IEEE ©1998 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

Return to Table of Contents


A Tutoring Strategy Supporting System for Distance Learning on Computer Networks

Gwo-Jen Hwang


Abstract - With the fast development of computer networks, people can access information and communicate with others without being constrained by space and time. Through network communications, people can discuss things with others to solve their problems. Therefore, how to implement distance cooperative CAL systems on computer networks becomes an interesting and challenging issue. In this paper, we present our work of applying a fuzzy expert system to support the tutoring strategies of a distance cooperative Computer-Assisted Learning environment. We analyze the special needs for developing CAL's in the distance cooperative environment, and define a set of parameters, which is concerned with effectiveness of learning through networks. We also model the behaviors of the students on networks through their on-line operations. A theoretical framework and some methodologies are proposed based upon the concepts and the techniques of fuzzy expert systems to cope with the problems. A fuzzy expert system is then implemented to guide the students during the learning processes and to help the CAL system to present feasible subject materials to the students. By evaluating our approach with a small group of college freshmen, we conclude that this research is worth further studying.


I. Introduction

Education is the foundation of a country. In the past decade, people have tried to refine subject materials and to develop new skills and tools for helping the progress of education. Some researchers have concentrated on the development of Computer-Assisted Learning (CAL) systems, which are designed to provide an individual learning environment for each student. Earlier CAL systems focused on the interactions between computer and single student. Some systems can even play the role of learning partners. The purpose of providing some learning partners is to encourage students learning through the interactions among the students and those computer partners. However, as human reactions are very complex, it is usually difficult to simulate human behaviors by computers. Moreover, when a student fails to understand some unit of the subject materials, he might have to stop learning if no instant help is available [3]. Therefore, it is more desirable to have human partners than computer partners [5]. Unfortunately, it is usually difficult for a student to find a suitable learning partner at the time and place he prefers to learn.

With the fast development of computer networks, people can access information and communicate with others without being constrained by space and time. This provides a good opportunity to cope with the learning problems. One of the most important features of computer network is the fast communication ability. Through network communications, people can discuss things with others to solve their problems. Therefore, how to implement distance cooperative CAL systems on computer networks is an interesting and challenging issue. In this paper, we present a knowledge-based system, which enables an intelligent tutoring method in the distance cooperative learning environment. To achieve this goal, the subject materials have been decomposed into pieces of modules to enable dynamic combination of each learning unit. We also describe learning status and teaching expertise with fuzzy membership functions and inference rules, which is useful in finding best combination of learning modules for each student.


II. Background

ITES is an intelligent tutoring and evaluation system originating from the CORAL (Cooperative Remotely Accessible Learning) in Taiwan [9]. It was built as a WWW server to manage learning requests from students through computer networks. The entire project was initiated by a research group at National Chaio Tung University in Taiwan, which consists of eight sub-tasks:

 

The goal of the entire research group is to accomplish a CAL system to help the learning progress via the development of an intelligent tutoring environment on computer network. The whole idea is depicted in Fig. 1.

 

 

Fig. 1. The strategy-support system of ITES


III. Design Considerations for Distance Learning

In a distance cooperative learning environment, the background and learning status of the students are usually significantly different; therefore, a learning unit may need to be represented in several ways. Subject materials and the tutoring procedures should be constructed according to the requirement of each student. Moreover, as network facilities enable the students to discuss or chat with each other during the learning processes, some special features or students reactions on computer networks also need to be taken into consideration in developing the CAL system.

In our approach, the tutoring expertise elicited from the experienced teacher is represented by a set of fuzzy rules stored in the knowledge base. The system will consider the network characteristics and the learning behaviors of the learner in order to generate a personalized teaching script according to these pre-defined tutoring strategies. As the subject contents in ITES were written in HTML format for WWW browsers, each learning unit can be considered as a set of HTML files. As shown in Fig. 2, three types of subject materials have been defined as follows [1][2]:

These subject materials represent curriculums, i.e., the basic contents of the courses. Several versions with different levels of difficulty are offered. In the example in Fig. 2(a), three versions of formal subject materials are given. Different students may receive different versions of the same course according to their past learning performance.

