See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223888893

Consensus models for AHP group decision making under row geometric mean
prioritization method

Article  in  Decision Support Systems · June 2010

DOI: 10.1016/j.dss.2010.03.003 · Source: DBLP

CITATIONS READS

246 697

4 authors, including:

Wei-Chiang Hong Yinfeng Xu

Xi'an Jiaotong University
116 PUBLICATIONS   3,150 CITATIONS   
192 PUBLICATIONS   2,187 CITATIONS   
SEE PROFILE
SEE PROFILE

Some of the authors of this publication are also working on these related projects:

[Forecasting] Introduction and Welcome from the Editor-in-Chief (ISSN 2571-9394) View project

New models in Linguistic Decision Making and Computing with Words View project

All content following this page was uploaded by Yucheng Dong on 26 March 2018.

The user has requested enhancement of the downloaded file.

 Decision Support Systems 49 (2010) 281–289

Contents lists available at ScienceDirect

Decision Support Systems

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / d s s

Consensus models for AHP group decision making under row geometric mean
prioritization method
Yucheng Dong a,b,⁎, Guiqing Zhang a, Wei-Chiang Hong b, Yinfeng Xu a,c
a
Management School, The Key Lab of the Ministry of Education for Process Control and Efficiency Projects, Xi'an Jiaotong University, Xi'an 710049, China
b
Department of Information Management, Oriental Institute of Technology, No. 58, Section 2, Sichuan Road, Panchiao, Taipei 220, Taiwan
c
State Key Lab for Manufacturing Systems Engineering, Xi'an 710049, China

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

Article history: The consistency measure is a vital basis for consensus models of group decision making using preference
Received 29 August 2009 relations, and includes two subproblems: individual consistency measure and consensus measure. In the
Received in revised form 3 March 2010 analytic hierarchy process (AHP), the decision makers express their preferences using judgement matrices
Accepted 21 March 2010
(i.e., multiplicative preference relations). Also, the geometric consistency index is suggested to measure the
Available online 27 March 2010
individual consistency of judgement matrices, when using row geometric mean prioritization method
(RGMM), one of the most extended AHP prioritization procedures. This paper further defines the consensus
Keywords:
Group decision making
indexes to measure consensus degree among judgement matrices (or decision makers) for the AHP group
AHP decision making using RGMM. By using Chiclana et al.'s consensus framework, and by extending Xu and
Row geometric mean prioritization method Wei's individual consistency improving method, we present two AHP consensus models under RGMM.
Geometric consistency index Simulation experiments show that the proposed two consensus models can improve the consensus indexes
Consensus of judgement matrices to help AHP decision makers reach consensus. Moreover, our proposal has two
desired features: (1) in reaching consensus, the adjusted judgement matrix has a better individual
consistency index (i.e., geometric consistency index) than the corresponding original judgement matrix; (2)
this proposal satisfies the Pareto principle of social choice theory.
© 2010 Elsevier B.V. All rights reserved.

1. Introduction consensus degree in [27,29,30]. The consistency measure is used to

measure the difference among decision makers, and is a vital basis of
Preference relations are widely used in group decision making. consensus models. For group decision making using preference
According to element forms in preference relations, there are often relations, the consistency measure itself includes two subproblems [31]:
two kinds of preference relations: linguistic preference rela-
tions [17,27,28] and numerical preference relations (i.e., multiplica- 1. When can a decision maker, considered individually, be said to be
tive preference relations [26,37,42] and fuzzy preference consistent and
relations [11,12,20,22,40]). In general, group decision making pro- 2. When can a whole group of decision makers be considered
blems using preference relations are faced by applying two different consistent.
models (or processes) before a final solution can be given [29,30]: 1)
the selection model and 2) the consensus model. The selection model In this paper, we call the first subproblem individual consistency
obtains the final solution according to the preferences given by the measure, and the second subproblem consensus measure. Generally,
decision makers. It involves two different steps: aggregation of at the beginning for each group decision making problem, decision
individual preferences and exploitation of the collective preference. makers’ opinions may differ from each other substantially. And,
The consensus model is an important aspect in group decision consensus models are decision aid tools to help decision makers to
making [5,6,9,10,13,16,27,29,30,44]. Classically, consensus is defined reach consensus, based on the established consistency measure.
as the full and unanimous agreement of all the decision makers In the analytic hierarchy process (AHP) [37], multiplicative
regarding all the possible alternatives. However, some researchers preference relations are called judgement matrices, and adopted to
consider that complete agreement is not necessary in real life. This has express the decision makers’ preferences. Many researchers
led to the use of the consistency measure, which also is called “soft” [1,7,18,23,32,34–36,39,41,46] focus on the selection model in AHP
group decision making (i.e., aggregation rules and prioritization
⁎ Corresponding author. Department of Management Science, Management School,
methods). Two of the methods that have been found to be the most
Xi'an Jiaotong, University, Xi'an 710049, China. useful in AHP group decision making are the aggregation of individual
E-mail address: [email protected] (Y. Dong). judgments (AIJ) and the aggregation of individual priorities (AIP). AIJ

0167-9236/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.dss.2010.03.003
282 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289

follows the common resolution scheme of selection models, as Crawford and Williams [14] have shown that the solution to the
mentioned above. In the aggregation phase of AIJ, the weighted above problem is unique and can be simply found as the geometric
geometric mean method is used to aggregating individual judgement means of the rows of matrix A:
matrices to obtain a collective judgement matrix. In the exploitation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
phase of AIJ, prioritization methods, such as eigenvalue method (EM) ∏nj = 1 aij
1=n

[38] and row geometric mean method (RGMM) [14], are used to wi =
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi :
derive a priority vector to order collective judgement matrix. AIP has ∑ni = 1 1=n ∏nj = 1 aij
differences in the common resolution scheme among other selection
models. The weighted geometric mean method is employed to The RGMM is also called the logarithmic least squares method.
aggregate individual priorities derived using prioritization methods, Dong et al. [15] and Herman and Koczkodaj [25] show that the effects
to obtain the best alternative(s). of these two prioritization methods are very similar. Moreover, the
The consistency problem is also a critical step in AHP group decision computational times of EM and RGMM are o(n2) and o(n), respectively.
making [2,8,19,21,24,33,34,38,43,45]. A number of studies focus on Thus, RGMM has a less computational time than EM, and we select
individual consistency in AHP. Saaty [37] develops an individual RGMM as the prioritization method in this paper.
consistency index based on the EM. Crawford and Williams [14] propose
an individual consistency index based on RGMM. Aguarón et al. [2] call 2.2. AIJ and AIP
Crawford and Williams's consistency index the geometric consistency
index, and provide the thresholds associated with it. Many approaches Consider an AHP group decision making problem. Let D={d1,d2,...,dm}
[8,21,24,38,45] have been developed to aid the AHP decision makers to be the set of decision makers, and λ={λ1,λ2,...,λm} be the weight vector of
revise the individual inconsistency in judgement matrices. There are decision makers, where λk N 0,k =1,2,...,m, ∑m k = 1 λk =1. Let A
(k)
=
fewer studies about the consensus building for AHP group decision (a(k))
ij n × n
be the judgment matrix provided by the decision maker dk
making. Bryson [7], Moreno-Jiménez et al. [34] and Van den Honert [41] (k =1,2,...,m). As mentioned above, two of the methods that have been
also introduce several interesting approaches for consensus building in found to be the most useful in AHP group decision making are AIJ and AIP.
AHP. In AIJ and AIP, the weighted geometric mean method is generally used as
In general, research progress in group decision making using the aggregation procedure.
preference relations can benefit research in AHP. Recently, the
(1) The aggregation of individual judgments
consensus problem has become a hot topic in group decision making
For AIJ, the decision makers use the weighted geometric mean
using preference relations. In particular, Chiclana et al. [13] present a
method to aggregate individual judgement matrices to obtain a
framework for integrating individual consistency into consensus model.
collective judgement matrix, A(c) = (a(c)
ij )n × n, where
Inspired by Chiclana et al.'s study [13], this paper develops AHP
consensus models under RGMM. This paper is organized as follows. In m

