situation with respect to the following two viewpoints separately:
ratio of derived steps for a whole answering process and extent of discussion.
By monitoring the learning situation under these points of view, the coordinator
is able to grasp the learning situation easier and generate advice timely.
In particular, it is necessary and sufficient to manage the learning situation
of the group globally, and not individually of each student.
The learner acts as a network client in place of the corresponding human
student in the webbased virtual learning space. This provides not
only the interaction interface for virtual learning space attached to the
corresponding student, but also the function of indirect interaction among
students, so as to judge their understanding levels or personalities, which
we call the focus function. According to the focus function, students select
the opinions of particular students whom they evaluate as key persons.
In order to realize the focus function, the learner needs to have the knowledge
about the corresponding student and exchange it with other learners. Therefore,
the personal learning history is prepared for the learner, which represents
the understanding level and personality of corresponding student. The learner
constructs and maintains the personal learning history according to the
current situation. Exchange of personal learning history is onetoone interaction so that public communication
is not necessary for the focus function.
Figure 3: Resolution derivation scenario and indicators
Therefore, we introduce the mobile agents called mediators as children
of the learner who take responsibilities for the exchange of personal learning
histories among learners. The mediator moves among learners by requesting/carrying
the personal learning history on the private talking line.
The coordinator grasps the learning situation from two viewpoints: ratio
of derived steps for a whole answering process and extent of discussion.
For the ratio of derived steps, which corresponds to the xaxis of
answer space in Figure 2, we have already proposed
the resolution derivation scenario which represents the phases of deriving
answer stepwisely [5, 6, 7].
The scenario is generated by means of projecting the answer space onto
xaxis and consists of ordered states which correspond to individual
phases of deriving answer. Grasping an approximate learning situation makes
it possible for the coordinator to generate advice timely and effectively
because each state corresponds to the individual ratio of derived step.
In our scenario structure, the current learning state is pointed by the
indicator current, which points out the currently discussing stage. The
coordinator infers the current state from student inputs and moves the
indicator to the corresponding state. However, the utilization of only
one current discussion indicator is not enough to manage the learning state
of a group sufficiently. In addition to current, indicators upper and lower
are prepared for the representation of current understanding levels of
a learning group. Upper points out the state of understanding level which
is estimated that best understanding student reached to and lower points
out the state of worst understanding student. The coordinator is able to
grasp the learning situation on the basis of the relationship among these
3 indicators (Figure 3).
On the other hand, the extent of discussion is estimated by the number
of derived answering paths with different discussion viewpoints. The difference
of discussion viewpoints among answering paths is defined as the ratio
common and uncommon answering steps. That is, if two answering paths
contain large number of answering steps as common part, they are regarded
as more similar paths; but if they have many different answering steps,
they are judged as different paths. Common answering steps mean that the
answering methods which are used to derive those steps are the same. Once
two answering paths were diverged, the following answering steps may be
derived based on different answering methods so that they are regarded
as being uncommon. From such viewpoint, the coordinator holds an answer
tree which was transformed from the whole answering paths as a tree structure.
Figure 4 shows the construction of an answer tree,
derived from the answer space in Figure 2. The answering
steps after the divergence are regarded as uncommon steps so that they
are copied as different objects (Figure 4a). Then,
the answer tree is transformed by means of collecting common answering
steps for the purpose of grasping the difference among the answering paths.
The nodes in the tree are generated as a collection of answering steps
that are common to particular answering paths and the path from root node
to particular leaf node corresponds to each answering path. When the answer
has been derived, the coordinator specifies derived/underived answering
paths, calculates the differences between the derived answering path and
other answering paths based on the answer tree, and estimates the extent
By grasping the learning situation from these aspects, the coordinator
is able to handle the changeable learning situation and generate appropriate
advice at the right time.
The learner is situated on each student's computer and acts as a pseudo
student in the webbased virtual learning environment. The learner
provides the interface to the human student and controls the private talking
among students such as focus function. Since the learner only connects
the private talking line according to the corresponding student's request,
it behaves independently with the coordinator that manages the public communication.
