Applied Soft Computing 24 (2014) 543–558

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

A solid transportation problem with type-2 fuzzy variables

Pei Liu, Lixing Yang ∗ , Li Wang, Shukai Li
State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing 100044, China

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

Article history: This paper focuses on generating the optimal solutions of the solid transportation problem under fuzzy
Received 27 April 2013 environment, in which the supply capacities, demands and transportation capacities are supposed to
Received in revised form 28 June 2014 be type-2 fuzzy variables due to the instinctive imprecision. In order to model the problem within the
Accepted 3 August 2014
framework of the credibility optimization, three types of new defuzzification criteria, i.e., optimistic value
Available online 12 August 2014
criterion, pessimistic value criterion and expected value criterion, are proposed for type-2 fuzzy variables.
Then, the multi-fold fuzzy solid transportation problem is reformulated as the chance-constrained pro-
Keywords:
gramming model with the least expected transportation cost. To solve the model, fuzzy simulation based
Fuzzy sets
Fuzzy possibility space
tabu search algorithm is designed to seek approximate optimal solutions. Numerical experiments are
Type-2 fuzzy variable implemented to illustrate the application and effectiveness of the proposed approaches.
Solid transportation problem © 2014 Elsevier B.V. All rights reserved.

1. Introduction

Traditional transportation problem (TP), which only considers two types of constraints associated with sources and destinations, is one
of the well-known network optimization problems. In practice, as the conveyance capacity is another critical parameter in transporta-
tion activities, numerous researchers have concentrated their studies on the solid transportation problem by simultaneously considering
restrictions of sources, destinations and conveyances. Clearly, the traditional transportation problem is a special case of the solid trans-
portation problem if only one type of conveyance is taken into account. With this concern, the solid transportation problem can be regarded
as a generalization of the traditional transportation problem (see Haley [1,2] for the differences between these two problems).
Due to the complex environment during the transportation activities, some critical parameters in the solid transportation problem are
always treated as uncertain variables to meet the practical situations. For instance, if one needs to make a transportation plan for the next
month, the supply capacity at each source, the demand at each destination and the conveyance capacity are often required to be estimated
by professional judgments or probability statistics because of no precise a priori information. In this case, it is more suitable to investigate
this problem by using fuzzy or stochastic optimization methodologies. (The relevant theoretical methods can be referred to probability
theory and fuzzy set theory, which can also be included in the framework of Granular Computing proposed by Pedrycz [23,24].) Up to
now, a variety of researches in literature have handled the solid transportation problems under the uncertain environments. For instance,
Jiménez and Verdegay [3] formulated a balanced fuzzy solid transportation problem by considering fuzzy supplies, fuzzy demands and
fuzzy conveyance capacities. Liu [4] investigated a non-balanced solid transportation problem in condition that the supply capacities,
demands, and conveyance capacities are all convex fuzzy numbers. Yang and Liu [5] proposed the fuzzy programming models for the fixed
charge solid transportation problem by assuming that the direct costs, fixed charges, supplies, demands and conveyance capacities are fuzzy
variables. Kundu et al. [6] discussed a multi-objective multi-item solid transportation problem with fuzzy coefficients. Ojha et al. [7] treated
the source availabilities, destination demands and conveyance capacities as different types of fuzzy numbers, such as general, trapezoidal
and triangular fuzzy numbers. Molla et al. [8] investigated a fixed charge solid transportation problem under a fuzzy environment and
employed three meta-heuristics to solve the problem. Yang and Feng [9] studied a bi-criteria solid transportation problem, where the
parameters are random variables, and designed a hybrid algorithm to generate an approximate optimal solution. Ojha et al. [10] studied a
stochastic discounted multi-objective solid transportation problem for breakable items by using analytical hierarchy process.
In reality, uncertainty may exist with multi-fold. For example, in estimating the membership degree of a fuzzy demand, it is not easy to
deduce a crisp membership function because of the instinctive imprecision. To handle this case, the type-2 fuzzy set theory was introduced

∗ Corresponding author. Tel.: +86 10 51683970.

E-mail address: lxyang@bjtu.edu.cn (L. Yang).

http://dx.doi.org/10.1016/j.asoc.2014.08.005
1568-4946/© 2014 Elsevier B.V. All rights reserved.
544 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

by Zadeh [11] in 1975, and further improved by Liu and Liu [12] in 2007 through using the credibility theory. As for the further development
of type-2 fuzzy theory, Qin et al. [13] presented three kinds of critical values for a regular fuzzy variable and proposed different methods
for the reduction of a type-2 fuzzy variable; Coupland and John [14] presented a new method based on geometric representations and
operations for defuzzifing a type-2 fuzzy set; Wu and Mendel [27] defined the concepts of uncertainty measures (including cardinality,
fuzziness, variance and skewness) for interval type-2 fuzzy sets; Kundu et al. [25] discussed two fixed charge transportation problems with
type-2 fuzzy variables; Greenfield et al. [26] provided a novel approach to improve the speed of defuzzification for discretized generalized
type-2 fuzzy sets.
For the solid transportation problem with fuzzy information, we have two motivations to explore this problem within the framework
of type-2 fuzzy set theory. Firstly, it is more general and commonsense to treat some critical parameters as type-2 fuzzy variables because
of the practical difficulties of determining their crisp membership functions. Secondly, when some parameters are assumed to be type-
2 fuzzy variables, designing an effective method to handle the optimization problem is also a challenging issue. With this concern, we
are particularly interested in how to formulate the transportation model, and then design effective algorithms to produce the optimal
transportation strategies. To this end, this study proposes three new defuzziness methods for type-2 fuzzy variables based on the credibility
measure, and then formulates the problem as a chance-constrained expected value programming model. Fuzzy simulation based tabu
search algorithm is also designed to seek an approximate optimal solution for the model. Numerical experiments show the application
and efficiency of the proposed approaches.
This paper is organized as follows. In Section “Problem statements”, we present a detailed description for the solid transportation
problem. In Section “Type-2 fuzzy variables”, we introduce some fundamental knowledge with respect to type-2 fuzzy variables. Then, three
types of defuzzification methods are proposed for type-2 fuzzy variables in Section “Defuzzification methods for type-2 fuzzy variables”,
where we deduce the possibility distributions of the reduced type-1 fuzzy parameters for type-2 trapezoidal fuzzy variables. Section
“Reformulation of the solid transportation problem with type-2 fuzzy variables” reformulates the solid transportation problem with type-
2 fuzzy parameters as a chance-constrained expected value model. In Section “Numerical experiments”, several numerical examples are
implemented to illustrate the efficiency of the model and algorithm. Finally, a conclusion is made in the last section.

