Expert Systems With Applications 93 (2018) 39–49

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Solving the interval linear programming problem: A new algorithm for

a general case
Hasan Ali Ashayerinasab, Hasan Mishmast Nehi, Mehdi Allahdadi∗
Mathematics Faculty, University of Sistan and Baluchestan, Zahedan, Iran

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

Article history: Based on the binding constraint indices of the optimal solution to the linear programming (LP) model, a
Received 19 June 2017 feasible system of linear equations can be formed. Because an interval linear programming (ILP) model
Revised 7 October 2017
is the union of numerous LP models, an interval linear equations system (ILES) can be formed, which
Accepted 8 October 2017
is the union of these conventional systems. Hence, a new algorithm is introduced in which an arbitrary
Available online 9 October 2017
characteristic model of the ILP model is chosen and solved. The set of indices of its binding constraints is
Keywords: then obtained. This set is used to form and solve an ILES using the enclosure method. If all the compo-
Optimal solution set nents of the interval solutions to this system are strictly non-negative, the optimal solution set (OSS) of
Interval linear programming the ILP model is determined as the subscription of the zone created by reversing the signs of the binding
Interval linear equations system constraints of the worst model and the binding constraints of the best model. The solutions to several
Monte Carlo simulation problems obtained by the new algorithm and a Monte Carlo simulation are compared. The proposed al-
gorithm is applicable to large-scale problems. To this end, an ILP model with 270 constraints and 270
variables is solved.
© 2017 Published by Elsevier Ltd.

1. Introduction
Many real world system parameters are inexact and are deter-
mined as interval numbers. Therefore, interval linear programming
(ILP) is an efficient method to characterize inexact parameters in
decision-making problems. Recently, obtaining the optimal solution
set (OSS) and the optimal range of the objective values of the ILP
problem has become important to researchers. Allahdadi and Mish-
mast Nehi (2013) determined the OSS of the ILP model from the
worst and best model constraints when all of the components of
the optimal solutions to the ILP model are strictly non-negative.
In fact, by assuming the positivity of all of the components of
the feasible point of the best model, if the number of constraints
and variables are equal, the exact OSS of the ILP model is par-
allel to the region that is the subscription of the feasible zone
of the best model and the feasible zone created by reversing the
inequality signs of the worst model constraints. Several methods
have been proposed to solve the ILP model. Some of the methods
transform the ILP model into two sub-models (Allahdadi, Mish-
mast Nehi, Ashayerinasab, & Javanmard, 2016; Fan & Huang, 2012;
Huang, Baetz, & Patry, 1995; Huang & Moore, 1993; Lu, Cao, Wang,
Fan, & He, 2014; Tong, 1994; Wang & Huang, 2014; Zhou,Huang,
∗
Corresponding author.
E-mail addresses: hmnehi@hamoon.usb.ac.ir
m_allahdadi@math.usb.ac.ir (M. Allahdadi).
Chen, & Guo, 2009). The solution area of these methods is formed
by solving these sub-models and obtaining their optimal solutions.
One solution, the BWC method (Tong, 1994), obtains the largest in-
terval of the objective function values. In addition, some of these
methods, such as a novel ILP (Huang & Moore, 1993), a two-step
method (TSM) (Huang et al., 1995), and another solution (SOM-2)
(Lu et al., 2014) have been proposed. However, part of the solu-
tion area of these methods may be infeasible. Several techniques
have been developed to remove the infeasible part of the solution
area of these methods, such as the modified ILP (MILP) (Zhou et al.,
2009), improved TSM (ITSM) (Wang & Huang, 2014), and robust
TSM (RTSM) (Fan & Huang, 2012). Allahdadi et al. (2016), intro-
duced two improvement methods, namely, IMILP and IILP, to delete
the non-optimal solutions to the solution areas of the MILP and ILP
methods, respectively.
An arbitrary point is a feasible solution to the ILP model if it
belongs to the feasible zone of the best model, and it is optimal
if it is an optimal solution to an arbitrary characteristic problem
of the ILP model. If the optimal solution to a linear programming
(LP) model exists, then it lies in the extreme point set or bounds of
the feasible region of the LP model. In fact, the optimal solution to
the LP problem is the solution to a linear equations system (LES),
whose equations are the binding constraints of the LP model. By
considering the variables (such as dual variables) corresponding to
the binding constraints of the LP model and by multiplying them
(H.M. Nehi),
in the columns of the technological matrix, an LES can be formed