ð cÞ ðkÞ λk
Section 2, we introduce some preliminary knowledge of AHP. In aij = ∏ aij :
k=1
Section 3, we define the consensus indexes (i.e., the geometric cardinal
consensus index and the geometric ordinal consensus index) for AHP, Then, we use certain prioritization method to derive a collective
and export Chiclana et al.'s consensus framework to AHP. In Section 4, priority vector w(c) = (w(c) (c) (c) T
1 ,w2 ,...,wn ) from A
(c)
to order the
we develop two AHP consensus models under Chiclana et al's alternatives.
framework and RGMM, and show some desired properties of the (2) The aggregation of individual priorities
proposed consensus models. In Section 5, an illustrative example is Let w(k) = (w(k) (k) T
1 ,...,wn ) be the individual priority vector derived
provided. Concluding remarks and future research agenda are presented from individual judgment matrix A(k) using certain prioritization
in Section 6. method. Then, the collective priority vector obtained using AIP is
w(c) = (w(c) (c) (c) T
1 ,w2 ,...,wn ) , where,
2. Preliminary knowledge

m ðkÞ λk
ð cÞ
∏k = 1 wi
2.1. Prioritization method wi =
 :
ðkÞ λk
∑ni = 1 ∏m k = 1 wi

Let A = (aij)n × n, where aij N 0 and aij × aji = 1, be a judgement

When selecting EM as the prioritization method, some researchers
matrix. The prioritization method refers to the process of deriving a
have some disputes on the use of the AIJ and the AIP. Ramanathan and
priority vector w = (w1,...,wn)T, where wi ≥ 0 and ∑ni = 1 wi = 1, from
Ganesh [35] observe that AIJ violates the Pareto principle of social choice
the judgement matrix A. Two most commonly used prioritization
theory [3], and suggest using AIP in AHP group decision making. Forman
methods (EM and RGMM) are listed below.
and Peniwati [23] argue that AIJ and AIP are philosophically different
(1) The eigenvalue method circumstances [8], and whether AIJ or AIP should be used depended
Saaty [37,38] proposes the principal eigenvector of A as the on whether the group intends to behave as a synergistic unit or as a
desired priority vector w, which can be obtained by solving the collection of individuals. However, when selecting RGMM as the priori-
linear system: tization method, Barzilai and Golany [4] have shown the equivalence
between AIJ and AIP (see Lemma 1).
T
Aw = λw; e w = 1;
Lemma 1. Let w(c) = (w1(c),w2(c),...,wn(c))T and w(c′) = (w1(c′),w2(c′),…,wn(c′))T
be two collective priority vectors, derived using AIJ and AIP under RGMM,
where λ is the principal eigenvalue of A. respectively. Then wi(c) = wi(c′), for i = 1,2,...n.
(2) Row geometric mean method
The RGMM uses the L2 metric in defining an objective function AIP doesn't violate the Pareto principle of social choice theory, so,
of the following optimization problem: based on Lemma 1, we have that AIJ satisfies the Pareto principle of
8 h


i2 social choice theory under RGMM (see Corollary 1).
< min∑n
i = 1 ∑j N i ln aij − lnðwi Þ− ln wj
: Corollary 1. Let w(k) = (w1(k),...,wn(k))T and w(c) = (w1(c),w2(c),...,wn(c))T be
:
s:t: wi ≥0; ∑ni = 1 wi = 1 as before. Then, if wi(k) ≥ wj(k) for k = 1,2,...,m, then wi(c) ≥ wj(c).
 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289 283

 
3. The consensus problem in AHP GOCI for GOCI(A(k)). If ∀k GOCI AðkÞ ≤GOCI, we conclude that the
acceptably ordinal consensus are reached among the decision makers.
3.1. Consistency measures in AHP
Remark 1. Let Q = {(w1,w2,...,wn)T|0 ≤ wi ≤ 1,∑ni= 1 wi = 1}. According
As indicated in previous sections, the problem of consistency to Definition 1 and Definition 2, we have that GCI ðAÞ =
2 min     2
measure itself includes two subproblems: individual consistency and ∑ibj log aij − log ðwi Þ + log wj ≤GCCIðAÞ. Thus,
ðn−1Þðn−2Þ w∈Q
consensus measure. For the RGMM, Crawford and Williams [14] and 2
we suggest GCCI≥GCI. Moreover, if GOCI(A) ≠ 0, then GOCIðAÞ≥ . So,
Aguarón et al. [2] have developed a consistency index to measure 2
n
individual consistency, namely the geometric consistency index (GCI) we also suggest GOCI≥ . It may be an interesting future research issue
n
(see Definition 1). to investigate the thresholds of GCCI and GOCI in details.

Definition 1. (Crawford and Williams [14]). Let A = (aij)n × n be a

judgement matrix, and let w = (w1,w2,...,wn)T be the priority vector 3.2. Exporting Chiclana et al.'s consensus framework to AHP
derived from A using the RGMM. The geometric consistency index
(GCI) is given by Recently, the consensus problem has become a hot topic in group



2 decision making using preference relations. In particular, Chiclana et
2 al. [13] present a framework for integrating individual consistency
GCI ðAÞ = ∑ log aij − log ðwi Þ + log wj :
ðn−1Þðn−2Þ ibj into consensus model. When exporting Chiclana et al.'s consensus
framework to the AHP group decision making under RGMM, the
When GCI(A) = 0, we consider A fully consistent. Aguarón et al. [2] implementation of this framework deals with a two-step procedure
also provide the thresholds (GCI) for GCI: GCI = 0:31 for n = 3; (see Fig. 1).
GCI = 0:35 for n = 4 and GCI = 0.37 for n N 4. When GCIðAÞbGCI, we
consider that the matrix A is of acceptably individual consistency. (1) Individual consistency improving method. The individual
In general, the computation of the consensus indexes for group consistency improving method firstly uses GCI to measure
decision making using preference relations is done by measuring the the individual consistency degree of judgement matrices A(k),
distance between individual preference values and collective prefer-  
k = 1,2,...,m. If GCI AðkÞ ≤GCI, it doesn't adjust A(k). Otherwise, it
ence values [5,6,13,16]. Inspired by this idea, we define the geometric helps the decision maker dk improve the GCI values of A(k).
cardinal consensus index for AHP group decision making using RGMM (2) Consensus model. Once all judgement matrices are of
(see Definition 2). acceptably individual consistency, we apply a consensus
model to reach acceptable consensus. The consensus model
Definition 2. Let A(k) = (aij)(k)
n × n be the judgement matrix provided by
firstly uses GCCI (or GOCI) to measure the consensus degree of
the decision makers dk, and w(c) = {w(c) (c) (c) T
1 ,w2 ,...,wn } be the collective (k)
 ðkÞ 
(1) (2) (m)
priority vectors, derived from {A , A ,...,A } using AIJ under  ðkÞ  matrices A ,k = 1,2,...,m. If GCCI A (k)≤GCCI (or
judgement
GOCI A ≤GOCI), k = 1,2,...,m, it doesn't adjust A , k = 1,2,...,
RGMM. Then, the geometric cardinal consensus index (GCCI) of Ak is
m. Otherwise, it helps the decision makers to reach acceptable
defined by
consensus. Finally, all judgement matrices A(k),k = 1,2,...,m, are