2. Problem statements

For the completeness of this research, we shall give a detailed description for the solid transportation problem (STP) in this section. As
an extension of the traditional transportation problem which only considers two factors (i.e., source and destination), the STP is related to
how to transport products from sources to destinations by conveyances so as to minimize the total transportation cost. For any STP, when
the total supply of products, the total demand of products and the total transportation capacity of conveyances are equal to each other, we
call such a problem as a balanced solid transportation problem. However, when the balance condition cannot be established, we have to
investigate the non-balanced solid transportation problem, in which it is required that the total available resources and the total capacities
of conveyances are not less than the total demands of destinations. The non-balanced solid transportation problem can be formulated as
follows:
I J K
min Z = i=1 j=1
c x
k=1 ijk ijk
J K
st. j=1
x
k=1 ijk
≤ ai , i = 1, 2, ..., I
I K
x ≥bj , j = 1, 2, ..., J (1)
i=1 k=1 ijk
I J
i=1
x
j=1 ijk
≤ ek , k = 1, 2, ..., K

xijk ≥0, ∀ i, j, k,

where the meanings of notations are listed as follows: i(= 1, 2, . . . I) is the index of source; j(= 1, 2, . . . J) is the index of destination; k(= 1, 2,
. . . K) is the index of conveyance; ai is the total supply of products in source i; bj is the total demand of products in destination j; ek is the
total transportation capacity of conveyance k; cijk is the unit transportation cost from source i to destination j by conveyance k; xijk is the
decision variable, which represents the transportation quantity from source i to destination j by conveyance k.
In this model, the first constraint states that the total amount of products transported from source i are no more than the total supply
in source i; the second constraint implies that the total amount of products transported from different sources should be greater than the
demands of destination j; the third constraint requires that the total amount of products transported from different sources to different
destinations by conveyance k are not greater than its transportation capacity. In addition, an implicit non-balance constraint for this model
I J K J
is a ≥ j=1 bj and
i=1 i
e ≥ j=1 bj . Apparently, the formulated non-balanced solid transportation problem is a linear programming
k=1 k
model.
Fig. 1 is given to clearly illustrate the transportation process in solid transportation problems, in which the products are required to be
transported from two sources to three destinations by three types of conveyances. Thus, the aim of this problem is to generate an optimal
transportation plan from source nodes to destination nodes by different conveyances (i.e., train, truck and ship). In transportation planning
process, three constraints need to be considered. (1) For destination j (say node 1), the total amount of products transported from source 1
and source 2 by trains and ships should be greater than the demands of this node. (2) For source i (say node 1), the total amount of products,
which are transported to three destinations by three types of conveyances, should be less than the supply capacities of this source. (3) For
conveyance trains, when we transport products from sources 1 and 2 to destinations 1 and 3, the total amount of products should not be
beyond the transportation capacities of trains, and the similar constraints are also applicable to ships and trucks.
The above mathematical model is formulated with certain parameters taking fixed values, which can be effectively solved by some
heuristics, for instance, tabu search algorithm (Yang and Feng [9]), analytical hierarchy process (Ojha et al. [10]), genetic algorithm (Vignaux
and Michalewicz [15]), etc. However, in reality, it is difficult to figure out the crisp values of some parameters due to the complexity of
the decision environment and the incompleteness of a prior information. In this case, the relevant parameters can be endowed with
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 545

Sources Destinations
1 train
truck
1 ship
2

3
Fig. 1. An illustration of solid transportation problem.

uncertain characteristics, and stochastic or fuzzy methods can be expectedly adopted to handle this problem effectively. As for stochastic
optimization, it is often required that the sample size (historical data) is sufficiently large to estimate the probability distributions. If
the sample size is too small (or even no sample is available) owing to information incompleteness, one may turn to the fuzzy theory
by estimating their membership functions. Practically, it is widely recognized that fuzziness often occurs with multi-fold forms in real-
world decision environments. Then, in this study we are particularly interested in the formulation and solution methods for the solid
transportation problem when some parameters are considered as type-2 fuzzy variables. In order to effectively formulate the non-balanced
STP under two-fold fuzzy environment, we firstly introduce some fundamental knowledge with regard to type-2 fuzzy variables and some
defuzzification methods in Sections “Type-2 fuzzy variables” and “Defuzzification methods for type-2 fuzzy variables”.

3. Type-2 fuzzy variables

As an extension of the traditional fuzzy set theory, Zadeh [11] introduced the concept of type-2 fuzzy sets in 1975, in which a type-2
fuzzy variable is described by the fuzzy membership function. To set up an axiomatic foundation for type-2 fuzzy theory, Liu and Liu [12,16]
defined the concept of fuzzy possibility space and proposed different defuzzification methods. The following discussion is mainly based
on the results proposed by Liu and Liu [12].

3.1. Possibility space

Hereinafter, we firstly introduce the concepts of possibility space, fuzzy variables and credibility measure. Let  be a nonempty set,
P() the power set of  and Pos a mapping from P() to [0,1]. A triplet (, P(), Pos) is called a possibility space if it satisfies the following
conditions:

(I) Pos {} =1 and Pos {˚} =0.

(II) For any {Ai |i ∈ I} ⊂ P(), Pos{ ∪ Ai } = supPos{Ai }.
i∈I i∈I

A fuzzy variable  is defined as a function from the possibility space (, P(), Pos) to the set of real numbers to describe fuzzy
phenomena, where the possibility measure (Pos) of fuzzy event { ∈ B} is defined by

Pos{ ∈ B} = sup (x).

x∈B

Here  (x) represents the possibility distribution (membership function) of .

The necessity measure (Nec) is defined as

Nec{ ∈ B} = 1 − Pos{ ∈ Bc } = 1 − sup  (x),

x ∈ Bc

i.e., the necessity of fuzzy event { ∈ B} is defined as the impossibility of the opposite event, and the credibility measure (Liu and Liu
[17]) of { ∈ B} is defined as
1
Cr{ ∈ B} = (Pos{ ∈ B} + Nec{ ∈ B}).
2

3.2. Regular fuzzy variable

If a fuzzy variable with the membership function  (x) satisfies sup (x) = 1, we call this fuzzy variable as a regular fuzzy variable
x∈
(RFV). In Liu and Liu [16], they also give a formal definition of RFV on interval [0,1].

Definition 3.2.1. (Liu and Liu [16]) Let (, P(), Pos) be a possibility space. A regular fuzzy variable  on interval [0,1] is defined as a
measurable mapping from  to the space [0, 1] such that for every t ∈ [0, 1],

{ ∈ |() ≤ t} ∈ P().
546 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

If the possibility distribution of  is as follows:

 
r1 r2 r3 ··· rn
∼ ,
1 2 3 ··· n

where ri ∈ [0, 1], i > 0 for each i = 1, 2, ..., n and maxni=1 i = 1, then  is a discrete RFV.
If  = [r1 , r2 ] with 0 ≤ r1 < r2 ≤ 1, then  is an interval RFV on interval [0,1].
If  = (r1 , r2 , r3 ) with 0 ≤ r1 < r2 < r3 ≤ 1, then  is a triangular RFV on interval [0,1].
If  = (r1 , r2 , r3 , r4 ) with 0 ≤ r1 < r2 < r3 < r4 ≤ 1, then  is a trapezoidal RFV on interval [0,1].
This paper focuses the discussion on regular fuzzy variables, but for an irregular fuzzy variable, we can firstly transform it into a RFV.
Here, we take the discrete irregular fuzzy variable as an example to illustrate the regularization method.

• Regularization method:
Let  be a discrete fuzzy variable,
 
r1 r2 r3 ... rn
∼ .
1 2 3 ... n

If maxni=1 i < 1, then  is called an irregular fuzzy variable. In order to convert an irregular fuzzy variable  into a regular fuzzy variable,
we give the following regularization steps:
• Step 1: select the maximum max from i (i = 1, 2, . . . n);
• Step 2: calculate regularization multiplier 1/max ;
• Step 3: multiply by 1/max for all i (i = 1, 2, . . ., n).

By using the above procedure, we finally produce the following possibility distribution, which is obviously a regular fuzzy variable:
  ⎛ ⎞
r1 r2 r3 ... rn r1 r2 r3 ... rn
⇒ ⎝ 1 2 3 n ⎠.
1 2 3 ... n ...
max max max max

We give an example to illustrate the process of regularization as follows:

Example 3.2.1. Let  be a discrete RFV with the following possibility distribution:

0.5 0.6 0.7 0.9

∼ .
0.2 0.4 0.8 0.6

Clearly, we have max4i=1 i = 0.8 < 1, i.e., max = 0.8. Thus, for each possibility measure, we multiply it by regularization multiplier
(1/max = 1/0.8) and produce the following possibility distribution. Then,  is transformed into a regular fuzzy variable  on interval [0,1].

