Soft Comput

DOI 10.1007/s00500-014-1426-2

METHODOLOGIES AND APPLICATION

Consistency issues of interval pairwise comparison matrices

Yucheng Dong · Xia Chen · Cong-Cong Li ·
Wei-Chiang Hong · Yinfeng Xu

© Springer-Verlag Berlin Heidelberg 2014

Abstract In this paper, we investigate the consistency Keywords Multiple attribute decision making · Interval
issues of interval pairwise comparison matrices in detail. pairwise comparison matrix · Consistency · Linear
Using logarithmic Manhattan distance to define the devi- programming
ation degree of a pairwise comparison matrix to consis-
tent pairwise comparison matrices, we propose a new con-
sistency index of pairwise comparison matrices. Based on 1 Introduction
this consistency index of pairwise comparison matrices, we
develop a consistency index of interval pairwise comparison The analytic hierarchy process (AHP) is proposed by Saaty
matrices. Several desired properties of the proposed consis- (1977, 1980), and has been used in almost all the applications
tency indexes are presented. Furthermore, linear program- related to multiple criteria decision making. In AHP, pairwise
ming (LP) models are developed to compute the consistency comparison matrices are used to express decision makers’
indexes. Then, we propose a LP-based consistency improv- opinions over alternatives. The consistency measure is an
ing model, which optimally preserves original pairwise com- classical issue in AHP, and is used to ensure that the decision
parison information in improving consistency. Meanwhile, maker is neither random nor illogical in his/her pairwise com-
considering the uncertainty plays an important role in the parisons. Numerous researchers focused on the AHP consis-
consistency index of interval pairwise comparison matrices, tency measure problem (Alonso and Lamata 2006; Brunelli
this consistency improving model is extended to simultane- et al. 2013; Koczkodaj 1993; Moreno-Jiménez et al. 2008;
ously manage the uncertain degree in interval pairwise com- Peláez and Lamata 2003; Srdjevic and Srdjevic 2013; Stein
parison matrices. Finally, we discuss the consistency-based and Mizzi 2007). Saaty (1980) developed a famous consis-
prioritization method, and propose the strong consistency tency index based on the eigenvalue prioritization method.
index of interval pairwise comparison matrices. Crawford and Williams (1985) and Aguarón and Moreno-
Jiménez (2003) presented the geometric consistency index
based on row geometric mean prioritization method, one of
the most extended AHP’s prioritization procedures. Mean-
while, numerous approaches (Bozóki et al. 2011; Cao et al.
Communicated by V. Loia. 2008; Dong et al. 2010, 2013a, b; Finan and Hurley 1997;
Saaty 2003; Xu et al. 1999) have been developed to aid the
Y. Dong (B) · X. Chen · C.-C. Li · Y. Xu AHP decision makers to revise the inconsistency in pairwise
Business School, Sichuan University, Chengdu 610065,
People’s Republic of China
comparison matrices. When pairwise comparison matrices
e-mail: [email protected] are of acceptable consistency, AHP decision makers use pri-
oritization methods (Choo and Wedley 2004; Dong et al.
W.-C. Hong 2008a, b, 2011; Saaty 2003; Srdjevic 2005) to obtain priori-
Department of Information Management, Oriental Institute
of Technology, No. 58, Section 2, Sichuan Road, Panchiao,
ties from pairwise comparison matrices.
Taipei 220, Taiwan However, due to the complexity and uncertainty involved
e-mail: [email protected] in real world decision problems, it is sometimes unrealistic

123
 Y. Dong et al.

or impossible to acquire exact judgments. Saaty and Vargas interval pairwise comparison matrices based on Saaty’s
(1987) have initiated the study of interval pairwise compar- consistency index, which is introduced in Sect. 2.3. How-
ison matrices. Specifically, an interval pairwise comparison ever, it is difficult to solve the nonlinear programming
matrix has the following structure: model presented in Wang et al. (2005b).
 = (
V vi j )n×n
⎛ In this paper, we will define new consistency indexes, and
− + − + ⎞
1  [v12 , v12 ] ... [v1n , v1n ] replace Saaty’s consistency index with the new indexes.
⎜ v− , v+ 1 ... −
[v2n + ⎟
, v2n ]⎟
=⎜
⎝...
21 21
⎠,
Then, we show that both Finan and Hurley’s model and Wang
 − + ... ... ... et al.’s consistency-based prioritization method can be mod-
− +
vn1 , vn1 [vn2 , vn2 ] ... 1 eled by LP models.
− + − + − +
The paper is organized as follows. Section 2 introduces the
where v i j = [vi j , vi j ], 0 < vi j ≤ vi j , and vi j × v ji = 1. research background and presents the problems need to be
Interval pairwise comparison matrices are also called interval further investigated. Section 3 provides a new consistency
judgement matrices in AHP. There are a number of studies index of pairwise comparison matrices, and then develops
on how to obtain priorities from interval pairwise compar- a consistency index of interval pairwise comparison matri-
ison matrices (Arbel 1989, 1991; Arbel and Vargas 1993; ces. Following this, Sect. 4 discusses how to deal with the
Chandran et al. 2005; Cox 2007; Haines 1998; Islam et al. inconsistency and uncertainty in interval pairwise compar-
1997; Kress 1991; Lan et al. 2009; Mikhailov 2002, 2004; ison matrices. Subsequently, Sect. 5 provides a LP version
Podinovski 2007; Salo and Hämäläinen 1992; Sugihara et of the consistency-based prioritization method, based on the
al. 2004; Wang et al. 2005a; Wang and Elhag 2007). Mean- proposed consistency index, and also presents the strong con-
while, similar to pairwise comparison matrices, the consis- sistency index of interval pairwise comparison matrices. Two
tency measure of interval pairwise comparison matrices is illustrative examples are provided in Sect. 6. Finally, con-
performed to ensure the derived priorities reliable. However, cluding remarks are included in Sect. 7.
there are fewer studies about consistency issues of interval
pairwise comparison matrices. Conde and Pérez (2010) pro-
posed an interesting consistency index to measure the incon-
2 Background
sistency degree of the interval pairwise comparison matrix.
Wang et al. (2005b) provided an interesting property to judge
In this section, we introduce two existing consistency
whether or not an interval pairwise comparison matrix is
indexes, Finan and Hurley’s consistency improving model,
consistent without solving any mathematical programming
and Wang et al.’s consistency-based prioritization method.
model. Wang et al. (2005b) further proposed an interesting
Then, we present the open problems to be studied in this
nonlinear programming model to generate interval priorities
paper.
from interval pairwise comparison matrices, based on Saaty’s
consistency index. Liu (2009) proposed a consistent index for
interval pairwise comparison matrices. Liu’s method is based 2.1 Two existing consistency indexes of pairwise
on the use of two pairwise comparison matrices obtained comparison matrices
from boundary information of the interval pairwise compar-
ison matrix. In AHP, pairwise comparison matrices and completely con-
In this paper, we will investigate the consistency issues of sistent pairwise comparison matrices are presented as Defi-
interval pairwise comparison matrices in detail. We mainly nitions 1 and 2 (Saaty 1980; Saaty and Vargas 1987).
focus on the use of LP models in dealing with the following
Definition 1 A = (ai j )n×n is called a pairwise comparison
open problems:
matrix if ai j > 0 and ai j × a ji = 1 for i, j = 1, 2, . . . , n.

(1) Finan and Hurley (1997) proposed an classical optimiza- Definition 2 The pairwise comparison matrix A = (ai j )n×n
tion model to improve consistency of pairwise compari- is considered completely consistent, if aik × ak j = ai j for
son matrices based on Saaty’s consistency index, which i, j, k = 1, 2, . . . , n.
is introduced in Sect. 2.2. However, it is an open prob-
lem to obtain optimum solution(s) to Finan and Hurley’s Saaty (1980) proposed the principal eigenvector of A as
model. Meanwhile, it is necessary to extend Finan and the desired priority vector w = (w1 , w2 , . . . , wn )T , which
Hurley’s model to improve the consistency of interval can be obtained by solving the linear system:
pairwise comparison matrices. n
(2) Wang et al. (2005b) proposed an interesting nonlinear ai j × w j = λ(A)
max × wi , (1)
programming model to generate interval priorities from j=1

123
Interval pairwise comparison matrices

(A)
where λmax is the principal eigenvalue of A, wi ≥ 0 and 2.2 Finan and Hurley’s consistency improving model
n
i=1 wi = 1. This prioritization method is called the eigen-
value prioritization method. Finan and Hurley (1997) proposed an interesting model to
In general, it is difficult to obtain completely consistent improve consistency of A, based on Saaty’s consistency
pairwise comparison matrices. Together with the eigenvalue index. The main idea of Finan and Hurley’s consistency
prioritization method, Saaty (1980) proposed a consistency improving model is to find out a n × n pairwise compari-
index to measure the inconsistency degree of pairwise com- son matrices A with a better consistency degree, and can be
parison matrices (see Definition 3). described as follows:

min|A − A||
Definition 3 Let A = (ai j )n×n be a pairwise comparison A , (4)
matrix. Saaty’s consistency index is defined as follows: s.t. CI(Saaty) (A) ≤ β