 2



2 of acceptably individual consistency and are of acceptable
ðkÞ ðkÞ ð cÞ ð cÞ
GCCI A = ∑ log aij − log wi + log wj : consensus.
ðn−1Þðn−2Þ ibj

If GCCI(Ak) = 0, then the kth decision maker is of fully cardinal In the AHP discipline, individual consistency improving methods
consensus with the collective preference. Otherwise, the smaller the are very widely investigated [8,21,24,38,45]. The RGMM version of Xu
value of GCCI(Ak), the more the cardinal consensus. According to the and Wei's individual consistency improving method [45] is discussed
actual situation, the decision in Appendix A.1, and is adopted to deal with individual inconsistency
  makers establish the thresholds GCCI for of judgement matrices in this paper. However, only a few researchers
GCCI(A(k)). If ∀k GCCI AðkÞ ≤GCCI, we conclude that the acceptably
cardinal consensus are reached among the decision makers. have discussed AHP consensus models. This paper mainly focuses on
Let w(k) =(w(k) (k) T AHP consensus models under RGMM. When the AHP decision makers
1 ,...,wn ) be the individual priority vector derived from
the judgment matrix A(k) using RGMM, and let v(k) =(v1(k),...,vn(k))T, where reach consensus under Chiclana et al.'s framework and RGMM, we
vi(k) is the position of the ith alternative in w(k). For example, if w(k) =
(0.3,0.2,0.5)T, then v(k) =(2,3,1)T. Let w(c) =(w1(c),w2(c),...,wn(c))T be the
collective priority vector, derived from {A(1),A(2),...,A(m)} using AIJ under
RGMM, and let v(c) =(v1(c),...,vn(c))T, where vi(c) is the position of the ith
alternative in w(c). Herrera-Viedma et al. [29,30] propose the comparison
approach of positions of alternatives between two preferences vectors to
measure the consensus degree. Inspired by this idea, we define the
geometric ordinal consensus index for AHP group decision making using
RGMM (see Definition 3).

Definition 3. Let w(k), w(c), v(k) and v(c) be as before. The geometric
ordinal consensus index (GOCI) of Ak is defined by

 1 n
= ∑ j vi −vi j :
ðkÞ ðkÞ ðcÞ
GOCI A
n i=1

If GOCI(Ak) = 0, then the kth decision maker is of fully ordinal

consensus with the collective preference. Otherwise, the smaller the
value of GOCI(Ak), the more the ordinal consensus. According to the
actual situation, the decision makers also may establish the thresholds Fig. 1. AHP consensus model under Chiclana et al.'s consensus framework.
284 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289

argue that the corresponding consensus model need satisfy the Step 4 Without loss of generality, suppose that
following properties:

 n o
ðτÞ ðkÞ
(1) it can improve the consensus indexes (i.e., CCCI and COCI) of GCCI Az = max GCCIðAz :
k
judgement matrices (or decision makers) to reach consensus,
(2) in reaching consensus, the adjusted judgement matrix has a
Let A(k) (k)
z + 1 = (aij,z + 1)n × n, where
better individual consistency index (i.e., GCI) than the
corresponding original judgement matrix, 8 0 1
> ðcÞ ð1−θÞ
(3) and it doesn't violate the Pareto principle of social choice theory. >
ðkÞ θ @wi;z A
>
>
< aij;z k=τ
ðkÞ ðcÞ
aij;z + = wj;z ;
1 >
>
4. Consensus models in AHP >
>
: aðkÞ k≠τ
ij;z
Under Chiclana et al.'s framework, once judgement matrices are of
acceptably individual consistency, we need to further apply consensus and z = z + 1. Then, go to Step 2.
ðkÞ ðcÞ
models to help AHP decision makers reach consensus. In this section, we Step 5 Let AðkÞ = Az and wðcÞ = wz . Output the adjusted
judgement

propose two AHP consensus models under Chiclana et al.'s framework matrix A , its cardinal consensus index GCCI AðkÞ , the
ðkÞ

and RGMM. collective priority vector wðcÞ and the number of iterations z.

4.1. Consensus models 4.1.2. Ordinal consensus model

Input: Judgment matrices {A(1), A(2),..., A(m)}, the weight vector of
Let {A(1), A(2),..., A(m)} and {λ1,λ2,...,λm} be as before. Let w(k) = (w1(k),..., decision makers {λ1,λ2,...,λm}, the threshold GOCI, and the
wn(k))T be the individual priority vector derived from the judgment established maximum number of iterations zmax ≥ 1 and
matrix A(k) using RGMM. Let w(c) = (w1(c),w2(c),...,wn(c))T be the collective 0 b θ b 1. n o
priority vector, derived from {A(1),..., A(m)} using AIJ under RGMM. In the Output: Adjusted judgement matrices
 Að1Þ ; Að2Þ ; :::; AðmÞ , the ordi-
subsection, we respectively propose two consensus models, based on nal consensus index GOCI AðkÞ (k = 1,2,...,m), the collective
 T
GCCI and GOCI. Without loss of generality, suppose that the judgment ðcÞ ðcÞ ðcÞ
matrix Aτ has a largest cardinal (or ordinal) consensus index. The main priority vector wðcÞ = w1 ; w2 ; :::; wn , and the number
step of the proposed consensus models is to construct a new judgement
of iterations z.
matrix
Aτ according to original judgement matrix Aτ. When structuring
Aτ = aτij , we suggest that Step 1 Let z = 0 and A(k) (k) (k)
z = (aijz )n × n = (aij )n × n;

Step 2 Let wz(k) = (w(k) (k) T

1,z ,...,wn,z ) be the individual priority vector
0 1ð1−θÞ

 ðcÞ derived from judgment matrix A(k) z using RGMM, and v(k) z =
ðτÞ ðτÞ θ @wi A
aij = aij ðcÞ
; (v(k)
1,z ,...,v(k) T
n,z ) , where v (k)
i,z is the position of the ith alternative in
wj
w(k) (c) (c) (c) (c) T
z . Let wz = (w1,z ,w2,z ,...,wn,z ) be the collective priority
vector, derived from the collective judgement matrix A(c) z ,
ðcÞ ðkÞ λk (c) (c)
where aijz = ∏m k = 1 ðaijz Þ ; using RGMM, and vz = (v1,z ,...,
where 0 b θ b 1. Follow this procedure until judgment matrices with (c) T (c)
vn,z ) , where vi,z is the position of alternative xi in w(c).
acceptable consensus are obtained or the established maximum number
Step 3 Calculate the ordinal consensus index
of iterations is obtained. Next, we describe these two consensus models
in details.
ðkÞ 1 n ðkÞ ðcÞ
GOCIðAz Þ = ∑ j vi;z −vi;z j ;
n i=1
4.1.1. Cardinal consensus model
Input: Judgment matrices {A(1), A(2),..., A(m)}, the weight vector of
for k = 1,2,...,m. If ∀k, GOCIðAðkÞ
z Þ≤GOCI or z ≥ zmax, then go to
decision makers {λ1,λ2,...,λm}, the threshold GCCI, the estab-
Step 5; otherwise, continue with the next step;
lished maximum number of iterations
n zmax ≥ 1 ando 0 b θ b 1. Step 4 Without loss of generality, suppose that
Output: Adjusted judgement matrices Að1Þ ; Að2Þ ; :::; AðmÞ , the cardinal


 n
o
consensus index GCCI AðkÞ (k = 1, 2,...,m), the collective ðτÞ
GOCI Az
ðkÞ
= max GOCI Az :

T k
ðcÞ ðcÞ ðcÞ
priority vector wðcÞ = w1 ; w2 ; :::; wn , and the number
of iterations z. Let A(k) (k)
z + 1 = (aij,z + 1)n × n, where

8 0 1ð1−θÞ
Step 1 Let z = 0 and A(k) (k) (k)
z = (aijz )n × n = (aij )n × n; >
>
>
θ wðcÞ
(c) (c) (c) (c) T
Step 2 Let wz = (w1,z ,w2,z ,...,wn,z ) be the collective priority vector, >
< aij;z @
ð kÞ i;z A k=τ
ðkÞ ð cÞ
derived from the collective judgement matrix A(c) (c)
z = (aijz ), aij;z + 1 = wj;z ;
ðcÞ ðkÞ λk >
>
where aijz = ∏m ða Þ , using RGMM. >
>
k = 1 ijz : ðkÞ
aij;z k≠τ
Step 3 Calculate the cardinal consensus index of A(k) z :

and z = z + 1. Then, go to Step 2.