0.5 0.6 0.7 0.9

∼ .
0.25 0.5 1 0.75

3.3. Fuzzy possibility space

In essence, fuzzy variables defined on the possibility space are referred to as type-1 fuzzy variables, in which the membership value of an
element reflects the degree belonging to this type-1 fuzzy variable, and moreover, the membership values are accurately predetermined.
However, as an extension of the type-1 fuzzy variable, the type-2 fuzzy variable aims to fuzzify membership values, where the membership
value is no longer an exact value but a type-1 regular fuzzy variable on [0,1]. The following discussion focuses on the introduction of type-2
fuzzy variables proposed by Liu and Liu [16].

Definition 3.3.1. (Liu and Liu [12]) Let P() be an ample field on the universe , F([0, 1]) the set of all regular fuzzy variables on [0,1],
and P̃os : P() → F([0, 1]) a set function on P() such that {P̃os(A)|atom A ∈ P()} is a family of mutually independent RFVs. We call P̃os
a fuzzy possibility measure if it satisfies the following conditions:

(P̃1) P̃os{˚} = 0̃;

(P̃2) for any subclass {Ai |i ∈ I} of P() (finite, countable or uncountable),

P̃os{ ∪ Ai } = supP̃os{Ai }.
i∈I i∈I

Moreover, if P̃os() (1) = 1, then we call P̃os a regular fuzzy possibility measure.

Definition 3.3.2. (Liu and Liu [12]) The triplet (, P(), P̃os) is referred to as a fuzzy possibility space (FPS), in which a mapping ˜ :  →
˜
is called a type-2 fuzzy variable if for any r∈ , the set { ∈ |() ˜
≤ r} is an element of P(), i.e., { ∈ |() ≤ r} ∈ P().
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 547

Definition 3.3.3. (Liu and Liu [16]) Let ˜ be a type-2 fuzzy variable defined on an FPS (, P(), P̃os). The secondary possibility distribution
˜ ˜ (x) of ˜ is defined as a mapping → F([0, 1]) such that 
function  ˜
˜ ˜ (x) = P̃os{ ∈ |() = x}, x ∈ , while the type-2 possibility distribution
˜
function  ˜ (x, u) of  is a mapping × Jx → [0, 1] such that


˜ (x, u) = Pos{

˜ ˜ (x) = u}, (x, u) ∈ × Jx , (2)

˜ ˜ (x), and Jx ⊂ [0, 1] is the support of 

where Pos is the possibility measure induced by the distribution of  ˜  (x), i.e., Jx = {u ∈ [0, 1]|˜ (x, u) >
0}.
The support of a type-2 fuzzy variable ˜ is denoted as

supp ˜ = {(x, u) ∈ × [0, 1]|˜ (x, u) > 0}. (3)

The type-2 triangular fuzzy variable is an important concept proposed by Liu and Liu [16], which has a secondary possibility distribution
depicted below.

Definition 3.3.4. (Liu and Liu [16]) If ˜ is a type-2 triangular fuzzy variable, denoted by ˜ = (r1 , r2 , r3 ; l , r ), the secondary possibility
distribution 
˜ ˜ (x) is defined as a type-1 regular triangular fuzzy variable, i.e.,

x − r1 x − r1 r2 − x x − r1 x − r1 x − r1 r2 − x

˜ ˜ (x) =
 − l min , , , + r min , (4)
r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1
for any x ∈ [r1 , r2 ], and
r3 − x r3 − x x − r2 r3 − x r3 − x r3 − x x − r2

˜ ˜ (x) =
 − l min , , , + r min , (5)
r3 − r2 r3 − r2 r3 − r2 r3 − r2 r3 − r2 r3 − r2 r3 − r2
for any x ∈ (r2 , r3 ], where  l ,  r ∈ [0, 1] are two parameters characterizing the degree of uncertainty of  taking the value x.
Motivated by the concept of type-2 triangular fuzzy variables, we especially extend the definition of the secondary possibility distribu-
tions for type-2 trapezoidal fuzzy variables in the following:

Definition 3.3.5. If ˜ is a type-2 trapezoidal fuzzy variable, denoted by ˜ = (r1 , r2 , r3 , r4 ; l , r ), the secondary possibility distribution 
˜ ˜ (x)
is defined as a type-1 regular triangular fuzzy variable, i.e.,
x − r1 x − r1 r2 − x x − r1 x − r1 x − r1 r2 − x

˜ ˜ (x) =
 − l min , , , + r min , (6)
r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1
for any x ∈ [r1 , r2 ], and for any x ∈ (r2 , r3 ], the secondary possibility distribution is 1, and
r4 − x r4 − x x − r3 r4 − x r4 − x r4 − x x − r3

˜ ˜ (x) =
 − l min , , , + r min , (7)
r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3
for any x ∈ (r3 , r4 ], where  l ,  r ∈ [0, 1] are two parameters characterizing the degree of uncertainty of  taking the value x.

Remark 3.3.1. Obviously, for a type-2 triangular fuzzy variable, the membership value of an element x in [r1 , r3 ] does not take a crisp
value but a regular triangular fuzzy variable. Likewise, for a type-2 trapezoidal fuzzy variable, the membership value of an element x in [r1 ,
r2 ] ∪ [r3 , r4 ] also takes a regular triangular fuzzy variable.

4. Defuzzification methods for type-2 fuzzy variables

For application purpose, some detailed defuzzification methods for type-2 fuzzy variables will be introduced in this section, which can
be conceived as a simplification process for two-fold uncertain information. Based on this, a type-2 fuzzy variable can be easily converted
into a type-1 fuzzy variable with the aid of reduction methods.

4.1. The critical values of regular fuzzy variables

In order to define the defuzzification methods for type-2 fuzzy variables, we firstly introduce three types of critical values (CVs) for a
regular fuzzy variable (RFV) by adopting the criteria proposed by Liu and Liu [17], Liu [18], including optimistic value criterion, pessimistic
value criterion and expected value criterion.
Let  be a RFV. Then,

(I) the optimistic value of , denoted by CV∗˛ [], is defined as:

CV∗˛ [] = sup{r|Cr{≥r}≥˛}; (8)

(II) the pessimistic value of , denoted by CV˛

∗ [], is defined as:

CV˛
∗ [] = inf {r|Cr{ ≤ r}≥˛}; (9)

(III) the expected value of , denoted by E[], is defined as:

 +∞  0
E[] = Cr{≥r}dr − Cr{ ≤ r}dr. (10)
0 −∞
548 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

Cr{ r} Cr{ r}

1 1
0.8
0.7
0.6
0.4
0.3
0.2

0 0.2 0.5 0.7 0.9 r 0 0.2 0.5 0.7 0.9 r

Fig. 2. Credibility functions of discrete fuzzy variable .

Cr{ r} Cr{ r}

1 1

0.5 0.5

0 r1 r2 r3 r r4 r 0 r1 r2 r3 r r4 r

Fig. 3. Credibility functions of trapezoidal fuzzy variable .

Next, an example is given here to clearly illustrate the different critical values defined above.

Example 4.1.1. Let  be a discrete RFV with the following possibility distribution:

0.2 0.5 0.7 0.9

∼ .
0.4 0.8 1 0.6

Fig. 2 gives the involved credibility functions with respect to credibility measure.
The optimistic value and pessimistic value of this fuzzy variable can be easily calculated by
⎧ ⎧
⎪
⎪ 0.9, 0 < ˛ ≤ 0.3 ⎪
⎪ 0.2, 0 < ˛ ≤ 0.2
⎨ 0.7, 0.3 < ˛ ≤ 0.6 ⎨ 0.5, 0.2 < ˛ ≤ 0.4
CV∗˛ [] = sup{r|Cr{≥r}≥˛} = CV˛
∗ [] = inf{r|Cr{ ≤ r}≥˛} =
⎪ 0.5, 0.6 < ˛ ≤ 0.8
⎪ ⎪ 0.7, 0.4 < ˛ ≤ 0.7
⎪
⎩ ⎩
0.2, 0.8 < ˛ ≤ 1. 0.9, 0.7 < ˛ ≤ 1.