https://doi.org/10.1016/j.eswa.2017.10.020
0957-4174/© 2017 Published by Elsevier Ltd.
40 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49

(similar to the dual problem constraints). In the optimal solution to 

n

an LP model, the objective gradient lies in the cone created by the Minimization Z ± = c±j x±j
gradients of the binding constraints. This optimality qualification is j=1

equivalent to the strong feasibility of the LES presented above. An 

n
(1)
subject to: a± x± ≥ b±
ij j i
, i = 1, . . . , m,
ILP model is the union of numerous characteristic models, which
j=1
are LP models. Hence, an interval linear equations system (ILES) x±j ≥ 0, j = 1, · · · , n.
can be formed, which is the union of the conventional systems de-
scribed above. In model (1) by considering certain values through interval param-
In this article, the exact OSS of the ILP model will be ob- eters, a conventional LP model that is called a characteristic model
tained. A new algorithm is introduced, and an arbitrary character- is obtained as follows:
istic problem of the ILP problem is chosen and solved. The set of
indices of its binding constraints is then obtained. With this index 
n
Minimization Z o = coj x j
set, an ILES is formed and solved using the enclosure method. If
j=1
all of the components of the interval solutions to this system are 
n
(2)
strictly nonnegative, the OSS of the ILP model can be obtained. subject to: aoi j x j ≥ boi , i = 1, . . . , m,
This algorithm has several advantages. (1) It obtains the exact j=1
OSS of the general ILP model when the related ILES is strongly fea- x j ≥ 0, j = 1, · · · , n.
sible. (2) The solution process of this algorithm is not complex. (3)
The algorithm obtains all of the optimal solutions to the ILP model where coj ∈ c±j = [c−j , c+j ], aoi j ∈ a±
ij
= [a−
ij
, a+
ij
], and boi ∈ b±
i
= [b−
i
, b+
i
].
when the solution set of the related ILES is strictly nonnegative,
whereas the solution area of previous solution methods only can Definition 2.3. An interval hyperplane H ± in IRn is defined by the
obtain some results. (4) The OSS obtained in Allahdadi and Mish- set {x± : p± x± = k± }, where p± is an interval vector in IRn , the
mast Nehi (2013) is expressed for special cases of the ILP model, interval vector p± does not consist of the zero vector, and k ± is an
but this algorithm obtains the exact OSS of the ILP model in the interval number. A set H = {x : px = k} ∈ H ± where p ∈ p± , k ∈ k± ,
general case. Hladik (2016) presented a discussion of the topolog- is called a characteristic hyperplane. An interval hyperplane H ± is
ical affections of the robust OSS, and several definitions related to the union of all conventional hyperplanes H.
interval numbers and a new method for solving the interval bilevel
Definition 2.4. A zone Xˆ of the feasible region of the ILP model
LP problem were studied by Ren, Wang, and Xue (2017). Several es-
(1) is called an extreme zone if each of its points is an extreme
sential definitions and theorems are expressed and proved to illus-
point of a characteristic model of the ILP model (1). In other words,
trate the algorithm described above. In addition, numerical exam-
a zone Xˆ is an extreme zone if it is the intersection of some n
ples are introduced to study the method. Finally, the solutions ob-
linearly independent defining interval hyperplanes of the feasible
tained using the algorithm and the Monte Carlo simulation (MCS)
region of the ILP model (1).
method are compared to illustrate the applicability of the algo-
rithm. A large-scale ILP model with 270 constraints and 270 vari- Theorem 2.1. Tong (1994) In the ILP model (1), the largest and
n
j=1 ai j x j ≥ bi , ∀i, x j ≥ 0, ∀ j and
ables is studied to demonstrate that the proposed algorithm is ap- + −
smallest feasible regions are
plicable. n
j=1 ai j x j ≥ bi , ∀i, x j ≥ 0, ∀ j, respectively.
− +