(A) where β < CI(Saaty) (A) and ||.|| is a distance metric between
(Saaty) λmax − n
CI (A) = , (2) two matrices. Finan and Hurley’s model optimally preserves
n−1 original pairwise comparison information in improving con-
(A) sistency (in a distance sense). However, it is still an open
where λmax is the principal eigenvalue of A.
problem to solve Finan and Hurley’s model.
In consistency improving approaches of pairwise compar-
Crawford and Williams (1985) proposed the geometric
ison matrices, the distance between original pairwise com-
consistency index (see Definition 4) for measuring the incon-
parison matrices and adjusted pairwise comparison matrices
sistency degree of pairwise comparison matrices.
is also used as the criteria of modification effectiveness (Cao
et al. 2008; Xu et al. 1999).
Definition 4 Let Q n = {w = (w1 , w2 , . . . , wn )T |wi >
n
0, i=1 wi = 1}. The geometric consistency index (GCI)
2.3 Consistency and priorities in interval pairwise
is defined as follows:
comparison matrices
n n
1
GCI(A) = 2 min (log(ai j ) − log(wi ) + log(w j ))2 . Interval pairwise comparison matrices and completely con-
n w∈Q n
i=1 j=1 sistent interval pairwise comparison matrices are presented
(3) as Definitions 5 and 6 (Arbel 1989, 1991; Haines 1998; Wang
et al. 2005a, b; Wang 2006; Wang and Elhag 2007).
Crawford and Williams (1985) have shown that the solu-  = (
Definition 5 V vi j )n×n is called an interval pairwise
tion to model (3) is unique and can be found simply as the − + − +
comparison matrix if v i j = [vi j , vi j ], 0 < vi j ≤ vi j , and
geometric means of the rows of matrix A:
vi−j × v +
ji = 1 for i, j = 1, 2, . . . , n.

n n
j=1 ai j Definition 6 The interval pairwise comparison matrix V 
wi =  .
n n is considered completely consistent, if ∃{w1 , w2 , . . . , wn },
i=1 ( j=1 ai j )
n
wi
where wi > 0 and i=1 n
wi = 1, satisfies vi−j ≤ w j
≤
This prioritization method is called the row geometric mean vi+j , for i, j = 1, 2, . . . , n.
prioritization method.
Wang et al. (2005b) proposed an interesting nonlinear pro-
Note 1. The original definition of the geometric con-
n n gramming model to generate interval priorities from interval
sistency index is GCI(A) = (n−1)(n−2)1
min i=1 j=1
w∈Q n pairwise comparison matrices, based on Saaty’s consistency
(log(ai j ) − log(wi ) + log(w j ))2 , but this definition is not index. Before introducing Wang et al.’s model, we define the
suitable for 2 × 2 pairwise comparison matrices. Therefore, concept of pairwise comparison matrices of V  (see Defini-
this paper makes a minor change on the original geometric tion 7).
consistency index, which does not change the essence of the
Definition 7 Let V  = (vi j )n×n be an interval pairwise com-
geometric consistency index.
Saaty (1980) proposed the consistency ratio (CR) rule (i.e., parison matrix. A = (ai j )n×n is a pairwise comparison
 if v − ≤ ai j ≤ v + and ai j × a ji = 1.
matrix of V
CR ≤ 0.1) to obtain the thresholds, denoted by CI(Saaty) , of ij ij
CISaaty . Aguarón and Moreno-Jiménez (2003) discussed the We denote N V as the set of the pairwise comparison matri-
thresholds, denoted by GCI, of GCI. When CI(Saaty) (A) (or ces of V.
GCI(A)) is smaller than the established threshold CI(Saaty) The main idea of Wang et al.’s consistency-based prior-
(or GCI), A is of acceptable consistency. itization method is to find out interval priorities from the

123
 Y. Dong et al.

acceptably consistent pairwise comparison matrices of V . 3.1 Consistency index of pairwise comparison matrices
Wang et al.’s model can be described as follows.
⎧ Before proposing a new consistency index of interval pair-
⎪ min/max wi
⎪
⎪ wise comparison matrices, we discuss the consistency index
⎪
⎪
⎪
⎨ s.t. A = (ai j )n×n ∈ N V of pairwise comparison matrices.
(A) ,(5)
⎪
⎪ CI(Saaty) (A) = λmax −n
≤ (Saaty)
⎪
⎪ n−1 CI Definition 8 Let A = (ai j )n×n be a pairwise comparison
⎪
⎪
⎩ n (A) matrix and let Mn be the set of n × n consistent pairwise
j=1 ai j × w j = λmax × wi , i = 1, 2, . . . , n
comparison matrices. Then, we define the deviation degree
where λ(A)
max is the principal eigenvalue of A. The purpose between A and Mn as follows:
(A)
λmax −n
of imposing constraint condition, CI(Saaty) (A) = n−1 ≤ d(A, Mn ) = min d(A, P), (6)
P∈Mn
CI(Saaty) , on model (5) is to derive the priorities from the
acceptably consistent pairwise comparison matrices of V . where d(A, P) is the distance between two pairwise com-
The optimal objective values of the above pair of model (5) parison matrices A and P.
consist of the possible interval of wi , which we denote by
[wiL , wiR ]. Repeating the above solution process for each pri- We set d(A, Mn ) as the consistency index (CI) of the pair-
ority wi (i = 1, . . . , n), all the priority intervals that meet the wise comparison matrix A, namely,
requirement of the acceptable consistency can be obtained. CI(A) = d(A, Mn ). (7)

2.4 Presentation of the problems When the distance between two pairwise comparison matri-
ces uses the logarithmic Euclidean distance metric, it can be
The above subsections introduce Finan and Hurley’s consis- shown that
tency improving model and Wang et al.’s consistency-based 1
n n
prioritization method. These studies are all based on Saaty’s CI(A) = min (log(ai j ) − log( pi j ))2 . (8)
n 2 P∈Mn
consistency index. However, there are several open problems i=1 j=1
that need to be further investigated: When the distance between two pairwise comparison matri-
ces uses the logarithmic Manhattan distance, it can be shown
(1) It is still an open problem to obtain optimum solution(s) that
to Finan and Hurley’s model. n n
(2) Finan and Hurley’s model focuses on pairwise compar- 1
CI(A) = min |log(ai j ) − log( pi j )|. (9)
ison matrices. It is necessary to extend this model to n 2 P∈Mn
i=1 j=1
improve the consistency of interval pairwise comparison
matrices. The above consistency index has a definite physical impli-
(3) Wang et al.’s consistency-based prioritization method is cation and reflects the deviation degree (in the distance
a nonlinear programming model. However, it is difficult sense) between the pairwise comparison matrix A and consis-
to solve this nonlinear programming model. tent pairwise comparison matrices. The smaller the value of
CI(A), the more consistent the pairwise comparison matrix
A. If CI(A) = 0, then A is a consistent pairwise comparison
In the rest sections of this paper, we will focus on these
matrix.
open problems. The main work is to define new consistency
indexes, and replace Saaty’s consistency index with the new Proposition 1 When the distance between two pairwise
indexes. Then, we show that both Finan and Hurley’s model comparison matrices uses the logarithmic Euclidean dis-
and Wang et al.’s consistency-based prioritization method tance metric, we have CI(A) = GCI(A).
can be modeled by LP models. As is known, LP models
can be solved in very little computational time using readily The proof of Proposition 1 is provided in Appendix A.
available softwares, such as Lindo and Matlab. Proposition 1 shows that the proposed consistency index
is equivalent to the famous geometric consistency index pre-
3 New consistency indexes sented in Crawford and Williams (1985) and Aguarón and
Moreno-Jiménez (2003), when the distance between two
In this section, we propose new consistency indexes for the pairwise comparison matrices uses the logarithmic Euclid-
pairwise comparison matrices and interval pairwise compar- ean distance metric. However, using the logarithmic Man-
ison matrices, and discuss the desired properties of the new hattan distance obtains a new consistency index (i.e., model
consistency indexes. (9)). This new consistency index will reduce computational

123
Interval pairwise comparison matrices

complexity in analyzing consistency issues of interval pair- 2007) consider V  consistent, if ∃{w1 , w2 , . . . , wn } ∈ Q n
wi
wise comparison matrices (see Sects. 3.3, 4.2, 5.1). Thus, this satisfies vi j ≤ w j ≤ vi+j , f or i, j = 1, 2, . . . , n. Property 2
−

paper adopts this new consistency index (i.e., model (9)). shows that the proposed consistency index of interval pair-
Similar to Proposition 1, we easily prove Proposition 2. wise comparison matrix has a good linkage to this classical
definition, and guarantees that V  is a consistent interval pair-
Proposition 2 When the distance between two pairwise ) = 0.
wise comparison matrix iff ICI(V
comparison matrices uses the logarithmic Manhattan dis-
tance metric, Property 3 Let F  = ( f i j )n×n , where  f i j = [ f i−j , f i+j ], be
1
n n another interval pairwise comparison matrix, and let 0 ≤
CI(A) = min |log(ai j ) − log(wi ) + log(w j )|. α ≤ 1. Let C  = (
ci j )n×n be the weighted geometric mean
n 2 w∈Q n
i=1 j=1 collective matrix of V  and F,  where c − + −
i j = [ci j , ci j ], ci j =
(10) − α − (1−α)
(vi j ) ( f i j ) + + α + (1−α)
and ci j = (vi j ) ( f i j ) . Then, C  is an
interval pairwise comparison matrix, and
3.2 Consistency index of interval pairwise comparison
matrices  ≤ max{ICI(V
ICI(C) ), ICI( F)}.