In addition to formal subject materials, some assistant materials are designed according to the student's on-line learning performance (e.g., concentration, willingness and patience as shown in Fig. 2(b)).

A set of control descriptions to depict what portions of subjects materials should be taken for a student and what is the order for presenting subject contents. A meta subject material does not hold any contents to be learned; instead, it describes which and how subject materials and assistant materials are going to be presented to the students (as shown in Fig. 2(c)).

 

Fig. 2(a) An example of Formal Subject Materials

 

Fig. 2(b) An example of Assistant Subject Materials

 

Fig. 2(c) An example of Meta Subject Materials

 

To reflect the learning status of each student, the meta subject materials are determined according to the students' personal records and their behaviors on the network. Currently, the following parameters are taken into consideration:

Nc:

Number of Communication sessions (initiated by the student during a learning session).

 

Ct:

Communication time. Total time for the communication sessions initiated by the student during a learning session.

 

Ect:

Effective communication time. Total time for the student to sends/receive effective messages to/from others during the communication sessions.

 

Nct:

Non-effective communication time. Total time for the student to send/receive non-effective messages during the communication sessions.

 

Ni:

Number of interrupt sessions (initiated by others during the student's learning session).

 

It:

Interrupt time. Total time for the interrupt sessions during the student's learning session.

 

Eit:

Effective interrupt time. Total time for the student to send/receive effective messages to/from others during the interrupt sessions.

 

Nit:

Non-effective interrupt time. Total time for the student to send/receive non-effective messages during the interrupt sessions.

 

Elt:

Effective Learning time. Total time that a student invoked the CAI during a session.

 

Est:

Effective session time. Est = Ect + Elt + Eit.

 

Nu:

Number of users.

 

Nr:

Number of resources (including video, audio and text).

 

Ni:

Number of interrupt. Times for a student being interrupted by others.

 

Id:

Idle time. Total time that the student was in idle status.

 

Nst:

Non-Effective session time. Nst = Nct + Nit +Id.

 

St:

Session time. The time interval from when the student logged in until he (or she) logged out. St = Est + Nst.

 

Efl:

Efficiency of Learning. Efl= Est/St.

 

Ave:

Average session time for a subject unit.

 


IV. A Strategy Support System with Fuzzy Approach

In a fuzzy system, the reasoning procedure involves three primary processes: fuzzification, fuzzy inference, and defuzzification. In the following subsections, we shall introduce the process of fuzzy reasoning and the formulation of the fuzzy tutoring knowledge base [4].

A. Fuzzy Reasoning

Fuzzification operations are used to combine a real time input value (e.g., temperature, speed, etc.) with stored membership function information to produce fuzzy input values. Fuzzy inference concerns the match of fuzzified input facts to the premise patterns of fuzzy rules. Defuzzification is the process of combining all fuzzy outputs into a specific composite result. In addition, membership functions provide the means of translation between linguistic expressions such as "Tom is very young" and numerical input facts such as "Tom is 18 years old". Consider the fuzzy rules given as follows:

RULE 1:

If a person is young and tall

Then he is suitable for playing basketball

 

RULE 2:

If a person is old

Then he is suitable for jogging

 

Fig. 3. The membership function of age

 

Assume that the input fact is "John is 67 years old" and the membership function of "age" is depicted in Fig. 3. The vertical reference line of Fig. 3 intersects the young membership function and old membership function at 0.46 and 0.92 respectively, which imply that the expression "John is young" is true with a higher degree (0.46) and the expression "John is old" is true with a lower degree (0.92). By Fuzzification operations, fuzzy input values are produced:

Fact 1: John is young

with degree 0.46

 

Fact 2: John is old

with degree 0.92

Each rule antecedent consists of a set of fuzzy patterns, which might be matched by the fuzzified input facts. In the fuzzy inference process, the system takes the minimum of the rule antecedent terms as the matching degree of the rule. Assume that we have "John is tall with degree 0.8", the match degree of rule 1 is minimum(0.46,0.8)=0.46, fuzzy outputs will be:

Output 1: John is suitable to play basketball

with degree 0.46

 

Output 2: John is suitable for jogging

with degree 0.92

In some fuzzy systems, the fuzzy output with maximum truth value is taken as the system output, which was called maximum defuzzification. With this approach, "John is suitable for jogging" will be the final result.