Additionally, the expected value is calculated by

4
E[] = ωi ri = 0.2 × 0.2 + 0.2 × 0.5 + 0.3 × 0.7 + 0.3 × 0.9 = 0.62.
i=1

In the following, we shall deduce general formulations of critical values for trapezoidal fuzzy variables.

Example 4.1.2. Let  be a trapezoidal RFV, denoted by  = (r1 , r2 , r3 , r4 ). Then, according to the membership function  (x), we have the
involved credibility functions in Fig. 3:

The optimistic value of  is


2˛r3 + (1 − 2˛)r4 , 0 < ˛ ≤ 0.5,
CV∗˛ [] = sup{r|Cr{≥r}≥˛} = (11)
2(1 − ˛)r2 − (1 − 2˛)r1 , 0.5 < ˛ ≤ 1.

Similarly, the pessimistic value of  is


(1 − 2˛)r1 + 2˛r2 , 0 < ˛ ≤ 0.5,
CV˛
∗ [] = inf {r|Cr{ ≤ r}≥˛} = (12)
(2˛ − 1)r4 + 2(1 − ˛)r3 , 0.5 < ˛ ≤ 1,
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 549

and the expected value of  is

1
E[] = (r1 + r2 + r3 + r4 ). (13)
4
Remark 4.1.1. Typically, interval fuzzy variables and triangular fuzzy variables are special forms of trapezoidal fuzzy variables. According
to Example 4.1.2, the critical values of interval fuzzy variables and triangular fuzzy variables can be easily inferred.

4.2. Possibility distributions of generated type-1 fuzzy variables

Obviously, type-2 fuzzy variables can be used to characterize two-fold forms in realistic decision environment. However, one of the
core problems is that the computational complexity will increase when some parameters are regarded as type-2 fuzzy variables in the
optimization models. Because of this, we shall focus on converting type-2 fuzzy variables into type-1 fuzzy variables with the aid of
defuzzification methods proposed in Section “The critical values of regular fuzzy variables”. That is, the fuzzy membership function will be
represented by its critical values. The following discussion is mainly on the possibility distributions of the deduced type-1 fuzzy variables.
Firstly, for a type-2 trapezoidal fuzzy variable, we have the following results.

Theorem 4.2.1. Let ˜ = (r1 , r2 , r3 , r4 ; l , r ) be a type-2 trapezoidal fuzzy variable. We assume that ˛∗ , ∗˛ and  E , respectively, are deduced
type-1 fuzzy variables by the optimistic value criterion, pessimistic value criterion and expected value criterion. We then have

(I) type-1 fuzzy variable ˛∗ deduced by the optimistic value criterion has the following possibility distributions:
⎧ (x − r )(1 +  − 2˛ )  r +r 
⎪
⎪
1 r r
, x ∈ r1 ,
1 2
⎪
⎪ r2 − r1 2
⎪
⎪
⎪
⎪ (x − r1 ) + (1 − 2˛)r (r2 − x)
⎪
⎪ r1 + r2
⎪
⎪ , x∈( , r2 ]
⎨ r2 − r1 2
∗ (x) = 1, x ∈ (r , r ] 0 < ˛ ≤ 0.5 ; (14)
˛
⎪
⎪
2 3
⎪
⎪ − + − − r3 + r4
⎪ (r
⎪ 4
x) (1 2˛)r (x r3 )
, x ∈ (r3 ,
⎪
⎪ −
]
⎪
⎪
r4 r3 2
⎪
⎪
⎩ (r4 − x)(1 + r − 2˛r ) , x∈(
r3 + r4
, r4 ]
r4 − r3 2
⎧ (x − r )(1 +  − 2˛ )  r +r 
⎪
⎪
1 l l
, x ∈ r1 ,
1 2
⎪
⎪ r2 − r1 2
⎪
⎪
⎪
⎪
⎪
⎪ (x − r1 ) + (1 − 2˛)l (r2 − x) r1 + r2
⎪
⎪ , x∈( , r2 ]
⎨ r2 − r1 2
∗ (x) = 1, x ∈ (r , r ] 0.5 < ˛ ≤ 1 ; (15)
˛
⎪
⎪
2 3
⎪
⎪ (r4 − x) + (1 − 2˛)l (x − r3 ) r3 + r4
⎪
⎪ , x ∈ (r3 ,
⎪
⎪ r4 − r3
]
⎪
⎪
2
⎪
⎪
⎩ (r4 − x)(1 + l − 2˛l ) , x∈(
r3 + r4
, r4 ]
r4 − r3 2
(II) type-1 fuzzy variable ∗˛ deduced by the pessimistic value criterion has the following possibility distributions:
⎧
⎪ (x − r1 )(1 − l + 2˛l ) r1 + r2
⎪
⎪ , x ∈ [r1 , ]
⎪
⎪ r2 − r1 2
⎪
⎪
⎪
⎪ (x − r1 ) − (1 − 2˛)l (r2 − x) r1 + r2
⎪
⎪ , x∈( , r2 ]
⎪
⎨ r2 − r1 2
∗˛ (x) = 1, x ∈ (r2, r3 ] 0 < ˛ ≤ 0.5 , (16)
⎪
⎪
⎪
⎪ (r4 − x) − (1 − 2˛)l (x − r3 ) r3 + r4
⎪
⎪ , x ∈ (r3 , ]
⎪
⎪ r − r 2
⎪
⎪
4 3
⎪
⎪ (r4 − x)(1 − l + 2˛l ) , r3 + r4
⎩ x∈( , r4 ]
r4 − r3 2
⎧ (x − r )(1 −  + 2˛ )  r +r 
⎪
⎪
1 r r
, x ∈ r1 ,
1 2
⎪
⎪ r2 − r1 2
⎪
⎪
⎪
⎪
⎪
⎪ (x − r1 ) − (1 − 2˛)r (r2 − x) r1 + r2
⎪
⎪ , x∈( , r2 ]
⎨ r2 − r1 2
∗˛ (x) = 1, x ∈ (r , r ] 0.5 < ˛ ≤ 1 ; (17)
⎪
⎪
2 3
⎪
⎪
⎪ (r4 − x) − (1 − 2˛)r (x − r3 ) , x ∈ (r3 , r3 + r4 ]
⎪
⎪
⎪ r4 − r3
⎪
⎪
2
⎪
⎪ − − + +
⎩ (r4 x)(1 r 2˛r )
, x∈(
r3 r4
, r4 ]
r4 − r3 2
550 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

(III) type-1 fuzzy variable  E deduced by the expected value criterion has the following possibility distributions:
⎧ x−r 1 x − r1 r1 + r2
⎪
⎪
1
− (l − r ) , x ∈ [r1 , ]
⎪
⎪ r2 − r1 4 r2 − r1 2
⎪
⎪
⎪
⎪ x − r1 1 r2 − x r1 + r2
⎪
⎪ − (l − r ) , x∈( , r2 ]
⎪
⎨ r2 − r1 4 r2 − r1 2
E (x) = 1, x ∈ (r2 , r3 ] (18)
⎪
⎪
⎪
⎪ r4 − x 1 x − r3 r3 + r4
⎪
⎪ − (l − r ) , x ∈ (r3 , ]
⎪
⎪ r4 − r3 4 r4 − r3 2
⎪
⎪
⎪
⎩ r4 − x 1 r4 − x r3 + r4
− (l − r ) , x∈( , r4 ].
r4 − r3 4 r4 − r3 2

Remark 4.2.1. The theorem given above illustrates the possibility distributions of the deduced type-1 fuzzy variables with respect to
x ∈ [r1 , r2 ] and x ∈ (r3 , r4 ]. In addition, the secondary possibility distribution of a type-2 trapezoidal fuzzy variable is 1 for any x ∈ (r2 , r3 ]
according to the Definition 3.3.5. That is, for any x ∈ (r2 , r3 ], the possibility distributions of the deduced type-1 fuzzy variables via three
defuzzification methods are ∗ (x) = ∗˛ (x) = E (x) = 1, 0 < ˛ ≤ 1.
˛

We next give a detail proof of the possibility distributions of the deduced type-1 fuzzy variables.