This subsection proposes a new consistency index of interval The proof of Property 3 is provided in Appendix A.
pairwise comparison matrices. Property 3 shows that the inconsistency of the collec-
tive interval pairwise comparison matrix is smaller than the
 = (
Definition 9 Let V vi j )n×n be an interval pairwise com- largest individual inconsistency. The corresponding consis-
parison matrix. We set min A∈N  CI(A) as the consistency tency property of the collective pairwise comparison matrix
V
.
index of V is presented in Xu (2000) and Escobar et al. (2004).

 as
In this paper, we symbolize the consistency index of V Property 4 Let F f i j )n×n , where 
 = ( f i j = [ f i−j , f i+j ] be
), i.e.,
ICI(V another interval pairwise comparison matrix. If f i−j ≤ vi−j
) = min CI(A).
ICI(V (11) and f i+j ≥ vi+j , then
A∈N 
V
 ≤ ICI(V
ICI( F) ).
This consistency index of the interval pairwise compari-
son matrix V  reflects the minimum deviation degree (in the The proof of Property 4 is provided in Appendix A.
distance sense) between the pairwise comparison matrices of Property 4 shows that it is easier to obtain consistency,
 and the consistent pairwise comparison matrices. It also
V with the uncertain degree of interval pairwise comparison
reflects the “best” consistency degree of the pairwise com- matrices increasing. A formal definition of the uncertain
parison matrices of V . The smaller the value of ICI(V
), the degree of interval pairwise comparison matrices will be pro-
more consistent the interval pairwise comparison matrix V . vided in Sect. 4.2.
The proposed consistency index of interval pairwise com-
parison matrices has several desired properties. 3.3 Computing consistency indexes
) = CI(V
Property 1 ICI(V ) if V
 degenerates to a pairwise
In this subsection, we show how to obtain the values of the
comparison matrix.
proposed two consistency indexes, respectively, using LP
Proof According to (9), (11) and Definition 7, Property 1 is models to solve model (9) and model (11).
obtained. 

Proposition 3 Using two transformed decision variables:
Property 1 shows that the consistency index of interval ci j = log(ai j ) − log( pi j ) and di j ≥ |ci j |. Model (9) can be
pairwise comparison matrices can degenerate to one of the transformed into a linear programming model P1 :
pairwise comparison matrices. n n
1
Property 2 ICI(V ) = 0 if and only if ∃{w1 , w2 , . . . , wn } ∈ min
n2
di j (12)
i=1 j=1
Q n satisfies
wi s.t. log( pik ) + log( pk j ) = log( pi j ), i, j, k = 1, 2, . . . , n
vi−j ≤ ≤ vi+j , for i, j = 1, 2, . . . , n. (13)
wj
log( pi j ) + log( p ji ) = 0, i, j = 1, 2, . . . , n (14)
The proof of Property 2 is provided in Appendix A.
In general, the researchers (Arbel 1989, 1991; Haines ci j = log(ai j ) − log( pi j ), i, j = 1, 2, . . . , n (15)
1998; Wang et al. 2005a, b; Wang 2006; Wang and Elhag ci j ≤ di j , i, j = 1, 2, . . . , n (16)

123
 Y. Dong et al.

−ci j ≤ di j , i, j = 1, 2, . . . , n (17) To preserve the information in V  as much as possible,

we hope that the distance between V  and V ∗ is minimal.
where, log( pi j ), ci j , di j (i, j = 1, 2, . . . , n) are decision
When the distance between two interval pairwise comparison
variables.
matrices uses the logarithmic Manhattan distance, we have
n n
The proof of Proposition 3 is provided in Appendix A. 1
min (|log(vi+j ) − log(vi∗+ − ∗−
j )| + |log(vi j ) − log(vi j )|).
∗
V 2n 2
i=1 j=1
Proposition 4 Using two transformed decision variables:
ci j = log(ai j ) − log( pi j ) and di j ≥ |ci j |. Model (11) can be (27)
transformed into a linear programming model P2 :
n n
∗ has the established consistency index, i.e.,
Meanwhile, V
1
min di j (18)
n2 ∗ ) ≤ ICI.
ICI(V (28)
i=1 j=1

s.t. log(ai j ) ≤ log(vi+j ), i, j = 1, 2, . . . , n (19)

As a result, an optimization model to improve the consis-
−log(ai j ) ≤ −log(vi−j ), i, j = 1, 2, . . . , n (20)  can be constructed as follows:
tency level in V
log(ai j ) + log(a ji ) = 0, i, j = 1, 2, . . . , n (21) ⎧ n n + ∗+ − ∗−
j=1 (|log(vi j )−log(vi j )|+|log(vi j )−log(vi j )|)
1
⎪
⎨ min
log( pik ) + log( pk j ) = log( pi j ), i, j, k = 1, 2, . . . , n ∗ 2n
V
2 i=1
∗ ∈ K n
s.t. V , (29)
⎪
⎩
(22) ICI(V∗ ) ≤ ICI
log( pi j ) + log( p ji ) = 0, i, j = 1, 2, . . . , n (23)
ci j = log(ai j ) − log( pi j ), i, j = 1, 2, . . . , n (24) where K n is the set of n × n interval pairwise comparison
matrices.
ci j ≤ di j , i, j = 1, 2, . . . , n (25)
−ci j ≤ di j , i, j = 1, 2, . . . , n (26)
4.2 Extension of the proposed consistency improving model
where, log(ai j ), log( pi j ), ci j , di j (i, j = 1, 2, . . . , n) are
decision variables. Property 4 shows that it is easier to obtain consistency, with
The proof of Proposition 4 is provided in Appendix A. the uncertain degree of interval pairwise comparison matri-
Propositions 3 and 4 guarantee that we can compute the ces increasing. Therefore, this subsection further expands the
CI value and the ICI value, using LP models. proposed consistency improving model (i.e., model (29)) to
From Property 1, Propositions 3 and 4, Corollary 1 is simultaneously manage the inconsistency and uncertainty in
obtained. interval pairwise comparison matrices. To do so, we first pro-
pose an uncertain index (UI) of interval pairwise comparison
Corollary 1 If V degenerates to a pairwise comparison matrices.
matrix, then P2 degenerates to P1 .
 = (
Definition 10 Let V vi j )n×n be an interval pairwise
comparison matrix. We define the uncertain index of V as
4 Consistency improving model in interval pairwise  n n + −
UI(V ) = n 2 i=1 j=1 (log(vi j ) − log(vi j )).
1
comparison matrices
The above uncertain index reflects the average uncertain
Based on the idea from Finan and Hurley’s model, in this
degree of pairwise comparison data in an interval pairwise
section we develop a consistency improving model in interval ), the less
comparison matrix. The smaller the value of UI(V
pairwise comparison matrices. . When UI(V ) = 0, V is a pairwise
the uncertain degree of V
comparison matrix. Clearly, 0 ≤ UI(V ) ≤ 2log(9), when
4.1 The proposed consistency improving model
using 1–9 scale.
 = ( − + − + If the decision maker hopes that the uncertain index of the
Let V vi j )n×n , where v
i j = [vi j , vi j ], 0 < vi j ≤ vi j , ∗ is smaller
adjusted interval pairwise comparison matrix V
and vi−j × v +
ji = 1, be an interval pairwise comparison matrix than L, then
provided by a decision maker. To improve the consistency
level in V, the decision maker wants to adjust the elements ∗ ) ≤ L .
UI(V (30)
. In fact, the main work of improving the consistency
of V
 is to find a suitable interval pairwise comparison
level in V Based on (27), (28) and (30), an optimization model to
∗ = (v ∗ 
∗ ∗− ∗+
matrix, V i j )n×n where vi j = [vi j , vi j ], with the simultaneously manage the inconsistency and uncertainty in
established consistency level ICI (< ICI(V )).  can be constructed as follows:
V

123
Interval pairwise comparison matrices

⎧ n n + ∗+ − ∗− and V∗ is a consistent pairwise comparison matrix; (2) If

⎪
⎪ min 2n12 i=1 j=1 (|log(vi j ) − log(vi j )| + |log(vi j ) − log(vi j )|)
⎪
⎨ V∗  degenerates to a pairwise comparison matrix, L = 0 and
V
∗ ∈ K n
s.t. V .
(31) ∗ is a pairwise comparison matrix with the
⎪
⎪ UI( ∗ ) ≤ L
V ICI = CI, then V
⎪
⎩
ICI(V∗ ) ≤ ICI established consistency index CI.
Clearly, if we set L ≥ 2log(9), model (31) reduces to The proof of Proposition 6 is provided in Appendix A.
model (29). So, in the rest of this section, we only focus on Note 2. According to the actual situation, the decision
the method to solve model (31). Before solving the model makers can set different L values and ICI. In general, the
(31), we propose a LP model P3 (i.e., model (32)–(51)): decision makers hope to decrease the inconsistency degree
1
n n and uncertain degree in V . In this situation, we set L <
min (h i j + m i j ) (32)  
UI(V ) and ICI < ICI(V ). Establishing the thresholds of
2n 2
i=1 j=1 the AHP consistency indexes is an open question (Apos-
s.t. log(vi∗− ∗+
j ) ≤ log(vi j ), i, j = 1, 2, . . . , n (33) tolou and Hassell 2002; Chu anad Liu 2002), and there is
disagreement about how to establish these thresholds (e.g.,
log(vi∗− ∗+
j ) + log(v ji ) = 0, i, j = 1, 2, . . . , n (34)
Golden and Wang 1989; Lane and Verdini 1989). In Appen-
log(ai j ) ≤ log(vi∗+
j ), i, j = 1, 2, . . . , n (35) dix B, we establish the ICI thresholds based on Golden and
− log(ai j ) ≤ − log(vi∗− j ), i, j = 1, 2, . . . , n (36) Wang’s approach, which decision makers use as a reference
in improving consistency level.
log(ai j ) + log(a ji ) = 0, i, j = 1, 2, . . . , n (37)
Note 3. Model (31) optimally preserves original pairwise
log( pik ) + log( pk j ) = log( pi j ) i, j, k = 1, 2, . . . , n comparison information in reaching acceptable consistency
(38) (in logarithmic Manhattan distance sense) for interval pair-
log( pi j ) + log( p ji ) = 0, i, j = 1, 2, . . . , n (39) wise comparison matrices, and can be solved by the LP model
n n P3 . When V  degenerates to a pairwise comparison matrix
1
di j ≤ ICI, i, j = 1, 2, . . . , n (40) and L = 0, we denote model (31) as P3∗ . Clearly, P3∗ has its
n2 origin in Finan and Hurley’s consistency improving model.
i=1 j=1
ci j = log(ai j ) − log( pi j ), i, j = 1, 2, . . . , n (41) Comparing with Finan and Hurley’s model, P3∗ is based on
the new consistency index and can be solved by a LP model.
ci j ≤ di j , i, j = 1, 2, . . . , n (42)

