Information Sciences 179 (2009) 688698

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

An iterative approach for estimation of student performances

based on linguistic evaluations
Efendi N. Nasibov *, A. vg Kinay
Department of Statistics, Faculty of Arts and Sciences, Dokuz Eylul University, Tinaztepe Campus, 35160 Buca, Izmir, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: In group decision analysis, numerous approaches have been suggested in an attempt to
Received 10 August 2007 solve the problem of aggregation of individual fuzzy opinions to form a group consensus
Received in revised form 8 October 2008 as the basis of a group decision. In this study, an optimization model, which reects differ-
Accepted 18 October 2008
ent points of view of many decision makers by weighting fuzzy opinions, is proposed for
the evaluation of student performances in student-centered learning. An iterative algo-
rithm is provided for the solution of this model, and the consequent theorem is proved.
Experimental results show that the proposed iterative algorithm yields more efcient
Keywords:
Performance evaluation
results than do the classical optimization methods. Moreover, the WABL (weighted averag-
Fuzzy aggregation ing based on the levels) method produces more accurate results than do the other fre-
Weighted averaging based on the levels quently used defuzzication methods, such as COA (center of area) and MOM (mean of
Least squares maxima).
Lagrange multipliers 2008 Elsevier Inc. All rights reserved.

1. Introduction

Various research approaches exist concerning the construction of a common opinion that reects the opinions of all
decision makers in a decision-making problem. These studies can be applied to different areas, such as the evaluation of
a workers performance, selection of the most suitable worker, and determination of a students success.
Chen [3] analyzed a multi-criteria personnel selection problem with multiple decision makers using the TOPSIS
(technique for order performance by similarity to ideal solution) procedure and vertex method with fuzzy information.
The TOPSIS procedure can briey be explained as a concept in which the chosen alternative has the shortest distance from
the positive ideal solution, while it has the farthest distance from the negative one. Saghaan and Hejazi [18] proposed a
modied TOPSIS procedure for the multi-criteria decision-making problem with multiple decision makers. Kuo et al. [6] pro-
posed a new method of analysis of multiple criteria based on the incorporated efcient model and concepts of TOPSIS to
solve decision-making problems with multiple judges and multiple criteria in real-life situations. Other studies have been
carried out by applying the AHP (analytical hierarchy process) approach, suggested by Saaty [17], to fuzzy numbers [2,5].
Bardossy et al. [1] suggested ve combination techniques for aggregating individual fuzzy opinions and dened seven
characteristics of the combination techniques. These combination techniques are crisp weighting, fuzzy weighting, minimal
fuzzy extension, convex fuzzy extension, and mixed linear extension. Hsu and Chen [4] proposed an aggregation method,
also called the similarity aggregation method. In this method, pairwise similarities of experts opinions are rst calculated.
Then, the average of these pairwise similarities is calculated for each expert. These average values represent their corre-
sponding experts agreement degrees. Finally, an aggregation of the experts opinions is obtained by combining the weighted
averages. Lee [7] proposed an iterative procedure for performing the aggregation of expert opinions. Wang and Parkan [23]

* Corresponding author. Tel.: +90 536 509 79 69; fax: +90 232 453 42 65.
E-mail addresses: efendi_nasibov@yahoo.com (E.N. Nasibov), ovgu.tekin@deu.edu.tr (A.. Kinay).

0020-0255/$ - see front matter 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.ins.2008.10.026
 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698 689