Proof.

(I) According to the Definition 3.3.5, Example 4.1.2 and Remark 4.1.1, type-1 fuzzy variable ˛∗ has the following possibility distributions:

⎧ x − r1
 x − r1 x − r1 r2 − x
⎪
⎨ 2˛ r − r + (1 − 2˛) r − r + r min r − r , r − r , 0 < ˛ ≤ 0.5
2 1 2 1 2 1 2 1
∗ (x) = 
˛
⎪
⎩ x − r1 x − r1 x − r1 r2 − x
2(1 − ˛) − (1 − 2˛) − l min , , 0.5 < ˛ ≤ 1
r2 − r1 r2 − r1 r2 − r1 r2 − r1

for any x ∈ [r1 , r2 ], and

⎧ r4 − x
 r4 − x r4 − x x − r3
⎪
⎨ 2˛ r4 − r3 + (1 − 2˛) r4 − r3 + r min r4 − r3 , r4 − r3 , 0 < ˛ ≤ 0.5
˛
 (x) = 
1 ⎪
⎩ r4 − x r4 − x r4 − x x − r3
2(1 − ˛) − (1 − 2˛) − l min , , 0.5 < ˛ ≤ 1
r4 − r3 r4 − r3 r4 − r3 r4 − r3

for any x ∈ (r3 , r4 ].

(II) Type-1 fuzzy variable ∗˛ has the following possibility distributions:
⎧ x − r1
 x − r1 x − r1 r2 − x
⎪
⎨ 2˛ r − r + (1 − 2˛) r − r − l min r − r , r − r , 0 < ˛ ≤ 0.5
2 1 2 1 2 1 2 1
∗˛ (x) = 
⎪
⎩ x − r1 x − r1 x − r1 r2 − x
2(1 − ˛) − (1 − 2˛) + r min , , 0.5 < ˛ ≤ 1
r2 − r1 r2 − r1 r2 − r1 r2 − r1

for any x ∈ [r1 , r2 ], and

⎧ r4 − x
 r4 − x r4 − x x − r3
⎪
⎨ 2˛ r4 − r3 + (1 − 2˛) r4 − r3 − l min r4 − r3 , r4 − r3 , 0 < ˛ ≤ 0.5
∗˛ (x) = 
⎪
⎩ r4 − x r4 − x r4 − x x − r3
2(1 − ˛) − (1 − 2˛) + r min , , 0.5 < ˛ ≤ 1
r4 − r3 r4 − r3 r4 − r3 r4 − r3

for any x ∈ (r3 , r4 ].

(III) Type-1 fuzzy variable  E has the following possibility distributions:

1 x − r1 x − r1 r2 − x x − r1 x − r1 x − r1 x − r1 r2 − x
E (x) = − l min , + + + + r min ,
4 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1 r2 − r1

for any x ∈ [r1 , r2 ], and

1 r4 − x r4 − x x − r3 r4 − x r4 − x r4 − x r4 − x x − r3
E (x) = − l min , + + + + r min ,
4 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3 r4 − r3

for any x ∈ (r3 , r4 ].

After simplifying the above equations, we can prove Theorem 4.2.1.

Next, an example will be given to illustrate the calculation of the reduced type-1 fuzzy variables.
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 551

Fig. 4. Possibility distribution of ˛∗ .

Example 4.2.1. Let ˜ = (3, 4, 5, 6; l = 0.5, r = 1) be a type-2 trapezoidal fuzzy variable. By using Theorem 4.2.1, we infer the possibility
functions for each generated type-1 fuzzy variable as follows:

⎧
⎪
⎪ (x − 3)(2 − 2˛), x ∈ [3, 3.5]
⎪
⎪ (x − 3) + (1 − 2˛)(4 − x), x ∈ (3.5, 4]
⎨
∗ (x) = 1, x ∈ (4, 5] 0 < ˛ ≤ 0.5 ,
˛
⎪
⎪
⎪
⎪ (6 − x) + (1 − 2˛)(x − 5), x ∈ (5, 5.5]
⎩
(6 − x)(2 − 2˛), x ∈ (5.5, 6]

⎧
⎪
⎪ (x − 3)(1.5 − ˛), x ∈ [3, 3.5]
⎪
⎪ (x − 3) + (0.5 − ˛)(4 − x), x ∈ (3.5, 4]
⎨
∗ (x) = 1, x ∈ (4, 5] 0.5 < ˛ ≤ 1 ,
˛
⎪
⎪
⎪
⎪ (6 − x) + (0.5 − ˛)(x − 4), x ∈ (4, 4.5]
⎩
(6 − x)(1.5 − ˛), x ∈ (4.5, 5]

⎧
⎪
⎪ (x − 3)(0.5 + ˛), x ∈ [3, 3.5]
⎪
⎪
⎨ (x − 3) − (0.5 − ˛)(4 − x), x ∈ (3.5, 4]
∗˛ (x) = 1, x ∈ (4, 5] 0 < ˛ ≤ 0.5 ,
⎪
⎪
⎪
⎪ (6 − x) − (0.5 − ˛)(x − 5), x ∈ (5, 5.5]
⎩
(6 − x)(0.5 + ˛), x ∈ (5.5, 6]

⎧
⎪
⎪ 2˛(x − 3), x ∈ [3, 3.5]
⎪
⎪
⎨ (x − 3) − (1 − 2˛)(4 − x), x ∈ (3.5, 4]
∗˛ (x) = 1, x ∈ (4, 5] 0.5 < ˛ ≤ 1 ,
⎪
⎪
⎪
⎪ (6 − x) − (1 − 2˛)(x − 5), x ∈ (5, 5.5]
⎩
2˛(6 − x), x ∈ (5.5, 6]

⎧
⎪
⎪ (9/8)(x − 3), x ∈ [3, 3.5]
⎪
⎪
⎨ (7/8)x − (5/2), x ∈ (3.5, 4]
E (x) = 1, x ∈ (4, 5]
⎪
⎪
⎪
⎪ (43/8) − (7/8)x, x ∈ (5, 5.5]
⎩
(9/8)(6 − x), x ∈ (5.5, 6].

To understand the membership functions more straightforwardly, Figs. 4–6, respectively, are given to depict geometric graphs for the
obtained possibility distribution functions. Moreover, Fig. 7 is also given to depict the support of the type-2 fuzzy variable.

Remark 4.2.2. For simplicity, the above discussion only considers the type-2 trapezoidal fuzzy variable with the form ˜ =
(r1 , r2 , r3 , r4 ; l, r ). However, since the type-2 triangular fuzzy variable is in essence a special form of the type-2 trapezoidal fuzzy variable
when r2 = r3 , the deduced type-1 fuzzy variables by three reduction criteria for type-2 triangular fuzzy variables can also be calculated
similarly.
552 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

Fig. 5. Possibility distribution of ∗˛ .

Fig. 6. Possibility distribution of  E .

Fig. 7. Support of type-2 fuzzy variable.