−ci j ≤ di j , i, j = 1, 2, . . . , n (43) 5 Further discussion

n n
1
f i j ≤ L , i, j = 1, 2, . . . , n (44) In this section, we propose a LP version of Wang et
n2
i=1 j=1
al.’s consistency-based prioritization method, and propose
f i j = log(vi∗+ ∗−
j ) − log(vi j ) i, j = 1, 2, . . . , n (45) a strong consistency index for interval pairwise comparison
gi j = log(vi+j ) − log(vi∗+
j ) i, j = 1, 2, . . . , n (46) matrices. Moreover, we also show the existence of the opti-
mal solution to the LP models proposed in this paper.
gi j ≤ h i j , i, j = 1, 2, . . . , n (47)
−gi j ≤ h i j , i, j = 1, 2, . . . , n (48)
5.1 A LP version of Wang et al.’s consistency-based
li j = log(vi−j ) − log(vi∗−
j ) i, j = 1, 2, . . . , n (49) prioritization method
li j ≤ m i j , i, j = 1, 2, . . . , n (50)
−li j ≤ m i j , i, j = 1, 2, . . . , n (51) In Sect. 2.3, we introduce Wang et al.’s consistency-based pri-
oritization method. Inspired by Wang et al. (2005b), this sub-
Lemma 1 Let V ∗ be an interval pairwise comparison section provides a LP version of Wang et al.’s consistency-
matrix. Then, ICI(V ∗ ) ≤ ICI if and only if ∃ A = (ai j )n×n ∈ based prioritization method, based on the proposed consis-
N V∗ and P = ( pi j )n×n ∈ Mn satisfy constraints (40)–(43). tency index. This new version can be described as follows.
The proof of Lemma 1 is provided in Appendix A. min/max wi (52)
Proposition 5 Model (31) can be equivalently described as s.t. log(ai j ) ≤ log(vi+j ), i, j = 1, 2, . . . , n (53)
P3 . − log(ai j ) ≤ − log(vi−j ), i, j = 1, 2, . . . , n (54)
The proof of Proposition 5 is provided in Appendix A. log(ai j ) + log(a ji ) = 0, i, j = 1, 2, . . . , n (55)
n n
 degenerates to a pairwise compari-
Proposition 6 (1) If V 1
| log(ai j ) − log(wi ) + log(w j )| ≤ CI (56)
son matrix, L = 0 and ICI = 0, then P3 degenerates to P1 n2
i=1 j=1

123
 Y. Dong et al.

n
5.3 A strong consistency index
wi ≥ 0, wi = 1, (57)
i=1
As stated in previous sections, the proposed consistency
where, wi (i = 1, 2, . . . , n), and ai j (i, j = 1, 2, . . . , n) are index of the interval pairwise comparison matrix V  reflects
decision variables. In model (52)–(57), constraints (53)–(55) the “best” consistency degree of pairwise comparison matri-
guarantee A ∈ N V . According to Proposition 2, constraint , and has a good linkage to the classical consistency
ces of V
(56) guarantees CI(A) ≤ CI. Thus, model (52)–(57) derives definition of interval pairwise comparison matrices. This sub-
the priorities from the acceptably consistent pairwise com-  (see
section investigates the “worst” consistency degree of V
parison matrices of V. Definition 11).
Proposition 7 Using two transformed decision variables:
Definition 11 Let V = ( vi j )n×n be an interval pairwise
ci j = log(ai j ) − log(wi ) + log(w j ) and di j ≥ |ci j |. Model
(52)–(57) can be transformed into a linear optimization comparison matrix. We set max A∈N  CI(A) as the strong
V
model P4 : . We symbolize the strong consistency
consistency index of V
 as SICI(V
index of V ), that is
min / max wi (58)
s.t. log(ai j ) ≤ log(vi+j ), i, j = 1, 2, . . . , n (59) ) = max CI(A).
SICI(V (67)
A∈N 
V
− log(ai j ) ≤ − log(vi−j ), i, j = 1, 2, . . . , n (60)
The smaller the value of SICI(V ), the more the strong
log(ai j ) + log(a ji ) = 0, i, j = 1, 2, . . . , n (61)
n n consistency of V. If SICI(V
) = 0, we consider V strong con-
1 ) are presented
di j ≤ CI (62) sistent. Several desired properties of SICI(V
n2 below.
i=1 j=1
ci j = log(ai j ) − log(wi ) + log(w j ), i, j = 1, 2, . . . , n ) = CI(V
) if V
 degenerates to a pairwise
Property 5 SICI(V
(63) comparison matrix.
ci j ≤ di j , i, j = 1, 2, . . . , n (64)
Property 6 Let F f i j )n×n , where 
 = ( f i j = [ f i−j , f i+j ], be
−ci j ≤ di j , i, j = 1, 2, . . . , n (65)
n another interval pairwise comparison matrix. If f i−j ≤ vi−j
wi ≥ 0, wi = 1, (66) and f i+j ≥ vi+j , then
i=1
 ≥ SICI(V
SICI( F) ).
Proof Similar to Proposition 5, we can prove Proposition 7.

 ) ≤ CI(A) ≤ SICI(V
Property 7 For any A ∈ N V , ICI(V ).
The optimal objective values of the above pair of P4 con-
sist of the possible interval of wi , which are denoted by Property 5 shows that the strong consistency index of inter-
[wiL , wiU ]. Repeating the above solution process for each val pairwise comparison matrices can degenerate to one of the
priority wi (i = 1, . . . , n), all the priority intervals can be pairwise comparison matrices. Different with the consistency
obtained. index of interval pairwise comparison matrices, Property 6
Note 4. P4 has its origin in the consistency-based prior- shows that it is more difficult to obtain strong consistency,
itization method presented in Wang et al. (2005b). Wang et with the uncertain degree increasing. Property 7 shows that
SICI(V ) provides a upper bound of consistency degree of
al.’s method is based on Saaty’s consistency ratio and formu-
pairwise comparison matrices of V , and ICI(V) provides a
lated as a nonlinear programming model. The proposed new
version is based on our consistency index and can be solved lower bound of consistency degree of pairwise comparison
matrices of V .
by a LP model.
In particular, similar to Proposition 4, using two trans-
5.2 The existence of the optimal solution to the LP models formed decision variables: ci j = log(ai j ) − log( pi j ) and
di j ≥ |ci j |. Model (67) can be transformed into a maxmin
In the above sections, consistency issues of interval pair- linear optimization model P5 :
wise comparison matrices are modeled as LP models (i.e., n n
1
P1 , P2 , P3 , P4 ). Here, we show the existence of the optimal max min
ai j pi j n2
di j (68)
solutions to these LP models (see Proposition 8). i=1 j=1

Proposition 8 The optimal solutions to P1 , P2 , P3 and P4 s.t. log(ai j ) ≤ log(vi+j ), i, j = 1, 2, . . . , n (69)

exist. − log(ai j ) ≤ − log(vi−j ), i, j = 1, 2, . . . , n (70)
The proof of Proposition 8 is provided in Appendix A. log(ai j ) + log(a ji ) = 0, i, j = 1, 2, . . . , n (71)

123
Interval pairwise comparison matrices

log( pik ) + log( pk j ) = log( pi j ), i, j = 1, 2, . . . , n (72) 

V (∗)
⎛ ⎞
log( pi j ) + log( p ji ) = 0, i, j = 1, 2, . . . , n (73) 1 [0.9189, 1.6613] [1.0924, 1.79] [2.312, 3.3032]
⎜ [0.602, 1.0883] 1 [2.5743, 4.109] [4.1534, 4.8272]⎟
⎜
=⎝ ⎟.
ci j = log(ai j ) − log( pi j ), i, j = 1, 2, . . . , n (74) [0.5586, 0.9154] [0.2434, 0.3885] 1 [5.1732, 7.2625]⎠
[0.3027, 0.4325] [0.2072, 0.2408] [0.1377, 0.1933] 1
ci j ≤ di j , i, j = 1, 2, . . . , n (75)
−ci j ≤ di j , i, j = 1, 2, . . . , n (76) If the decision maker sets L = 0.2, then Table 1 shows
ICI = ICI(0.2) = 0.1887. Applying P3 obtains an adjusted
Note 5. Using ICI(V ) and SICI(V), we may totally eval- interval pairwise comparison matrix with the established con-
.
uate the range of the consistency degree of V sistency level:

V (∗∗)
⎛ ⎞
1 [0.9421, 1.338] [1.1143, 1.6284] [2.5033, 3.1759]
6 Illustrative examples ⎜[0.7474, 1.0615] 1 [2.7304, 3.6446] [4.2135, 4.7713]⎟
=⎜
⎝[0.6141, 0.8974]
⎟.
[0.2744, 0.3662] 1 [5.587, 6.9526] ⎠
[0.3149, 0.3995] [0.2096, 0.2373] [0.1438, 0.179] 1
To show how these theoretical results work in practice, let us
consider the following examples. If the decision maker sets L = 0.1, then Table 1 shows
ICI = ICI(0.1) = 0.2115. Applying P3 obtains an adjusted
6.1 Example 1 interval pairwise comparison matrix with the established con-
sistency level:
Consider the interval pairwise comparison matrix presented  (∗∗∗)
V
in Islam et al. (1997) and Wang et al. (2005a): ⎛
1 [0.9716, 1.1324] [1.2647, 1.5243] [2.7069, 3.0623]
⎞