B. Fuzzy Strategy-Support Rules

In building the fuzzy expert system for network-based CALs, we first analyze the expertise needed to enable intelligent tutoring. For the determination of a feasible version of formal subject materials, the student's knowledge level and intelligent quality are taken into considerations. The knowledge level is decided according to the score of each unit test. The intelligent quality is determined by a standard test (usually the test is done when a student enters a school). The membership functions of knowledge level and IQ are shown in Fig. 4. Assume that there are 3 versions of formal subject materials with different difficulties. The fuzzy rules for version selection of appreciated formal subject materials are given as follows:

If

Knowledge-level is good and IQ is high

 

Then

Version is definitely 1

 

If

Knowledge-level is good and IQ is average

 

Then

Version is possibly 1 or 2

 

If

Knowledge-level is good and IQ is low

 

Then

Version is definitely 2

 

If

Knowledge-level is average and IQ is high

 

Then

Version is possibly 1 or 2

 

If

Knowledge-level is average and IQ is average

 

Then

Version is definitely 2

 

If

Knowledge-level is average and IQ is low

 

Then

Version is possibly 2 or 3

 

If

Knowledge-level is poor and IQ is high

 

Then

Version is definitely 3

 

If

Knowledge-level is poor and IQ is average

 

Then

Version is possibly 2 or 3

 

If

Knowledge-level is poor and IQ is low

 

Then

Version is 3

 

 

Fig. 4: The membership functions of Knowledge level and IQ

 

 

The system then examines the behaviors of the student on the network to select a proper set of assistant subject materials which will be inserted into the formal subject materials to construct a personalized tutoring procedure. The student behaviors taken into considerations are as follows:

The fuzzy variable for "concentration" is defined by X= (Ect + Elt)/St and the membership function is given in Fig. 5. Assume that there are T frames in a unit, we have the following fuzzy rules:

If

concentration is high

 

Then

do nothing

 

If

concentration is average

 

Then

insert INT(T_0.5) corresponding concentration frames

 

If

concentration is high

 

Then

insert T corresponding concentration frames

 

Fig. 5: Membership functions for "concentration"

 

The fuzzy variable for "willingness" is defined by X=Efl= Est/St and the membership function is given in Fig. 6. The related fuzzy rules are

If

willingness is high

 

Then

do nothing

 

If

willingness is average

 

Then

insert small set of corresponding willingness frames

 

If

willingness is high

 

Then

insert super set of corresponding willingness frames

 

Fig. 6: Membership functions for "willingness"

  

The fuzzy variable for "patience" is X=St and the membership function is given in Fig. 7.

Fig. 7: Membership functions for "patience"

The related fuzzy rules are

If

patience is high

 

Then

do nothing

 

If

patience is average

 

Then

insert patience frames

 

If

patience is high

 

Then

ask the user to take a rest

 

These selected frames are then inserted into the formal teaching script. Each concentration frame is appended to the corresponding frame of formal subject materials, which ask some easy-to-answer questions to attract attentions from the students. Willingness frames are placed randomly among those formal frames to promote the interests of the students. Patience frames are placed at the end of the unit to ask the students to take a rest. That is, a complete teaching script is composed of the formal subject materials and those assistant subject materials, and the generation process is completely achieved by considering the responses of the students.

Consider the example given in Fig. 8, a student with average knowledge level, average IQ, low level of patience, average level of concentration and low degree of willingness. The system first selects Version 2 of the formal subject materials by considering two factors, i.e., knowledge level and IQ. This curriculum is composed of six frames of subject contents. Later, the behaviors of the students, i.e., concentration, willingness, and patience, are analyzed one by one, and the corresponding assistant subject materials are selected respectively according to the degree of those features --- 4 frames for "concentration", 1 frame for "willingness" and 1 frame for "patience".

