Information Fusion 43 (2018) 13–26

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

An approach to quality function deployment based on probabilistic T

linguistic term sets and ORESTE method for multi-expert multi-criteria
decision making
Wu Xinglia, Liao Huchanga,b,*
a
Business School, Sichuan University, Chengdu 610064, China
b
Department of Computer Science and Artificial Intelligence, University of Granada, E-18071 Granada, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: The quality function deployment (QFD) is an effective tool to translate the customer requirements (CRs) to the
Multi-expert multi-criteria decision making design requirements (DRs) of a product. The process of selecting the optimal innovative product design to
Quality function deployment maximize customer satisfaction is full of uncertainty and fuzziness regarding to the users’ preferences, the re-
Probabilistic linguistic term set lationships between CRs and DRs and the merits of product designs. This study proposes a multi-expert multi-
Innovative product design selection
criteria decision making method to solve the innovative product design selection problem by developing an
ORESTE
enhanced QFD method combined with the complicated fuzzy linguistic representation model, the probabilistic
Shared cars
linguistic term set (PLTS), and the ranking method, ORESTE. Firstly, we propose a probability aggregation
method to integrate the individuals’ subjective evaluations into group ones expressed as PLTSs. On this basis, we
extend the QFD into the probabilistic linguistic context to get the DRs’ fuzzy weights. Then, based on a new
distance measure between PLTSs, a probabilistic linguistic global preference score function and three kinds of
probabilistic linguistic preference intensity formulas are proposed. Furthermore, we develop a PL-ORESTE
method to obtain the preference, indifference and incomparability relations between the alternatives. For the
facility of application, we develop the procedure of the QFD-based PL-ORESTE method. Given that the “shared
cars” is a new industry appeared in Chinese market, we finally illustrate the applicability of the proposed method
by a case study concerning the selection of innovative designs of Panda shared cars.

1. Introduction
Economic globalization has promoted technical innovation con-
stantly and has shortened the lifecycle of products. Customer require-
ments (CRs) are diversifying and changing fast. If the enterprises fail to
meet the CRs, the internal conflicts will be arisen and the development
cycle time will be increased simultaneously, while the market pene-
tration will be decreased with poor product quality, lower customer
satisfaction and revenues [3]. Therefore, to make survival in the fierce
competition, it becomes an important and inevitable task for the en-
terprises to analyze their customers’ needs and then choose an in-
novative design which maximizes the customers’ satisfaction. General,
however, it is impossible to make evaluation on new product design
schemes directly according to the CRs but the design requirements
(DRs) [4]. Practically, the innovative product design selection is a
special multi-expert multi-criteria decision making (MEMCDM) pro-
blem involving different groups of individuals, such as users, designers,
researchers and investors. There are three problems to be solved: ① how
to translate the CRs exactly into the DRs that act as the criteria for
evaluating the design schemes; ② how to express the individuals’ eva-
luations comprehensively and aggregate the group opinions correctly; ③
how to select the optimal design effectively based on the assessments of
the alternatives.
The quality function deployment (QFD) is a useful technique that is
extensively employed when the innovative products need to capture the
voice of customers to maximize the customers’ satisfaction [2]. In QFD,
the importance of the CRs, the strength relationships between the CRs
and the DRs, and the merits of the design schemes regarding to the DRs,
are difficult to be expressed by crisp data because the available in-
formation is usually subjective and imprecise [26]. Temponi et al. [30]
developed a fuzzy logic-based extension of QFD as a tool to capture the
imprecise requirements of customers. It is testified that the fuzzy QFD is
a good method to deal with imprecise information [13,17,35]. How-
ever, there are some serious defects existed in the fuzzy QFD methods:
(1) They are limited in representing a full picture of each individual's

*
Corresponding author at: Business School, Sichuan University, Chengdu 610064, China.
E-mail address: liaohuchang@scu.edu.cn (H. Liao).

https://doi.org/10.1016/j.inffus.2017.11.008
Received 30 September 2017; Received in revised form 7 November 2017; Accepted 13 November 2017
Available online 22 November 2017
1566-2535/ © 2017 Elsevier B.V. All rights reserved.
X. Wu, H. Liao
view given that the single linguistic term is used to express the
evaluation of an object. It is observed that people usually make
assessments with some hesitancy due to the limited knowledge and
complicated reality. For example, one may describe a product's
quality as “between high and very high” rather than simple term
“high” or “very high”.
(2) They fail to integrate the group opinions into the overall assess-
ments. Calculating the group opinions with simple averaging op-
erators [2,3,35] would loss original information.
(3) Less information is maintained in the DRs’ weights that are re-
presented as crisp numbers after the calculations in the MEMCDM
process [10,17,35,36]. This would lead to the fuzzy information
loss badly.
To solve the first defect of the fuzzy QFD, Onar et al. [23] in-
troduced the hesitant fuzzy linguistic QFD which uses the hesitant fuzzy
linguistic term set (HFLTS) to capture the vagueness more accurately.
The HFLTS [27], which uses more than one linguistic terms to represent
the individuals’ qualitative assessments, is a powerful tool to describe
the imprecision and vagueness. It has received widespread attention in
recent years [19,21,22,32]. However, the HFLTS can only express the
assessments of single individual. Although some aggregation methods
[27,28,31] of HFLTS can be used to integrate the hesitant fuzzy lin-
guistic evaluations of a group, the information loss is serious. In fact,
there is a probability distribution of the evaluations given by a group.
Different linguistic terms in a hesitant fuzzy linguistic element (HFLE,
the element of a HFLTS) [22] should assign different weights (prob-
abilities) to depict the preferences of the decision makers (DMs)
[37,38,39]. In this case, Pang et al. [24] extended the HFLTS and in-
troduced the probabilistic linguistic term set (PLTS) by adding prob-
ability to each linguistic term. The PLTS is an effective representation
form to model group opinions. However, there is no mathematical
formula to calculate the weights of linguistic terms for the collective
probabilistic linguistic element (PLE, the element of a PLTS) corre-
sponding to a group. In this paper, we firstly propose a probability
aggregation method to integrate the individuals’ assessments expressed
as the HFLEs into the collective ones that are expressed as the PLEs. In
addition, given that the weights of DRs, which are derived from the
weights of CRs and the impact degrees of DRs over CRs subjectively, is
uncertain and imprecise, we then try to extend the QFD to probabilistic
linguistic context to get the DRs’ fuzzy weights that are represented as
PLEs. This is the first focus of this paper.
Furthermore, an appropriate ranking method is supposed to select
the optimal design to maximize the customers’ satisfaction based on the
DRs’ fuzzy weights. There are some ranking methods under the hesitant
fuzzy environment, such as the HFL-TOPSIS [19], the HFL-VIKOR [21],
and the HFL-ELECTRE [32]. But all of them are based on the crisp
weights of criteria. The ORESTE (organísation, rangement et Synthèse
de données relarionnelles, in French), proposed by Roubens [29], is a
general ranking method which does not require crisp weights of cri-
teria. Besides, the detailed relationships between alternatives dis-
tinguished as preference, indifference and incomparability relations can
be derived by ORESTE [25], which makes it easy to select a reliable
alternative. However, the ORESTE cannot cope with the imprecise
evaluations that are expressed as PLEs. In ORESTE, Besson's ranks are
employed to express the evaluations. Considering the usefulness of the
PLTS, less information is involved in Besson's ranks. In this paper, we
try to improve the ORESTE by combining it with the PLTS and propose
a MEMCDM method named as the PL-ORESTE. To do so, we firstly
develop a novel distance measure between PLTSs; then we propose a
probabilistic linguistic global preference score function to integrate the
criteria fuzzy weights and the criteria values of the alternatives; fur-
thermore, we propose three kinds of intensity formulas and analyze the
values of thresholds under the probabilistic linguistic context; finally,
the procedure of the PL-ORESTE method is proposed. This method not
only captures the fuzzy information but avoids the complicated
14
Information Fusion 43 (2018) 13–26
calculation process in the traditional ORESTE. This consists the second
focus of this paper.
Based on the above investigation, to handle the innovation product
design selection problems, this paper then develops a synthesis
MEMCDM algorithm by combing the enhanced QFD method with the
PL-ORESTE. The HFLEs are employed to express each individual's
vague evaluations, while the PLEs are used to represent the group
opinions. The improved QFD focuses on translating the fuzzy weights of
CRs into the fuzzy weights of DRs, while the PL-ORESTE aims to rank
the alternatives based on the assessment of the DRs. The main con-
tributions of this paper are summarized as follows:
(1) We propose a novel probability aggregation method to obtain the
collective evaluations of a group.
(2) We extend the QFD to probabilistic linguistic context to get the DRs’
fuzzy weights that are represented as PLEs. We then develop a
distance measure between PLTSs. Based on that, we propose a
probabilistic linguistic global preference score function to integrate
the DRs’ fuzzy weights and the DRs’ values of the alternatives.
(3) We introduce a PL-ORESTE method to solve the general MEMCDM
problems with hesitant fuzzy linguistic evaluations and fuzzy cri-
terion weights. The relations between alternatives are distinguished
in terms of preference, indifference and incomparability but not just
simple rankings.
(4) We further propose a QFD-based PL-ORESTE method to cope with
the MEMCDM problems concerning the innovative design selection.
It captures the voice of customers by integrating the classical QFD
with the PL-ORESTE under probability linguistic context.
(5) We illustrate the applicability of the QFD-based PL-ORESTE method
by a case study concerning the selection of innovative designs of
Panda shared cars. We provide a helpful reference for the en-
terprises related to “shared cars” to make survive by providing sa-
tisfying cars to customers.
The remainder of this paper is organized as follows. Section 2 re-
views the classical QFD and ORESTE methods as well as some concepts
and operations of HFLEs and PLEs. Section 3 proposes the PL-ORESTE
method to solve the general probabilistic linguistic MEMCDM pro-
blems. The QFD-based PL-ORESTE method is developed in Section 4 to
handle the innovative design selection problems. Section 5 presents a
case study concerning the Panda shared cars’ innovative design selec-
tion. The paper ends with some concluding remarks in Section 6.
2. Preliminaries
For the facility of presentation, in this section, we introduce the QFD
and ORESTE methods firstly. Then we review some concepts related to
HFLTS and PLTS.
2.1. The QFD method
The QFD was originated in Japan [1] to handle the product design
issue for converting the customers’ needs into specifications [3,14]. The
MEMCDM problem for innovative product design selection consists of a
finite set of m innovative product design alternatives (A1, ..., Ai, ..., Am),
a set of U CRs (CR1, ..., CRu, ..., CRU) whose weight vector is
WCR = (w1, ...,wu, ...,wU )T , a set of n DRs {DR1, ..., DRj,..., DRn}whose
weight vector is WDR = (ω1, ...,ωj , ...,ωn )T , and two groups of individuals
including a group of Q experts G1 = {e1, ...,eq, ...,eQ} and a group of Z
customers G2 = {f1 , ...,fz , ...,fZ } . This model aims to select the preferred
design which meets the customers’ needs with the highest degree. It is
impossible to evaluate the designs’ comprehensive performance on CRs
directly but can be obtained based on the DRs. So the key issue of this
problem is to translate the assessments on CRs into those on DRs.
The QFD diagram, also called the house of quality (HOQ), is the hub
of the QFD method. It focuses on translating the weights of CRs into
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

essential in product component design. The DRs’ correlations are

Correlation specified in a matrix which is located in the roof of the “house of
Matrix (E) quality” and identified as area (E) in Fig. 1. As this paper mainly
studies the decision problems concerning the selection of innovative
Design requirements product schemes which are designed in advance, it is not necessary
(DRs) (C) for us to consider the relationship between DRs in this paper.
Importance degrees of Step 6. Calculate the importance degrees of DRs by
U U
Requirements(CRs)

∼ = 1 ⎛∑ w
ω ∼ r͠ / ∑ w ∼ ⎞, j = 1, 2, ..,n
(A) Customer

j u ju u⎟
U ⎜u=1 (1)
CRs (B)

⎝ u=1 ⎠
Relationships matrix
(D)