improved Lees study by suggesting two methods, both of which are based on the weighted distances between experts opin-
ions. They indicated that one common opinion could be obtained from the decision makers opinions on various subjects. Ma
and Zhou [9] proposed a group decision support system for assessing students learning outcomes. Yong and Wen-Kang [22]
obtained the consensus degree coefcient, which uses the relative weight agreement degrees by weighting the fuzzy opin-
ions of experts. Tsabadze [20] proposed a specic kind of fuzzy aggregation operator, as well as a method of fuzzy aggrega-
tion for solving group decision-making problems by using the coordination index and coefcient of similarity, which were
proposed in [19]. Pasi and Yager [15] presented two denitions of a majority opinion related to a linguistic quantier. The
rst possible denition of a majority-based aggregation was proposed and formalized by means of an induced ordered
weighted averaging (IOWA) operator. The second proposal formalizes the concept of the majority opinion as a fuzzy subset.
Xu and Chen [25] developed an interactive method to solve fuzzy multi-attribute group decision-making problems. They
used the hybrid weighted averaging (HWA) operator to aggregate the decision information given by each decision maker
instead of the weighted arithmetic averaging (WAA) and the ordered weighted averaging (OWA) operators. Xu [24] inves-
tigated group decision-making problems with multiple types of linguistic preference relations, and introduced two proce-
dures. In the rst procedure, the linguistic weighted arithmetic averaging (LWAA) operator is used to fuse all of the
individual additive linguistic preference relations into the collective additive linguistic preference relation. Then, the proce-
dure uses the linguistic arithmetic averaging (LAA) operator to derive the averaged preference degree of one alternative over
all of the other alternatives. Similarly, in the second procedure, the linguistic weighted geometric averaging (LWGA) operator
is used for multiplicative linguistic preference relation operations. This procedure used the linguistic geometric averaging
(LGA) operator instead of the LAA operator.
In the student-centered learning system, which is considered in this paper, a students performance is based on the eval-
uation of a set of criteria where each criterion has a different level of importance for each lecturer. Moreover, points awarded
for any level (such as absent, few, middle, good, and strong) in each criterion may vary among lecturers. In our study, aggre-
gated weight values, which reect the opinions of lecturers on the importance of each different criterion, are obtained from
the relationship between linguistic evaluations and grade evaluations. These aggregated weight values are computed
through an iterative algorithm. The use of nal weight values provides consistency among different lecturers in assessing
student performances. In other words, we propose a method for obtaining aggregate weight values that can reect the dif-
ferent points of view of lecturers, in order to evaluate student performances in a student-centered learning system. In our
iterative algorithm, we optimize the defuzzication parameter and the weight values. The objective function in the optimi-
zation problem is based on the least square errors method. Our study differs from Lees [7] in that we use both the defuzz-
ied values and the least square method.
This paper is organized as follows. In Section 2, some background to Lees and Wang and Parkans approaches, along with
the three defuzzication methods, is presented. In Section 3, we dene our problem. Section 4 provides a detailed explana-
tion of our solution to the problem outlined in Section 3. Section 5 demonstrates the superiority of the WABL method and the
effectiveness of the iterative procedure supported by numerical results. Finally, our conclusions are presented in Section 6.

2. Preliminaries

e a1 ; a2 ; a3 ; a4 and B
Let A e B
e b1 ; b2 ; b3 ; b4 be two trapezoidal fuzzy numbers and S2 A; e be the similarity measure be-
tween A e and B. e Unlike Hsus similarity measure, Lees similarity measure includes a distance metric between fuzzy numbers,
which was also used by Tong and Bonissone [21]. The dissimilarity measure is dened as c  S2 A; e B,
e where c > 1. The value
of the parameter c affects the aggregation of experts opinions.
Lee [7] suggested minimizing the sum of the weighted dissimilarities between the aggregated opinion and each experts
opinion. Re i i 1; 2; . . . ; n represents the i-th experts opinion, and R e represents the aggregated opinion. To nd the R
e value,
the equation given below must be solved, where m > 1 is an integer, c > 1, and the wi values are weight coefcients:
   
e Pwi m c  S2 R
n
Z W; R e i ; R
e ! min
i1
s:t:   1
P
n
M WjW w1 ; w2 ; . . . ; wn ; wi P 0; wi 1 :
i1

The solution to this optimization problem was introduced in Lees study [7] as follows:

Xn
eP 1
R we i m0 R ei; 2
n m0
e
i1 w i i1
h  i1=m0 1
1= c  S2 R e i ; R
e
ei
w i1=m0 1 ; 3
Pn h  e i ; R
e
j11= c  S2 R

e and w
R e i values can be obtained only by an iterative procedure.
690 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698

As mentioned before, Wang and Parkan [23] suggested two methods, which are called LSDM (least squares distance
method) and DLSM (defuzzication-based least squares method). These methods are based on the solution of an optimiza-
tion problem by minimizing the sum of the squared distances between all pairs of the weighted opinions. In LSDM, fuzzy
opinions are used, whereas in DLSM, defuzzied values are used in calculations.
e i ri1 ; . . . ; r im and R
Let R e j r j1 ; . . . ; rjm be fuzzy numbers. The integer m denes the shape of a fuzzy number. For in-
stance, if m is 3, the fuzzy number will be triangular; if it is 4, the fuzzy number will be a trapezoidal fuzzy number. wi and wj
represent weight values.
The minimization problem for the LSDM can be shown as follows:
m 
P
n P
n P
J wi r ik  wj rjk 2 ! min
i1 j1 k1
ji
4
s:t:
Pn
wi 1; wi P 0; i 1; . . . ; n:
i1

The solution of this optimization problem is given below:

Theorem 1 [23]. Let W w1 ; . . . ; wn T be the optimum solution of the problem (4). Then,