 

Fig. 8: An example to describe the concept of our strategy


V. Implementation and Evaluation

ITES has been successfully implemented on TANet (Taiwan Academic Network). Its network address is http://www.ites.ncnu.edu.tw (the home page is presented in Fig. 9). Currently, three courses are available in ITES; including "Introduction to Computer Science", "Chemistry" and "Mathematics". Fig. 10 is an example of the "Introduction to Computer Science" course. If the behavior of the student is abnormal in browsing the subject materials, a monitoring program will try to send messages to the student to detect his (or her) learning status. For example, if a student stays on a frame for a very long time, the monitoring program will ask the student to give some response to see if the student is still browsing the frame (as shown in Fig. 11). To dynamically adjust the subject materials for each student, each learning process is divided into several learning sessions. For example, if the concentration degree of a student is low, ITES will immediately insert some assistance frames for concentration into the learning path of next session (see Fig. 12). In the followings, we shall discuss some technical problems of implementing ITES. Some experimental results are also given to evaluate its performance.

 

Fig. 9: The homepage of ITES

 

 

Fig. 10: An illustrative example of the course "Introduction to Computer Science"

 

Fig. 11: An example to illustrate how ITES detects the status of the student

 

Fig. 12. Some questions are depicted to the student if the concentration degree is low

A. Dynamic Linking and On-Line Monitoring

To implement the dynamic linking and on-line monitoring features, each HTML file invokes a monitoring program in the beginning and a time-recording function in the ending of the file (see Fig. 13). The monitoring program is a Java applet, which is responsive of recording the entrance time and watching the behaviors of the student. To correctly record and modify the functional dependency of the HTML materials, each link in the original HTML file is started with a special character string "$$$" so that the learning path can be reconstructed.

 

Call Java Applet: Initiate the monitoring program

Record the entrance time

Descriptions of subject contents

...

<A HREF="$$$F001-003.htm"></A>

...

 

Call Java Applet: Record the ending time

Fig. 13. An illustrative example of a HTML file

B. Fuzzy Reasoning

In implementing a fuzzy reasoning system, several conditions needed to be taken into considerations. For example, consider the following facts and rules:

Fact 1:

John is very young

CF=0.64

 

Fact 2:

John is tall

CF=0.5

 

Rule 1:

If X is young and X is tall

Then Choose X as a candidate

 

(CF=1.0)

 

Rule 2:

If X is young and X is not very tall

Then Choose X as a candidate

 

(CF=0.8)

 

None of Rule 1 and Rule 2 will be activated by applying pattern matching algorithms of conventional expert system unless some domain-dependent rules are added to indicate the relationships among very young, young, more or less young, not young, not very young, not old, not more or less old, etc. Another way to cope with these problems is to apply a set of domain-independent control rules to generate equivalent fuzzy facts from original input facts. That is, the following qualitative fuzzy facts are generated by control rules to replace the above fact 'John is very young CF=0.64' and then the same way must be applied to any input fact:

John is very young

CF=0.64

 

John is young

CF=0.8

 

John is more or less young

CF=0.894

 

John is NOT very young

CF=0.36

 

John is NOT young

CF=0.2

 

John is NOT more or less young

CF=0.106

 

John is old

CF=0.2

 

John is very old

CF=0.04

 

John is more or less old

CF=0.447

 

John is NOT old

CF=0.8

 

John is NOT very old

CF=0.96

 

John is NOT more or less old

CF=0.533

 

It is obvious that many redundant facts will be generated with this approach; therefore, another approach was proposed in the followings to meet our requirement [7][8]. Without loss of generality, we illustrate the development of the fuzzy reasoning features on a well-known expert system shell, CLIPS [10]. We have divided the knowledge base into two groups of rules, the domain-dependent fuzzy rules and the domain-independent control rules. The domain-dependent fuzzy rule looks like:

If

the student's record is good and his IQ is high

 

Then

course A is a good choice for him

 

Where good record means (S-function with gama=80 and beta=20) and high IQ means (S-function with gama=130 and beta=20).