2.2. The ORESTE method

The ORESTE is an attractive method to handle the multi-criteria

Importance degrees of
DRs (F)
Fig. 1. The structure of HOQ.
appropriate weights of DRs of the new products according to the in-
fluence degrees of DRs over CRs [2,14]. The HOQ typically contains six
parts as shown in Fig. 1. The importance degrees of CRs are the key
input and the importance degrees of DRs are the output. The influence
degrees of DRs over CRs are the body of HOQ [11].
Since vagueness and imprecision is inherent in QFD, the fuzzy QFD
method [30] has received much attention, especially in product design
decision making. Chen and Weng [3] employed the fuzzy QFD for a
writing instrument design selection to maximize customer satisfaction
and minimize cost and technical difficulty with respect to each DR.
Zhang et al. [35] combined the fuzzy QFD with the fuzzy Delphi and
gave a case study for the integrated Kitchen stove innovative design
selection. Li [17] used the fuzzy QFD with the fuzzy weighted rating
method to solve the fuzzy automatic washing machine design selection
problem. Zhang and Chu [36] integrated the fuzzy QFD with two op-
timization models, namely, the logarithmic least square model and the
weighted least square model, for horizontal directional drilling machine
selection. Haq and Boddu [10] presented a MEMCDM approach based
on the fuzzy QFD, TOPSIS and AHP. Li [16] extended the QFD with the
2-tuple linguistic representation model, which can cope with multi-
granular linguistic evaluation information. The QFD method has also
been extensively researched in Refs. [6,13]. The fuzzy QFD method
involves the following steps [4]:
Algorithm 1: (The fuzzy QFD method).
Step 1. The DMs are supposed to pre-capture what benefits are
needed by customers for a certain product or service according to
the market and experience, and then determine the elements of CRs,
i.e., CRu(u = 1, 2, ...,U ) , shown as area (A) in Fig. 1.
Step 2. Identify the importance degrees of CRs. Z customers are
invited to give their judgments on the importance degrees of CRs.
The evaluation of the customer fz on CRu is denoted as w ∼(z ) . The
u
collective opinions on CRu of the customer group is calculated by the
average operator where w ∼ = 1 ∑Z w ∼(z )(u = 1, 2, ...,U ) .
u z=1 u
Z (m − 1) n2
Step 3. Determine the elements of DRs. The DMs need to define the
engineering characteristics, i.e., DRj (j = 1, 2, ...,n) , based on the
elements of CRs.
Step 4. Identify the relationships between DRs and CRs, shown as
area (D) in Fig. 1. The expert groups are invited to make evalua-
tions. The influence degree of DRj over CRu evaluated by the expert
(q)
eq is denoted as r͠ ju . Then we can get the group assessments by the
1 Q (q)
average operator r͠ ju = Q ∑q = 1 r͠ ju (j = 1, 2, ...,n; u = 1, 2, ...,U ) .
Step 5. Identify the relationship between DRs. The interdependent
physical relationship between the product structures, i.e., DRs, is
decision making (MCDM) problems in which the crisp weights of cri-
teria are difficult or unable to access. Huylenbroeck [12] pointed out
that the most interesting part of ORESTE is to separate preference (P),
indifference (I) and incomparability (R) relations of alternatives
through the conflict analysis, which makes the results more easily ac-
cepted by the DMs. The ORESTE has been applied to various MCDM
problems with vague information, such as selecting nuclear waste
managerial alternatives [5], ranking information and communication
technology research centers [7], and choosing Radar detection strate-
gies [15].
In ORESTE, the importance degree of the criterion cj is represented
by a Besson's rank rj rather than a crisp value. The merit of the alter-
native Ai with respect to the criterion cj is also represented by a Besson's
rank rj(Ai) (For more details about Besson's rank, please refer to Ref.
[25]). The ORESTE method consists of two main tasks: computing the
weak ranking and building the PIR structure. It involves the following
steps.
Algorithm 2: (Classical ORESTE method).
∼
Step 1. Calculate the global preference score D (aij ) where
∼
D (aij ) = ς (r j )2 + (1 − ς )(r j (Ai ))2 (2)
ς is a coefficient to weight the rank of the criterion cj, i.e., rj, and that
of the alternative Ai with respect to the criterion cj, i.e., rj(Ai). The
value of ς is given by the DMs.
Step 2. Determine the global weak ranking r(aij). The Besson's rank r
∼
(aij) is determined according to the values of D (aij )
(i = 1, 2, ...,m ; j = 1, 2, ...,n ) in descending order.
Step 3. Calculate the weak rank R(Ai), where
n
R (Ai ) = ∑ r (aij)
j=1 (3)
Step 4. Compute the preference intensities. The average preference
intensity between Ai and At is defined as
n
∼ ∑ j = 1 max[(r (atj ) − r (aij )), 0]
T (Ai , At ) =
(4)
The net preference intensity between Ai and At is defined as
∼ ∼ ∼
ΔT (Ai , At ) = T (Ai , At ) − T (At , Ai ) (5)
Step 5. Set up the PIR structure. The principle of the indifference
and incomparability test (conflict analysis) is:
∼ ∼, then
(1) If ΔT (Ai , At ) ≤μ

15
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

∼ ∼
⎧ Ai I At , if T (Ai , At ) ≤ ξ͠ and T (At , Ai ) ≤ ξ͠ of HFLEs (for more detail about the HFLEs’ union, please refer to Ref.
∼ ∼ [27]) is expressed as hS (quality ) = {s0, s1, s2, s3} . Since the same weight
⎨ Ai R At , if T
⎩ (Ai , At ) > ξ͠ or T (At , Ai ) > ξ͠
is assigned to each linguistic term, this expression cannot reflect the
teacher's real instruction quality given that most of students provide
∼ ∼, then their judgments as “a little good”. That is to say, although the union of
(2) If ΔT (Ai , At ) >μ
∼ ∼ ∼
HFLEs can be used to aggregate the evaluations of a group, the
⎧ Ai R At , if min(T (Ai , At ), T (At , Ai ))/ ΔT (Ai , At ) ≥ λ͠
⎪ ∼ ∼ ∼ ∼ ∼ information loss is serious.
A P At , if min(T (Ai , At ), T (At , Ai ))/ ΔT (Ai , At ) < λ͠ and T (Ai , At ) > T (At , Ai )
⎨ i ∼ ∼ ∼ ∼ ∼
⎪ At P Ai , if min(T
⎩ (Ai , At ), T (At , Ai ))/ ΔT (Ai , At ) < λ͠ and T (Ai , At ) < T (At , Ai ) Pang et al. [24] extended the HFLTS and proposed the PLTS by
adding probability to each linguistic term. The PLTS can avoid the loss
∼, ξ͠ and λ͠ are three different thresholds (for details about of original information of a group and thus has gained many scholars’
where μ
attention [18,38,39].
these thresholds, please refer to Ref. [25]).
Step 6. Determine the strong ranking. The final result is a joint
Definition 3. [24]. Let S = {sα α = −τ , …, −1, 0, 1, …, τ } be a LTS, a PLTS
decision based on the weak ranking and the PIR structure.
is defined as:

2.3. HFLTS and PLTS K

⎧ ⎫
hS (p) = s (k ) (p(k ) ) s (k ) ∈ S, p(k ) ≥ 0, k = 1, 2, ...,K ,
⎨
∑ p(k) ≤1
⎬
The subscript-symmetric linguistic term set (LTS) ⎩ k=1 ⎭ (6)
S = {sα α = −τ , ...,−1, 0.1, ...,τ } is a commonly used linguistic evalua-
tion scale, where sα represents the possible value of a linguistic variable where s (k ) (p(k ) ) is the k-th linguistic term s (k ) associated with the
and the specific meaning of sα is given by DMs in practical. Especially, probability p(k), and K is the number of all different linguistic terms in
s0 expresses an assessment of “indifference”, and s−τ and sτ respectively hS(p).
denotes the lower and upper limits of the values of the linguistic vari- Pang et al. [24] proposed some operations of PLTSs. But they are
able. τ is a positive integer. Some basic operations of S are: (1) limited due to the fact that the operated values may exceed the bounds
sα ⊕ sβ = sα + β ; (2) λsα = sλα , λ ∈ [−1, 1]; (3) if α > β, then sα > sβ; (4) of the given LTS and consequently the linguistic information is lost in
the negation operator is neg (sα ) = s−α . Given that the LTS S is in dis- results. To avoid this defect, Gou et al. [8,9] redefined more logical
crete form, Xu [33,34] extended the LTS S to continuous interval form operations of PLTSs based on two equivalent transformation functions.
S = {sα α ∈ [−τ , τ ]} .
The HFLTS has been proved to be an effective tool to express single Definition 4. [8,9]. Let S = {sα α = −τ , …, −1, 0, 1, …, τ } be a LTS, hS(p),
individual's judgments that admits in more than one linguistic terms. hS1 (p) and hS2 (p) be three PLTSs, and θ ∈ [0, 1].
hS (p) = {s (k ) (p(k ) ) k = 1, 2, ...,K } with α(k) being the subscript of the
Definition 1. [27]. Let S0 = {s0, s1, ...,s τ } be a LTS. A HFLTS, HS0 , is an linguistic term s (k ) . η(k) ∈ g(hS), η1(k1) ∈ g (hS1) , η2(k2) ∈ g (hS2) and
ordered finite subset of the consecutive linguistic terms of S0. k = 1, 2, ...,K , k1 = 1, 2, ...,K1, k2 = 1, 2, ...,K2 , where g and g −1 are the
equivalent transformation functions.
Liao et al. [22] extended and formalized the definition of HFLTS
mathematically as follows. g : [−τ , τ ] → [0, 1], g (hS (p)) = {(α (k )/2τ + 1/2)(p(k ) )} = hθ (p), θ

Definition 2. [22]. Let xi ∈ X (i = 1, 2, ...,N ) be fixed and ∈ [0, 1];

S = {sα α = −τ , …, −1, 0, 1, …, τ } be a LTS. A HFLTS on X, HS, is in
mathematical form of g −1: [0, 1] → [−τ , τ ], g −1 (hθ (p)) = {((2θ − 1) τ )(p(θ) ) θ∈[0, 1]} = hS (p).
HS = { < x i , hS (x i ) > x i ∈ X }
where hS(xi) is a set of some values in S and can be expressed as (1)
⎛ ⎞
hS (x i ) = {sφl (x i ) sφl (x i ) ∈ S, l = 1, …, Li } hS1 (p) ⊕ hS2 (p) = g −1 ⎜ ⋃ (k1)
{(η1 + η2(k2) − η1(k1) η2(k2) )(p1(k1) p2(k2) )} ⎟ ;
⎜ η (k1) ∈ g (h 1 (p)), η (k2) ∈ g (h 2 (p)) ⎟
⎝ 1 S 2 S ⎠
with Li being the number of linguistic terms in hS(xi). hS(xi) denotes the (2)
possible degrees of the linguistic variable xi to S. For convenience, hS(xi) ⎛ ⎞
is called the HFLE. hS1 (p) ⊗ hS2 (p) = g −1 ⎜ ⋃ (k ) (k ) (k ) (k )
{(η1 1 η2 2 )(p1 1 p2 2 )} ⎟.
(k1) 1 (k 2) 2
⎝ η1 ∈ g (hS (p)), η2 ∈ g (hS (p)) ⎠
Remark 1. Note that, in Definition 2, the linguistic terms are chosen in The PLTS has gained many scholars’ attention. Zhang et al. [39]
discrete form from S, and the subscripts of sφl (x i ) , φl, belong to introduced the concept of the probabilistic linguistic preference relation
{ −τ , …, −1, 0, 1, …, τ } . In order not to lose much information in (PLPR) whose elements are PLTSs, and then used it to assess the in-
calculation, we consider from now on the extension φl ∈ [−τ , τ ], vestment risk. Zhang et al. [38] investigated the consensus reaching
which is much general and flexible [20]. process for group decision making with PLPRs. Liao et al. [18] proposed
The HFLE is flexible in expressing the assessment of an individual. the probabilistic linguistic linear programming method to handle the
In the sense that it takes the linguistic terms with equal importance, the multi-criteria decision making problems and gave a case study about
HFLE has some flaws in representing the collective opinions of a group, evaluating the hospital levels in China.
which can be illustrated by Example 1. To compare PLTSs, Pang et al. [24] defined the score and the de-
Example 1. Let S = {s−3 = very bad, s−2 = bad, s−1 = a little bad, s0 = viation degrees of a PLTS. For a PLTS hS (p) = {s (k ) (p(k ) ) k = 1, 2, ...,K }
medium , s1 = a little good, s2 = good,s3=very good} be a LTS. Twenty with α(k) being the subscript of the linguistic term s(k), the score of hS(p)
students are invited to evaluate the quality of a teacher's instruction. is
Suppose that ten students think it is “a little good”, expressed as {s1};
K K
five hold it is “between medium and a little good”, denoted as {s0, s1}; two E (hS (p)) = sα , α = ∑ α (k) p(k) / ∑ p(k)
sure it is “at least a little good”, i.e., {s1, s2, s3}; one student insists it is k=1 k=1 (7)
“medium”, i.e., {s0}, and two students cannot give any evaluation. In
this case, the overall evaluation of all students calculated by the union The deviation degree of hS(p) is