 2 ðkÞ ðcÞ
ðkÞ
GCCI Az = Step 5 Let AðkÞ = Az , and wðcÞ = wz . Output the adjusted judge-
ðn−1Þðn−1Þ




  ment matrix AðkÞ , its ordinal consensus index GOCI AðkÞ , the
ðkÞ ð cÞ ð cÞ 2
∑ log aijz − log wi;z + log wj;z :
ibj collective priority vector wðcÞ and the number of iterations z.

ðkÞ
If ∀k, GCCIðAz Þ≤GCCI or z ≥ zmax, then go to Step 5; otherwise, Remark 2. The ordinal consensus model and the cardinal consensus
continue with the next step; model have similarity with each other. The main difference between
 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289 285

these two consensus models is using different consensus indexes. Step 1 We randomly generate m n × n judgement matrices {A(1),...,A(m)},
These two consensus models both are inspired by Xu and Wei's whose entries are uniformly randomly selected from {1/9,1/
individual consistency method [45]. They reduce into the RGMM 8,...,1,...,8,9}.
version of Xu and Wei's method, when the number of the decision Step 2 We random generate a weight vector, λ = {λ1,λ2,...,λn}, where
makers m = 1. λi(i = 1,2,...n) is uniformly distributed on [0,1]. Then we
λ
normalize λ = {λ1,λ2,...,λn}, that is λi = n i .
Remark 3. The proposed consensus models are artificial consensus ∑i = 1 λi

improving methods. We also may allow the decision makers to Step 3 Using the RGMM version of Xu and Wei's individual consistency
participate in consensus models. For example, if the decision maker improving method, we transform {A(1),...,A(m)} into the judge-
with the worst consensus index does not wish to change his/her original ment matrices with acceptably individual consistency. For
preference values, we consider the decision maker with the second simplicity, we still denote these judgement matrices with
worst consensus index and repeat the process. acceptably individual consistency as {A(1),...,A(m)}.
Step 4 We take {A(1),...,A(m)}, λ = {λ1,λ2,...,λm}, GCCI, zmax and 0 b θ b 1
4.2. Properties of consensus models as the inputs of the cardinal consensus model. Applying the
cardinal consensus model, we obtain the adjusted judgement
In this subsection, we introduce some desired properties of the
AðkÞ (k = 1,2,...,n), and the cardinal consensus indexes
matrices
proposed consensus models. Before proposing these properties, we GCCI AðkÞ (k = 1,2,...,n), and the number of iteration z.


introduce Lemma 2, presented in Escobar et al. [19]. Step 5 If max ðGCCI AðkÞ Þ≤GCCI, then p = 1. Otherwise p = 0. Output
k
Lemma 2. (Escobar et al. [19]). Let {A(1), A(2),..., A(m)}, {λ1,λ2,...,λm}, z and p.
   

and A(c) be as before. Then, GCI AðcÞ ≤ max GCI AðkÞ .

k n o In simulation method I, we replace the cardinal consensus model,
Theorem 1. L e t {A ( 1 ) ,A ( 2 ) , ... ,A ( m ) }, Að1Þ ; Að2Þ ; :::; AðmÞ and GCCI, GCCI by the ordinal consensus model, GOCI and GOCI, respectively.
n o
  ðkÞ 
ð1Þ ð2Þ
A ; A ; :::; A ðmÞ be as before. Then, GCI A ðkÞ ≤GCI A and We call the modified simulation method, simulation method II.

  ðkÞ  When setting different input parameters of simulation method I,
GCI A ðkÞ ≤GCI A , for k = 1,2,...,m. we run this simulation method 1000 times, obtaining the average
The proof of Theorem 1 is provided in Appendix A.2. values of z and p. We also set different input parameters for simulation
Theorem 1 guarantees the adjusted judgement matrix has a better method II, and run it 1000 times to obtain the average values of z and
individual consistency index (i.e., GCI) than the corresponding original p. Tables 1 and 2 respectively show the corresponding average values
judgement matrix, when using the proposed consensus models. of z and p, under the different input parameters for simulation method
From Theorem 1, we have Corollary 2. I and simulation method II.

Corollary 2. Let AðcÞ and AðcÞ be the collective judgement matrices of Note 1. The average value of p has a definite physical implication, and
fAð1Þ ; Að2Þ ; :::; AðmÞ g a n d fAð1Þ ; Að2Þ ; :::; AðmÞ g, r e s p e c t i v e l y . T h e n , reflects the success ratio of reaching consensus in the simulation
experiments. Moreover, in running the simulation experiments, we
GCIðAðcÞ Þ≤GCI a n d GCIðAðcÞ Þ≤GCI u n d e r t h e c o n d i t i o n t h a t
set zmax = 300, and approximatively consider that the decision makers
GCIðAðkÞ Þ≤GCI for k = 1,2,...,m.
are of fully cardinal consensus when GCCI(A(k)) ≤ 0.01 for k = 1,2,...,m.
Theorem 2. Let {A(1),A(2),...,A(m)}, w(k) = {w(k) (k) (k)
1 ,w2 ,...,wn } (k = 1,2,...,
n o From Tables 1 and 2, we have the following observations:
ðcÞ ðcÞ ðcÞ ðcÞ ðcÞ ðcÞ
m), wðcÞ = w1 ; w2 ; :::; wn and wðcÞ = fw1 ; w2 ; :::; wn g be as
ðcÞ ðcÞ
before. Then, if wi(k) N wj(k), for k = 1,2,...,m, we have that wi Nwj and (1) Simulation method I/Simulation method II definitely help
ðcÞ ðcÞ decision makers reach cardinal/ordinal consensus, under the
wi Nwj .
established input parameters.
(2) With the increase of n and m, the number of the iterations z is
The proof of Theorem 2 is provided in Appendix A.2. increasing. Moreover, the smaller the parameter θ, the smaller
Theorem 2 guarantees the proposed consensus models satisfy the the number of the iterations z.
Pareto principle of social choice theory.
Based on Theorem 1, Theorem 2 and the above simulation analysis,
4.3. Discussion on convergence of the consensus models we find that these two consensus models satisfy the properties
presented in Section 3.2.
Naturally, we hope that the proposed consensus models can
improve the consensus indexes (i.e., GCCI and GOCI) of judgement Table 1
matrices to help the decision makers reach consensus. In the Average values of z and p for simulation method I under the different input parameters.
following, we use simulation methods to study this issue. The main
n m GCCI θ = 0.1 θ = 0.3 θ = 0.5 θ = 0.8
idea of the simulation methods is to randomly generate decision
makers’ judgement matrices and the corresponding weights of the p z p z p z p z
decision makers. Using the RGMM version of Xu and Wei's individual 3 3 0.01 1 5.49 1 7.67 1 11.76 1 31.09
consistency improving method [45], we transform these judgement 0.31 1 2.02 1 2.76 1 3.55 1 9.07
matrices into ones with acceptably individual consistency. Then, we 3 7 0.01 1 12.42 1 17.96 1 27.6 1 76.59
0.31 1 5.68 1 6.9 1 9.2 1 22.63
take these judgement matrices with acceptably individual consistency 4 5 0.01 1 9.35 1 13.05 1 19.53 1 54.16
as the inputs of the proposed consensus models to study the 0.35 1 4.08 1 4.61 1 6.33 1 14.88
convergence of the proposed consensus models. 4 9 0.01 1 15.7 1 22.63 1 35.19 1 98.13
Next, we describe the simulation methods in details. 0.35 1 7.73 1 8.65 1 11.81 1 28.11
7 7 0.01 1 12.11 1 16.35 1 25.45 1 70.62
0.37 1 6.03 1 6.2 1 7.21 1 15.53
Simulation method I 7 9 0.01 1 15.25 1 20.96 1 32.66 1 91.01
Input: n, m, GCCI, zmax and 0 b θ b 1. 0.37 1 7.98 1 8.19 1 9.68 1 20.77
9 20 0.37 1 18.58 1 19.14 1 19.62 1 39.6
Output: z and p
286 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289