Since most expert system shells do not support fuzzy knowledge representation and reasoning, we propose a new rule format to enable fuzzy reasoning on these shells. The rule of the example given above will be translated into the following domain-dependent fuzzy rules:

(defrule

rule_1

(record good ?degree_1)

(IQ high ?degree_2)

=>

(assert (rule1 min ?degree_1 course-A ))

(assert (rule1 min ?degree_2 course-A )))

 

(defrule

rule_1a

(record ?record)

=>

(assert (record goof S-function 80 20 ?record)))

 

(defrule

rule_1b

(IQ ?iq)

=>

(assert (IQ high S-function 130 20 ?iq)))

 

Some domain-independent fuzzy control rules are also generated for performing fuzzy reasoning:

(defrule

Close-function

?id <- ($?fact Close ?B ?G ?input )

=>

( retract ?id) ;remove fact

( assert (?fact =(/ 1 (+ 1 (** (/ (- ?input ?G) ?B) 2)))) ))

 

(defrule

min-rule

(?rule-id min ?degree1 $?goal )

( not (?rule-id min ?degree2&:(< ?degree2 ?degree1) $?))

=>

( assert (all ?degree1 $?goal)))

 

(defrule

max-rule

(all ?degree1 $?goal)

(not (all ?degree2&:( ?degree2 ?degree1) $?))

=>

(assert ( $?goal )))

The rule_1a and rule_1b are used to perform fuzzification operations with the help of the membership mapping function. The rule_1 and Minimum selection rule concern the process of fuzzy inference. And the Maximum selection rule is for defuzzification. With the cooperation of domain-dependent rules and domain-independent control rules, fuzzy reasoning can be performed on pure conventional expert system shells directly. The process of fuzzy reasoning must be guided under the control of some phase control rules, so that the phase changes in the order of membership-function mapping phase, minimum-selection phase and maximum-selection phase. The example only shows the most primitive case in constructing fuzzy reasoning mechanism on CLIPS. Some other fuzzy operations such as linguistic and certainty factor can also be performed with the help of more complicated control rules. A more detailed fuzzy reasoning architecture in our system is given in Fig. 14.

Fig. 14: The architecture of the fuzzy inference engine

 

There are two parts in the architecture of fuzzy reasoning. The left part is a flowchart of generated facts and the right one consists of required rules during the process of fuzzy reasoning. The Linguist Parsing rules divide the linguistic facts given by users into linguistic words and general words. And then the Domain dependent Fuzzy Rules are activated to find all possible matched facts. Later the M.F. Mapping Rules calculate the fuzzy degrees of all matched facts according to the membership function in fuzzy knowledge base. The Linguistic Operation Rules and the Certainty Factor Operation Rules are responsible for applying appropriate operations to fuzzy degrees of facts respectively (e.g. linguistic operations includes concentration, dilation and so on according to linguistic words generated previously). Finally, the Min Selection Rules set the fuzzy degree of each rule as the minimum degree of all its facts and the Max Selection Rules choose the rule with maximum fuzzy degree to fire and complete the process of fuzzy reasoning.

C. Small Group Evaluation

To evaluate the performance of ITES, we had an evaluation based upon a group of freshmen from English department [6]. The evaluation was treated as a course assignment of "Introduction to computer Science". All of the forty-one students had experience in using WINDOWS and computer networks before doing the test. The results of the test are summed up as follows:

 

1. THE PROGRAM IS HELPFUL FOR ASSIST LEARNING.

5%

73%

15%

5%

2%

5

4

3

2

1

Strongly Agree

 

 

 

Strongly Disagree

2. THE PROGRAM IS INTERESTING.

3%

29%

51%

17%

0%

5

4

3

2

1

Strongly Agree

 

 

 

Strongly Disagree

3. WILLING TO USE THIS PROGRAM AGAIN.

1.5%

44%

39%

12%

2.5%

5

4

3

2

1

Strongly Agree

 

 

 

Strongly Disagree

4. DIFFICULTY OF COURSE CONTENT

5%

10%

24%

54%

10%

5

4

3

2

1

Too Easy

Appropriate

No Comment

Average

Too Hard

5. SELF-PATH HYPERLINK FEATURE

71%

15%

7%

5%

Helpful

Wouldn't Affect Learning

No Comment

Get Lost

6. ENCOUNTER PROBLENS WHILE LEARNING?

54%

15%

7%

17%

7%

Find the answer by myself

Discuss with friends on-line

Talk to whomever on-line

Find the answer later

Skip the part Incomplete

7.OVERALL FUNCTIONS

17%

61%

10%

7%

5%

5

4

3

2

1

Complete

 