G1 e
W P 0; 5
eT G1 e
where eT 1; . . . ; 1 is the transpose of e, and G1 is the inverse of G, elements of which are dened as
8
>
> P
m
< n  1
> r2ik ; i; j 1; . . . ; n; i j;
k1
g ij 6
>
> P
m
>
: r ik r jk ; i; j 1; . . . ; n; i j:
k1

The minimization problem for the DLSM can be shown as

P
n P
n
J wi zi  wj zj 2 ! min
i1 j1
ji
7
s:t:
Pn
wi 1; wi P 0; i 1; . . . ; n;
i1

where zi represents defuzzication values and is as follows:

1 Xm
zi r ik ; i 1; . . . ; n: 8
m k1

The solution of this optimization problem is given in the following theorem.

Theorem 2 [23]. Let W w1 ; . . . ; wn T be the optimum solution for the problem (7). The optimum solution can be given as
follows:
1=zi
wi Pn ; i 1; . . . ; n: 9
k1 1=zk

In our study, three defuzzication procedures are used: WABL (weighted averaging based on the levels), COA (center of
area), and MOM (mean of maxima). The mathematical form of the WABL method was dened by Nasibov [11,12] and given
as follows:
Z 1
IA cL LA a cR RA a  pada;
0
Z 1
10
pada 1;
0
cL P 0; cR P 0; cL cR 1; p : 0; 1 ! E :
WABL parameters cL and cR represent the weights of the left and right sides of the fuzzy number, respectively. LA a and RA a
are the left and right side functions of the fuzzy number, respectively. LA a is a non-decreasing function, and RA a is a non-
increasing function, with both being left continuous. pa is a distribution function of the importance of the level sets. By
using the distribution function, WABL adds all level sets into the defuzzication process [13].
 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698 691

In literature, the COA and MOM methods are the most frequently used methods [10,16]. The COA method determines the
center of the area of the membership function and is dened as follows:
R1
xlA xdx
COAA R1
1 : 11
l xdx
1 A

In the MOM method, the defuzzied value is the mean of the xi elements that have the maximum membership values.
MOMs mathematical form is shown below:
Xm
xi
MOMA : 12
i1
m

In this study, the WABL, COA, and MOM methods that were mentioned above will be used to calculate the defuzzication of
the fuzzy information.

3. Denition of the problem

In the student-centered learning system, after each problem-based learning session, each students performance is as-
sessed by using a predened set of evaluation criteria. These evaluation criteria are specied as leadership, research skill,
responsibility, discussion skill, and creativity, and they were dened by all lecturers in our department (Table 1). The impor-
tance, i.e., the weight, of each evaluation criterion could be different for each lecturer. Consequently, even if the fuzzy answer
is the same for any evaluation criterion, its reection, as a defuzzied value, might be different for each lecturer because of
the lecturers different points of view. Because of this, the WABL method is used for defuzzication. The fundamental reason
to use the WABL method is that it can be adjusted for a dened task to produce more accurate results when compared to the
other known methods [1114]. However, unlike the studies [1114], the focus for this study is to develop an algorithm for
obtaining the optimal parameters and to prove the consequent theorem.

Table 1
The evaluation form for each problem-based learning session.

STUDENT NAME-SURNAME GRADE

p
EVALUATION CRITERIA
1. LEADERSHIP w1 Absent Middle Strong
2. CREATIVITY w2 Absent Middle Strong
3. RESEARCH SKILL w3 Absent Few Middle Good Strong
4. RESPONSIBILITY w4 Absent Few Middle Good Strong
5. DISCUSSION SKILL w5 Absent Few Middle Good Strong

( x)

ABSENT MIDDLES STRONG

10 90 x
0 50 100

Fig. 1. Membership function for the criteria leadership and creativity.

692 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698

( x)

ABSENT FEW MIDDLE GOOD STRONG

0 10 30 50 70 90 100
x

Fig. 2. Membership function for the criteria research skill, responsibility and discussion skill.

The fuzzy values of the evaluation criteria are triangular or trapezoidal fuzzy numbers as shown in Figs. 1 and 2:
In general, the membership function of the trapezoidal fuzzy number A a; b; c; d is
8 xa
> ; x 2 a; b;
>
> ba
< 1; x 2 b; c;
lA x dx 13
> dc ;
> x 2 c; d;
>
:
0; otherwise:

In (13), a trapezoidal fuzzy number transforms into a triangular fuzzy number when b c. Similarly, a triangular fuzzy num-
ber can be denoted as a special case of a trapezoidal fuzzy number.
The defuzzication values of the fuzzy numbers R~ ij can be obtained from any of the defuzzication methods as explained
in [8,10,16]. However, for the reason indicated above, we have concentrated on the use of the WABL method, which gives
results that are more accurate.
The WABL values that are used in our study for the defuzzication of the triangular and trapezoidal fuzzy numbers are
calculated below.
Theorem 3 [14]. Let A a; b; c be a triangular fuzzy number. pa q 1aq , q P 0, and cL , cR can be any values that satisfy
normality and positivity conditions (10). The WABL value of the fuzzy number A can be calculated as follows:
   
q1 q1
IA cR c  c  b cL a b  a : 14
q2 q2

Theorem 4 [14]. Let A a; b; c; d be a trapezoidal fuzzy number with membership function (13). pa q 1aq , q P 0, and
cL , cR can be any values that satisfy normality and positivity conditions (10). The WABL value of the fuzzy number A can be cal-
culated as follows:
   
q1 q1
IA cR d  d  c cL a b  a : 15
q2 q2

Assume that the performances of n students will be evaluated by using m evaluation criteria. The evaluation result
of a student i with respect to criterion j is indicated as Re ij , where i 1; . . . ; n; j 1; . . . ; m. The numerical values of
students grades after evaluation of each respective student are denoted by p1 ; p2 ; . . . ; pn . A vector W w1 ; w2 ; . . . ; wm T
represents the unknown weights of the m evaluation criteria. Rij values are the deffuzzied values of the fuzzy numbers
e ij . Our objective function in this present study is to minimize the sum of squared distances between all grade
R
evaluations and the linguistic evaluations of each criterion. Therefore, our nonlinear optimization problem can be
constituted as follows:

!2
P
n P
m
LW wj Rij  pi ! min
i1 j1

s:t: 16
Pm
wj 1:
j1

Problem (16) is investigated in the next section.

 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698 693

4. Solution of the problem

We want to determine the wj , j 1; . . . ; m values that minimize the expression LW. There is no sign restriction and no
upper limit for the value of wj ; j 1; . . . ; m. Increasing values of wj ; j 1; . . . ; m show their effectiveness and the signs show
their positive or negative effect, as given below in more detail.
Theorem 5. Vector W w1 ; w2 ; . . . ; wm T , which represents the optimum solution for the problem (16), is

G1 e1  eT G1 RT P
W G1 RT P; 17
eT G1 e
where eT 1;    ; 1 is the transpose of e and G1 is the inverse of G. G and RT matrices and W and P vectors are, respectively, as
follows:
2 3
P
n P
n P
n
6 R2i1 Ri1 Ri2  Ri1 Rim 7
6 i1 i1 i1 7
6 7
6P n P
n Pn 7
6 Ri1 Ri2
6 R2i2  Ri2 Rim 7
7
6
G 6 i1 i1 i1 7 ; 18
7
6 .. .. .. .. 7
6 . . . . 7
6 7
6 n 7
4P P
n P 2 5
n
Ri1 Rim Ri2 Rim  Rim
i1 i1 i1 mm
2 3
R11 R21  Rn1
6 7
6 R12 R22  Rn2 7
6 7
RT 6 . .. .. .. 7 ; 19
6 .. . . . 7
4 5
R1m R2m    Rnm mn
3
2 2 3
w1 p1
6 7 6 7
6 w2 7 6 p2 7
6 7 6 7
W6 . 7 ; P6 . 7 : 20
6 .. 7 6 . 7
4 5 4 . 5
wm m1 pn n1

Proof. The Lagrange multiplier method can be applied to the problem (16) as below:
!2 !
X
n X
m X
m
LW wj Rij  pi  2k wj  1 : 21
i1 j1 j1

@LW
Letting @wz
0 for each z 1; . . . ; m we obtain
!
@LW Xn Xm
2 wj Rij  pi  Riz  2k 0; z 1; . . . ; m 22
@wz i1 j1

that can be simplied as

!
X
n X
m
wj Rij Riz  pi Riz  k 0; z 1; . . . ; m: 23
i1 j1

Expression (23) can be rewritten in matrix form as

GW  RT P  ke 0; 24
T
where G and R are matrices as dened in (18) and (19), respectively. As dened in problem (16), the sum of all weight val-
ues must be equal to 1, and this restriction can be rewritten as

eT W 1: 25
Consequently, (24) and (25) can be solved together as

1  eT G1 RT P
k 26
eT G1 e
694 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698

and
 
G1 e 1  eT G1 RT P

W G1 RT P: 27
eT G1 e
As a result, weight values can be calculated for each evaluation criterion, as allowed by Eq. (27). However, these calculated
weight values are valid for an optimism degree b cR , which is predened before the calculations explained above. From
this perspective, we can expand problem (16) by using Eq. (15) as shown in (28). In other words, when we take the WABL
defuzzication values Rij into account, a solution of the optimization problem given below will provide us with both the opti-
mal optimism degree b and optimal weight values wj for the obtained b :
!2
P
n P
m 
Lb; W wj bAij 1  bBij  pi ! min
i1 j1