Table 2 Table 3
Average values of z and p for simulation method II under the different input parameters. The GCI, GCCI and GOCI values of A(k) (k = 1,2,...,5) and A(c).

n m GOCI θ = 0.1 θ = 0.3 θ = 0.5 θ = 0.8 A(1) A(2) A(3) A(4) A(5) A(c)

p z p z p z p z GCI(•) 0.1349 0.2358 0.1194 0.1657 0.2208 0.056

GCCI(•) 0.4285 0.8707 0.4113 0.3774 2.1944 Undefined
3 3 0 1 2.28 1 2.90 1 3.76 1 9.73
GOCI(•) 0 0 0 0 1 Undefined
2/3 1 0.83 1 0.94 1 1.07 1 2.33
3 7 0 1 6.58 1 8.43 1 11.96 1 31.47
2/3 1 2.64 1 3.24 1 3.93 1 9.33
4 5 0 1 6.22 1 8.16 1 11.54 1 28.47
1/2 1 3.72 1 4.68 1 5.98 1 14.42
4 9 0 1 11.94 1 15.44 1 22.44 1 60.43
Let w(c) = (w1(c),w2(c),...,wn(c))T be the collective priority vector, derived
1/2 1 6.98 1 9.12 1 13.11 1 30.95
7 7 0 1 14.09 1 19.47 1 28.74 1 56.22 from A(c) using RGMM. A(c) and w(c) are listed below.
2/7 1 11.35 1 14.95 1 21.77 1 54.9 0 1
7 9 0 1 18.82 1 24.69 1 37.88 1 99.95 1 2:227 3:4756 5:0937
2/7 1 14.56 1 19.46 1 28.69 1 75.55 B 0:4490 1 1:9663 4:0005 C
AðcÞ =B
@ 0:2877
C;
9 20 2/9 1 38.44 1 54.7 1 84.54 1 235.93 0:5086 1 2:4623 A
0:1963 0:2500 0:4061 1

wðcÞ = f0:4984; 0:2727; 0:1541; 0:0747gT :

5. Numerical examples
We compute the GCI, GCCI and GOCI values of A(k)(k = 1,2,...,5) and
In order to show how the consensus models work in practice, let us A(c). These values are listed in Table 3.
consider the following example. Suppose we have a set of five decision Table 3 shows that GCIðAðkÞ ÞbGCI = 0:35 (k = 1,2,...,5), so A(1),...,A(5)
makers providing the following judgement matrices {A(1),...,A(5)} on a are of acceptably individual consistency.
set of four alternatives. Let w(k) = (w1(k),...,w4(k))T be the individual When setting GCCI = 0:35 and θ = 0.8, we apply the cardinal
priority vector derived from judgment matrix A(k) using RGMM. A(k) consensus model to adjust judgement matrices. The adjusted
and w(k) (k = 1,2,...,5) are listed below. judgement matrices and the collective judgement matrix are
0 1 AðkÞ ; ðk = 1; 2; :::; 5Þ and AðcÞ , respectively. The corresponding prior-
ðkÞ ðkÞ
1 4 6 7 i t y v e c t o r a r e wðkÞ = ðw1 ; :::; w4 ÞT ( k = 1 , 2 , . . . , 5 ) a n d
B1=4 1 3 4C
Að1Þ B
=@ C; ðcÞ ðcÞ
wðcÞ = ðw1 ; :::; w4 ÞT , respectively. The cardinal consensus model
1= 6 1= 3 1 2A
1= 7 1= 4 1= 2 1 ends in the 4th iteration. We find that AðkÞ = AðkÞ and wðkÞ = wðkÞ for

wð1Þ = f0:6145; 0:2246; 0:0985; 0:0624gT : k = 1,2,...,4. And, Að5Þ , AðcÞ , wð5Þ , and wðcÞ are listed below.
0 1
0 1 1:0000 1:1297 2:2014 4:3352
1 5 7 9 B 0:8852 1:0000 1:1044 3:4108 C
B1=5 1 4 6C Að5Þ =B
@ 0:4543
C;
Að2Þ B
=@ C; 0:9055 1:0000 2:6158 A
1= 7 1= 4 1 2A
0:2307 0:2932 0:3823 1:0000
1= 9 1= 6 1= 2 1
ð5Þ T
w = f0:3954; 0:2949; 0:2222; 0:0875g :
wð2Þ = f0:6461; 0:227; 0:0793; 0:0476gT :
0 1 0 1
1 3 5 8 1:0000 2:8439 4:4040 6:4243
B1=3 5C B 0:3516 1:0000 2:4940 4:1575 C
Að3Þ =B
1 4 C; AðcÞ B
=@ C;
@1= 5 1= 4 1 2A 0:2271 0:4010 1:0000 2:1677 A
1= 8 1= 5 1= 2 1 0:1557 0:2405 0:4613 1:0000

w
ð3Þ
= f0:5693; 0:2764; 0:0967; 0:0575g :
T wðcÞ = f0:5540; 0:2556; 0:1233; 0:0671gT :

0 1
1 4 5 6
B1=4 1 3 3C We compute the GCI and GCCI values of AðkÞ ðk = 1; 2; :::; 5Þ and AðcÞ .
Að4Þ =B
@1= 5 1= 3
C;
1 2A These values are listed in Table 4.
1= 6 1= 3 1= 2 1 When setting GOCI = 0 and θ = 0.5, we apply the ordinal
ð4Þ T consensus model to adjust judgement matrices. The adjusted
w = f0:5967; 0:2208; 0:1089; 0:0736g :
judgement matrices and the collective judgement matrix are
0 1
1 1= 2 1 2 AðkÞ ; ðk = 1; 2; :::; 5Þ and AðcÞ , respectively. The corresponding priority
B 2 1 1=2 3C ðkÞ ðkÞ
Að5Þ =B C; vector are wðkÞ = ðw1 ; :::; w4 ÞT ( k = 1 , 2 , . . . , 5 ) and
@ 1 1= 2 1 4A ðcÞ ðcÞ
1=2 1= 3 1= 4 1 wðcÞ = ðw1 ; :::; w4 ÞT , respectively. The ordinal consensus model

wð5Þ = f0:2247; 0:2958; 0:378; 0:1015gT :