16
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

1/2
K K method called the probability aggregation method to obtain the prob-
⎛ ⎞
σ (hS (p)) = ⎜ ∑ (p(k ) (α (k ) − α ))2⎟ / ∑ p(k ) ability of each linguistic term.
⎝ k=1 ⎠ k=1 (8) Let E = {eq q = 1, 2, ...,Q} be a set of experts whose weight vector is
For two PLTSs hS1 (p) and hS2 (p) , if E (hS1 (p)) > E (hS2 (p)) , then (γ(1), γ(2), ..., γ(Q))T, S = {sα α = −τ , …, −1, 0, 1, …, τ } be a LTS. Suppose
hS1 (p) > hS2 (p) ; if E (hS1 (p))=E (hS2 (p)) , then (1) if σ (hS1 (p)) < σ (hS2 (p)), that hS(q) = {sφ(ql ) sφ(ql ) ∈ S, l = 1, ...,L(q) } (q = 1, 2, ...,T ) are T HFLEs on S
hS1 (p) > hS2 (p) , (2) if σ (hS1 (p)) = σ (hS2 (p)) , hS1 (p) = hS2 (p) . given by T experts (there are Q − T experts who do not give any eva-
luation). Let hS = {sφl sφl ∈ S, l = 1, ...,L} be the group HFLE. Then, the
3. The probabilistic linguistic ORESTE (PL-ORESTE) method weight of sφ(ql ) in hS(q) is defined as
1
This section aims to propose a PL-ORESTE method to solve the ⎧ L(q) , if φl is included in hS(q)
vφ(ql ) = , l = 1, …, L(q) ; q = 1, 2, ...,T
general MEMCDM problems where the assessments of alternatives and ⎨ 0,
⎩ if φl is not included in hS(q)
the weights of criteria are represented by HFLEs or PLTSs. First of all,
we describe the probabilistic linguistic MEMCDM problem. Then we (10)
propose a probability aggregation method to get the collective eva- The weight of hS(q) given by the expert eq is γ(q), q = 1, 2, ...,T . Then,
luations expressed as PLTSs. A novel distance measure between PLTSs is the weight of sφl in the group HFLE is
developed which acts as the basis for combining the ORESTE method
T
with PLTSs. Finally, we develop the PL-ORESTE method. p(φl) = ∑ vφ(ql ) γ (q), l = 1, …, L
q=1 (11)
3.1. Description of the probabilistic linguistic MEMCDM problem
Thus, the group opinions can be expressed as the PLTS
Compared with the MCDM problems, the MEMCDM models are T
excellent with multiple experts involving in evaluation, which makes hS (p) = ⎧s (φl) (p(φl) ) s (φl) ∈ S, p(φl) = ∑ vφ(ql ) γ (q), l = 1, …, L⎫
the judgments more believable. In a MEMCDM process, the alternatives ⎨
⎩ q=1 ⎬
⎭ (12)
{A1, ..., Ai, ..., Am}, the criteria {c1, ..., cj, ..., cn} and their weights (ω1,
..., ωj, ..., ωn) are defined. The object of the alternative Ai with respect to Remark 2. With Eq. (12), the judgments given by the individuals in
the criterion cj is denoted as aij, where i = 1, 2, ...,m , j = 1, 2, ...,n . A HFLEs can be aggregated into the PLTS. If the individuals’ judgments
group of individuals {e1, ..., eq, ..., eQ} including experts, investors, are represented in PLTSs, with Eq. (12), we can also get the collective
customers, etc., are invited to make judgments, which are further ag- PLTS in case we use the original probabilities to replace the value of
gregated into the group opinions. vφ(ql ) . If the weights of the experts are not given, we can suppose that
When the experts are asked to make evaluation, both the assess- γ (q) = 1/ Q , q = 1, 2, ...,T .
ments on aij (i = 1, 2, ...,m ; j = 1, 2, ...,n) and the importance of the
Remark 3. To represent information flexibly, we denote φl by a real
criteria cj (j = 1, 2, ...,n) are straightforwardly given with simple or
number rather than an integer. For example, if one states that the
complex linguistic expressions. The LTSs for criteria are established by
design of a car is “between good and very good but more close to very
DMs in advance and different criteria may be associated with different
good”, then we can express it as {s1.5, s2} to reflect the opinion of “more
scales and semantics. Subsequently, the hesitant fuzzy linguistic eva-
close to good” which cannot be implied in the traditional HFLE {s1, s2}.
luations on aij and cj given by the experts are converted to HFLEs based
In this sense, the experts are allowed to express their opinions in a
on the context-free grammar [27]. Then a HFL judgment matrix of the
richer way. If one is sure that the design of a car is “between good and
expert eq can be built as
very good but 80% close to good”, it can be denoted as {s1, s1.2}. The
11(q)
A1 ⎡ hS ⋯ hS1j (q) ⋯ hS1n (q) ⎤ HFLEs whose subscripts are real numbers can be determined according
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥ to the closeness degree to the linguistic terms.
⎢ i1(q) ij (q) in (q)
⎥
H (q) = Ai ⎢ hS ⋯ hS ⋯ hS ⎥ Example 2. (Continued to Example 1). All experts’ opinions in
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥ Example 1 are integrated as hS (quality )={{s1}(10), {s0, s1}(5), {s1, s2,
Am ⎢ ⎥
⎢ hSm1(q) ⋯ hSmj (q) ⋯ hSmn (q) ⎥ (9) s3}(2), {s0}(1)}. The weight of each expert is γ (q) = 1/20 . Then the
⎣ ⎦
aggregated PLTS of the group is hS (quality ) =
with the fuzzy criterion weight vector W = (ω1(q) , ...,ωj(q) , ...,ωn(q) ) where {s0 (0.18), s1 (0.66), s2 (0.03), s3 (0.03)} . It reflects the actual group
ωj(q) = hSj (q) . assessment fully.
There are mainly three kinds of aggregation methods of the HFLEs
3.2. The probability aggregation method to obtain the collective judgments
to get the collective evaluations of the DMs, including the union method
[27], the arithmetic mean aggregation (AMA) operator [28] and the
In this subsection, we develop a probability aggregation method to
weighted averaging (WA) operator [31]. In the following, we compare
integrate the individuals’ opinions which are expressed as the HFLEs to
these three aggregation methods with our proposed method to illustrate
the collective judgments of the group which are expressed as PLTSs.
its effectiveness.
The PLTS improves the HFLEs from two aspects: ① the expert gives
the probability of each linguistic term when making evaluation. For
(1) The drawback of the union method has been illustrated in
example, one may express that he/she is 60% sure that the quality of a
Example 1. In addition, the union method is suitable for integrating
product is “high”, 20% sure that it is “very high”. Then his/her assess-
the evaluations of a small number of experts who are easy to reach
ment can be expressed as a PLTS 〈s1(0.6), s2(0.2)〉; ② assign the prob-
consensus but is limited to aggregate the opinions of a large group.
ability to each linguistic term based on all individuals’ evaluations to
(2) The AMA operator of HFLEs is based on the envelopes of HFLEs. Let
express the group opinions (this is the main issue discussed in this
S be a LTS. For a set of HFLEs hS(q) = {sφ(ql ) sφ(ql ) ∈ S, l = 1, ...,L(q) }
paper). For the first case, generally, it is difficult for an expert to give
(q = 1, 2, ...,T ) given by the experts eq (q = 1, 2, ...,T ), the collective
the probabilities of linguistic terms as it does not conform to expressing
judgment of the group is
habits. As the HFLE is a special case of PLTS in which the probabilities
of the terms are equal, we suppose that the experts use the HFLEs to ⎡1
T
1
T
⎤
express their assessments and the group's opinions are represented by h͠ S = ⎢ ∑ sφ−l (q), ∑ sφ+l(q) ⎥
T q=1
T q=1 (13)
PLTSs after aggregation. In this sense, we introduce an aggregation ⎣ ⎦

17
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Example 5. LetS = {s−3, …, s0, …, s3} be a LTS, and