s:t: 28
Pm
wj 1;
j1

0 6 b 6 1;
e ij and
where aij ; bij ; cij ; dij are parameters of the trapezoidal fuzzy number R
q1
Aij dij  dij  cij ; 29
q2
q1
Bij aij bij  aij : 30
q2
By substituting cR with b, cL with 1  b, and using denitions (29,30), and Theorem 4, the defuzzied value Rij can be written
as
Rij bAij 1  bBij : 31
By derivating the objective function of the problem (28) with respect to b and equating to zero we obtain,
!
@Lb; W Xn Xm Xm
2 bwj Aij  wj Bij wj Bij  pi  wj Aij  Bij  0: 32
@b i1 j1 j1

Thus, we have
Pn Pm 
i1 j1 wj Aij
 Bij wj Bij  pi
b Pn Pm 2 2
 33
i1 j1 wj Aij  Bij

or in the matrix form

BW  PT A  BW
b : 34
WT A  BT A  BW
SSE values

Beta values

Fig. 3. Beta values versus SSE values for six data sets.
 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698 695

Lemma 4.1. If Aij > Bij ; i 1 . . . ; n; j 1; . . . ; m, then Lb; W is a convex function with respect to the parameter b.
Pm
Proof. Since j1 wj 1; there are some wk 0; 1 6 k 6 m, that holds
2
@ Lb; W
w1 A11  B11    wm A1m  B1m 2    w1 An1  Bn1    wm Anm  Bnm 2 > 0: 35
@b2
In matrix form, it could be shown as

A  BWT A  BW > 0; 36

where
2 3 2 3
A11  B11 A12  B12    A1m  B1m w1
6A  B A22  B22    A2m  B2m 7 6w 7
6 21 21 7 6 2 7
AB6
6 .. .. .. .. 7
7 ; W6
6 ..
7
7 :
4 . . . . 5 4 . 5
An1  Bn1 An2  Bn2    Anm  Bnm nm wm m1

In Fig. 3, SSE versus b for different data sets is shown. It can be seen that all of these curves are convex.
Any standard optimization procedure (Golden Section, Binary Section, etc.) can be used for the calculation of the b value
because of the convexity of the Lb; W function with respect to the b values. In our study, a more effective iterative proce-
dure, which is explained below, is suggested.
G1 e1eT G1 RT P
Initially, R and W0 eT G1 e
G1 RT P are calculated as the solution of the problem (28) by using an initial b0 .
T
Then, a new b  WBWP
T
ABW
ABT ABW
is calculated by using W0 and iterations are repeated. In other words, the optimal solution
of the problem (28) could be obtained related to the iteration b0 ! W0 ! b1 ! W1 ! b2 !    ! b ! W . The following
algorithm illustrates this procedure.
Algorithm 1

Step 0. Enter the linguistic evaluations like leadership, creativity, etc. for each student.
Step 1. Enter numerical grade values that coincide with the linguistic evaluations (P vector).
Step 2. Dene e > 0 certainty, and initialize b0 and k as 0.5 and 0, respectively.
Step 3. Calculate Rij for i 1; . . . ; n and j 1; . . . ; m for the bk value from (31).
Step 4. Calculate G from (18) and then obtain G1 .
Step 5. Calculate wj for j 1; . . . ; m from (27).
Step 6. Calculate A and B matrices from (29) and (30).
Step 7. Obtain the b value from (34).
Step 8. If jbk  b j < e, then go to Step 10.

Table 2
Five data sets obtained from ve lecturers.

Student No. Evaluation criteria Lecturers grade evaluations

L C R.S. R D.S. Lec-1 Lec-2 Lec-3 Lec-4 Lec-5
1 M M F F M 40 45 40 40 52
2 M S G G G 80 75 80 65 80
3 A A F F F 30 15 15 15 32
4 A M M M F 45 45 35 35 48
5 S S G G S 85 90 80 90 92
6 S S S S S 100 100 100 100 100
7 M A F F M 30 25 30 30 44
8 A A F M F 30 25 20 20 36
9 S M G G G 75 75 75 75 80
10 A M F M M 40 40 35 35 48
11 M M G G S 75 70 70 50 76
12 A A A A A 0 10 0 0 20
13 M M S S S 80 75 85 80 84
14 M M M M M 50 50 40 50 60
15 S S M M S 80 75 80 80 84
16 M S M M G 75 70 70 65 72
17 A S M G M 70 65 60 55 64
18 A A A F F 20 30 10 10 28
19 S S F F S 60 75 65 70 76
20 S M G M G 75 70 70 70 76

L = leadership, C = creativity, R.S. = research skill, R = responsibility, D.S. = discussion skill, A = absent, F = few, M = middle, G = good, S = strong.
696 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698

Step 9. Update k as k 1 and bk as b , and go to Step 3.