A personal learning history is the model of corresponding student which
is held by the learner. The personal learning history represents the understanding
level and the characteristic of corresponding student. Some data of personal
learning history are prepared by the human student beforehand and others
are gathered by the learner occasionally through the learning. Currently,
the picture and utterances of students are collected as a personal learning
history. The features of a student do not change through the learning,
so the picture is set by each student before the learning starts. Utterances
indicate the understanding level of students and also their attitudes toward
the learning; i.e. active or passive, understanding or notunderstanding,
and so on.
Figure 4: Construction of answer tree
Figure 5: Mechanism for acquiring personal learning information
They are gathered and added to the personal learning history
by the learner when the corresponding students send their opinions to the
public communication line.
In order to exchange the personal learning history through private talking
line, the learner generates mediators for each communication. The mediator
is constructed as a mobile agent which processes its task while moving
through the network autonomously . Figure 5 shows
the movement of mediator for acquiring the personal learning history of
other students. When the corresponding student requests to get the personal
learning histories of particular students, the mediators are generated
by the learner respectively. Once generated, the mediators move to the
target learners through the network and ask for the personal learning histories,
attended inherently to the target learners. After the acquisition of personal
learning histories, the mediators move back to their original learner and
disappear autonomously, since their roles are to acquire the personal learning
histories from target learners. Under such mechanism, students are able
to know other students' characteristics even in our webbased virtual
learning environment without any direct interaction.
We have implemented our prototype system on Internet using UDP protocol,
since UDP protocol is suitable to control the frequent interaction of short
messages. Figure 6 shows the interaction interface
in our system. Two communication tools are prepared: answerboard screen
and interaction space. The answerboard screen is a public communication
tool which is used to arrange the group's answering process. Only one student
is permitted to input on the answerboard screen at a time so that
the input right is set. On the answerboard screen, ID, student's name,
and contents of input are shown. The answerboard screen functions
as a blackboard in our real world. Descriptions on the answerboard
screen can be erased or modified, but once they are overwritten, you cannot
see them again. On the other hand, the interaction space is prepared for
free conversation so that all students are able to input freely. In order
for the coordinator of our system to grasp the learning situation easily,
commands that classify the opinions are introduced: Appreciate, Inquire,
Assert, and Confirm. Students choose the commands when they input their
opinions. In addition to the commands, students specify the target inputs
which trigger off their opinions for the purpose of grasping the flow of
conversation smoothly. Thus, in addition to the ID, student's name, and
contents of input, command and ID of target input are also displayed on
Figure 6: Interaction interface on student operation
As for the coordinator, we prepared several advice which indicate the
states of learning situation when the learning proceeds inappropriately.
Currently, the coordinator generates advice when it detects the following
- learning situation has not been changed for a long time,
- some students cannot understand the currently discussed stage, and
- students have not derived all viewpoints of solving the exercise.
The coordinator's objective is to activate the discussion, so the advice
is generated on the interaction space as the same style as all other students'
utterances. Figure 7 shows an example of advice generated by the coordinator.
As for the advice, the speaker's name is set as ``!!!!!'', the command
of advice is ``advice'', and the ID of target input is nothing because
the advice is generated for the learning group but not for individual students.
We have evaluated the generated advice based on this coordinator mechanism.
We made 5 groups of 2 to 6 students, who use the prototype system, to solve
a mathematical exercise.
Figure 7: Advice example of coordinator on interaction
|What is your impression of the exercise?
||Group with coordinator
||Group without coordinator
|Can you understand the exercise?