Table 4
The GCI and GCCI values of AðkÞ ðk = 1; 2; :::; 5Þ and AðcÞ .
Note 2. The judgment matrices A(1),A(2),...,A(4) are first provided in Xu
Að1Þ Að2Þ Að3Þ Að4Þ Að5Þ AðcÞ
[43].
Let λ = {0.1,0.3,0.1,0.2,0.3} be the weights of decision makers. Let A(c) GCI(•) 0.1349 0.2358 0.1194 0.1657 0.037 0.0688
GCCI(•) 0.1054 0.2579 0.1051 0.1150 0.3138 Undefined
be the collective judgement matrix derived from {A(1),...,A(5)} using AIJ.
 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289 287

ends in the 2nd iteration. We find that AðkÞ = AðkÞ and wðkÞ = wðkÞ for Education of China. Wei-Chiang Hong acknowledges the financial
k = 1, 2, ..., 4. And, Að5Þ , AðcÞ , wð5Þ , and wðcÞ are listed below. support of grants (nos. 98-2410-H-161-001 and 98-2811-H-161-001)
from NSC of Taiwan.
0 1
1 0:9560 1:7984 3:6521
B 1:0460 1 0:9406 3:3085 C Appendix A
B
=@ C;
Að5Þ 0:5560 1:0631 1 2:8719 A
0:2738 0:3023 0:3482 1 A.1. The RGMM version of Xu and Wei's method
ð5Þ T
w = f0:3028; 0:3021; 0:2985; 0:0966g :
Xu and Wei [45] propose an individual consistency improving
0 1 method based on the EM and Saaty's consistency index. In Xu and Wei’
1:0000 2:7049 4:1448 6:1022
B 0:3697 method, we replace EM by RGMM, and replace Saaty's consistency index
B 1:0000 2:3767 4:1197 C
C;
AðcÞ =@ by the geometric consistency index. In this way, we obtain the RGMM
0:2413 0:4207 1:0000 2:2293 A
0:1639 0:2427 0:4486 1:0000 version of this method.
Let z be the number of iterations, and 0 b θ b 1. The steps of the RGMM
T
wðcÞ = f0:5416; 0:2597; 0:1299; 0:0688g : version of Xu and Wei's method are described as follows:

We compute the GCI and GOCI values of AðkÞ ðk = 1; 2; :::; 5Þ and The RGMM version of Xu and Wei's method
AðcÞ . These values are listed in Table 5. Input: The judgement matrix A, the threshold GCI and 0 b θ b 1.
We can find that the results in this example are in accordance with Output: The adjusted judgement matrix B, its geometric consistency
Theorems 1 and 2. index GCI(B) and the number of iterations z.

6. Conclusion
Step 1 Let A(0) = (aij(0))n × n = (aij)n × n and z = 0;
Consensus models have been a hot topic in group decision making Step 2 Let w(z) = (w1(z),w2(z),...,wn(z))T be the priority vector derived
using preference relations. In general, research progress in group from A(z) using RGMM.
decision making using preference relations can benefit research in Step 3 We calculate the geometric consistency index:
AHP. This paper defines the consensus indexes (i.e., the geometric
2 2
cardinal consensus index and the geometric ordinal consensus index) GCIðAÞ = ∑ ðlogðaij Þ− logðwi Þ + logðwj ÞÞ :
to measure consensus degree among judgement matrices (or decision ðn−1Þðn−2Þ ibj
makers). Inspired by Chiclana et al.'s consensus framework [13] and
Xu and Wei's individual consistency improving method [45], we Step 4 If GCIðAðzÞ Þ≤GCI, then go to Step 6; otherwise, continue with
propose two new AHP consensus models (i.e., cardinal consensus the next step;
model and ordinal consensus model) under RGMM. Simulation results Step 5 Let A(z+ 1) = (aij(z+ 1))n × n, where
and field experiments show these consensus models can effectively
improve the consensus indexes (i.e., GCCI and GOCI) of judgement 0 1ð1−θÞ

θ wðzÞ
ðz + 1Þ ðzÞ @ i A
matrices to help AHP decision makers reach consensus. Comparing aij = aij :
ðzÞ
the existing AHP consensus models, the proposed two consensus wj
models have two desired features:
Step 6 Let B = A(z). Output the adjusted judgement matrix B, its
(1) in reaching consensus, the adjusted judgement matrix has a
geometric consistency index GCI(B) and the number of
better individual consistency index (i.e., geometric consistency
iterations z.
index) than the corresponding original judgement matrix,
(2) and they satisfy the Pareto principle of social choice theory.
Thus, our proposals are not only very effective to help AHP Theorem 3. Let {A(z)} be the judgement matrix sequence in the RGMM
decision makers reach consensus, but also providing more convinced version of Xu and Wei's method. Then, we have that GCI(A(z + 1)) b GCI(A(z))
alternatives. for each z, and lim ðGCIðAðzÞ ÞÞ≤GCI.
In addition, similar to individual consistency improving method of z→∞
The proof of Theorem 3 is provided in Appendix A.2.
AHP [8,21,45], the consensus models should only be considered as a
Theorem 3 guarantees that any judgement matrix with unacceptably
decision aid, which the decision makers use as a reference to modify
individual consistency can be transformed into one with acceptably
their own judgement matrices. Therefore, a feasible future research
individual consistency (in the GCI sense), by using the RGMM version of
issue could be focused on the effectiveness assessment of consensus
Xu and Wei's method.
models.

Acknowledgements A.2. The proofs of theorems

Yucheng Dong and Yinfeng Xu acknowledge the financial support of

Proof of Theorem 1. We first prove GCIðAðkÞ Þ≤GCIðAðkÞ Þ. Let A(k) z =
grants (nos. 70801048 and 70525004) from NSF of China and a grant
(a(k)
ijz )n × n (k = 1,2,...,m) be the adjusted judgment matrices in the zth
(no. 200806981067) form the Ph.D. Programs Foundation of Ministry of
iteration using the cardinal consensus model. Let wz(c) =(w1,z (c) (c) (c) T
,w2,z,...,wn,z )
be the collective priority vector, derived from the collective judgement
ðcÞ ðkÞ λk
matrix A(c) (c)
z = (aijz )n × n, where aijz = ∏k = 1 ðaijz Þ , using RGMM.
m
n o
Table 5 ðτÞ ðkÞ
Without loss of generality, suppose that GCCIðAz Þ = max GCCIðAz Þ .
The GCI and GOCI values of AðkÞ ðk = 1; 2; :::; 5Þ and AðcÞ . k
We consider two cases. □
Að1Þ Að2Þ Að3Þ Að4Þ Að5Þ AðcÞ

GCI(•) 0.1349 0.2358 0.1194 0.1657 0.1077 0.0659 Case A k = τ. In this case, according to the cardinal consensus
GCCI(•) 0 0 0 0 0 Undefined ðkÞ
model, we have that A(k) (k)
z + 1 = (aij,z + 1)n × n, where aij;z + 1 =
288 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289

ðkÞ
ðcÞ
wi;z ð1−θÞ
ðcÞ
wi;z For each z, GCI(A(z)) ≥ 0. Thus, the sequence {GCI(A(z))} is monotone
ðaij;z Þθ ð ðcÞ
Þ . Let Wz = ðwij;z Þ = ð ðcÞ
Þ. From Lemma 2, we
wj;z wj;z decreasing and has lower bounds. Then we have
have that GCI(A(k) (k) (k)
z+ 1 ) ≤ max{GCI(Az ),GCI(Wz)} = GCI(Az ).