 

 

Incomplete

8. Anxiety:

Mean: 62.073

 

Minimum: 44

 

Maximum: 81

9. Grade on post-test:

Mean: 65.829

 

Minimum: 35

 

Maximum: 98

 

From the test results, it can be seen that 71% of the students thought the self-path hyper-link feature of this program was helpful for learning; moreover, when difficulty was encountered, only 15% of the students took advantage of the on-line communication, and 54% of the students still tried to find the answers by themselves. From which we realize that the communication learning feature was accepted by only a part of students, and the development of the tutoring strategy supporting systems is very important and helpful since the students try to find the answers by themselves most of the time.


VI. Conclusion

Due to the advance of computer networks and communication techniques, interactive learning through computer networks receives a lot of attention. In this paper, a fuzzy expert system for supporting the tutoring strategies of a distance cooperative learning environment is proposed. We analyze the parameters concerning the effectiveness of learning through networks, and implement the strategy- support system with fuzzy approach. From the experience of the whole project, it can be seen that a lot of work is needed to enhance the distance cooperative learning system. We also believe that it is worthy to exert more efforts in this study. Now, a knowledge acquisition system for eliciting teaching experience and a test database for primary schools is being developed which will become a part of this research.


Acknowledgement

This study was supported by National Science Council of R.O.C., under the contract numbers NSC-85-2511-S-260-001-CL and NSC-87-2511-S-260-001-ICL. 


References

[1] Chen, F.R., Hwang, G.J., & Tseng S.S. (1993). A New Approach for the Development of Intelligent CAI Systems, International Conference on Computers in Education, Taiwan, 1993.

[2] Chen, F.R., Hwang, G.J. & Tseng, S.S. (1995). On the development of an Intelligent CAL system for Chemistry Courses, Proceedings of the National Science Council, Republic of China, 5 (1), 9-18.

[3] Mitchell, P.D. & Grogono, P.D. (1993). Modeling techniques for tutoring systems, Computer Education, 20(1), 55-61, 1993.

[4] Giarratano, J.R., & Riley, G. (1989). Expert Systems: principles and programming, PWS-KENT Publishing.

[5] Hopper, S. (1992). Cooperative learning and computer-based instruction, Educational Technology research & Development, 40(3), 21-38.

[6] Huang, K.G., Chang, Y.S. (1985). The evaluation of intelligence computer learning in a distant cooperative learning environment, proceedings of the Project Paper on Intelligent Computer Assistant Learning of National Science Council, R.O.C., 146-152.

[7] Hwang, G.J., Chen, F.R., & Tseng, S.S. (1993). Development of a knowledge acquisition system for computer-assisted instruction, International Conference on Computers in Education.

[8] Hwang, G.J. (1995). Knowledge acquisition for fuzzy expert systems, International Journal of Intelligent Systems, 10, 541-590.

[9] Sun, C.T., & Chou, C. (1996). Experiencing CORAL: design and implementation of distance cooperative learning, IEEE Transactions on Education, 39(3), 357-366.

[10] Tsai, W.J., Shyu I.J., & Hwang, G.J. (1993). An Integrated Development Environment for Fuzzy Expert Systems, National Computer Symposium, Taiwan, 1993.


Author Contact Information

Gwo-Jen Hwang
Information Management Department
National Chi Nan University
Puli, Nantou, Taiwan 545, R.O.C.
ITES system homepage
Voice phone numbers: 886-35-710431, 886-932-288932
Fax number: 886-49-915215
E-mail: gjhwang@oplympus.ncnu.edu.tw


Author Biography

Gwo-Jen Hwang was born on April 16, 1963, in Taiwan, Republic of China. In 1991, he received his Ph.D. degree from the Department of Computer Science and Information Engineering at National Chiao Tung University in Taiwan. He is now an associate professor of National Chi-Nan University. He is also the Director of Computer Center of that university. His research interests include expert systems, fuzzy reasoning, multimedia systems and computer-assisted instructions.