||Group with coordinator
||Group without coordinator
Table 1: Impression and understanding of the exercise
These groups were divided into two groups; one used the prototype system
which contained the coordinator mechanism and another studied with the
same interface but without the support of a coordinator. The groups were
selected randomly. After the learning, we asked about the impression and
the understanding of the exercise (Table 1). For the group with the coordinator,
80% students understood the exercise after all in spite of the number of
students who thought that the exercise was to some extent difficult. Furthermore,
no student was not able to understand the exercise, while one student could
not comprehend in the group without the coordinator. On the other hand,
Table 2 shows the results of the questions about the advice generated by
the coordinator and the questions asked by the students who participated
in the group with the coordinator. Most students who answered that the
advice is inappropriate for both questions answered that the exercise is
easy in Table 1. Therefore, the coordinator may detect inappropriate learning
situation precisely and promote understanding of students toward the exercise,
especially of those who did not understand well and really needed effective
||Timing of advice
||Content of advice
|not so appropriate
Table 2: Advice of the coordinator
The learners were implemented using AgentSpace 
as a middleware to control the behavior of a mediator. Figure
8(a) is an interface for generating requests. In the upper window,
the causality of utterances on interaction space is arranged based on corresponding
student's utterances. The arrangement of utterances on the upper window
helps to decide the focusing students for generating requests. Once a student
decides to focus students, he/she inputs IP addresses of focusing students,
because mediators need IP addresses where they will work beforehand
in our current version. Then, he/she specifies the file name of a focusing
student's personal learning history. If a student wants to know only the
particular utterances of focusing students, he/she sets the ID's of corresponding
utterances shown in the upper window. Figure 8(b) shows
the result windows of requests for personal learning history. When requests
have been completed successfully, the result windows are generated and
the personal learning histories of focusing students are shown individually.
Currently, the picture of a focusing student is shown in the upper window
and his/her utterances are shown in the lower window.
In this paper, we proposed a collaborative learning environment which
contains two different agents: coordinator and learner. The coordinator
monitors the public communication among learning groups and generates advice
so as to lead the groups to their learning goal.
Figure 8: Interface for handling requests
For this purpose, the coordinator grasps the learning situation globally
from two viewpoints: the ratio of derived steps for a whole answering process
and the extent of the discussion. Although the management structure of
the learning situation is simple, the coordinator may be able to find that
in the most cases students are not able to cope with inappropriate learning
situation by themselves. On the other hand, the learner controls the private
talking such as focus function. The learner holds the personal learning
history of the corresponding student as his/her characteristics and acquires
other students' personal learning histories by generating the mobile agents
called mediators. Currently, these agents function independently. However,
for our future work, the interactions among coordinator and learners are
necessary for the coordinator to generate more effective advice. In addition,
the evaluation of the interaction interface of our prototype system and
the preparation of more factors for personal learning history based on
the result of the evaluation are also necessary.
The authors are very grateful to Prof. T. Fukumura of Chukyo University,
and Prof. Y. Inagaki and Prof. J. Toriwaki of Nagoya University for their
perspective remarks, and also wish to thank our research members for their
many discussions and cooperations.
1. A. M. Florea: ``An Agentbased Collaborative
Learning'', Proc. of ICCE'99, Vol. 1, pp. 161-164 (1999).
2. H. Ogata, K. Imai, K. Matsuura and Y. Yano: ``Knowledge
Awareness Map for Openended and Collaborative Learning on World Wide
Web'', Proc. of ICCE'99, Vol. 1, pp. 319-326 (1999).
3. M. Nakamura and S. Otsuki: ``Group Learning Environment
Based on Hypothesis Generation and Inference Externalization'', Proc.
of ICCE'98, Vol. 2, pp. 535-538 (1998).
4. G. Liming, H. Minghua and Q. Yuhui: ``A Webbased
MultiAgent Collaboration Learning System'', Proc. of ICCE'98,
Vol. 1, pp. 205-210 (1998).
5. T. Kojiri and T. Watanabe: ``Cooperative Learning
Support Mechanism, Based on Scenario of Specifying Solving Process'', Proc.
of ICCE'98, Vol. 2, pp. 532-534 (1998).
6. T. Kojiri and T. Watanabe: ``Adaptable Learning
Environment for Supporting a Group of Unspecified Participants in Web'',
Proc. of SITE'99, pp. 1937-1942 (1999).
7. T. Kojiri and T. Watanabe: ``A Management Method
of Leaning Situation in Collaborative Learning'', Proc. of ICCE'99,
Vol. 1, pp. 386-393 (1999).
8. J. E. White: ``Telescript Technology: Mobile Agents'',
Software Agents (J. Bradshaw (ed.)), MIT Press, pp. 437-472 (1997).
9. I. Sato: ``AgentSpace'', http://islab.is.ocha.ac.jp/agent/