Step 10. Determine wj for j 1; . . . ; m and b as optimal parameters.
Step 11. Stop.

5. Application

Our suggested algorithm has been developed as a software application in the C++ programming language, and six data
sets were used in our experiments. Five lecturers were asked to assign a numerical grade for 20 students linguistic evalu-
ations on a set of criteria, and these results are presented in Table 2.

Table 3
The results of WABL, COA and MOM methods for six data sets.

w1 w2 w3 w4 w5 b SSE

SET-1
WABL q 0 0.00956 0.30578 0.48621 0.01862 0.21707 0.6957 632.7768
WABL q 10 0.07401 0.27346 0.33260 0.15891 0.16102 1.0000 509.2870
WABL q 20 0.06713 0.26621 0.30243 0.17903 0.18521 1.0000 548.9734
COA 0.13666 0.26764 0.40699 0.02919 0.49121 898.5181
MOM 0.00827 0.23748 0.26071 0.17357 0.31998 619.8824
SET-2
WABL q 0 0.16148 0.36840 0.19952 0.10789 0.16271 0.6395 545.0104
WABL q 10 0.20645 0.30856 0.07994 0.30636 0.09868 0.9536 326.3041
WABL q 20 0.20073 0.30014 0.06035 0.32416 0.11461 1.0000 334.6471
COA 0.04321 0.34304 0.14735 0.07361 0.39279 788.8690
MOM 0.15407 0.28467 0.03247 0.35967 0.23406 450.6970
SET-3
WABL q 0 0.00840 0.27971 0.61982 0.26598 0.37484 0.5020 742.3373
WABL q 10 0.08986 0.26215 0.51052 0.09134 0.22881 0.5975 374.5684
WABL q 20 0.09250 0.25883 0.49935 0.07637 0.22568 0.6077 351.8783
WABL q P 160 0.09359 0.25485 0.48591 0.06086 0.22652 0.6050 6331.0163
COA 0.05211 0.27876 0.65237 0.30875 0.42972 840.3809
MOM 0.08842 0.25155 0.47249 0.05657 0.24411 331.0723
SET-4
WABL q 0 0.29591 0.29923 0.27982 0.01364 0.11140 0.5230 948.9585
WABL q 10 0.33875 0.27143 0.16596 0.22569 0.00183 0.6473 402.4287
WABL q 20 0.33742 0.26682 0.15718 0.24153 0.00295 0.6817 370.6944
WABL q P 35 0.33615 0.26435 0.15254 0.24902 0.00205 0.7014 6356.3858
COA 0.24779 0.29292 0.31159 0.05250 0.20021 1155.5135
MOM 0.31929 0.25954 0.11957 0.26545 0.03615 357.4269
SET-5
WABL q 0 0.22653 0.23129 0.17829 0.17679 0.18710 0.8922 32.4472
WABL q 10 0.15137 0.17143 0.08667 0.39351 0.37036 1.0000 534.5341
WABL q 20 0.14007 0.16359 0.11166 0.41176 0.39625 1.0000 672.0539
COA 0.02683 0.14553 0.03286 0.19090 0.72326 1038.6621
MOM 0.06001 0.13449 0.10612 0.35066 0.56096 1055.4155
SET-ALL
WABL q 0 0.13724 0.29743 0.35300 0.00240 0.20993 0.6512 4959.6541
WABL q 10 0.18517 0.26073 0.23407 0.18722 0.13281 0.9796 3732.1042
WABL q 20 0.17762 0.25380 0.20779 0.20717 0.15361 1.0000 3790.1983
COA 0.01510 0.26558 0.29708 0.02517 0.44742 6145.2656
MOM 0.12601 0.23355 0.14284 0.21854 0.27906 4332.4927

Table 4
Number of iterations of the iterative q 0 and Golden Section methods.

Certainty level Method Iterations

SET-1 SET-2 SET-3 SET-4 SET- 5 SET-ALL
e 0:001 Iterative 9 9 2 6 9 9
Golden Section 15 15 15 15 15 15
e 0:0001 Iterative 12 12 6 10 11 12
Golden Section 20 20 20 20 20 20
e 0:00001 Iterative 15 15 10 14 14 15
Golden Section 24 24 24 24 24 24
e 0:000001 Iterative 18 18 14 17 17 18
Golden Section 29 29 29 29 29 29
 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698 697

Table 5
The results of iterative method for SET-ALL data set with q 10.