hS1 (p) = {s1 (0.3), s2 (0.5)} , hS2 (p) = {s−1 (0.2), s1 (0.3), s2 (0.5)} and
where sφ−l (q) = min sφ(ql ) and sφ+l (q) = max sφ(ql ) . hS3 (p) = {s−1 (0.6), s1 (0.2), s2 (0.1)} be three PLTSs. Suppose that the
l l
Example 3. (Continued to Example 1). The envelopes of the HFLEs evaluations on the three alternatives A1, A2, A3 are hS1 (p) , hS2 (p) ,
1 2 3
given by students are h͠ S = [s1, s1] of ten, h͠ S = [s0, s1] of five, h͠ S = [s1, s3] hS3 (p) , respectively. Based on the PLTS comparison method presented
͠ 4 in Section 2.3, we get hS1 (p) > hS2 (p) > hS3 (p) . In ORESTE, the Besson's
of two and hS = [s0, s0] of one. Then the collective judgment of the
ranks are r (A1 ) = 1, r (A2 ) = 2 and r (A3 ) = 3. Obviously, as for the
group is h͠ S (quality ) = [s0.67, s1.17]. It can reflect most students’ opinions
evaluations, A1 is extremely close to A2 but far behind A3. However, the
on the teacher's instruction that is “close to a little good”. However, the
Besson's ranks reflect the same degree of difference between these three
original linguistic items are lost. In addition, it cannot reflect the
alternatives and consequently weaken the information seriously.
distribution of group opinions and the situation that some experts
cannot give any evaluation. This subsection aims to propose the PL-ORESTE method to handle
the general MEMCDM problems. In PL-ORESTE, the PLTSs are em-
(3) Wang [31] proposed a WA operator to aggregate the group
ployed to express the group evaluations and the ORESTE is used to rank
opinions in the form of the extended HFLEs. For a set of HFLEs
the alternatives. The key to combine PLTSs and ORESTE is to use the
hS(q) = {sφ(ql ) sφ(ql ) ∈ S, l = 1, ...,L(q) } (q = 1, 2, ...,T ) given by the experts eq
distance measure between two PLTSs, which displays the relationships
(q = 1, 2, ...,T ) with the weight vector W = (ω1, ω2 , ...,ωQ )T , the collec-
between criteria and the relationships between alternatives, to replace
tive judgment is
the Besson's ranks in the classical ORESTE method. In the following, we
Q Q start by introducing the distance measure between PLTSs.
q=1 q=1
(
⊕ (ωq hS(q) ) = ⊕ ∪ sφ(q)∈ h (q) {ωq sφ(ql ) } =
l S ) It is observed that the numbers of linguistic terms in different PLTSs
are usually different. Thus we should add some linguistic terms to the
∪ sφ(1)∈ h (1),..., sφ(Q)∈ h (Q) {ω1 sφ(1)
l
⊕ ⋯⊕ωQ sφ(Ql ) } (14)
l S l S shorter ones to maintain the same numbers of linguistic terms. Below
Example 4. (Continued to Example 1). Suppose the weights of students we introduce a new method to add the linguistic terms. For two PLTSs
are the same (The students who do not give any evaluation are hS1 (p) = {sα1(k ) p1(k ) k = 1, 2, ...,K1} and hS2 (p) = {sα2(k ) p2(k ) k = 1, 2, ...,K2}
ignored). Based on the WA operator of HFLEs, the collective with K1 and K2 being the numbers of linguistic terms in hS1 (p) and hS2 (p) ,
judgment of the group is hS (quality ) = {s13/18, s14/18, s15/18,s16/18, s17/ respectively, if K1 > K2, then we add K1 − K2 linguistic terms to hS2 (p) .
The added linguistic terms are sα2(1) , and the sum of the probabilities of
18,s18/18,s19/18,s20/18,s21/18}. Similar to the AMA operator, it can reflect
the majority views of the group. However, the computation process is sα2(1) equals to p2(1). In other words, we divide sα2(1) p2(1) to multiple terms
complex and the result is hard to understand given that there are too as sα2(1) p12(1) , sα2(1) p22(1) , …, with the condition that
many similar elements. p12(1) + p22(1) + ⋯pd2(1) = p2(1) and p12(1) = p1(1) , p22(1) = p1(2) ,…,
pd2(1) ≤ p1(d) , pd2(1)+1 = 0…
Through the above comparative analyses, we can find that the Zhang et al. [39] introduced the distance measure between two
probability aggregation method is more reliable as it can express the PLTSs hS1 (p) = {sα1(k ) p1(k ) k=1, 2, ..., K} and
group opinions completely. Its advantages can be summarized as fol- 2
hS (p) = {sα p 2(k ) 2(k )
k = 1, 2, ...,K } as
lows:
K
α1(k ) − α1(k )
(1) It retains all the original linguistic terms given by the DMs. d (hS1 (p), hS2 (p)) = ∑ (p1(k) p2(k) )
k=1
2τ (16)
(2) It considers both the weights of experts and the hesitancy degree
(the number of linguistic terms) of a HFLE. The probability dis- where α 1(k)
and α 2(k)
are the subscripts of the k-th linguistic terms sα1(k )
tribution of the linguistic terms are reflected completely. and sα2(k ) , respectively.
(3) The calculation process is simple.
(4) The result is expressed as the PLTS, which is convenient for further Example 6. (Continued to Example 5). To compute the distance
computation. between hS1 (p) and hS2 (p) , hS1 (p) is translated to
hS1 (p) = {s1 (0.2), s1 (0.1), s2 (0.5)} . According to Eq. (16), we have
According to Eq. (12), the individuals’ evaluations are aggregated d (hS1 (p), hS2 (p)) = 0.01, d (hS1 (p), hS3 (p)) = 0.06 and d (hS2 (p), hS3 (p)) = 0 .
into collective judgments expressed as PLTSs. Then, a probabilistic The results are against our intuition. From Eq. (16), we can find that if
linguistic judgment matrix is established as one of p1(k), p2(k) is zero or α1(k ) = α1(k ) , the value of the k-th item in the
right side of Eq. (16) is zero. Thus, we need to develop a new distance
A1
11
⎡ hS (p) ⋯ hS1j (p) ⋯ hS1n (p) ⎤ measure between PLTSs.
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥
⎢ i1 ⎥ Definition 5. Let S = {sα α = −τ , …, −1, 0, 1, …, τ } be a LTS,
L (p) = Ai ⎢ hS (p) ⋯ hSij (p) ⋯ hSin (p) ⎥
⎢ ⋮ hS1 (p) = {sα1(k ) p1(k ) k = 1, 2, ...,K1} and hS2 (p) = {sα2(k ) p2(k ) k = 1, 2, ...,K2}
⋮ ⋱ ⋮ ⋱ ⋮ ⎥
Am ⎢ m1 ⎥ be two PLTSs in S, where K1 = K2 = K (if K1 ≠ K2, we should add
⎣ hS (p)
⎢ ⋯ hSmj (p) mn
⋯ hS (p) ⎥ (15)
⎦ some linguistic terms to the shorter one). The distance between hS1 (p)
The criterion weight vector (ω1, ..., ωj, ..., ωn) with ωj = T
hSj (p) is and hS2 (p) is defined as
also determined based on Eq. (12).
d (hS1 (p), hS2 (p))
2 2 1/2
3.3. The PL-ORESTE method ⎧⎛ 1 K ⎧ 2(k ) α1(k ) − α2(k ) 1(k )
⎪⎜ K ∑k = 1 p ⎛ ⎞ + (p1(k ) − p2(k ) )⎛ α ⎞ ⎫ ⎞⎟ , p1(k ) ≥ p2(k )
2τ
⎪ ⎨
⎩ ⎝ ⎠ ⎝ 2τ ⎠ ⎬ ⎭⎠
= ⎝ 1/2
The ORESTE is an excellent ranking method which accelerates the ⎨ 1(k ) 2(k ) 2 2(k ) 2
⎪⎛⎜ 1 ∑K ⎧p1(k ) ⎛ α − α ⎞ + (p2(k ) − p1(k ) )⎛ α ⎞ ⎫ ⎞⎟ , p1(k ) < p2(k )
decision making process as the DMs do not need to spend time to de- ⎪⎝ K k = 1 ⎨
⎩ ⎝ 2τ
⎠ ⎝ 2τ ⎠ ⎬ ⎭⎠
⎩
termine the criterion weights. By contrast, it enhances the credibility of
(17)
results by conflict analysis. However, the decision matrix composed of
Besson's ranks maintains less evaluation information, which may result where α and α
1(k) 2(k)
are the subscripts of the k-th linguistic terms sα1(k )
2(k )
in inaccurate results (see Example 5). Furthermore, the ORESTE is only and sα , respectively.
applicable in the MCDM with single DM since it is hard to determine the
unified ranks of multiple experts. Remark 3. If p1(k ) = p2(k ) = 1 (k = 1, 2, ...,K ) , Eq. (17) turns to the

18
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Euclidean distance between HFLEs. Obviously, Dij ∈ [0, 1].

The probabilistic linguistic preference score of the alternative Ai is
Property 1. d (hS1 (p), hS2 (p)) ∈ [0, 1], especially if hS1 (p) = hS2 (p), then
integrated by
d (hS1 (p), hS2 (p)) = 0 ; if p1(k) = p2(k ) = 1(k = 1) and α1(k ) − α 2(k ) = 2τ ,
then d (hS1 (p), hS2 (p)) = 1. 1
n
Di =
n
∑ Dij
Proof. It is obvious that α1(k ) − α 2(k ) ∈ [0, 2τ ] and |α1(k)| ∈ [0, τ]. j=1 (24)
Therefore, α1(k ) − α 2(k ) /2τ ∈ [0, 1] and |α1(k)|/2τ ∈ [0, 0.5]. Let
ϑ = max{ α1(k ) − α 2(k ) /2τ , α1(k ) /2τ } ∈ [0, 1]. Then, it follows The weak rank of Ai, r(Ai), is determined according to Di from the
1/2 smallest to the largest directly. The P and I relations are established in
d (hS1 (p),
1
hS2 (p))≤ K ( K
∑k = 1 {p2(k ) ϑ2 + (p1(k ) − p2(k ) )ϑ2} ) the following ways: if Di > Dt, then r(At) > r(Ai) and At P Ai; if Di = Dt ,
1/2 1/2
= ( 1
K
K
∑k = 1 {p1(k ) ϑ2} ) 1
( K
≤ K ∑k = 1 ϑ2 = ) ϑ . This completes the proof. then r (Ai ) = r (At ) and Ai I At (i, t ∈ (1, 2, ..., m)). The weak ranking of
alternatives is quite useful in some applications. Nevertheless, it is
Example 7. (Continued to Example 6). To compute the distance difficult for the DMs to accept the I relation between Ai and At in the
between hS1 (p) and hS2 (p) , hS1 (p) is translated to hS1 (p) = {s1 (0.2),
case that Di = Dt but their performance are significantly different under
s1 (0.1), s2 (0.5)} . Then we have d (hS1 (p), hS2 (p)) = 0.016. Similarly, we
some criteria, and the P relation in the case that Di is close to Dt. The
have d (hS1 (p), hS3 (p)) = 0.053 and d (hS2 (p), hS3 (p)) = 0.034 . Our distance
incomparability (R) relation is not considered in the above assumption.
measure effectively avoids the defect of Eq. (16).
In addition, under the fuzzy linguistic environment, the crisp preference
In the following, we employ the distance measure between PLTSs scores may lead to the loss of information to some extent. Thus, we
defined as Eq. (17) as the basis of the PL-ORESTE operations. After further divide the weak ranking relations P and I between alternatives
building the group probabilistic linguistic judgment matrix, the fol- into the PIR structure by introducing three preference intensities and
lowing thing we should do is to construct the weak ranking of the al- three thresholds to derive the reliable results.
ternatives. Under the probabilistic linguistic circumstance, the preference in-
Motivated by the HFL-TOPSIS method [19], the maximum PLTS tensity measuring the degree of dominance of Ai over At is calculated
hSj + (p) of Ai with respect to cj is defined as based on the probabilistic linguistic global preference score.
ij
⎧ max {hS (p)}, for the benefit criterion cj Definition 7. Let Dij and Dtj be the probabilistic linguistic global
i = 1,2, … , m
hSj + (p) = preference scores of Ai and At over the criterion cj, respectively. The
⎨ min {hSij (p)}, for the cost criterion cj
⎩ i = 1,2, … , m (18) preference intensity of Ai over At with respect to cj is defined as
The minimum PLTS hSj − (p) of Ai with respect to cj is defined as Tj (Ai , At ) = max {(Dtj − Dij ), 0} (25)
ij
⎧ min {hS (p)}, for the benefit criterion cj
i = 1,2, … , m
hSj − (p) =
⎨ max {hSij (p)}, for the cost criterion cj The average preference intensity of Ai over At is defined as
⎩ i = 1,2, … , m (19) n
1
The weight of the most important criterion satisfies T (Ai , At ) =
n
∑ max{(Dtj − Dij ), 0}
j=1 (26)
ω+ = max ωj = max {hSj (p)}
j = 1,..., n j = 1,..., n (20) The net preference intensity of Ai over At is defined as

The weight of the most unimportant criterion satisfies ΔT (Ai , At ) = T (Ai , At ) − T (At , Ai ) (27)

ω− = min ωj = min {hSj (p)} As pointed by Pastijn and Leysen [25], the preference (P) relation
j = 1,..., n j = 1,..., n (21) between Ai and At is confirmed when their net preference intensity is
Considering that different criteria may have different scales and big enough; otherwise, the indifference (I) or incomparability (R) re-
semantics, we need to conduct normalization. Inspired by the normal- lation is identified. Furthermore, the I relation is established when the
ization process adopted in VIKOR [21], we get the normalized value as net preference intensity is close to zero and their preference intensities
under all criteria are close to zero as well. By contrast, the R relation is
ij
hS (p) = d (hSj + (p), hSij (p))/ d (hSj + (p), hSj − (p)) (22) identified when their preference intensities are very big under some
criteria in case that their net preference intensity is close to zero.
where hSj + (p)
is the maximum PLTS and over the criterion cj, hSj − (p)
Accordingly, three thresholds are introduced to distinguish the PIR
respectively.
relations. The indifference threshold (δ) is used to differentiate I and R
Following Eq. (2) in the classical ORESTE, we define the probabil-
with respect to each criterion; the preference threshold (μ) is used to
istic linguistic global preference score function as follows:
separate P; the incomparability threshold (ξ) is used to distinguish I and
R in global. The value of these thresholds are derived from the global
Definition 6. The probabilistic linguistic global preference score
preference scores.
function of the alternative Ai with respect to the criterion cj is in the
mathematical form of
Definition 8. For two PLTSs hS1 (p) = {sα1(k ) p1(k ) k = 1, 2, ...,K1} and
2 2 1/2 hS2 (p) = {sα2(k ) p2(k ) k = 1, 2, ...,K2} , if d (hS1 (p), hS2 (p)) < ɛ , we suppose
⎡ d (hSj + (p), hSij (p)) ⎞ +
⎛ d (ω , ωj ) ⎞ ⎤
Dij = ⎢ψ ⎜⎛ j+ j− ⎟ + (1 − ψ) + , ω−) ⎥
⎜ ⎟ hS1 (p) is indifferent to hS2 (p) and ɛ is called the probabilistic linguistic
d (hS (p), hS (p)) ⎠ ⎝ d (ω ⎠⎦
⎣ ⎝ (23) indifference threshold.
where d (hSj + (p), hSij (p)) , d (hSj + (p), hSj − (p)) , d (ω+, ωj ) and d (ω+,
ω−) are Without loss of generality, analyzing the simplest case that
the distances calculated by Eq. (17). d (hSj + (p), hSij (p))/ d (hSj + (p),
hSj − (p)) p1(k ) = p2(k ) = 1 and K = 1, if sα1(k ) − sα2(k ) ≤ 0.5 (such as hS1 (p) = {s0 (1)}
expresses the normalized probabilistic linguistic distance of the and hS2 (p) = {s0.5 (1)} ), we deem that hS1 (p) is indifferent to hS2 (p) . In this
alternative Ai with respect to the criterion cj, which is abbreviated as situation, d (hS1 (p), hS2 (p)) ∈ [0, 0.5/2τ ]. Thus, we get ɛ ∈ [0, 0.5/2τ].
dij. d (ω+, ωj )/ d (ω+, ω−) denotes the normalized probabilistic linguistic Afterwards, we need to determine the value of δj, which implies that
distance of the criterion cj with respect to its importance degree, which if Tj (Ai , At ) = max {(Dtj − Dij ), 0} ≤ δj and Tj (At , Ai ) =
is abbreviated as dj. ψ ∈ [0, 1] expresses the relative importance max {(Dij − Dtj ), 0} ≤ δj , then At is indifferent to Ai on the criterion cj.
between dij and dj. Without loss of generality, we let ψ = 0.5. Considering that Dij is derived from dij and dj, we carry out the