Finally, we aim to show the reduction methods for type-2 discrete fuzzy variables. For this purpose, we consider the following type-2
discrete fuzzy variable:
⎧
⎨ 1, with possibility[0.1, 0.5],
˜ = 2, with possibility(0.2, 0.5, 0.6),
⎩
3, with possibility(0.1, 0.3, 0.6, 0.8).
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 553

Obviously, for this type-2 fuzzy variable, the possibility of each outcome takes a value represented by a type-1 fuzzy variable. That is,
˜ takes the value 1, 2 and 3 with fuzzy possibilities [0.1, 0.5], (0.2, 0.5, 0.6), (0.1, 0.3, 0.6, 0.8), respectively. By using the defuzzification
methods, we have
 
0.5, 0 < ˛ ≤ 0.5 0.1, 0 < ˛ ≤ 0.5
CV∗˛ [
˜ ˜ (1)] = , CV˛ ˜ ˜ (1)] =
∗ [ ,
0.1, 0.5 < ˛ ≤ 1 0.5, 0.5 < ˛ ≤ 1

 
0.6 − 0.2˛, 0 < ˛ ≤ 0.5 0.2 + 0.6˛, 0 < ˛ ≤ 0.5
CV∗˛ [
˜ ˜ (2)] = , CV˛
∗ [
˜ ˜ (2)] = ,
0.8 − 0.6˛, 0.5 < ˛ ≤ 1 0.4 + 0.2˛, 0.5 < ˛ ≤ 1

 
0.8 − 0.4˛, 0 < ˛ ≤ 0.5 0.1 + 0.4˛, 0 < ˛ ≤ 0.5
CV∗˛ [
˜ ˜ (3)] = , CV˛ ˜ ˜ (3)] =
∗ [ ,
0.5 − 0.4˛, 0.5 < ˛ ≤ 1 0.4 + 0.4˛, 0.5 < ˛ ≤ 1

˜ ˜ (1)] = 0.3,
E[ ˜ ˜ (2)] = 0.45,
E[ ˜ ˜ (3)] = 0.45.
E[

By the operations mentioned above, the generated type-1 fuzzy variables are not necessarily regular as the maximum value of possibility
measures for outcomes is not necessarily equal to 1. Then, we can employ the regularization method previously proposed to transform
an irregular fuzzy variable into a regular fuzzy variable. Take the type-1 fuzzy variable generated by the expected value criterion as an
example. Since the irregular fuzzy variable is

1 2 3
E = ,
0.3 0.45 0.45

we have max = 0.45 < 1 and regularization multiplier 1/max = 1/0.45. Then, the generated regular type-1 fuzzy variable turns out to be:

1 2 3
E = .
2/3 1 1

5. Reformulation of the solid transportation problem with type-2 fuzzy variables

By using the defuzzification methods proposed in Section “Defuzzification methods for type-2 fuzzy variables”, we shall reformulate
the solid transportation problem as a chance-constrained expected value programming model in the following.
In the solid transportation problem, due to the complexity of the decision environments and the incompleteness of a prior information,
it is not easy to precisely estimate the crisp values of some parameters. Then, in this study we assume that the supply capacities, demands,
conveyances capacities and unit transportation costs are all type-2 fuzzy variables pre-specified by professional judgments, denoted by ãi ,
b̃j , ẽk and c̃ijk , respectively. Then we have the following formulation:
I J K
min Z= i=1 j=1
c̃ x
k=1 ijk ijk
J K
st. j=1
x
k=1 ijk
≤ ãi , i = 1, 2, . . ., I
I K
x ≥b̃j , j = 1, 2, . . ., J (19)
i=1 k=1 ijk
I J
i=1
x
j=1 ijk
≤ ẽk , k = 1, 2, ..., K

xijk ≥0, ∀i, j, k.

Apparently, the formulation (19) is an uncertain model without specific mathematical meanings. Next, we are intended to use the
defuzzification criteria proposed above to generate the corresponding type-1 fuzzy variables, and then reformulate the model based on
the expected value operator of fuzzy variables and credibility measure. For simplicity, type-2 fuzzy variables c̃ijk , ãi , b̃j , ẽk are reduced into
type-1 variables c̄ijk , āi , b̄j , ēk , respectively.
In literature, the fuzzy optimization methods have been applied to a variety of real-world decision-making problems up to now (see
Liu and Liu [17], Yang et al. [19–22], Kundu et al. [25], Qin et al. [13]). Through minimizing the expected value of the total transportation
cost with chance constraints, we finally reformulate the solid transportation problem into the following fuzzy optimization model:
 J K 
I
min E[Z] = E i=1 j=1
c̄ x
k=1 ijk ijk
J K
st. Cr j=1
x
k=1 ijk
≤ āi ≥i , i = 1, 2, . . ., I
I K (20)
Cr i=1
x ≥b̄j
k=1 ijk
≥ˇj , j = 1, 2, . . ., J
I J
Cr i=1
x
j=1 ijk
≤ ēk ≥k , k = 1, 2, . . ., K

xijk ≥0, ∀i, j, k,

where i (i = 1, 2, . . ., I), ˇj (j = 1, 2, . . ., J) and  k (k = 1, 2, . . ., K) are predetermined credibility confidence levels. In this model, the objective
function is to minimize the expected total transportation cost. The chance constraints aim to require that the credibility measures of supply
554 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

capacity constraints, demand constraints and transportation capacity constraints, respectively, should not be less than the pre-specified
confidence levels i (i = 1, 2, . . ., I), ˇj (j = 1, 2, . . ., J) and  k (k = 1, 2, . . ., K), for guaranteeing the reliability of the optimized transportation
strategies.
For solution convenience, we here discuss an equivalent form to further simplify the expected value model. To this end, a proposition
is firstly given in the following.

Proposition 5.1. (Yang and Liu [5]) Suppose that  is a fuzzy number with continuous membership function  (x), g(x, ) = h(x) − .

(I) Let r0 = inf {r| (r) = 1}. Then we have Cr {g(x, ) ≤ 0} ≥ ˛ if and only if h(x) ≤ F (˛), where


sup{F|F = −1

(2˛)}, ˛ ≤ 0.5,
F (˛) = (21)
sup{F|F = −1

(2(1 − ˛)), F < r0 }, ˛ > 0.5.

(II) Let r0 = sup {r| (r) = 1}. Then we have Cr {g(x, ) ≥ 0} ≥ ˛ if and only if h(x) ≥ F (˛), where


inf {F|F = −1

(2˛)}, ˛ ≤ 0.5,
F (˛) = (22)
inf {F|F = −1

(2(1 − ˛)), F > r0 }, ˛ > 0.5.

Moreover, if c̄ijk (∀i, j, k) are independent fuzzy variables, we can produce an equivalent model below.

I J K
min E[Z] = i=1 j=1 k=1
E[c̄ijk ]xijk
J K
st. j=1
x
k=1 ijk
≤ F¯ i (i ), i = 1, 2, ..., I
I K
x ≥Fˇ¯ (ˇj ), j = 1, 2, ..., J (23)
i=1 k=1 ijk j
I J
i=1
x
j=1 ijk
≤ F¯ k (k ), k = 1, 2, ..., K

xijk ≥0, ∀i, j, k,

where, the first and third constraints are inferred by Eq. (21), and the second constraint is deduced by Eq. (22).
To solve the proposed model, a tabu search based solution algorithm, designed by Yang and Liu [5], will be employed in this study. The
algorithmic process is particularly given below for the completeness of this research.