Iteration numbers b wj ; j 1; 5 SSE

1 0.5000 0.11917 0.23974 0.15960 0.18827 0.29322 4451.0432
2 0.7369 0.15512 0.25160 0.18947 0.19484 0.20898 3927.0836
3 0.8637 0.17175 0.25678 0.21105 0.19263 0.16779 3777.8210
4 0.9260 0.17918 0.25900 0.22304 0.19012 0.14867 3742.0216
5 0.9552 0.18248 0.25997 0.22896 0.18862 0.13996 3734.1757
6 0.9685 0.18396 0.26039 0.23174 0.18787 0.13603 3732.5293
7 0.9746 0.18463 0.26058 0.23302 0.18752 0.13426 3732.1907
8 0.9773 0.18492 0.26067 0.23359 0.18736 0.13346 3732.1217
9 0.9786 0.18506 0.26070 0.23386 0.18728 0.13310 3732.1077
10 0.9791 0.18512 0.26072 0.23397 0.18725 0.13294 3732.1049
11 0.9794 0.18514 0.26073 0.23403 0.18723 0.13287 3732.1043
12 0.9795 0.18516 0.26073 0.23405 0.18723 0.13283 3732.1042
13 0.9796 0.18516 0.26073 0.23406 0.18723 0.13282 3732.1042

A common optimism degree and a weight value for each criterion are obtained by analyzing the ve lecturers relevant
data sets. The details of the process are presented in Table 3. In this table, the methods, which give the minimum least square
errors, are highlighted. As illustrated in the table, the value of q can be modied, and, therefore, WABL can produce better
results than do the MOM and the COA methods.
Optimum solutions found by using both our iterative method and the well-known Golden Section method are obtained
for different certainty levels to compare the effectiveness of the two methods (Table 4).
The comparison of our iterative method and the Golden Section method is performed with the sign test rather than the
Wilcoxon sign rank test because of the asymmetrical distribution of data. As shown in Table 4, when the results are inves-
tigated separately for each certainty level, the number of iterations in the Golden Section method remains constant. On the
other hand, the number of iterations changes in our iterative method. Although the number of iterations increases as the
certainty level decreases for both methods, our iterative method needs fewer iterations as compared to the Golden Section
method. The sign test showed that there is a statistically signicant difference between the number of iterations in our iter-
ative method and the Golden Section method at a 0:05 (p 0:0313. As a result, it can be said that our proposed method
can produce optimal solutions in fewer iterations when compared with Golden Section method.
A demonstration of the proposed iterative method applied on the merged data set SET-ALL for e 0:0001 can be seen in
Table 5. Starting from b 0:5, the suggested iterative procedure obtains the b and wj ; j 1; . . . ; 5 values as a target-driven
process.

6. Conclusions

When evaluating student performances, gradations can be carried out by different lecturers. Because each lecturer has
different opinions, the same student may receive different grades from different lecturers. Furthermore, these student eval-
uations were carried out in such a way that only the lecturer gave a single numerical value to the student based on his or her
subjective observations. Nevertheless, the student performance evaluation system is suggested as a result of the necessity of
carrying out such evaluations through linguistic terms, which are more inclined to the human way of thinking, and later
obtaining the numerical results of these evaluations. Hence, a mathematical model is suggested to create a common and
unbiased evaluation process by merging the opinions of all lecturers in order to estimate optimal parameters. The solution
to this model is provided using an iterative algorithm. With this model, the performance evaluations, calculated against lin-
guistic evaluations by using estimated parameters reecting the evaluation strategy of all of the lecturers, will take into con-
sideration the knowledge and experience of the lecturers. The result will be a common decision system and will reect not
only one but all decision makers opinions.
The main contribution of the study presented in this article is the development of an iterative algorithm for obtaining the
optimal parameters, along with the proof of the consequent theorem. Compared to other defuzzication methods, the WABL
method, used in calculations of the proposed iterative method as a basis for the defuzzication of the linguistic variables, has
been observed as superior by numerical results. Furthermore, it has been tested and shown that the proposed iterative meth-
od reaches the optimum parameter estimation in fewer steps when compared with the Golden Section method as the rep-
resentative of classical optimization methods.