19
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

approximate calculation to get δj based on dij and dj as follows (To average preference intensity T(Ai, At) by Eq. (26) and the net pre-
perform the calculations smoothly, let dj = 0 and ference intensity ΔT(Ai, At) by Eq. (27).
d (hSj + (p), hSj − (p)) = 1): Step 8. Determine the values of the thresholds and establish the PIR
structure according to Eq. (28).
Dtj − Dij = (1/2(dtj )2 + 1/2(dj )2)1/2 − (1/2(dij )2 + 1/2(dj )2)1/2 Step 9. Get the strong ranking of the alternatives based on both the
= 2 /2 dij − dtj weak ranking and the PIR structure.
2
= d (hSj + (p), hStj (p))/ d (hSj + (p), hSj − (p)) − d (hSj + (p), hSij (p))
2 4. The QFD-based PL-ORESTE method
2
/ d (hSj + (p), hSj − (p)) ≈ 2
d (hSij (p), hStj (p))
This section focuses on extending the classical QFD method into the
2
Thus, we get δj ∈ 2 [0, 0.5/2τ ]. In a MEMCDM, if we take the same probabilistic linguistic circumstance and combining it with the pro-
scale of LTS for each criterion, δj (j = 1, ...,n) are unified as δ. To sim- posed PL-ORESTE method to deal with a special MEMCDM problem for
plify the presentation, we suppose that the seven-valued LTS is used, innovative product design selection based on the CRs and the DRs.
i.e., 2τ = 6. Then, we can get ɛ ∈ [0, 0.8333] and the indifference ∼ , r͠ and
For the problems described in Section 2.1, the values of w u ju
∼ are expressed as fuzzy numbers (such as triangular fuzzy numbers
threshold δ ∈ [0, 0.0589] for each criteria. ω j
The values of the preference threshold μ and the incomparability [1,2,4,23] and 2-tuple linguistic representation model [6,13,16]) in the
threshold ξ are obtained based on δ through the analyses on the P and R fuzzy QFD methods. However, they are unable to express vague in-
relations: formation comprehensively as we have justified in the Introduction. In
addition, using the average operator to aggregate group judgments
(1) For simple dominance that Ai P At, it is the minimum condition that would loss information seriously. Therefore, we extend the fuzzy QFD
Dij and Dtj (j = 1, 2, ...,n) are equal for all criteria except one. That is to the probabilistic linguistic context in which the weights of CRs and
1 n−1 the relation between CRs and DRs are expressed as HFLEs and then
to say, for n − 1 criteria, n − 1 ∑ j = 1 max[(Dtj − Dij ), 0] = 0 ; for the
n-th criterion, |Tn(Ai, At)| ≥ δ. Then T (Ai , At ) − T (At , Ai ) = δ / n . aggregated to PLTSs as group opinions. In this way, the original fuzzy
As this is a minimum condition, we let Ai P At when information is retained.
T (Ai , At ) − T (At , Ai ) ≥ δ / n . Therefore, μ = δ / n is taken as the The PL-ORESTE proposed in Section 3 is effective in handling the
preference threshold. general MEMCDM problems, but it is unable to solve the innovative
(2) For the incomparability relation Ai R At, we should have design selection problem which aims to find a preferred design to
T (Ai , At ) − T (At , Ai ) < δ / n and Tj(Ai, At) ≥ δ for at least one maximize the CRs. Considering that the QFD method is excellent in
translating the CRs into the DRs, we combine it with the PL-ORESTE
criterion. In this case, if n is odd, T (Ai , At ) ≥
1
n ( n
2
δ )
+δ =
(n + 2) δ
2n
; if
method to propose a QFD-based PL-ORESTE method to handle the in-
n is even, T (Ai , At ) ≥
1
n ( )=
n
2
δ
δ
2
. Thus, the incomparability novative design selection problem.
threshold ξ = (n + 2) δ /2n if n is odd, and ξ = δ/2 if n is even. The algorithm of the QFD-based PL-ORESTE method is presented as
below and also illustrated in Fig. 2 intuitively.
According to the above analyses on the PIR relations, we establish
the rules shown as Eq. (28) for the PIR test according to the thresholds: Algorithm 4: (The QFD-based PL-ORESTE method).

⎧ ⎧ ΔT (Ai , At ) > 0, Ai P At Step 1. Identify the CRs, the DRs corresponding to the CRs, and the
⎪ If ΔT (Ai , At ) ≥ μ, then ⎨ ΔT (A , A ) > 0, At P Ai
⎪ ⎩ t i innovative product design alternatives. Then the customers (G2)
⎨ T (Ai , At ) < ξ and T (At , Ai ) < ξ , Ai I At make assessments over the importance of CRs (denoted as wu), while
⎪ If ΔT (Ai , At ) < μ, then ⎧ the experts (G1) evaluate each design Ai with respect to each DR
⎪ ⎨
⎩ T (Ai , At ) ≥ ξ or T (At , Ai ) ≥ ξ , Ai R At
⎩ (denoted as DRj) and the relation between the CRs and the DRs
(28) (denoted as rju). The linguistic evaluations are translated to the
To make it easy to be understood, we summarize the procedure of HFLEs hSu (z )(u=1, 2, ..., U; z = 1, 2, ...,Z ) , hSij (q) (i = 1, 2, ...,m ;
the PL-ORESTE method as follows. j = 1, 2, ...,n;q = 1, 2, ...,Q) and hSju (q)(j = 1, 2, ...,n;u = 1,2, ..., U;
q = 1, 2, ...,Q) , respectively.
Algorithm 3: (The PL-ORESTE method). Step 2. Aggregate the evaluations of all individuals to group opi-
nions in PLTSs by Eq. (12). Then, we obtain the CRs’ weight vector
Step 1. Collect and translate each individual’ evaluations into WCR = (w1, ...,wu, ...,wU )T where wu = hSu (p)(u = 1, 2, ...,U ) , the rela-
HFLEs, including the assessments of alternatives, hSij (q) , and the tion matrix R between CRs and DRs, i.e.,
weights of criteria, hSj (q) , where i = 1, 2, ...,m ; j = 1, 2, ..., CR1 ⋯ CRu ⋯ CRU
n; q = 1, 2, ...,Q . Then, build the HFL judgment matrices H(q) as
DR1 r ⋯ r1u ⋯ r1U
Eq. (9) and the fuzzy criterion weight vectors ⎡ 11 ⎤
⋮ ⎢⋮ ⋱ ⋮ ⋱ ⋮ ⎥
W = (ω1(q) , ...,ωj(q) , ...,ωn(q) ) of all the experts where q = 1, 2, ...,Q . ⋯ ⋯ r jU ⎥
R = DRj ⎢ r j1 r ju
Step 2. Aggregate the group opinions represented as the PLTSs ⎢⋮
⋮ ⋱ ⋮ ⋱ ⋮ ⎥
hSij (p) and hSj (p)(i = 1, 2, ...,m ; j = 1, 2, ...,n) by Eq. (12). Then, es- ⎢r ⋯ rnu ⋯ rnU ⎥
DRn ⎣ n1 ⎦ (29)
tablish the probabilistic linguistic judgment matrix L(p) by Eq. (15).
Step 3. Find the maximum and minimum PLTSs, including hSj + (p) ,
hSj − (p) , ω+ and ω− by Eqs. (18)-(21), respectively. where r ju = hSju (p) denotes the influence degree of DRj over CRu, the
Step 4. Calculate the distances d (hSj + (p), hSij (p)) , d (hSj + (p), hSj − (p)) , judgment matrix J
d (ω+, ωj ) and d (ω+, ω−) by Eq. (17). Then compute the normalized
probabilistic linguistic values of dij and dj.
Step 5. Compute the probabilistic linguistic global preference score
Dij by Eq. (23).
Step 6. Determine the weak rank r(Ai) by Eq. (24).
Step 7. Compute the preference intensity Tj(Ai, At) by Eq. (25), the

20
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Input the original

elements

A set of CRs A set of DRs A set of alternatives

Customer group Expert group

Part 1: Evaluation
CRs’ weights in Relation of DRs over Merits of alternatives
collection and
HFLEs CRs in HFLEs over DRs in HFLEs
expression

Part 1: Evaluation
Group opinion Group opinion Group opinion
group aggregation
in PLTSs in PLTSs in PLTSs

Global
scores

Part3: Ranking
alternatives
DRs’ weights Weak rankings PIR structure
in PLTSs

Strong rankings
Part 2: DRs’ weights
calculation

Fig. 2. The algorithm of the QFD-based PL-ORESTE method.

DR1 ⋯ DRj ⋯ DRn d (ω+, ωj )/ d (ω+, ω−) is denoted as dj. Then, construct the decision
A1 a ⋯ a1j ⋯ a1n matrix E where
⎡ 11 ⎤
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥ DR1 ⋯ DRj ⋯ DRn
J = Ai ⎢ a i1 ⋯ aij ⋯ ain ⎥
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥ A1 ⎡ d11 ⋯ d1j ⋯ d1n ⎤
⎢ ⎥ ⎢ ⋮
Am a ⋯ amj ⋯ amn ⎦ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥
⎣ m1 (30) ⎢d
E = Ai ⋯ dij ⋯ din ⎥
⎢ i1 ⎥
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥
where aij = hSij (p) represents the judgment of the alternative Ai with Am ⎢ d m1 ⋯ dmj ⋯ dmn ⎥
⎣ ⎦ (32)
respect to DRj.
Step 3. Calculate the DRs’ weights. The weight vector of DRs,
WDR = (ω1, ...,ωj , ...,ωn )T , is computed based on the CRs’ weights and Step 5. Compute the probabilistic linguistic global preference score
the influence degrees of DRs over CRs. To maintain the fuzzy in- Dij based on dij and dj by Eq. (23) and then we can obtain the matrix
formation completely, the final weights of DRs are still expressed in G where
the PLTSs with
DR1 ⋯ DRj ⋯ DRn
U
ωj = hSj (p) = ∑ wu rju, j = 1, 2, ...,n A1 ⎡ D11 ⋯ D1j ⋯ D1n ⎤
u=1 (31) ⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥
G = Ai ⎢D ⋯ Dij ⋯ Din ⎥
⎢ i1 ⎥
⋮ ⎢ ⋮ ⋱ ⋮ ⋱ ⋮ ⎥
Step 4. Find the maximum PLTS hSj + (p) and the minimum PLTS
Am ⎢ Dm1 ⋯ Dmj ⋯ Dmn ⎥
hSj − (p) under each DRj by Eq. (18) and Eq. (19), respectively. In ⎣ ⎦ (33)
addition, select the most important DR ω+ and the most unimportant
DR ω− by Eq. (20) and Eq. (21), respectively. After that, calculate The following steps are the same as Steps (6)–(9) in Algorithm 3 of
the distance d (hSj + (p), hSij (p)) , d (hSj + (p), hSj − (p)) , d (ω+, ωj ) and the PL-ORESTE method.
d (ω+, ω−) by Eq. (17). The normalized distance
d (hSj + (p), hSij (p))/ d (hSj + (p), hSj − (p)) is denoted as dij and