• Tabu Search Algorithm:

• Step 1: Randomly choose a solution x. Let present.sol = x, best.sol = x, present step = 0, best step = 0, then initialize tabu list;
• Step 2: Produce the neighborhood N (present.sol) of present.sol, let present step++;
• Step 3: Compute the objective values of the elements in N (present.sol) by fuzzy simulation;
• Step 4: Find the best solution x* in N (present.sol) such that x* is non-tabu or satisfies aspiration criterion;
• Step 5: If the objective value of solution x* is better than that of best.sol, then let present.sol = x* , best.sol = x* , best step = present step;
otherwise, let present.sol = x* ;
• Step 6: Update the tabu list;
• Step 7: If present step-best step ≤ M (the largest number of iteration that objective value of best.sol does not change), go to Step 2;
otherwise go to Step 8;
• Step 8: Output best.sol, stop.

In the algorithm, the meanings of different symbols are:

x: the index of a feasible solution;

present.sol: the best solution in the neighborhood;
best.sol: the optimal solution obtained up to the current iteration;
present step: the current iteration number;
best step: the iteration number when we get the optimal solution.

6. Numerical experiments

In this section, a set of numerical experiments will be implemented to illustrate the efficiency and effectiveness of the proposed
approaches.
In the experiments, assume that there are four sources (denoted by O), five destinations (denoted by D) and two conveyances. The
relevant unit transportation costs between different ODs by different conveyances are given in Tables 1 and 2, which are represented by
type-2 triangular or trapezoidal fuzzy variables.
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 555

Table 1
The unit transportation cost between different ODs by conveyance 1.

cij1 1 2 3 4 5

1 (5,6,7,8;0.5,1) (6,10,12,16;0.5,0.9) (8,9,10,11;0.6,1) (2,3,4,6;0.8,1) (5,6,10,11;0.6,1)

2 (2,3,4,5;0.5,0.7) (9,10,11,13;0.5,1) (3,4,5,6;0.2,0.6) (3,5,7,8;0.2,1) (8,10,11,12;0.3,0.7)
3 (1,3,5,6;0.3,0.7) (11,12,13,14;0.4,1) (3,5,6,8;0.4,0.8) (10,12,13,14;0.4,0.8) (4,5,9,10;0.5,1)
4 (2,4,6,8;0.4,0.6) (8,10,14,16;0.5,1) (4,5,8,9;0.4,0.6) (2,4,6,10;0.5,0.9) (6,8,10,11;0.6,0.8)

Table 2
The unit transportation cost between different ODs by conveyance 2.

cij2 1 2 3 4 5

1 (5,6,8;0.5,1) (9,10,11;0.6,1) (6,8,10;0.5,0.9) (4,6,7;0.8,1) (5,6,7;0.3,0.7)

2 (6,8,12;0.6,1) (9,10,11;0.5,1) (11,12,14;0.4,1) (3,5,9;0.4,0.8) (2,4,6;0.4,0.6)
3 (4,5,6;0.2,1) (2,4,6;0.5,0.9) (6,7,8;0.5,1) (10,12,16;0.3,0.7) (10,11,12;0.5,0.7)
4 (5,6,7;0.6,1) (3,5,7;0.2,0.6) (8,9,10;0.5,1) (7,11,13;0.4,0.8) (8,10,12;0.5,0.9)

Table 3
Comparison of optimal objective values with different parameters.

Test index M N Best step Opt. objective Relative error (%)

1 15 20 17 1182.6 3.63
2 15 40 10 1143.0 0.16
3 20 40 10 1143.0 0.16
4 20 100 27 1141.2 0.00
5 40 100 27 1141.2 0.00
6 40 120 30 1155.9 1.29
7 70 150 70 1141.2 0.00
8 100 160 5 1141.2 0.00
9 150 180 4 1141.2 0.00
10 180 240 109 1141.2 0.00

Average value 0.524

Variance 0.0121

In addition, the total supply of products, total demand of products and total capacity of conveyances, respectively, are listed as follows:

ã1 = (72, 74, 76, 77; 0.4, 0.8), ã2 = (74, 75, 79, 80; 0.5, 1), ã3 = (74, 76, 78, 79; 0.3, 0.7),

ã4 = (67, 68, 69, 70; 0.2, 0.6), b̃1 = (38, 39, 40, 42; 0.5, 1), b̃2 = (58, 60, 61, 62; 0.4, 0.8),

b̃3 = (45, 47, 49, 51; 0.3, 0.7), b̃4 = (42, 43, 44, 45; 0.4, 0.6), b̃5 = (54, 58, 59, 60; 0.2, 0.7),

ẽ1 = (138, 139, 140, 141; 0.5, 1), ẽ2 = (156, 158, 162, 166; 0.5, 1).

In the model, the predetermined confidence levels are set as i = 0.9, ˇj = 0.9 and  k = 0.8, respectively, where i = 1, 2, 3, 4, j = 1, 2, 3, 4, 5
and k = 1, 2.

6.1. The first set of experiments with expected values criterion

In the first set of experiments, the expected value criterion-based reduction method will be used to convert a type-2 fuzzy variable into
a type-1 fuzzy variable, where the generated type-1 fuzzy variable has possibility distributions with the forms given in Theorem 4.2.1.
Then, tabu search algorithm will be applied to solving the corresponding mathematical model.
In the algorithm, the length of the tabu tenure is a crucial parameter to determine the quality of the generated approximate optimal
solutions. However, a large parameter may potentially cost too much computational time in the searching process. To obtain a satisfactory
solution in an acceptable time, we set tabu length as 6 in the experiments. In addition, to show the solution characteristics with respect to
different algorithmic conditions, we particularly consider two parameters in the solving process. That is, (1) “M” denotes the given iteration
length when the optimal objective value does not change and (2) “N” denotes the number of elements in the neighborhood. By randomly
setting these two parameters with reasonable values, we list the returned approximate optimal objectives in Table 3.
In Table 3, we compare generated optimal objective values with different parameter settings. In these experiments, the relative errors
among different optimal objectives are computed by
Opt. objective − Best opt. objective
Relative error = × 100% (24)
Best opt. objective
Obviously, the computational results show that relative errors of the generated optimal objective values are no more than 3.63%, and
the average value and variance, respectively, are only 0.524% and 0.0121%. In particular, we find the best objective value 6 out of 10 times
in the experiments. These characteristics illustrate the robustness and effectiveness of the tabu search algorithm in solving expected value
reduction criterion-based model.
To show the detailed features of the solution searching process, we list the computational results in Fig. 8 for the fourth test since it
produces the best solution at iteration 27, where parameters M and N, respectively, are set as 20 and 100. Specifically, Fig. 8 demonstrates
556 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

Fig. 8. The change of optimal objective values up to each iteration.

Sources Destinations Sources Destinations

1 1
1 44.8 1
2 2
2 14.4 2
59.8
41.6 3 3
3 3
32.7
4 4
3.6 61.8
4 4
5 5
Conveyance 1 Conveyance 2

Fig. 9. Best solution by using the expected value reduction criterion.

Fig. 10. Optimal objective values with different ˛.

the change of the best objective values obtained up to each iteration. It is easy to see that for the initial several iterations, the best objective
encountered decreases drastically. After iteration 6, the best objective changes with a fairly slow ratio, and at iteration 27, the optimal
objective is found.
Next, we give the generated optimal solution in Fig. 9 to straightforwardly understand the transportation plan. In this solution,
conveyance 1 provides transportation service for 5 OD pairs with the transportation amount 137.1, while conveyance 2 only provides
transportation service for 2 OD pairs with transportation amount 121.6.