⎛ ⎞ ⎜[0.8831, 1.0292]
=⎜
1 [2.9023, 3.3265] [4.3028, 4.6914]⎟⎟.
1 [1, 2] [1, 2] [2, 3] ⎝ [0.656, 0.7907] [0.3006, 0.3445] 1 [5.897, 6.6086] ⎠
⎜ [4, 5]⎟
 = ⎜ [1/2, 1] [3, 5]
1 [0.3266, 0.3694] [0.2132, 0.2324] [0.1513, 0.1696] 1
V ⎟.
⎝ [1/2, 1] [1/5, 1/3] 1 [6, 8]⎠
If the decision maker sets L = 0, then Table 1 shows
[1/3, 1/2] [1/5, 1/4] [1/8, 1/6] 1
ICI(0) = CI = 0.2326. Applying P3 obtains an adjusted
We use P2 to obtain ICI(V) = 0.224. First, using model pairwise comparison matrix with the established consistency
(29) improves the consistency level. Without loss of gener- level:
⎛ ⎞
ality, if the decision maker sets ICI = 0.15 and ICI = 0.1, 1 1.0055 1.3943 2.9811
respectively, then the corresponding adjusted interval pair- ⎜0.9945 1 3.0158 4.5621⎟
V =⎜ ⎝0.7172 0.3316
⎟.
wise comparison matrices are 1 6.0353⎠
⎛ ⎞ 0.3354 0.2192 0.1657 1
1 [0.9087, 2] [1, 2] [2, 3.3954]
⎜[0.5, 1.1005] 1 [2.5805, 5] [4, 5] ⎟
∗ = ⎜
V ⎟.
⎝ [0.5, 1] [0.2, 0.3875] 1 [4.8075, 8]⎠ 6.2 Example 2
[0.2945, 0.5] [0.2, 0.25] [0.125, 0.208] 1
Consider the following interval pairwise comparison matrix,
and which has been investigated by Arbel and Vargas (1993),
⎛ ⎞ Haines (1998), Wang et al. (2005a, b) and Liu (2009), respec-
1 [0.8849, 2] [1, 2] [2, 3.5593]
⎜[0.5, 1.1301] 1 [2.2486, 5] [4, 5] ⎟
tively:
∗∗
V =⎝⎜ ⎟.
[0.5, 1] [0.2, 0.4447] 1 [3.9812, 8]⎠ ⎛ ⎞
[0.281, 0.5] [0.2, 0.25] [0.125, 0.2512] 1
1 [2, 5] [2, 4] [1, 3]
⎜[1/5, 1/2] 1 [1, 3] [1, 2] ⎟
V = ⎜ ⎝[1/4, 1/2] [1/3, 1]
⎟.
Next, based on the extension of model (29) (i.e., model 1 [1/2, 1]⎠
(31) or P3 ) and the established ICI thresholds (see Table 1 in [1/3, 1] [1/2, 1] [1, 2] 1
Appendix B), we simultaneously manage the inconsistency
and uncertainty in V . Since n = 4 and UI(V ) = 0.3517, We use P2 to obtain ICI(V ) = 0, so V is a consis-
)) ≤ 0.1637, and thereby V  is tent interval pairwise comparison matrix. Based on P5 , we
Table 1 shows that ICI(UI(V
use MATLAB software to find A ∈ N V and SICI(V ) =
of unacceptable consistency.
CI(A) = 0.3385. The matrix A is listed below:
Without loss of generality, if the decision maker sets
⎛ ⎞
the uncertain index L = 0.3, then Table 1 shows ICI = 1 5 2 1
ICI(0.3) = 0.1637. Applying P3 obtains an adjusted interval ⎜1/5 1 3 2⎟
pairwise comparison matrix with the established consistency A=⎜ ⎝1/2 1/3 1 1⎠ .
⎟
level: 1 1/2 1 1

123
 Y. Dong et al.

Therefore, V is not a strong consistent interval pairwise com- (3) We propose a LP-based model to decrease the incon-
parison matrix. For any X ∈ N V , CI(X ) ∈ [0, 0.3385]. sistency and uncertainty in interval pairwise compari-
Let B = (bi j )4×4 and C = (ci j )4×4 , where, son matrices. This proposal optimally preserves orig-
⎧ ⎧ inal pairwise comparison information in reaching the
⎪ + ⎪ −
⎨ vi j i < j ⎨ vi j i < j established consistency level (in logarithmic Manhat-
bi j = 1 i = j , ci j = 1i = j . tan distance sense). This proposed model has its origin
⎪
⎩v − i > j ⎪
⎩v + i > j
ij ij
in the consistency improving model presented in Finan
and Hurley (1997). Comparing with Finan and Hurley’s
The pairwise comparison matrices B and C are obtained from model, our model is based on the new consistency index,
boundary information of the interval pairwise comparison and can be used in interval pairwise comparison matrices.
matrix V , and listed below. Particularly, our model can be solved by a LP.
⎛ ⎞ ⎛ ⎞ (4) We propose a consistency-based prioritization method,
1 5 4 3 1 2 2 1
⎜1/5 based on the new consistency index. The proposed
1 3 2⎟ ⎜1/2 1 1 1 ⎟
B=⎜ ⎟ ⎜
⎝1/4 1/3 1 1⎠ , C = ⎝1/2 1 1 1/2⎠ .
⎟ method is a LP version of consistency-based prioritiza-
tion method.
1/3 1/2 1 1 1 1 2 1
(5) We propose a strong consistency index of interval pair-
Liu (2009) uses Saaty’s consistency index of B and C to wise comparison matrices. The consistency index pro-
evaluate the consistency degree of the interval pairwise com- vides a lower bound of the consistency degree of inter-
parison matrix V . When applying the consistency index pro- val pairwise comparison matrices, and the strong consis-
posed in this paper, we can find CI(B) = 0.2012, CI(C) = tency index provides a upper bound of the consistency
0.0866, and CI(B), CI(C) ∈ [0, 0.3385]. degree of interval pairwise comparison matrices. There-
fore, we can totally evaluate the consistency degree of
interval pairwise comparison matrices, using both the
consistency index and the strong consistency index.
7 Conclusions

In this paper, we propose the consistency index of interval

pairwise comparison matrices, develop the model to deal with In this paper, most of consistency issues of interval pair-
the inconsistency and uncertainty in interval pairwise com- wise comparison matrices are modeled as LP models. It is
parison matrices, and discuss a LP version of consistency- well known that the average time complexity is O(n) when
based prioritization method using the proposed consistency using the classical simplex methods, presented in Dantzig
index. The major contributions in this paper are as follows: (1963) to solve LP models. In general, LP models can be
solved in very little computational time using readily avail-
able softwares, such as Lindo and Matlab. Therefore, the
(1) By defining the distance of a pairwise comparison matrix results in this paper are also of simplicity and convenience
to the set of consistent pairwise comparison matrices, for the application of consistent interval pairwise comparison
we develop the consistency index of pairwise compari- matrices in decision-making problems. Meanwhile, we argue
son matrices. This consistency index is equivalent to the there are three interesting research avenues for the future:
famous geometric consistency index presented in Craw-
ford and Williams (1985) and Aguarón and Moreno-
Jiménez (2003), when the distance between two pair- (1) Pairwise comparison matrices are also called multiplica-
wise comparison matrices uses the logarithmic Euclidean tive preference relations. There are three kinds of pref-
distance metric. When using the logarithmic Manhattan erence relations: linguistic preference relations (Alonso
distance, we obtain a new consistency index used in this et al. 2009; Cabrerizo et al. 2013; Dong et al. 2013a, b;
paper. Xu and Wang 2014; Zhang et al. 2014), fuzzy preference
(2) Based on this new consistency index of pairwise compar- relations (Alonso et al. 2008; Barrenechea et al. 2014;
ison matrices, we propose the consistency index of inter- Chiclana et al. 1998, 2009; Herrera-Viedma et al. 2004;
val pairwise comparison matrices. The proposed consis- Wu and Chiclana 2014; Tapia García et al. 2012; Xu et
tency index of interval pairwise comparison matrices has al. 2013, in press), and multiplicative preference relations
a good linkage to the classical consistency definition, and (Saaty 1977; Dong et al. 2014, in press). The transforma-
can degenerate to the one of pairwise comparison matri- tion functions among different kinds of preference rela-
ces. Several desired properties of the proposed consis- tions have been presented (Chiclana et al. 2001; Dong et
tency index are presented, and a LP model is developed al. 2009), so the proposed consistency methods may be
to compute this consistency index. exported to other kinds of preference relations.

123
Interval pairwise comparison matrices

(2) In some situations, it is difficult for the decision makers Without loss of generality, let P = ( pi j )n×n ∈ Mn , satisfy-
to give all elements in pairwise comparison matrices (or ing
preference relations), which leads to the uses of incom- n n
1
plete pairwise comparison matrices (or incomplete pref- CI(A) = (log(ai j ) − log( pi j ))2 . (79)
erence relations). Some researchers focused on calculat- n2
i=1 j=1
ing the missing values of incomplete pairwise compari-
Let w = (w1 , w2 , . . . , wn )T ∈ Q n be the priorities, obtained
son matrices (Cabrerizo et al. 2010b; Fedrizzi and Giove
from P using the row geometric mean prioritization method.
2007; Herrera-Viedma et al. 2007). It will be interest-
Since P ∈ Mn , according to the row geometric mean priori-
ing in any future research to discuss the management of wi
tization method, we have pi j = w . Consequently,
incomplete interval pairwise comparison matrices, based j

on the consistency indexes presented in this paper. n n

1
(3) In group decision making based on pairwise comparison CI(A) = (log(ai j ) − log(wi ) + log(w j ))2
n2
matrices (or preference relations), two related but differ- i=1 j=1
ent measures are widely studied: consistency and consen- 1
n n
sus. Consistency is performed to ensure that the decision ≥ min (log(ai j ) − log(wi ) + log(w j ))2
n 2 w∈Q n
maker is being neither random nor illogical in his/her i=1 j=1
pairwise comparisons, and consensus means the prefer- = GCI(A). (80)
ences among a group of decision makers are similar. It
Based on (78) and (80), we have CI(A) = GCI(A) when the
will be interesting in any future research to discuss the
distance between two pairwise comparison matrices uses the
uses of our proposal in different consensus models and
logarithmic Euclidean distance metric.
frameworks (Alonso et al. 2013; Cabrerizo et al. 2010a;
Chiclana et al. 2013; Herrera-Viedma 2014; Pérez et al. 