21
X. Wu, H. Liao
Compared with the classical fuzzy QFD, the QFD-based PL-ORESTE
method has the following advantages:
(1) The hesitant and fuzzy linguistic evaluations of each individual are
completely expressed by HFLEs.
(2) The collective evaluations are represented as PLTSs which main-
tains the group opinions completely.
(3) The information loss issue is avoided since the DRs’ fuzzy weights
do not need to translate to crisp weights.
(4) The ranking of alternatives are more convincing with the help of the
PIR structure.
5. A case study: the innovative design selection of the Panda
shared cars
In this section, we employ the QFD-based PL-ORESTE method to
select the innovative design of Panda shared cars. Some comparative
analyses are conducted with the QFD-based PL-ORESTE method as well
as the classical ORESTE method to illustrate the effective of our pro-
posed method.
5.1. Case description about the Panda shared cars in China
The concept of "shared economy" has continuously permeated and
changed the traditional lifestyle of ourselves over the past several years.
As a most representative form of “shared economy”, “shared travel” can
improve the efficiency of transport and solve the problem of scarcity of
transportation resources. Therefore, it has been strongly advocated at
the national level. Aimed at solving the travel problems in the last mile
of mankind, the “shared bikes” are growing rapidly. Recently, to aid the
medium and long distance travel, the “shared cars” began to appear in
Chinese market. Shared cars mainly refers to time-sharing leasing
mode. That is to say, a certain company who puts a lot of cars in the
market provides users with timely travel services of the cars, including
scheduling, insurance, parking and maintenance. It takes anywhere
self-service rental car business model to bring users high-quality travel
experience. It can meet the needs of traveling, shopping, party needs
and so on. Compared with private cars, the “shared cars” are eco-
nomical. It can ease traffic jams and reduce environmental pollution.
Predictably, it will become an important trend in future travel. In
China, the “shared cars” is at the initial stage of industrial development
but has shown great potential for market development because China is
a large country with very limited social resources per capita. It not only
produces opportunities but also brings challenges for Chinese “shared
cars” enterprises. As an emerging industry, only those “shared cars”
that meet people's need and preferences can survive for a long time.
“Panda” is a Chinese shared cars intelligent travel platform with
new energy vehicles. It focuses on the form of "internetwork + vehicle
networking + energy Internet + auto service". By right of the unique
mode of electric vehicles and excellent smart travel experience, in just
over a year, “Panda” has set up branches in Chongqing, Hangzhou,
Chengdu and other 8 cities in China. The total number of new energy
vehicles put into operation have reached to more than 5000 units with
Table 1
The linguistic scales used on different items.
Linguistic term Linguistic scale for the importance Linguistic scale for the influence degree
of CRs of DRs over CRs
s−3 None (N) None (N)
s−2 Strongly Unimportant (SU) Extremely Weak (EW)
s−1 Unimportant (U) Weak (W)
s0 Medium (M) Medium (M)
s1 Important (I) Strong (S)
s2 Strongly Important (SI) Extremely Strong (ES)
s3 Perfect (P) Perfect (P)
22
Information Fusion 43 (2018) 13–26
the number of users increased by an average of 50% per month.
“Panda” has become the fastest time sharing leasing business and the
largest shared project of new energy vehicles in a single city in China.
Nowadays, there are more than 100 shared car brands in Chinese
market, the larger of which include “Car2Share”, “EVcard”, “Gofun”,
“Car2Go”, “Panda”, “TOGO”, etc. The investment are constantly ex-
panding to seize the market. To success in the competition of the
“shared cars” market, two challenges must be solved by Panda: ① gain
the favor of users (only by maximizing the user's needs can Panda ob-
tain the market); ② control costs (the downside is that the “shared cars”
maintain expensive production costs, maintenance costs, and operating
costs. Only by controlling costs can “Panda” operate normally).
Consequently, it is urgent for “Panda” to develop new models that
maximize people's travel demand and keep down costs. This is a typical
innovative product design selection problem. We attempt to apply the
proposed QFD-based PL-ORESTE method to select the most excellent
innovative shared car design for “Panda”.
Suppose that there are five innovative design alternatives {A1, A2,
A3, A4, A5} proposed by the design team. Through interviews and
surveys on the users, five major CRs are identified to represent the
biggest concerns for shared cars.
- Price (CR1): It includes deposit and lease charge. This is the focus of
the user's choice of using a shared car rather than a private car.
- Comfortability (CR2): This is related to the interior design, seat
material, air flow, etc.
- Safety (CR3): This refers to the safety problem determined by the
vehicle itself, such as body material, body structure, chassis, etc.
- Convenience (CR4): It is an important concern of users on shared car
compared to private car, which includes the flow of borrowing the
car, charge, driving process, etc.
- Space (CR5): It includes storage space, riding space, leg space, etc.
The Panda's APP design, positioning, unlocking and other online
operations are so mature that we do not need to further investigate. For
the offline of Panda cars, there is not required to rebuild the chassis
system, power system, body structure of the car. Accordingly, with the
help of the design teams, six DRs are selected regarding to the five CRs
of Panda shared cars.
- Cost (DR1): Unlike “shared bicycles”, shared cars have high costs,
including vehicle production costs, maintenance costs, APP oper-
ating costs, etc.
- Car body material (DR2): Steel plate material, light metal material
and composite material are widely used in car body, such as roof,
door, inlettake, engine cover and radiator, etc.
- Seat material (DR3): There mainly are three kinds of seat materials,
including cortical material, fabric material and Alcantara, which are
associated with different prices, advantages and disadvantages.
- Car internal decoration (DR4): As a crucial part of the passenger
experience, it is an important innovative design objective, including
mats, perfume, accessories, interior decoration, storage box, etc.
- On-board system (DR5): It consists of GPS navigation, monitoring,
Linguistic scale for the alternatives Linguistic scale for the alternatives on
on Cost other criteria
Very Low (VL) Very Bad (VB)
Low (L) Bad (B)
A Little Low (ALL) A Little Bad (ALB)
Medium (M) Medium (M)
A Little High (ALH) A Little Good (ALG)
High (H) Good (G)
Very High (VH) Very Good (VG)
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Table 2
Assessments using HFLEs by User 1 and Expert 1.

CRs weights DRs

DR1 DR2 DR3 DR4 DR5 DR6

CR1 Between SI and P ES {s2} W {s−1} M {s0} N {s−3} W {s−1} M {s0}

{s2, s3}
CR2 U {s−1} EW {s−2} N {s−3} S {s1} Between M and S{s0, S {s1} Between M and S{s0,
s1} s1}
CR3 I {s1} ES {s2} ES {s2} EW{s−2} N {s−3} N {s−3} N {s−3}
CR4 I {s1} W {s−1} N {s−3} N {s−3} N {s−3} N {s−3} N {s−3}
CR5 SU {s−2} N {s−3} N {s−3} N {s−3} M {s0} N {s−3} N {s−3}

Alternatives
A1 L {s−2} G {s2} ALB {s−1} Between ALB and M ALG {s1} ALG {s1}
{s−1, s0}
A2 Between L and ALL Between M and ALG ALB {s−1} Between ALB and B Between ALB and M Between ALB and M
{s−2, s−1} {s0, s1} {s−2, s−1} {s−1, s0} {s−1, s0}
A3 Between H and VH M {s0} Between ALG and G ALB {s−1} G {s2} G {s2}
{s2, s3} {s1, s2}
A4 Between ALH and H Between B and ALB ALG {s1} Between ALG and G Between ALB and B Between B and ALB
{s1, s2} {s−2, s−1} {s1, s2} {s−2, s−1} {s−2, s−1}
A5 H {s2} ALG {s1} Between ALG and G ALB {s−1} ALG {s1} ALG {s1}
{s1, s2}

Bluetooth and so on. Especially, the navigation and positioning DR1 DR2 DR3 DR 4 DR5 DR6
capabilities for the “shared cars” are extremely important. CR1 ⎡ {s2 (1)} {s−1 (0.8)} {s0 (1)} {s−3 (1)} {s−1 (0.8)} {s0 (1)} ⎤
- Air-conditioning system (DR6): It aims to realize the refrigeration, CR2 ⎢ {s−2 (1)} {s−3 (1)} {s1 (1)} {s0 (0.3), s1 (0.7)} {s1 (1)} {s0 (0.4), s1 (0.6)} ⎥
⎢ ⎥
R = CR3 ⎢ {s2 (1)} {s2 (1)} {s−2 (1)} {s−3 (1)} {s−3 (1)} {s−3 (1)} ⎥
heating, ventilation and air purification of the air inside the car, CR 4 ⎢ {s−1 (1)} {s−3 (1)} {s−3 (1)} {s−3 (1)} {s1 (1)} {s−3 (1)} ⎥
which can provide a comfortable ride environment for customers. ⎢
CR5 {s−3 (1)} {s−3 (1)} {s−3 (1)} ⎥
⎣ {s0 (1)} {s−3 (1)} {s−3 (1)} ⎦

Then, twenty users fz (z = 1, 2, ...,20) , including ten lead users and

ten ordinary users with the same weight, are interviewed about the and the judgment matrix as

DR1 DR2 DR3 DR 4 DR5 DR6

A1 ⎡ {s−3 (0.3), s−2 (0.7)} {s2 (1)} {s−1 (1)} {s−1 (0.25), s0 (0.75)} {s1 (1)} {s1 (1)} ⎤
A2 ⎢ {s−2 (0.1), s−1 (0.9)} {s0 (0.3), s1 (0.7)} {s−1 (1)} {s−2 (0.2), s−1 (0.8)} {s−1 (0.85), s0 (0.15)} {s−1 (0.7), s0 (0.3)} ⎥
⎢ ⎥
J = A3 ⎢ {s2 (0.4), s3 (0.6)} {s0 (0.5), s1 (0.5)} {s1 (0.7), s2 (0.3)} {s−1 (1)} {s1 (0.9), s2 (0.1)} {s1 (0.8), s2 (0.2)} ⎥
A 4 ⎢ {s1 (0.8), s2 (0.2)} {s−2 (0.3), s−1 (0.7)} {s0 (0.8), s1 (0.2)} {s1 (0.4), s2 (0.6)} {s−2 (0.3), s−1 (0.7)} {s−1 (0.7), s0 (0.3)} ⎥
A5 ⎢ {s2 (0.5), s3 (0.5)} {s0 (0.7), s1 (0.3)} {s1 (0.8), s2 (0.2)} {s−1 (0.8)} {s1 (1)}
⎥
{s1 (0.8), s2 (0.2)} ⎦
⎣

importance of the CRs. Five experts eq (q = 1, 2. ..,5), including one Step 3. According to the CRs’ weight vector WCR and the matrix R,
product development manager, two product designers, one sales man- we compute the DRs’ weights by Eq. (31). The results are
ager and one salesperson with the same weight, are invited to assess the
impact degrees of the DRs over the CRs and the merits of alternatives ω1 = {s2 (0.27), s2.4 (0.33), s2.5 (0.09), s2.7 (0.11)},
over the DRs. A questionnaire is conducted to collect the individuals’ ω2 = {s−0.1 (0.22), s0.1 (0.07), s1.1 (0.26), s1.2 (0.09)},
opinions. The linguistic scales they used on different objects are shown ω3 = {s0.4 (0.27), s0.6 (0.33), s0.7 (0.09), s0.9 (0.11)},
in Table 1. The evaluation results are shown in Table 2 (Given the
ω4 = {s−1.6 (0.09), s−1.3 (0.21), s−1.1 (0.21), s−0.9 (0.49)}
limited space, we present only one user's and one expert's evaluation
information). ω5 = {s1.1 (0.6), s1.3 (0.2)},
ω6 = {s0.1 (0.3), s0.3 (0.45), s0.5 (0.1), s0.7 (0.15)}
5.2. Solve the case by the QFD-based PL-ORESTE method