Definition 2.5. A point xˆ = (xˆ1 , xˆ2 , · · · , xˆn ) is said to be a feasible


point of the ILP model (1) if nj=1 a+ xˆ ≥ b−
ij j i
, ∀i, and xˆ j ≥ 0, ∀ j.

2. Preliminaries Theorem 2.2. Rohn (1993) If the set Ax = b, where A ∈ IRn×n and
b ∈ IRn is solved, the interval vector r is obtained by the enclosure
An interval number a ± is defined as [a− , a+ ], where a− ≤ a+ . method, then for i = 1, · · · , n:
If a− = a+ , then a ± will degenerate, and the interval number a ±
 1

transforms into a crisp number.
ri− = min −x∗i + (xci + |xci | )Mii , (−x∗i + (xci + |xci | )Mii ) ,
2Mii − 1
 1

Definition 2.1. Suppose that, A− = (a−
ij
), A+ = ( a+
ij
) ∈ Rm×n , m, n ∈ ri+ = max x∗i + (xci − |xci | )Mii , ( x∗i + (xci − |xci | )Mii ) ,
N are two matrices that for all i, j; (a− ) ≤ ( a+ ). Thus, an interval 2Mii − 1
ij ij
matrix is given as follows: where M = (I − |(Ac )−1 |A )−1 , xc = (Ac )−1 bc , x∗ = M ( |xc | +
|(Ac )−1 |b ), and Ac is a non-singular matrix and ρ (|(Ac )−1 |A ) < 1.
A± = [A− , A+ ] = {A ∈ Rm×n | A− ≤ A ≤ A+ }.
3. Review of previous solution methods
The radius and centre of A ± are A± = 12 (A+ − A− ) and Ac =
The following subsections present a review of previous solu-
1
2 ( A − + A+ ), respectively. Thus, A± = [A− , A+ ] = [Ac −  , Ac +
A±
tion methods. First, the BWC method is reviewed, and the best and
A± ]. The union of all m × n interval matrices is denoted by IRm×n .
worst models are examined to determine the exact OSS of the ILP
An interval vector v ± is introduced as the set v± = {v | v− ≤ v ≤
model when all of the components of the feasible solution are pos-
v+ } , where v− , v+ ∈ Rn are crisp vectors (Fiedler, Nedoma, Ramik,
itive. Then, an introduction to various solution methods and their
Rohn, & Zimmermann, 2006).
solution areas is presented.

Definition 2.2. An interval number a ± is non-negative (positive) if

3.1. BWC method
a− ≥ 0 (a− > 0 ) and is non-positive (negative) if a+ ≤ 0 (a+ < 0 ).

Tong transformed the ILP problem (1) into worst and best sub-
A general form of the ILP model is defined as follows: problems, which are summarized as follows (Tong, 1994):
 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49 41

 
The best sub-problem: Minimization Z − = c−j x−j + c−j x+j
j∈A1 j∈A2

n
Minimization Z − = c−j x j subject
 to: 
j=1 |a±i j |+ sign(a±i j )x−j + |a±i j |− sign(a±i j )x+j ≥ b−i , i = 1, . . . , m, (5)

n
(3) j∈A1 j∈A2
subject to: a+ x
ij j
≥ b−
i
, i = 1, . . . , m, x−j ≥ 0, f or j ∈ A1 ,
j=1 x+j ≥ 0, f or j ∈ A2 .
x j ≥ 0, j = 1, · · · , n.
The optimal solutions to sub-problem 1 are x−j , (j ∈
opt
The worst sub-problem:
A1 ), x+j , ( j ∈ A2 ), and its objective function is −
Zopt . Sub-problem
opt