6.2. The second set of experiments with optimistic and pessimistic value criteria

In this subsection, we firstly use optimistic value criterion to reduce type-2 fuzzy parameters with different confidence level ˛. Then,
the same algorithm is also designed to search for the corresponding optimal solutions for each model. To capture the characteristics of
the optimal solutions, eight experiments with ˛ = 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, respectively, are implemented on a personal computer
with a 2.50 GHz CPU and 4 GB memory. In these experiments, we set the parameters in tabu search algorithm as M = 20, N = 100, and tabu
length = 6.
Fig. 10 is given to clearly depict the sensitivity of optimal objectives with respect to ˛. Obviously, as shown, the optimal objective
is gradually reduced with the increase of parameter ˛. In particular, when ˛ is not greater than 0.5, the curve associated with optimal
objectives is an approximately linear line.
The following discussion focuses on the display of approximate optimal solutions generated via the optimistic and pessimistic value
criteria with the same parameters settings, where parameters M = 20, N = 100 and ˛ = 0.8.
 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558 557

Sources Destinations Sources Destinations

1 1
1 1
44.7 27.8
2 2
2 42.3 2
32.0
3 3
12.8
3 3 61.7
28.7 4 4
4 8.2 4
5 5
Conveyance 1 Conveyance 2

Fig. 11. Best solution by using the optimistic value reduction criterion.

Sources Destinations Sources Destinations

1 1
1 1
44.9 27.4
2 2
2 41.7 2
32.4
3 3
12.4
3 3 61.9
29.3
4 4
4 9.0 4
5 5
Conveyance 1 Conveyance 2

Fig. 12. Best solution by using the pessimistic value reduction criterion.

(I) Solution with the optimistic value reduction criterion

If we use optimistic value criterion to reduce type-2 fuzzy parameters, the generated type-1 fuzzy variables have possibility distribu-
tions with the forms in Theorem 4.2.1. Using the designed algorithm, the optimal solution is listed in Fig. 11 with the optimal objective
value 1153.1.
(II) Solution with the pessimistic value reduction criterion
On the other hand, if pessimistic value criterion is used to reduce type-2 fuzzy parameters, the generated type-1 fuzzy variables have
possibility distributions with the forms listed in Theorem 4.2.1. Using the designed algorithm to solve the model, we give the optimal
solution in Fig. 12, and the optimal objective value is 1157.7.

7. Conclusions

The solid transportation problem is an important network optimization problem in the traffic engineering. To characterize the uncer-
tainty in the practical decision environment, this paper treated the supply capacities, demands and conveyance capacities as type-2 fuzzy
variables. To give a modeling framework for optimization problems with multi-fold uncertainty, different reduction methods were pro-
posed to transform a type-2 fuzzy variable into a type-1 fuzzy variable, including optimistic value criterion, pessimistic value criterion
and expected value criterion. Then the solid transportation problem was reformulated as a chance-constrained expected value model
enlightened by the credibility optimization methods. To solve the model, tabu search algorithm was employed to search an approximate
optimal solution. The numerical experiments illustrated the application and effectiveness of the proposed approaches.

Acknowledgements

This research was supported by the National Natural Science Foundation of China (No. 71271020), Research Foundation of State Key
Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University (No. RCS2014ZT02), and National Basic Research Program of China
(No. 2012CB725400).

References

[1] K.B. Haley, The solid transportation problem, Oper. Res. 10 (1962) 448–463.
[2] K.B. Haley, The multi-index problem, Oper. Res. 11 (1963) 368–379.
[3] F. Jiménez, J.L. Verdegay, Solving fuzzy solid transportation problems by an evolutionary algorithm based parametric approach, Eur. J. Oper. Res. 117 (1999) 485–510.
[4] S. Liu, Fuzzy total transportation cost measures for fuzzy solid transportation problem, Appl. Math. Comput. 174 (2006) 927–941.
[5] L. Yang, L. Liu, Fuzzy fixed charge solid transportation problem and algorithm, Appl. Soft Comput. 7 (2007) 879–889.
[6] P. Kundu, S. Kar, M. Maiti, Multi-objective multi-item solid transportation problem in fuzzy environment, Appl. Math. Model. 37 (2012) 2028–2038.
[7] A. Ojha, B. Das, S. Mondal, M. Maiti, An entropy based solid transportation problem for general fuzzy costs and time with fuzzy equality, Math. Comput. Model. 50 (2009)
166–178.
558 P. Liu et al. / Applied Soft Computing 24 (2014) 543–558

[8] S. Molla-Alizadeh-Zavardehi, S. Sadi Nezhadb, R. Tavakkoli-Moghaddamc, M. Yazdani, Solving a fuzzy fixed charge solid transportation problem by metaheuristics, Math.
Comput. Model. 57 (2013) 1543–1558.
[9] L. Yang, Y. Feng, A bicriteria solid transportation problem with fixed charge under stochastic environment, Appl. Math. Model. 31 (2007) 2668–2683.
[10] A. Ojha, B. Das, S. Mondal, M. Maiti, A stochastic discounted multi-objective solid transportation problem for breakable items using Analytical Hierarchy Process, Appl.
Math. Model. 34 (2010) 2256–2271.
[11] L.A. Zadeh, Concept of a linguistic variable and its application to approximate reasoning I, Inform. Sci. 8 (1975) 199–249.
[12] Z. Liu, Y.K. Liu, Fuzzy possibility space and type-2 fuzzy variable, in: Proceedings of the 2007 IEEE Symposium on Foundations of Computational Intelligence (FOCI),
2007.
[13] R. Qin, Y.K. Liu, Z. Liu, Methods of critical value reduction for type-2 fuzzy variables and their applications, J. Comput. Appl. Math. 235 (2011) 1454–1481.
[14] S. Coupland, R. John, A fast geometric method for defuzzification of type-2 fuzzy sets, IEEE Trans. Fuzzy Syst. 16 (2008) 929–941.
[15] G.A. Vignaux, Z. Michalewicz, A genetic algorithm for the liner transportation problem, IEEE Trans. Syst. Man Cybern. 21 (1991) 445–452.
[16] Z. Liu, Y.K. Liu, Type-2 fuzzy variables and their arithmetic, Soft Comput. 14 (2010) 729–747.
[17] B. Liu, Y.K. Liu, Expected value of fuzzy variable and fuzzy expected value models, IEEE Trans. Fuzzy Syst. 10 (2002) 445–450.
[18] B. Liu, Theory and Practice of Uncertain Programming, Physica-Verlag, New York, 2002.
[19] L. Yang, K. Li, Z. Gao, Train timetable problem on a single-line railway with fuzzy passenger demand, IEEE Trans. Fuzzy Syst. 17 (2009) 617–629.
[20] L. Yang, K. Li, Z. Gao, X. Li, Optimizing trains movement on a railway network, Omega 40 (2012) 619–633.
[21] L. Yang, X. Zhou, Constraint reformulation and a Lagrangian relaxation-based solution algorithm for a least expected time path problem, Transport. Res. Part B 59 (2014)
22–44.
[22] L. Yang, X. Zhou, Z. Gao, Credibility-based rescheduling model in a double-track railway network: a fuzzy reliable optimization approach, Omega 48 (2014) 75–93.
[23] W. Pedrycz, Granular Computing: Analysis and Design of Intelligent Systems, CRC Press/Francis Taylor, Boca Raton, 2013.
[24] W. Pedrycz, Allocation of information granularity in optimization and decision-making models: towards building the foundations of granular computing, Eur. J. Oper.
Res. 232 (2014) 137–145.
[25] P. Kundu, S. Kar, M. Maiti, Fixed charge transportation problem with type-2 fuzzy variables, Inform. Sci. 255 (2014) 170–186.
[26] S. Greenfield, F. Chiclana, R. John, S. Coupland, The sampling method of defuzzification for type-2 fuzzy sets: experimental evaluation, Inform. Sci. 189 (2012) 77–92.
[27] D. Wu, J.M. Mendel, Uncertainty measures for interval type-2 fuzzy sets, Inform. Sci. 177 (2007) 5378–5393.