2010, 2013, 2014).
The proof of Property 2 If {w1∗ , w2∗ , . . . , wn∗ } ∈ Q n satis-
wi∗
Acknowledgments Yucheng Dong would like to acknowledge the fies vi−j ≤ w∗j ≤ vi+j , for i, j = 1, 2, . . . , n. Let W ∗ =
financial support of a Grant (No. 71171160) from NSF of China, and a w∗
Grant (No. skqx201308) from Sichuan University. Yinfeng Xu would (wi∗j )n×n = ( wi∗ )n×n . Then, we have that W ∗ ∈ Mn and
j
like to acknowledge the financial support of Grants (Nos. 71071123 and
W ∗ ∈ N V . According to Definition 9, we obtain ICI(V) = 0.
61221063) from NSF of China and acknowledge the financial support
of program for Changjiang Scholars and Innovative Research Team in
 ∗
Similarly, if ICI(V ) = CI(W ) = 0, we can prove that
wi
University (IRT1173) from Ministry of Education of China. ∃{w1 , w2 , . . . , wn } ∈ Q n satisfy vi−j ≤ w j
≤ vi+j , for
i, j = 1, 2, . . . , n. This completes the proof of Property 2.


Appendix A: Proofs
The proof of Property 3 According to Definition 5, we eas-
 is an interval pairwise comparison matrix. Let
ily find that C
The proof of Proposition 1.
A = (ai j )n×n ∈ N V , P ∗ = ( pi∗j )n×n ∈ Mn , and
∗ ∗

Proof Without loss of generality, let w ∗ = (w1∗ , w2∗ , . . . , wn∗ )T 1

n n
∈ Q n , and ) =
ICI(V | log(ai∗j ) − log( pi∗j )|. (81)
n2
i=1 j=1
n n
1 Let A∗∗ = (ai∗∗ ∗∗ = ( p ∗∗ )
GCI(A) = (log(ai j ) − log(wi∗ ) + log(w ∗j ))2 . j )n×n ∈ N F
, P i j n×n ∈ Mn , and
n2
i=1 j=1 n n
(77)  = 1
ICI( F) | log(ai∗∗ ∗∗
j ) − log( pi j )|. (82)
n2
i=1 j=1
wi∗
Let W ∗ = (wi∗j )n×n = ( ), we easily find that W ∗ ∈
w∗j Based on A∗ and A∗∗ , we construct a new pairwise compari-
Mn . From (3), we obtain son matrix Ac = (aicj )n×n , where aicj = (ai∗j )α × (ai∗∗
j )
(1−α) .
∗
We easily prove that A ∈ NC. Based on P and P , we
c ∗∗
n n
1
GCI(A) = (log(ai j ) − log(wi∗j ))2 construct a new pairwise comparison matrix P c = ( picj )n×n ,
n2
i=1 j=1 where picj = ( pi∗j )α × ( pi∗∗j )(1−α) . We easily prove that
1
n n P c ∈ Mn . According to Definition 9, we have
≥ 2 min (log(ai j ) − log( pi j ))2 n n
n P∈Mn 1
i=1 j=1  ≤
ICI(C) | log(aicj ) − log( picj )|
= CI(A). (78) n2
i=1 j=1

123
 Y. Dong et al.

guarantee that h i j = |gi j | = | log(vi+j ) − log(vi∗+

n n
1 j )|. Sim-
= |α(log(ai∗j ) − log( pi∗j ))
n2 ilarly, constraints (49)–(51) guarantee that m i j = |li j | =
| log(vi−j ) − log(vi∗−
i=1 j=1
j )|. As a result,
+(1 − α)(log(ai∗∗ ∗∗
j ) − log( pi j ))| (83) 1
n n
1
n n
(h i j + m i j ) = (| log(vi+j ) − log(vi∗+
j )|
2n 2 2n 2
From (81)–(83), we have that i=1 j=1 i=1 j=1

 ≤ αICI(V) + (1 − α)ICI( F)
 +| log(vi−j ) − log(vi∗−
j )|). (85)
ICI(C)
 
≤ max{ICI(V ), ICI( F)}, (84)
Thus, model (31) can be equivalently described as P3 . This
which completes the proof of Property 3. 
 completes the proof of Proposition 5. 


The proof of Property 4 For any A ∈ N V , we find A ∈ N F The proof of Proposition 6 From constraints (44)–(45), V ∗
under the condition that f i−j ≤ vi−j and f i+j ≥ vi+j . Thus, degenerates to a pairwise comparison matrix when L = 0.
N V ⊆ N F. According to Definition 9, we have ICI( F)  ≤ According to constraints (35)–(37), we have V ∗ = A =
∗− ∗+
ICI(V). This completes the proof of Property 4. 
 (ai j )n×n , i.e., vi j = vi j = ai j . When ICI = 0, con-
straints (40)–(43) guarantee that V ∗ is a consistent pair-
The proof of Proposition 3 In P1 , constraints (13)–(14) guar- wise comparison matrix, that is V ∗ = P = ( pi j )n×n .
antee that P = ( pi j )n×n ∈ Mn . Constraints (15)–(17) Further, according to constraints (46)–(48) and the objec-
enforce that di j ≥ |ci j | = | log(ai j ) − log( pi j )|. Accord- tive function (i.e., Eq. (32)), we have h i j = |gi j | =
ing to the objective function (i.e., Eq. (12)), any feasible | log(vi+j ) − log( pi j )|. According to constraints (49)–(51)
solutions with di j > |ci j | are not the optimal solution to and the objective function (i.e., Eq. (32)), we also have
P1 . Thus, constraints (15)–(17) guarantee that di j = |ci j | = m i j = |li j | = | log(vi−j ) − log( pi j )|. Since V  is a pair-
n n
| log(ai j ) − log( pi j )| and CI(A) = min n12 i=1 j=1 di j .
− +
wise comparison matrix, we have vi j = vi j , and h i j = m i j .
As a result, model (9) can be equivalently described as P1 . ) = min 1 2 i=1 n n
As a result, CI(V j=1 (h i j + m i j ). In
This completes the proof of Proposition 3. 
 2n
this way, P3 degenerates to P1 . This completes the proof of
Proposition 6. 

The proof of Proposition 4 In P2 , constraints (19)–(21) guar-
antee that A = (ai j )n×n ∈ N V , and constraints (22)–(23)
guarantee that P = ( pi j )n×n ∈ Mn . Constraints (24)–(26) The proof of Proposition 8 We only prove the existence of
enforce that di j ≥ |ci j | = | log(ai j ) − log( pi j )|. According the optimal solution to P1 . The existence of the optimal solu-
to the objective function (i.e., Eq. (18)), any feasible solu- tions to P2 , P3 and P4 can be similarly proved. Any n × n
tions with di j > |ci j | are not the optimal solution to P2 . consistent pairwise comparison matrices satisfy all the con-
Thus, constraints (24)–(26) guarantee that di j = |ci j | = straints of P1 , and therefore represent feasible solutions. So,
) = min 12 i=1
| log(ai j ) − log( pi j )| and ICI(V n n
j=1 di j .
P1 has a non-empty feasible region. A closed bounded fea-
n
As a result, model (11) can be equivalently described as P2 . sible region for P1 would satisfy the assumption of Weier-
This completes the proof of Proposition 4. 
 strass theorem. This could be achieved, by introducing an
upper bound for di j . Let w = (w1 , w2 , . . . , wn )T ∈ Q n
The proof of Lemma 1 Constraints (41)–(43) enforce that be the priorities, obtained from A using the row geometric
wi
di j ≥ |ci j | = | log(ai j ) − log( pi j )|. According to Defi- mean prioritization method, and P = ( pi j )n×n = ( w j
)n×n .
nition 9, ICI(V ∗ ) ≤ 12 i=1 n n
j=1 di j . Thus, constraint
According to the row geometric
√ mean prioritization method,
n n n
 ∗
(40) guarantees ICI(V ) ≤ ICI. This completes the proof we have that pi j = √ k=1 aik
, and P ∈ Mn satisfies all
n n
k=1 a jk
of Lemma 1. 
 the constraints of P1 . Since ci j = log(ai j ) − log( pi j ), a suit-
able inequality that does not affect the optimal solution could
The proof of Proposition 5 In P3 , constraints (33)–(34) guar-
be
antee that V ∗ is an interval pairwise comparison matrix. Con-
straints (35)–(37) guarantee that A = (ai j )n×n ∈ N V∗ , and di j ≤ | log(ai j ) − log( pi j )|
constraints (38)–(39) guarantee that P = ( pi j )n×n ∈ Mn .  
 n 
From Lemma 1, constraints (40)–(43) guarantee ICI(V ∗ ) ≤  1 
= log(ai j ) − log(aik ) − log(a jk ) . (86)
ICI. Moreover, constraints (44)–(45) guarantee UI(V ∗ ) ≤  n 
k=1
L, and constraints (46)–(48) enforce that h i j ≥ |gi j | =
| log(vi+j ) − log(vi∗+j )|. According to the objective function
According to Weierstrass theorem, we prove the existence
(i.e., Eq. (32)), any feasible solutions with h i j > |gi j | are of the optimal solution to P1 , which completes the proof of
not the optimal solution to P3 . Thus, constraints (46)–(48) Proposition 8. 