Acknowledgments

The authors are very grateful to Professor Witold Pedrycz as well as the anonymous referees for their insightful and con-
structive comments and suggestions that have led to an improved version of this paper.
698 E.N. Nasibov, A. vg Kinay / Information Sciences 179 (2009) 688698

References

[1] A. Bardossy, L. Duckstein, L. Bogardi, Combination of fuzzy numbers representing expert opinions, Fuzzy Sets and Systems 57 (1993) 173181.
[2] C.G.E. Bonder, J.G. Graan, F.A. Lootsma, Multicriteria decision analysis with fuzzy pairwise comparisons, Fuzzy Sets and Systems 29 (1989) 133143.
[3] C.-T. Chen, Extensions of the TOPSIS for group decision-making under fuzzy environment, Fuzzy Sets and Systems 114 (2000) 19.
[4] H.M. Hsu, C.T. Chen, Aggregation of fuzzy opinions under group decision-making, Fuzzy Sets and Systems 79 (1996) 279285.
[5] C. Kahraman, D. Ruan, I. Dogan, Fuzzy group decision-making for facility location selection, Information Sciences 157 (2003) 135153.
[6] M.-S. Kuo, G.-H. Tzeng, W.-C. Huang, Group decision-making based on concepts of ideal and anti-ideal points in fuzzy environment, Mathematical and
Computer Modelling 45 (2007) 324339.
[7] H.S. Lee, Optimal consensus of fuzzy opinions under group decision-making environment, Fuzzy Sets and Systems 132 (2002) 303315.
[8] C.-T. Lin, C.S.G. Lee, Neural Fuzzy Systems, Prentice-Hall, 1996.
[9] J. Ma, D. Zhou, Fuzzy set approach to the assessment of student-centered learning, IEEE Transactions on Education 43 (2) (2000) 237241.
[10] J.M. Mendel, Uncertain Rule-Based Fuzzy Logic Systems, Prentice-Hall, 2001.
[11] E.N. Nasibov, Certain integral characteristics of fuzzy numbers and a visual interactive method for choosing the strategy of their calculation, Journal of
Computer and System Sciences International 41 (4) (2002) 584590.
[12] E.N. Nasibov, Aggregation of fuzzy values in linear programming problems, Automatic Control and Computer Sciences 37 (2) (2003) 111.
[13] E.N. Nasibov, R. Shikhlinskaya, Adjustment of the parameters of WABL-aggregation for locating the center of gravity of polynomial-type fuzzy number,
Automatic Control and Computer Sciences 37 (6) (2003) 3442.
[14] E.N. Nasibov, A. Mert, On WABL and COA defuzzications for polynomial shape triangular and trapezoidal fuzzy numbers, in: B. Turksen, R.A. Aliev, S.V.
Ulyanov (Eds.), Proc. 3rd Int. Conf. on Soft Computing, Computing with Words and Perceptions in System Analysis, Decision and Control, Antalya,
Turkey, 2005, pp. 149158.
[15] G. Pasi, R.R. Yager, Modeling the concept of majority opinion in group decision making, Information Sciences 176 (2006) 390414.
[16] W. Pedrycz, F. Gomide, Introduction to Fuzzy Sets, MIT Press, Cambridge, MA, 1998.
[17] T.L. Saaty, How to make a decision: the analytic hierarchy process, European Journal of Operational Research 48 (1990) 926.
[18] S. Saghaan, S.R. Hejazi, Multicriteria group decision-making using a modied fuzzy TOPSIS procedure, in: M. Mohammadian (Ed.), Proc. Int. Conf. on
Computational Intelligence for Modelling, Control and Automation, and Int. Conf. on Intelligent Agents, Web Technologies and Internet Commerce,
Vienna, Austria, 2005, pp. 215221.
[19] T. Tsabadze, The coordination index of nite collection of fuzzy sets, Fuzzy Sets and Systems 107 (1999) 177185.
[20] T. Tsabadze, A method for fuzzy aggregation based on group expert evaluations, Fuzzy Sets and Systems 157 (2006) 13461361.
[21] R.M. Tong, P.P. Bonissone, A linguistic approach to decision-making with fuzzy sets, IEEE Transactions on Systems, Man Cybernetics 10 (1980) 716
723.
[22] D. Yong, S. Wen-Kang, Aggregating fuzzy opinions under group decision-making, Journal of Computer and Systems Sciences International 42 (5) (2003)
727731.
[23] Y.-M. Wang, C. Parkan, Two new approaches for assessing the weights of fuzzy opinions in group decision analysis, Information Sciences 176 (2006)
35383555.
[24] Z. Xu, Group decision making based on multiple types of linguistic preference relations, Information Sciences 178 (2008) 452467.
[25] Z.-S. Xu, J. Chen, An interactive method for fuzzy multiple attribute group decision making, Information Sciences 177 (2007) 248263.