Now we use Algorithm 4 to select the most excellent innovative Step 4. Using Eqs. (20) and (21), we find the maximum weight is ω1
shared car design for the “Panda” company. As Step 1 is given above, and the minimum weight is ω4. Similarly, the maximum PLTSs
we start our calculation from Step 2. under different DRs are {s−3 (0.3), s−2 (0.7)}, {s2(1)}, {s1(0.7), s2(0.3)},
{s1(0.4), s2(0.6)}, {s1(0.9), s2(0.1)}, {s1(0.8), s2(0.2)}, respectively,
Step 2. All individuals’ evaluations are aggregated to the while the minimum PLTSs under different DRs are
group opinions expressed as PLTSs. By Eqs. (10)–(12), we {s2 (0.4), s3 (0.6)}, {s−2 (0.3), s−1 (0.7)}, {s−1 (1)},{s−2 (0.2),s−1 (0.8)},
get the CRs’ weight vector as WCR = {s−2 (0.3), s−1 (0.7)},{s−1 (0.7,s0 (0.3)} , respectively. Calculate the dis-
({s2 (0.75), s3 (0.25)}, {s−1 (1)}, {s−1 (0.36), s1 (0.44)}, {s1 (1)} , the relation tances between the DRs’ weights, d(ω1, ωj) (j = 1, 2, ...,6) , by
Eq. (17). The result is d+={0, 0.162, 0.130, 0.243, 0.108,0.151}. Using
, {s−2 (0.3), s−1 (0.7)})T
the formula d (ω+, ωj )/ d (ω+, ω−) , we can calculate the normalized
matrix R of the CRs over the DRs as
distances are d1 = 0 , d2 = 0.667 , d3 = 0.535, d4 = 1, d5 = 0.444 ,
d6 = 0.621. Table 3 presents the normalized distances of dij.
Step 5. The probabilistic linguistic global preference scores Dij

23
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Table 3 Step 3. Compute the DRs’ weights by Eq. (31) based on the opera-
The normalized distances of dij. tions of HFLEs [9]. The hesitant fuzzy linguistic weights of DRs are
∼ = {s , s } , ω
ω ∼ = {s , s } , ω ∼ = {s , s } , ω
∼ = {s , s , s } ,
1 2.2 2.4 2 0.5 0.6 3 0.5 0.7 4 −1.3 −1.1 −0.9
Distance DR1 DR2 DR3 DR4 DR5 DR6 ∼ ∼
ω5 = {s1.1, s1.2} , ω6 = {s0.1, s0.3, s0.5} .
A1 0 0 1 0.683 0.143 0.233 Step 4. Let S = {sα α = −τ , ...,−1, 0, 1, ...,τ } be a LTS,
A2 0.346 0.412 1 1 0.895 1 ∼ = h j = s j s j ∈ S, l = 1, …, L be the hesitant fuzzy linguistic
ω j S { φl φl j}
A3 1 0.473 0 0.96 0 0
weights of DRj. The crisp weights of DRs can be calculated based on
A4 0.671 1 0.486 0 1 1
A5 0.963 0.527 0.301 0.889 0.143 0
the expected values of the hesitant fuzzy linguistic weights of the
DRs as

1 Lj τ + φl j ⎞ n ⎛1 Lj τ + φl j ⎞
Table 4 ωj = ⎜⎛ ∑l=1 ⎟/∑ ⎜ ∑l=1 ⎟ , j = 1, 2, ...,n
2τ ⎠ j=1 L 2τ ⎠
⎝ Lj
The probabilistic linguistic global preference scores.
⎝ j
Global score DR1 DR2 DR3 DR4 DR5 DR6 (34)

A1 0 0.472 0.762 0.856 0.330 0.469

A2 0.245 0.554 0.802 1 0.706 0.832 Then, we obtain ω1 = 0.26, ω2 = 0.17 , ω3 = 0.17 , ω4 = 0.09, ω5 = 0.2
A3 0.707 0.578 0.378 0.980 0.314 0.439 and ω6 = 0.11.
A4 0.474 0.850 0.511 0.707 0.774 0.832
Step 5. The overall utility values of Ai (i = 1, 2, 3, 4, 5) obtained by
A5 0.681 0.601 0.434 0.946 0.330 0.439
the WA operator of HFLEs as Eq. (14) are hS (A1 ) = {s1.1, s1.2, s1.3} ,
hS (A2 ) = {s−0.3, s−0.2, s−0.1, s0} ,
(i = 1, 2, ...,5; j = 1, 2, ...,6) are computed by Eq. (23) and listed in hS (A3 ) = {s−0.2, s−0.1, s0, s0.1, s0.2, s0.3, s0.4, s0.5} ,
Table 4. hS (A 4 ) = {s−0.8, s−0.7, s−0.6, s−0.5, s−0.4} , hS (A5 ) = {s−0.3, s−0.2, s−0.1, s0} .
Step 6. Using Eq. (24) to obtain the preference score of each al- Step 6. The ranking of the Panda's “shared cars” innovative designs
ternative, we have D1 = 0.4815, D2 = 0.6898, D3 = 0.5660 , based on the expected function and variance function of HFLEs [21]
D4 = 0.6913, D5 = 0.5718. Thus, the weak ranking of the alternatives is A1 ≻A3 ≻A2 = A5 ≻A 4 .
is A1≻A3≻A5≻A2≻A4 .
Step 7. Based on Eqs. (25)–(27) and Table 4, the preference in- We can find that the results of the QFD-based PL-ORESTE method
tensity, the average preference intensity and the net preference in- and the HFL-QFD method combined with the WA operator of HFLEs are
tensity between two alternatives are computed. Table 5 presents the similar. Both of them hold that A1 is the only preferred alternative.
average preference intensities between pairwise alternatives. However, the latter method only distinguishes the preference and in-
Step 8. According to reasonable ranges of the indifference threshold difference relations between alternatives. Besides, information loss is
δ, we let δ = 0.05 in this case. Then μ = δ / n = 0.008, serious in the later method. This fact can be reflected in the following
ξ = δ/2 = 0.025. Since T (A3 , A5 ) − T (A5 , A3 ) = 0.006 < μ , T aspects: ① The original linguistic terms and their probability distribu-
(A3 , A5 ) = 0.016 < ξ and T (A5 , A3 ) = 0.01 < ξ , we have A3 I A5. tion are lost when aggregating the group opinions by the WA operator.
Since T (A2 , A 4 ) − T (A 4 , A2 ) = 0.002 < μ , T (A2 , A 4 ) = 0.099 > ξ Although the information is still expressed as HFLEs in the calculation
and T (A 4 , A2 ) = 0.097 > ξ , we deem A2 R A4. The PIR relations of process, the linguistic terms in a HFLE are very close to each other, that
the Panda's “shared cars” innovative designs are shown in Table 5. is to say, the uncertain and fuzzy information is lost to some extent. ②
According to Table 5, we get the strong ranking of the alternatives as The DR's fuzzy weights are refined as crisp weights. ③ When we fuse the
A1 ≻A3 = A5 ≻A2 = A 4 ( A3 = A5 represents A3 I A5; A2 = A 4 re- preference values of an alternative with respect to all criteria, the
presents A2 R A4). Thus, A1 is the only preferred alternative. standardization is ignored but it is essential.

5.3. Solve the case by the HFL-QFD method combined with the WA operator
of HFLEs
In this subsection, we compare our proposed method with the HFL-
QFD method combined with the WA operator of HFLEs.
Step 1. Same to Step 1 of the QFD-based PL-ORESTE method.
Step 2. The WA operator of HFLEs is employed to integrate the
users’ evaluations on the importance of the CRs, the experts’ judg-
ments on the impact degrees of the DRs over CRs and the merits of
the alternatives over the DRs. The collective group decision in-
formation expressed as HFLEs are shown in Table 6 (we take the
negation of HFLEs for the cost criteria DR1).
5.4. Solve the case by the classical ORESTE method
Below the classical ORESTE method given as Algorithm 2 is used to
rank the alternatives. According to the alternatives’ normalized dis-
tances in Table 3, we get the Besson's ranks rj(1, 2, ..., 5) of the criteria
for the importance degrees and the Besson's ranks r j (Ai )(i = 1, 2, ...,5) of
the alternatives with respect to each DR for their merits. These Besson's
ranks consist the decision data in ORESTE, shown in Table 7.
∼
The global preference scores D (aij ) calculated by Eq. (2) and the
global weak ranking r(aij) are shown in Table 8 (let ς = 0.5).
According to Eq. (3), we get the weak ranking of the alternatives as
A1≻A3≻A5≻A2≻A4.
Based on Eqs. (4) and (5), we get the average preference intensities

Table 5
The average preference intensities between pairwise alternatives.

A1 A2 A3 A4 A5

T(A1, Ai) Relation T(A2, Ai) Relation T(A3, Ai) Relation T(A4, Ai) Relation T(A5, Ai) Relation

A1 – – 0 < 0.075 < 0.067 < 0.063 <

A2 0.212 I – – 0.205 > 0.097 R 0.199 >
A3 0.156 > 0.081 < – – 0.084 < 0.01 I
A4 0.277 > 0.099 R 0.210 > – – 0.194 >
A5 0.15 > 0.081 < 0.016 I 0.074 < – –

24
X. Wu, H. Liao Information Fusion 43 (2018) 13–26

Table 6
The collective group decision information derived from the WA operator.

CR CRs’ weights DR1 DR2 DR3 DR4 DR5 DR6

CR1 {s2, s2.5} {s2} {s−1} {s0} {s−3} {s−1} {s0}

CR2 {s−1} {s−2} {s−3} {s1} {s0.4, s1} {s1} {s0.2, s1}
CR3 {s0.1} {s2} {s2} {s−2} {s−3} {s−3} {s−3}
CR4 {s1} {s−1} {s−3} {s−3} {s−3} {s1} {s−3}
CR5 {s−1.6, s−1} {s−3} {s−3} {s−3} {s0} {s−3} {s−3}

Alternatives
A1 {s3, s2.4} {s2} {s−1} {s−0.4, s0} {s1} {s1}
A2 {s1.2, s1} {s0.4, s1} {s−1} {s−1.4, s−1} {s−1, s−0.8} {s−1, s−0.4}
A3 {s−2.2, s−3} {s0, s1} {s1, s1.6} {s−1} {s1, s1.2} {s1, s1.4}
A4 {s−1, s−1.4} {s−1.6, s−1} {s0, s0.4} {s1.2, s2} {s−1.6, s−1} {s−1, s−0.4}
A5 {s−2.4, s−2.6} {s0, s0.6} {s1, s1.4} {s−1} {s1} {s1, s1.4}

Table 7 Table 9
The decision data of classical ORESTE. The average preference intensities between pairwise innovative designs of the ORESTE
method.
Besson's ranks DR1 DR2 DR3 DR4 DR5 DR6
1 5 3 6 2 4 Intensity A1 A2 A3 A4 A5

A1 1 1 4.5 2 2.5 3 A1 – 0 0.149 0.073 0.097

A2 2 2 4.5 5 4 4.5 A2 0.264 – 0.260 0.108 0.222
A3 5 3 1 4 1 1.5 A3 0.167 0.115 – 0.111 0.059
A4 3 5 3 1 5 4.5 A4 0.247 0.118 0.267 – 0.181
A5 4 4 2 3 2.5 1.5 A5 0.142 0.104 0.087 0.052 –

Table 8
The global preference scores and the global weak rankings of the classical ORESTE method.