 n
o
lim ðzÞ ðzÞ
GCI A = inf GCI A :
z→∞
Case B k ≠ τ. In this case, A(k) (k) (k) (k)
z + 1 = Az . So, GCI(Az + 1) = GCI(Az ). lim  ðzÞ 
Let A = A∞ , then
GCI(A(k) z→∞
Summarizing Case A and Case B, we have that z + 1) ≤ GCI
 ∞ n
o
(A(k)
z ) for k = 1,2,...,m. GCI A = inf GCI A
ðzÞ
:
ðkÞ
Consequently, GCIðAðkÞ Þ≤GCIðA0 Þ = GCIðAðkÞ Þ for k = 1,2,...,m.
Similarly, we can prove GCIðAðkÞ Þ≤GCIðAðkÞ Þ. This completes the
Suppose that GCIðA∞ Þ N GCI. By applying the RGMM version of Xu
proof of Theorem 1.
and Wei's method to continue improving the consistency of A∞, we
can obtain the adjusted judgement matrix A∞ . Obviously, we have
∞
ðcÞ ðcÞ GCIðA∞ ÞbGCIðA
Þ:  

Proof of Theorem 2. We first prove wi Nwj . Let A(k) (k)

z = (aijz )n × n
Thus, GCI A∞ b inf GCI AðzÞ , which contradicts the definition of
(k = 1,2,...,m) be adjusted judgment matrices in the zth iteration using (z)
inf{GCI(A )}.
the cardinal consensus model. Let wz(k) = (w1,z (k) (k) T
,...,wn,z ) be individual
This completes the proof of Theorem 3. □
priority vector derived from judgment matrix A(k) z using RGMM. Let
w(c) (c) (c) (c) T
z = (w1,z ,w2,z ,...,wn,z ) be the collective priority vector, derived
References
from the collective judgement matrix A(c) (c)
z = (a ijz ) n × n , where
ðcÞ ðkÞ λk
aijz = ∏m k = 1 ða ijz Þ , using RGMM. □ [1] J. Aguarón, M.T. Escobar, J.M. Moreno-Jiménez, Consistency stability intervals for a
judgement in AHP decision support systems, European Journal of Operational
Research 145 (2) (2003) 382–393.
By induction over z, the proof can be completed. For z = 0, we have [2] J. Aguarón, J.M. Moreno-Jiménez, The geometric consistency index: Approximated
that w(k)
z =w
(k)
. Since wi(k) N wj(k) (k = 1,2,...,m), we have that wi,0 (k)
N thresholds, European Journal of Operational Research 147 (2003) 137–145.
(k) (c) (c) [3] K.J. Arrow, Social Choice and Individual Values, 2nd edition, Wiley, New York, 1963.
wj,0 for k = 1,2,...,m. According to Corollary 1, we have that wi,0 N wj,0 . [4] J. Barzilai, B. Golany, AHP rank reversal normalization and aggregation rules, INFOR
Suppose that it is true for z, i.e., w(k) (k) (c)
i,z N wj,z for k = 1,2,...,m, and wi,z N 32 (1994) 57–64.
(c)
wj,z . [5] D. Ben-Arieh, T. Easton, Multi-criteria group consensus under linear cost opinion
ðτÞ elasticity, Decision Support Systems 43 (2007) 713–721.
Without loss of generality, suppose that GCCIðAz + 1 Þ = max
ðkÞ k [6] D. Ben-Arieh, T. Easton, B. Evans, Minimum cost consensus with quadratic cost
fGCCIðAz + 1 Þg: For z + 1, we consider two cases. functions, IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and
Humans 39 (2009) 210–217.
[7] N. Bryson, Group decision-making and the analytic hierarchy process: exploring
Case A k = τ. In this case, according to the cardinal consensus model, we the consensus-relevant information content, Computers & Operations Research
ðcÞ 23 (1996) 27–35.
ðkÞ ðkÞ wi;z ð1−θÞ
have that Az(k) (k)
+ 1 =(aij,z + 1)n × n, where aij;z + 1 = ðaij;z Þθ ð ðcÞ
Þ . [8] D. Cao, L.C. Leung, J.S. Law, Modifying inconsistent comparison matrix in analytic
wj;z hierarchy process: a heuristic approach, Decision Support Systems 44 (2008)
(k) (k) (c) (c)
Since wi,z N wj,z and wi,z N wj,z , according to Corollary 1, we have 944–953.
(k) (k) [9] Y.L. Chen, L.C. Cheng, Mining maximum consensus sequences from group ranking
that wi,z + 1 N wj,z + 1. data, European Journal of Operational Research 198 (2009) 241–251.
[10] A.K. Choudhury, R. Shankar, M.K. Tiwari, Consensus-based intelligent group
decision-making model for the selection of advanced technology, Decision Support
Systems 42 (2006) 1776–1799.
Case B k ≠ τ. In this case, Az(k) (k) (k) (k) (k)
+ 1 = A z and wz + 1 = w z . Since w i,z N [11] F. Chiclana, F. Herrera, E. Herrera-Viedma, Integrating three representation models in
(k) (k) (k)
wj,z , we have wi,z + 1 N wj,z + 1. fuzzy multipurpose decision making based on fuzzy preference relations, Fuzzy Sets
and Systems 97 (1998) 33–48.
Summarizing Case A and Case B, we have that w(k) (k)
i,z + 1 N wj,z + 1for [12] F. Chiclana, F. Herrera, E. Herrera-Viedma, Integrating multiplicative preference
k = 1,2,...,m. From Corollary 1, we have that w(c)
i,z + 1N w(c)
j,z + 1. relations in a multipurpose decision-making model based on fuzzy preference
ðcÞ ðcÞ relations, Fuzzy Sets and Systems 122 (2001) 277–291.
Consequently, wi Nwj . [13] F. Chiclana, F. Mata, L. Martínez, E. Herrera-Viedma, S. Alonso, Integration of a
ðcÞ ðcÞ
Similarly, we can prove wi Nwj . This completes the proof of consistency control module within a consensus decision making model, Interna-
tional Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 16 (2008)
Theorem 2.
35–53.
[14] G. Crawford, C. Williams, A note on the analysis of subjective judgement matrices,
Proof of Theorem 3. Let Q = {(w1,w2,...,wn)T|0 ≤ wi ≤ 1,∑in= 1 wi = 1}. Journal of Mathematical Psychology 29 (1985) 387–405.
According to Definition 1, we have that [15] Y.C. Dong, Y.F. Xu, H.Y. Li, M. Dai, A comparative study of the numerical scales and the
prioritization methods in AHP, European Journal of Operational Research 186 (2008)
229–242.

 2



 2 [16] Y.C. Dong, Y.F. Xu, H.Y. Li, B. Feng, The OWA-based consensus operator under
z + 1 ðz + 1Þ ðz + 1Þ ðz + 1Þ
GCI A = ∑ log aij −log wi + log wj linguistic representation models using position indexes, European Journal of
ðn−1Þðn−2Þ ibj
Operational Research 203 (2010) 455–463.