123
Interval pairwise comparison matrices

Appendix B: Establishing the ICI thresholds based on Table 1 The values of ICI(L) for different n when α = 0.05
Golden and Wang’s approach L n=3 n=4 n=5 n=6 n=7 n=8 n=9

Property 4 shows that it is easier to obtain consistency, with 0 0.0343 0.2326 0.4343 0.5151 0.6194 0.6897 0.7186
the uncertain degree of interval pairwise comparison matri- 0.1 0.0075 0.2115 0.3999 0.4658 0.5800 0.6683 0.7014
ces increasing. In the following, we establish the ICI thresh- 0.2 0.0000 0.1887 0.3498 0.4245 0.5431 0.6233 0.6537
olds of interval pairwise comparison matrices, according to 0.3 0.0000 0.1637 0.3175 0.4100 0.5399 0.5629 0.6045
different uncertain indexes. When the uncertain index is L, 0.4 0.0000 0.1380 0.2969 0.3817 0.4516 0.5506 0.5981
the ICI threshold is denoted as ICI(L). As mentioned earlier, 0.5 0.0000 0.1182 0.2640 0.3565 0.4253 0.5317 0.5384
establishing the thresholds of the AHP consistency indexes is 0.6 0.0000 0.0932 0.2306 0.3025 0.3820 0.4590 0.5135
an open question (Apostolou and Hassell 2002; Chu anad Liu 0.7 0.0000 0.0682 0.1903 0.2876 0.3637 0.3949 0.4210
2002), and there is disagreement about how to establish these 0.8 0.0000 0.0466 0.1778 0.2486 0.3427 0.3771 0.4430
thresholds (e.g., Golden and Wang 1989; Lane and Verdini 0.9 0.0000 0.0248 0.1521 0.2002 0.3072 0.2975 0.3773
1989), so the proposed ICI(L) should only be considered as 1.0 0.0000 0.0000 0.1121 0.1801 0.2657 0.3041 0.3281
a decision aid which decision makers use as a reference to 1.1 0.0000 0.0000 0.0876 0.1228 0.2539 0.2462 0.3058
obtain the acceptable consistency. 1.2 0.0000 0.0000 0.0636 0.1013 0.1649 0.2266 0.2741
The main idea of establishing ICI(L) is inspired by the 1.3 0 0.0000 0.0368 0.0747 0.1528 0.1798 0.2207
work of Golden and Wang (1989). When using Saaty’s con- 1.4 0 0.0000 0.0167 0.0740 0.1162 0.1756 0.1768
sistency ratio rule (i.e., ≤ 0.1), Golden and Wang (1989) 1.5 0 0.0000 0.0000 0.0635 0.1100 0.1271 0.1400
have found that the ratio of the acceptably consistent ones
among random pairwise comparison matrices is over 20 %
for n = 3, and about 4 % for n > 3. It is impossible to obtain
a matrix consistency for n > 5. The detailed analysis also
can be found in Peláez and Lamata (2003). Then, let vik− k+
j = min{t1 , t2 } and vi j = max{t1 , t2 }.
Here, the established thresholds ICI(L) guarantee the ratio
Step 3: Compute ICI( V (k) ), k = 1, 2, . . . , m, using
of the acceptably consistent ones among random interval
P2 .
pairwise comparison matrices with the uncertain index L is
constant and small, for different matrix sizes. Step 4: By sorting {ICI(V (1) ), . . . , ICI(V (m) )} in ascend-

The procedure of establishing ICI thresholds is described ing order, we obtain {b , . . . b , . . . , b }, where b(k) is
(1) (k) (m)

as follows. the kth smallest element in {ICI( V (1) ), . . . , ICI(V (m) )}.
∗
Step 5: Let k = r ound(m × α), where round is usual
The procedure of establishing the ICI thresholds ∗
round operation. Output the threshold ICI(L) = b(k ) .
Input: The matrix size n, the number of random matrices Note 6. In this procedure, Step 2 ensures log(vik+ j ) −
m, the established uncertain index L ∈ [0, 2 log(9)], and the
parameters α log(v ) = L, which is a stronger condition than UI(
k−
ij V (k) ) =
Output: The ICI thresholds, ICI(L) L (k = 1, 2, . . . , m). Step 4 ensures that the ratio of the
Step 1: For each n, we, respectively, generate m pair- acceptably consistent interval pairwise comparison matrices
wise comparison matrices {A(1) , . . . A(k) , . . . , A(m) }, whose among { 
V (1) , . . . , V (m) } is approximately α. We argue that

entries are uniformly randomly generated using 1–9 scale. this ratio is small, so we suggest setting α = 0.05 according
(k)
Step 2: Based on A(k) = (ai j )n×n , we generate an inter- to the principle of small-probability events.
When setting α = 0.05, m = 5,000, different matrix size
val pairwise comparison matrix  V (k) = (v k) , where
i j n×n n and different L, we run this procedure to obtain the values
v k− k+
i j = [vi j , vi j ]. We consider three cases:
k
of ICI(L), which are shown in Table 1.
(k) (k)
Case A: log(9)−log(ai j ) ≥ L and log(ai j )−log(1/9) ≥ We conclude that the interval pairwise comparison matrix
 ) ≤ ICI(UI(V)).
L. In this case, let t1 = ai(k)
j , and t2 is uniformly randomly
V is of acceptable consistency, when ICI(V
(k) (k) 
Otherwise, we conclude that V is of unacceptable con-
selected from {e L+log(ai j ) , elog(ai j )−L }, where e is the base sistency. For example, if the matrix size of V  is 4 and
of natural logarithms.  
U I (V ) = 0.2, then V is of acceptable consistency when
(k) (k)
Case B: log(9)−log(ai j ) ≥ L and log(ai j )−log(1/9) < ICI(V) ≤ ICI(0.2) = 0.1887. Meanwhile, for explanatory
(k)
L. In this case, let t1 = ai(k)
j , and t2 = e
L+log(ai j )
. convenience, ICI(0) is denoted by CI in this paper.
(k) (k) When using the ICI thresholds, the decision maker first
Case C: log(9)−log(ai j ) < L and log(ai j )−log(1/9) ≥
(k)
sets a L value, and then the ICI threshold is determined by
L. In this case, let t1 = ai(k)
j , and t2 = e
log(ai j )−L
. ICL(L).