DR1 DR2 DR3 DR4 DR5 DR6

Score Global ranking Score Global ranking Score Global ranking Score Global ranking Score Global ranking Score Global ranking

A1 1 1 3.606 15.5 3.824 19.5 4.472 25 2.264 6.5 3.536 14

A2 1.581 2.5 3.808 17.5 3.824 19.5 5.523 30 3.162 13 4.257 22.5
A3 3.606 15.5 4.123 21 2.236 4.5 5.099 29 1.581 2.5 3.021 11.5
A4 2.236 4.5 5 28 3 10 4.301 24 3.808 17.5 4.257 22.5
A5 2.915 9 4.528 26 2.550 8 4.743 27 2.264 6.5 3.021 11.5

between pairwise alternatives shown in Table 9.
∼ = 0.03, ξ͠ = 0.15 and λ = 2 in this paper.
Similar to Ref. [25], let μ
Thus, we can get the strong ranking of the alternatives is
A1 = A3 = A5 ≻A2 = A 4 with A1 R A3, A3 I A5 and A2 R A4, which implies
that both A1 and A3 can be selected as the preferred alternatives.
The ranking results of the QFD-based PL-ORESTE method and the
classical ORESTE is different that there are more incomparability re-
lations among the alternatives in ORESTE. The reasons for the differ-
ences are as follows: ① The major difference is that the decision in-
formation in ORESTE is consisted by Besson's ranks which results in
information loss seriously, while the PL-ORESTE method takes the
distances between PLTSs as decision data, which reserves more in-
formation as illustrated in Examples 5 and 7. ② Translating the global
preference scores to the global Besson's ranks leads to the loss of in-
formation again in ORESTE. To avoid this defect, the PL-ORESTE
method derives the weak ranking and calculates the preference in-
tensity between pairwise alternatives based on the global preference
score directly. ③ It is hard to determine the values of thresholds in
ORESTE. This increases the difficulty of decision making. In PL-OR-
ESTE, we derive the values of thresholds based on the defined prob-
abilistic linguistic indifference threshold, and it is easy to select the
values of thresholds when making decision.
6. Conclusions
It is critical to select an optimal design that maximize user
satisfaction for companies to take the market quickly. This is a special
MEMCDM problem with imprecise information as well as multiple users
and more than one experts. To settle the problem, we have proposed a
QFD-based PL-ORESTE method by integrating three techniques: PLTS,
QFD and ORESTE. Firstly, to depict the group opinions effectively, we
have given a rational formula to calculate the weights of the linguistic
terms in the collective PLTS, and a novel distance measure between
PLTSs has been proposed to keep the operation results more reasonable.
In addition, given that ORESTE is an excellent ranking method with no
requirement of crisp criteria weights, we have extended the ORESTE
into probabilistic linguistic context to develop a new PL-ORESTE
method. Since QFD has been a successful method to translate the
weights of CRs into those of DRs based on which the design is easy to
make evaluations, we have introduced QFD into the PL-ORESTE
method and proposed a QFD-based PL-ORESTE method to evaluate the
design alternatives based on CRs . Finally, we have given a case study
about selecting innovative design of shared cars for the users of Panda
shared cars. This case have proven the feasibility and practicability of
the proposed method. It also has practical reference significance for
other shared car enterprises.
For further research, on the theoretic level, other ranking methods,
such as ELECTRE, VIKOR and TOPSIS, can be used to combine with
PLTSs to solve different kinds of MEMCDM problems. Furthermore, we
can demonstrate the applicability of the QFD-based PL-ORESTE method
by comparing it with other MEMCDM methods through simulated
scenarios. On the practice application level, we can further establish a

25
X. Wu, H. Liao
systematic CRs and DRs for shared cars innovative design selection.
Also, we can employed the QFD-based PL-ORESTE method to cope with
other innovative product design selection problems, such as computer
design, refrigerator design, etc., to meet customers’ satisfaction. We can
also use it in other fields such as supplier selection.
Acknowledgments
The authors would like to thank the editors and anonymous re-
viewers for their insightful and constructive commendations that have
led to an improved version of this paper. The work was supported by
the National Natural Science Foundation of China (71501135,
71771156), the China Postdoctoral Science Foundation (2016T90863,
2016M602698), the Scientific Research Foundation for Excellent Young
Scholars at Sichuan University(No. 2016SCU04A23), and the
International Visiting Program for Excellent Young Scholars of SCU.
References
[1] Y. Akao, Quality function deployment: integrating customer requirements into
product design, Brain Res. 498 (1) (2004) 135–139.
[2] M. Bevilacqua, F.E. Ciarapica, G. Giacchetta, A fuzzy-QFD approach to supplier
selection, J. Purchasing Supply Manage. 12 (1) (2006) 14–27.
[3] Y. Chen, R.Y.K. Fung, J. Tang, Rating technical attributes in fuzzy QFD by in-
tegrating fuzzy weighted average method and fuzzy expected value operator, Eur. J.
Oper. Res. 174 (3) (2006) 1553–1566.
[4] L.H. Chen, M.C. Weng, An evaluation approach to engineering design in QFD
processes using fuzzy goal programming models, Eur. J. Oper. Res. 172 (1) (2006)
230–248.
[5] C. Delhaye, J. Teghem, P. Kunsch, Application of the ORESTE method to a nuclear
waste management problem, Int. J. Prod. Econ. 24 (1-2) (1991) 29–39.
[6] M. Dursun, E.E. Karsak, An Integrated Approach Based On 2-Tuple Fuzzy
Representation and QFD for Supplier Selection 247 Springer, Netherlands, 2014,
pp. 621–634.
[7] M. Fasanghari, M. Mohamedpour, M.A. Mohamedpour, A novel method combining
ORESTE, fuzzy set theory, and TOPSIS method for ranking the information and
communication technology research centers of Iran, International Conference on
Information Technology: New Generations, 2009, pp. 165–170.
[8] X.J. Gou, H.C. Liao, Z.S. Xu, F. Herrera, Double hierarchy hesitant fuzzy linguistic
term set and MULTIMOORA method: a case of study to evaluate the implementation
status of haze controlling measures, Inf. Fusion 38 (2017) 22–34.
[9] X.J. Gou, Z.S. Xu, Novel basic operational laws for linguistic terms, hesitant fuzzy
linguistic term sets and probabilistic linguistic term sets, Inf. Sci. 372 (2016)
407–427.
[10] A.N. Haq, V. Boddu, Analysis of enablers for the implementation of leagile supply
chain management using an integrated fuzzy QFD approach, J. Intell. Manuf. 28
(2017) 1–12.
[11] J.R. Hauser, D. Clausing, The house of quality, Harvard Bus. Rev. 66 (3) (1988)
63–73.
[12] G.V. Huylenbroeck, The conflict analysis method: bridging the gap between
ELECTRE, PROMETHEE and ORESTE, Eur. J. Oper. Res. 82 (3) (1995) 490–502.
[13] W.C. Ko, Construction of house of quality for new product planning: a 2-tuple fuzzy
linguistic approach, Comput. Ind. 73 (C) (2015) 117–127.
[14] A.H.I. Lee, H.Y. Kang, C.Y. Yang, C.Y. Lin, An evaluation framework for product
planning using FANP, QFD and multi-choice goal programming, Int. J. Prod. Res. 48
(13) (2010) 3977–3997.
[15] I.D. Leeneer, H. Pastijn, Selecting land mine detection strategies by means of out-
ranking MCDM techniques, Eur. J. Oper. Res. 139 (2) (2002) 327–338.
26
Information Fusion 43 (2018) 13–26
[16] M. Li, The extension of quality function deployment based on 2-Tuple linguistic
representation model for product design under multigranularity linguistic en-
vironment, Math. Probl. Eng. 3 ((2012) (2012) 939–955.
[17] M. Li, The method for product design selection with incomplete linguistic weight
information based on quality function deployment in a fuzzy environment, Math.
Probl. Eng. 7 (2013) (2013) 87–118.
[18] H.C. Liao, L.S. Jiang, Z.S. Xu, J.P. Xu, F. Herrera, A probabilistic linguistic linear
programming method in hesitant qualitative multiple criteria decision making, Inf.
Sci. 415-416 (2017) 341–355.
[19] H.C. Liao, Z.S. Xu, Approaches to manage hesitant fuzzy linguistic information
based on the cosine distance and similarity measures for HFLTSs and their appli-
cation in qualitative decision making, Expert Syst. Appl. 42 (12) (2015) 5328–5336.
[20] H.C. Liao, Z.S. Xu, X.J. Zeng, Distance and similarity measures for hesitant fuzzy
linguistic term sets and their application in multi-criteria decision making, Inf. Sci.
271 (2014) 125–142.
[21] H.C. Liao, Z.S. Xu, X.J. Zeng, Hesitant fuzzy linguistic VIKOR method and its ap-
plication in qualitative multiple criteria decision making, IEEE Trans. Fuzzy Syst. 23
(5) (2015) 1343–1355.
[22] H.C. Liao, Z.S. Xu, X.J. Zeng, J.M. Merigó, Qualitative decision making with cor-
relation coefficients of hesitant fuzzy linguistic term sets, Knowl.-Based Syst. 76
(2015) 127–138.
[23] S. Ç. Onar, G. Büyüközkan, B. Öztayşi, C. Kahraman, A new hesitant fuzzy QFD
approach: an application to computer workstation selection, Appl. Soft Comput. 46
(2016) 1–16.
[24] Q. Pang, H. Wang, Z. Xu, Probabilistic linguistic term sets in multi-attribute group
decision making, Inf. Sci. 369 (2016) 128–143.
[25] H. Pastijn, J. Leysen, Constructing an outranking relation with ORESTE, Math.
Comput. Modell. 12 (10-11) (1989) 1255–1268.
[26] R.M. Rodríguez, B. Bedregal, H. Bustince, Y.C. Dong, B. Farhadinia, C. Kahraman,
L. Martínez, V. Torra, Y.J. Xu, Z.S. Xu, F. Herrera, A position and perspective
analysis of hesitant fuzzy sets on information fusion in decision making. Towards
high quality progress, Inf. Fusion 29 (C) (2016) 89–97.
[27] R.M. Rodríguez, L. Martı́nez, F. Herrera, Hesitant fuzzy linguistic terms sets for
decision making, IEEE Trans. Fuzzy Syst. 20 (2012) 109–119.
[28] R.M. Rodríguez, L. Martı́nez, F. Herrera, A group decision making model dealing
with comparative linguistic expressions based on hesitant fuzzy linguistic term sets,
Inf. Sci. 241 (12) (2013) 28–42.
[29] M. Roubens, Preference relations an actions and criteria in multicriteria decision
making, Eur. J. Oper. Res. 10 (1) (1982) 51–55.
[30] C. Temponi, J. Yen, W.A. Tiao, House of quality: A fuzzy logic-based requirements
analysis, Eur. J. Oper. Res. 117 (2) (1999) 340–354.
[31] J.Q. Wang, J. Wang, Q.H. Chen, H.Y. Zhang, X.H. Chen, An outranking approach for
multi-criteria decision-making with hesitant fuzzy linguistic term sets, Inf. Sci. 280
(2014) 338–351.
[32] H. Wang, Extended hesitant fuzzy linguistic term sets and their aggregation in
group decision making, Int. J. Comput. Intell. Syst. 8 (1) (2015) 14–33.
[33] Z.S. Xu, Deviation measures of linguistic preference relations in group decision
making, Omega 33 (3) (2005) 249–254.
[34] Z.S. Xu, H. Wang, On the syntax and semantics of virtual linguistic terms for in-
formation fusion in decision making, Inf. Fusion 34 (2017) 43–48.
[35] F. Zhang, M. Yang, W. Liu, Using integrated quality function deployment and theory
of innovation problem solving approach for ergonomic product design, Comput.
Ind. Eng. 76 (1) (2014) 60–74.
[36] Z. Zhang, X. Chu, Fuzzy group decision-making for multi-format and multi-granu-
larity linguistic judgments in quality function deployment, Expert Syst. Appl. 36 (5)
(2009) 9150–9158.
[37] S. Zhang, Z.S. Xu, Y. He, Operations and integrations of probabilistic hesitant fuzzy
information in decision making, Inf. Fusion 38 (2017) 1–11.
[38] Y.X. Zhang, Z.S. Xu, H.C. Liao, A consensus process for group decision making with
probabilistic linguistic preference relations, Inf. Sci. 414 (2017) 260–275.
[39] Y. Zhang, Z. Xu, H. Wang, H. Liao, Consistency-based risk assessment with prob-
abilistic linguistic preference relation, Appl. Soft Comput. 49 (2016) 817–833.