2 [17] Y.C. Dong, Y.F. Xu, S. Yu, Linguistic multiperson decision making based on the use
2 ðz + 1Þ of multiple preference relations, Fuzzy Sets and Systems 160 (2009) 603–623.
= min ∑ log aij − log ðwi Þ + log wj
ðn−1Þðn−2Þ w∈Q ibj [18] M.T. Escobar, J.M. Moreno-Jiménez, Aggregation of individual preference structures
in AHP group decision making, Group Decision and Negotiation 16 (2007) 287–301.
2



 2 [19] M.T. Escobar, J. Aguarón, J.M. Moreno-Jiménez, A note on AHP group consistency for
ðz + 1Þ ðzÞ ðzÞ
≤ ∑ log aij − log wi + log wj :
ðn−1Þðn−2Þ ibj the row geometric mean priorization procedure, European Journal of Operational
Research 153 (2004) 318–322.
[20] Z.P. Fan, S. Xiao, G. Hu, An optimization method for integrating two kinds of
wi
ðzÞ preference information in group decision-making, Computers & Industrial Engi-
ðz + 1Þ ðzÞ
Since aij = ðaij Þθ ð ðzÞ
Þð1−θÞ , we have that neering 46 (2004) 329–335.
wj [21] J.S. Fiana, W.J. Hurley, The analytic hierarchy process: does adjusting a pairwise
comparison matrix to improve the consistency ratio help? Computers & Operations
Research 24 (1997) 749–755.
z + 1 2θ ðzÞ ðzÞ ðzÞ 2
GCIðA Þ≤ ðlog ðaij Þ− logðwi Þ + logðwj ÞÞ [22] J. Fodor, M. Roubens, Fuzzy Preference Modelling and Multicriteria Decision Support,
ðn−1Þðn−2Þ Kluwer, Dordrecht, 1994.
[23] E. Forman, K. Peniwati, Aggregating individual judgments and priorities with the
z z
= θGCIðA ÞbGCIðA Þ: analytic hierarchy process, European Journal of Operational Research 108 (1998)
165–169.
 Y. Dong et al. / Decision Support Systems 49 (2010) 281–289 289

[24] P.T. Harker, Derivatives of the Perron root of a positive reciprocal matrix: with [44] Z.S. Xu, An automatic approach to reaching consensus in multiple attribute group
application to the analytic hierarchy process, Applied Mathematics and Computation decision making, Computers & Industrial Engineering 56 (4) (2009) 1369–1374.
22 (1987) 217–232. [45] Z.S. Xu, C.P. Wei, A consistency improving method in analytic hierarchy process,
[25] M.W. Herman, W.W. Koczkodaj, A Monte Carlo study of pairwise comparison, European Journal of Operational Research 116 (1999) 443–449.
Information Processing Letters 57 (1996) 25–29. [46] J.-M. Yeh, C. Lin, B. Kreng, J.-Y. Gee, A modified procedure for synthesising ratio
[26] F. Herrera, E. Herrera-Viedma, F. Chiclana, Multiperson decision-making based on judgements in the analytic hierarchy process, Journal of the Operational Research
multiplicative preference relations, European Journal of Operational Research 129 Society 50 (1999) 867–873.
(2001) 372–385.
[27] F. Herrera, E. Herrera-Viedma, J.L. Verdegay, A model of consensus in group decision Yucheng Dong is an Assistant Professor at the School of Management, Xi'an Jiaotong
making under linguistic assessments, Fuzzy Sets and Systems 78 (1996) 73–87. University, Xi'an, China. He is a research fellow of the Department of Information
[28] F. Herrera, L. Martínez, A 2-tuple fuzzy linguistic representation model for computing Management, Oriental Institute of Technology, Taipei, Taiwan. He received his Ph.D.
with words, IEEE Transactions on Fuzzy Systems 8 (2000) 746–752. degree in management from Xi'an Jiaotong University in 2008. His current research
[29] E. Herrera-Viedma, S. Alonso, F. Chiclana, F. Herrera, A consensus model for group interests include group decision making, computing with words, and on-line algorithm.
decision making with incomplete fuzzy preference relations, IEEE Transactions on His research results have been published in the European Journal of Operational Research,
Fuzzy Systems 15 (2007) 863–877. IEEE Transactions on Fuzzy Systems, System Engineering and Electronics, and Fuzzy Sets
[30] E. Herrera-Viedma, F. Herrera, F. Chiclana, A consensus model for multiperson and Systems, among others.
decision making with different preference structures, IEEE Transactions on Systems,
Man and Cybernetics Part A: Systems and Humans 32 (2002) 394–402. Guiqing Zhang is a PhD candidate at the School of Management, Xi'an Jiaotong University,
[31] E. Herrera-Viedma, F. Herrera, F. Chiclana, M. Luque, Some issues on consistency of Xi'an, China. She received her B.S. degree from the Department of Information and
fuzzy preference relations, European Journal of Operational Research 154 (2004) Computation Science, Chongqing University in 2004, and her M.S. degree from the
98–109. Department of Applied Mathematics, Chongqing University in 2007. Her research interests
[32] Y.-S. Huang, J.-T. Liao, Z.-L. Lin, A study on aggregation of group decisions, Systems include group decision making and decision support systems.
Research and Behavioral Science 26 (2009) 445–454.
[33] L.C. Leung, D. Cao, On consistency and ranking of alternatives in fuzzy AHP, European Wei-Chiang Hong is an Assistant Professor at the Department of Information Manage-
Journal of Operational Research 124 (2000) 102–113. ment, Oriental Institute of Technology, Taipei, Taiwan. He received his Ph.D. degree in
[34] J.M. Moreno-Jiménez, J. Aguarón, M.T. Escobar, The core of consistency in AHP-group management from Da Yeh University in 2008. He is the editor-in-chief of the International
decision making, Group Decision and Negotiation 17 (2008) 249–265. Journal of Applied Evolutionary Computation. His current research interests include
[35] R. Ramanathan, L.S. Ganesh, Group preference aggregation methods employed in decision analysis, evolutionary computation, and decision support systems. His research
AHP: an evaluation and intrinsic process for deriving members’ weightages, results have been published in Applied Mathematical Modelling, Electric Power Systems
European Journal of Operational Research 79 (1994) 249–265. Research, International Journal of Electrical Power & Energy Systems, Energy Conversion
[36] F.D. Robert, E.H. Forman, Group decision support with the analytic hierarchy process, and Management, and Applied Mathematics and Computation, among others.
Decision Support Systems 8 (2) (1992) 99–124.
[37] T.L. Saaty, The Analytic Hierarchy Process, McGraw-Hill, New York, 1980. Yinfeng Xu is a Professor at the School of Management, Xi'an Jiaotong University, Xi'an,
[38] T.L. Saaty, Decision-making with the AHP: why is the principal eigenvector China. He received his Ph.D. degree in operational research from the Academy of
necessary, European Journal of Operational Research 145 (2003) 85–91. Mathematics and Systems Science, Chinese Academy of Sciences, in 1992. His current
[39] B. Srdjevic, Linking analytic hierarchy process and social choice methods to support research interests include combinatorial optimization, theoretical computer science and
group decision-making in water management, Decision Support Systems 42 (2007) decision analysis. His research results have been published in Theoretical Computer
2261–2273. Science, Journal of Global Optimization, Information Processing Letters, Journal of
[40] T. Tanino, Fuzzy preference orderings in group decision making, Fuzzy Sets and Combinatorial Optimization, Discrete & Computational Geometry, European Journal of
Systems 12 (1984) 117–131. Operational Research, and IEEE Transactions on Fuzzy Systems, among others.
[41] R.C. Van den Honert, Stochastic group preference modelling in the multiplicative
AHP: a model of group consensus, European Journal of Operational Research 110
(1998) 99–111.
[42] Y.M. Wang, K.-S. Chin, G.K.K. Poon, A data envelopment analysis method with
assurance region for weight generation in the analytic hierarchy process, Decision
Support Systems 45 (2008) 913–921.
[43] Z.S. Xu, On consistency of the weighted geometric mean complex judgement matrix
in AHP, European Journal of Operational Research 126 (2000) 683–687.

View publication stats