123
 Y. Dong et al.

References Choo EU, Wedley WC (2004) A common framework for deriving pref-
erence values from pairwise comparison matrices. Comput Oper Res
Aguarón J, Moreno-Jiménez JM (2003) The geometric consistency 31:893–908
index: approximated thresholds. Eur J Oper Res 147:137–145 Conde E, Pérez M (2010) A linear optimization problem to derive rel-
Alonso JA, Lamata MT (2006) Consistency in the analytic hierarchy ative weights using an interval judgement matrix. Eur J Oper Res
process: a new approach. Int J Uncertain Fuzziness Knowl Based 201:537–544
Syst 14(4):445–459 Cox MAA (2007) Examining alternatives in the interval analytic hierar-
Alonso S, Chiclana F, Herrera F, Herrera-Viedma E, Alcalá-Fdez J, chy process using complete enumeration. Eur J Oper Res 180:957–
Porcel C (2008) A consistency-based procedure to estimate missing 962
pairwise preference values. Int J Intell Syst 23:155–175 Crawford G, Williams C (1985) A note on the analysis of subjective
Alonso S, Cabrerizo FJ, Chiclana F, Herrera F, Herrera-Viedma E (2009) judgement matrices. J Math Psychol 29:387–405
Group decision-making with incomplete fuzzy linguistic preference Chu P, Liu JKH (2002) Note on consistency ratio. Math Comput Model
relations. Int J Intell Syst 24:201–222 35:1077–1080
Alonso S, Pérez IJ, Cabrerizo FJ, Herrera-Viedma E (2013) A linguis- Dantzig GB (1963) Linear programming and extensions. Princeton Uni-
tic consensus model for Web 2.0 communities. Appl Soft Comput versity Press, New York
13:149–157 Dong YC, Fan Z, Yu S (2014) Consensus building in a local context
Apostolou B, Hassell JM (2002) Note on consistency ratio: a reply. for the AHP-GDM with the individual numerical scale and prioriti-
Math Comput Model 35:1081–1083 zation method. IEEE Trans Fuzzy Syst. doi:10.1109/TFUZZ.2014.
Arbel A (1989) Approximate articulation of preference and priority 2312974
derivation. Eur J Oper Res 43:317–326 Dong YC, Hong WC, Xu YF (2013) Measuring consistency of linguis-
Arbel A (1991) A linear programming approach for processing approx- tic preference relations: a 2-tuple linguistic approach. Soft Comput
imate articulation of preference. In: Korhonen P, Lewandowski A, 17(11):2117–2130
Wallenius J (eds) Multiple criteria decision support. Lecture Notes Dong YC, Hong WC, Xu YF, Yu S (2011) Selecting the individual
in Economics and Mathematical Systems, vol 356. Springer, Berlin, numerical scale and prioritization method in the analytic hierarchy
pp 79–86 process: a 2-tuple fuzzy linguistic approach. IEEE Trans Fuzzy Syst
Arbel A, Vargas LG (1993) Preference simulation and preference pro- 19:13–26
gramming: robustness issues in priority deviation. Eur J Oper Res Dong YC, Hong WC, Xu YF, Yu S (2013) Numerical scales generated
69:200–209 individually for analytic hierarchy process. Eur J Oper Res 229:654–
Barrenechea E, Fernandez J, Pagola M, Chiclana F, Bustince H (2014) 662
Construction of interval-valued fuzzy preference relations from igno- Dong YC, Xu YF, Li HY (2008) On consistency measures of linguistic
rance functions and fuzzy preference relations. Knowl Based Syst preference relations. Eur J Oper Res 189:430–444
58:33–44 Dong YC, Xu YF, Li HY, Dai M (2008) A comparative study of the
Bozóki S, Fülöp J, Koczkodaj WW (2011) An LP-based inconsistency numerical scales and the prioritization methods in AHP. Eur J Oper
monitoring of pairwise comparison matrices. Math Comput Model Res 186:229–242
54:789–793 Dong YC, Xu YF, Yu S (2009) Linguistic multiperson decision making
Brunelli M, Canal L, Fedrizzi M (2013) Inconsistency indices for based on the use of multiple preference relations. Fuzzy Sets Syst
pairwise comparison matrices: a numerical study. Ann Oper Res 160:603–623
211(1):493–509 Dong YC, Zhang GQ, Hong WC, Xu YF (2010) Consensus models for
Cabrerizo FJ, Herrera-Viedma E, Pedrycz W (2013) A method based AHP group decision making under row geometric mean prioritiza-
on PSO and granular computing of linguistic information to solve tion method. Decis Support Syst 49:281–289
group decision making problems defined in heterogeneous contexts. Escobar MT, Aguarón J, Moreno-Jiménez JM (2004) A note on AHP
Eur J Oper Res 230:624–633 group consistency for the row geometric mean priorization proce-
Cabrerizo FJ, Moreno JM, Pérez IJ, Herrera-Viedma E (2010a) Analyz- dure. Eur J Oper Res 153:318–322
ing consensus approaches in fuzzy group decision making: advan- Fedrizzi M, Giove S (2007) Incomplete pairwise comparison and con-
tages and drawbacks. Soft Comput 14:451–463 sistency optimization. Eur J Oper Res 183:303–313
Cabrerizo FJ, Pérez IJ, Herrera-Viedma E (2010b) Managing the con- Finan JS, Hurley WJ (1997) The analytic hierarchy process: does adjust-
sensus in group decision making in an unbalanced fuzzy linguistic ing a pairwise comparison matrix to improve the consistency ratio
context with incomplete information. Knowl Based Syst 23:169–181 help. Comput Oper Res 24:749–755
Chiclana F, Herrera F, Herrera-Viedma E (1998) Integrating three rep- Golden BL, Wang Q (1989) An alternate measure of consistency. In:
resentation models in fuzzy multipurpose decision making based on Golden BL, Wasil EA, Harker PT (eds) The analytic hierarchy
fuzzy preference relations. Fuzzy Sets Syst 97:33–48 process: applications and studies. Springer, Berlin
Chiclana F, Herrera F, Herrera-Viedma E (2001) Integrating multiplica- Haines LM (1998) A statistical approach to the analytic hierarchy
tive preference relations in a multipurpose decision-making model process with interval judgments. (I) distributions on feasible regions.
based on fuzzy preference relations. Fuzzy Sets Syst 122:277–291 Eur J Oper Res 110:112–125
Chiclana F, Herrera-Viedma E, Alonso S, Herrera F (2009) Cardinal Herrera-Viedma E, Cabrerizo FJ, Kacprzyk J, Pedrycz W (2014) A
consistency of reciprocal preference relations: a characterization of review of soft consensus models in a fuzzy environment. Inf Fusion
multiplicative transitivity. IEEE Trans Fuzzy Syst 17:14–23 17:4–13
Chiclana F, Tapia Garcia JM (2013) A statistical comparative study of Herrera-Viedma E, Chiclana F, Herrera F, Alonso S (2007) Group
different similarity measures of consensus in group decision making. decision-making model with incomplete fuzzy preference relations
Inf Sci 221:110–123 based on additive consistency. IEEE Trans Syst Man Cybern Part B
Cao D, Leung LC, Law JS (2008) Modifying inconsistent comparison Cybern 31(2):227–234
matrix in analytic hierarchy process: a heuristic approach. Decis Herrera-Viedma E, Herrera F, Chiclana F, Luque M (2004) Some issues
Support Syst 44:944–953 on consistency of fuzzy preference relations. Eur J Oper Res 154:98–
Chandran B, Golden B, Wasil E (2005) Linear programming models for 109
estimating weights in the analytic hierarchy process. Comput Oper Islam R, Biswal MP, Alam SS (1997) Preference programming and
Res 32:2235–2254 inconsistent interval judgments. Eur J Oper Res 97:53–62

123
Interval pairwise comparison matrices

Koczkodaj WW (1993) A new definition of consistency of pairwise Salo A, Hämäläinen RP (1992) Processing interval judgments in the
comparisons. Math Comput Model 18:79–84 analytic hierarchy process. In: Goicoechea A, Duckstein L, Zoints
Kress M (1991) Approximate articulation of preference and priority S (eds) Multiple criteria decision making. Proceedings of the ninth
derivation: a comment. Eur J Oper Res 52:382–383 international conference held in Fairfax. Springer, New York, pp
Lan JB, Lin J, Cao LJ (2009) An information mining method for deriving 359–372
weights from an interval comparison matrix. Math Comput Model Sugihara K, Ishii H, Tanaka H (2004) Interval priorities in AHP by
50:393–400 interval regression analysis. Eur J Oper Res 158:745–754
Lane EF, Verdini WA (1989) A consistency test for AHP decision mak- Stein WE, Mizzi PJ (2007) The harmonic consistency index for the
ers. Decis Sci 20:575–590 analytic hierarchy process. Eur J Oper Res 177(1):488–497
Liu F (2009) Acceptable consistency analysis of interval reciprocal Tapia García JM, Del Moral MJ, Martínez MA, Herrera-Viedma E
comparison matrices. Fuzzy Sets Syst 160:2686–2700 (2012) A consensus model for group decision-making problems
Mikhailov L (2002) Fuzzy analytical approach to partnership selection with interval fuzzy preference relations. Int J Inf Technol Decis Mak
in formation of virtual enterprises. Omega 30:393–401 11(4):709–725
Mikhailov L (2004) A fuzzy approach to deriving priorities from interval Wang YM, Yang JB, Xu DL (2005a) A two-stage logarithmic goal pro-
pairwise comparison judgments. Eur J Oper Res 159:687–704 gramming method for generating weights from interval comparison
Moreno-Jiménez JM, Aguarón J, Escobar MT (2008) The core of matrices. Fuzzy Sets Syst 152:475–498
consistency in AHP-group decision making. Group Decis Negotiat Wang YM, Yang JB, Xu DL (2005b) Interval weight generation
17:249–265 approaches based on consistency test and interval comparison matri-
Peláez JI, Lamata MT (2003) A new measure of consistency for positive ces. Appl Math Comput 167:252–273
reciprocal matrices. Comput Math Appl 46:1839–1845 Wang YM (2006) On lexicographic goal programming method for gen-
Pérez IJ, Cabrerizo FJ, Alonso S, Herrera-Viedma E (2014) A new con- erating weights from inconsistent interval comparison matrices. Appl
sensus model for group decision making problems with non homo- Math Comput 173:985–991
geneous experts. IEEE Trans Syst Man Cybern Syst 44:494–498 Wang YM, Elhag TMS (2007) A goal programming method for obtain-
Pérez IJ, Cabrerizo FJ, Herrera-Viedma E (2010) A mobile decision ing interval weights from an interval comparison matrix. Eur J Oper
support system for dynamic group decision making problems. IEEE Res 177:458–471
Trans Syst Man Cybern Part A Syst Hum 40:1244–1256 Wu J, Chiclana F (2014) A social network analysis trust-consensus
Pérez IJ, Wikström R, Mezei J, Carlsson C, Herrera-Viedma E (2013) A based approach to group decision-making problems with interval-
new consensus model for group decision making using fuzzy ontol- valued fuzzy reciprocal preference relations. Knowl Based Syst
ogy. Soft Comput 17:1617–1627 59:97–107
Podinovski VV (2007) Interval articulation of superiority and precise Xu Y, Li KW, Wang H (2013) Consistency test and weight generation
elicitation of priorities. Eur J Oper Res 180:406–417 for additive interval fuzzy preference relations. Soft Comput. doi:10.
Saaty TL (1977) A scaling method for priorities in hierarchical struc- 1007/s00500-013-1156-x
tures. J Math Psychol 15:234–281 Xu Y, Wang H (2014) A comment on “Incomplete fuzzy linguistic
Saaty TL (1980) The analytic hierarchy process. McGraw-Hill, New preference relations under uncertain environments”. Inf Fusion 20:2–
York 5
Saaty TL (2003) Decision-making with the AHP: why is the principal Xu ZS (2000) On consistency of the weighted geometric mean complex
eigenvector necessary. Eur J Oper Res 145:85–91 judgement matrix in AHP. Eur J Oper Res 126:683–687
Saaty TL, Vargas LG (1987) Uncertainty and rank order in the analytic Xu ZS, Wei CP (1999) A consistency improving method in analytic
hierarchy process. Eur J Oper Res 32:107–117 hierarchy process. Eur J Oper Res 116:443–449
Srdjevic B (2005) Combining different prioritization methods in the Zhang GQ, Dong YC, Xu YF (2014) Consistency and consensus mea-
analytic hierarchy process synthesis. Comput Oper Res 32:1897– sures for linguistic preference relations based on distribution assess-
1919 ments. Inf Fusion 17:46–55
Srdjevic B, Srdjevic Z (2013) Synthesis of individual best local priority
vectors in AHP-group decision making. Appl Soft Comput 13:2045–
2056

123