n 2 is introduced as follows:
Minimization Z + = c+j x j  
Minimization Z + = c+j x+j + c+j x−j
j=1

n
(4) j∈A1 j∈A2
subject to: a− x ≥ b+
ij j i
, i = 1, . . . , m, subject
 to: 
j=1 |a±i j |− sign(a±i j )x+j + |a±i j |+ sign(a±i j )x−j ≥ b+i , i = 1, . . . , m, (6)
x j ≥ 0, j = 1, · · · , n. j∈A1 j∈A2
x+j ≥ x−jopt , f or j ∈ A1 ,
The optimal solutions to sub-problems (3) and (4) form a box as 0 ≤ x−j ≤ x+jopt , f or j ∈ A2 ,
follows
 
a+ , a± ≥0 a− , a± ≥0
x± = ( x±
1 , x2 , · · · , xn ),
± ±
where |a± |+ = ij ij
, |a± |− = ij ij
, and
ij −a− , a± <0 ij −a+ , a± <0
 ij ij ij ij
where for all j, x±j = [x−j , x+j ]. This box is the solution area of the 1, a± ≥0
sign(a± )= ij
.
BWC method. ij −1, a±ij
<0
Compounding the optimal solutions to these sub-problems
Theorem 3.1. (Allahdadi and Mishmast Nehi 2013; Tong 1994) If Z∗ forms the solution area of the ILP method. The solution area and
is the optimal objective value of an arbitrary characteristic model as objective values of the ILP method are x±j = [x−j , x+j ] and Zopt
±
=
∗ ∗
a model (2), and Z − and Z + are the optimal objective values of the
opt opt opt
− +
∗ ∗ [Zopt , Zopt ], respectively.
model (3) and model (4) respectively, then Z ∗ ∈ [Z − , Z + ].
3.4. MILP method

3.2. Concept of the optimal solution set Define the index sets B1 = { j : j ∈ A1 , and a±
ij
≥ 0}, B2 = { j : j ∈
A1 , and a±
ij
≤ 0}, B3 = { j : j ∈ A2 , and a±
ij
≤ 0}, and B4 = { j : j ∈
In this subsection, the OSS of the ILP problem is determined by
A2 , and a± ≥ 0}. Zhou et al. (2009), introduced the MILP method
the best and worst constraints with the supposition that all of the ij

components of the feasible solution to the ILP model are positive. to guarantee that the solutions obtained using the ILP method are
This assumption will be obtained using the infeasibility test by re- absolutely feasible. This method uses sub-problem 1 as shown in
placing each variable with zero. model (5). In the second sub-problem, the following constraints are
added to the constraints of model (6):
Definition 3.1. If there is a characteristic problem such that x∗ is  
− ( |a±
δ j | x j − |aδ j | x jopt ) +
− + ± + −
(|a±δ j |+ x−j − |a±δ j |− x+jopt ) ≥ 0, (7)
its optimal solution, then x∗ is also an optimal solution to the ILP
j∈B2 j∈B4
model (1). The OSS of the ILP model (1) corresponds to the set
of all solutions that are optimal for a given characteristic problem where δ is the index of the binding constraints in model (5) as
among the solutions that compose the ILP model (1). follows:
 
|a±δ j |+ sign(a±δ j )x−jopt + |a±δ j |− sign(a±δ j )x+jopt = b−δ .
In Allahdadi and Mishmast Nehi (2013), the OSS of the ILP prob-
j∈A1 j∈A2
lem is determined when the numbers of constraints and variables
are equal. The feasibility of the solution area of the MILP method is ensured
using the additional constraints (7).
Theorem 3.2. Allahdadi and Mishmast Nehi (2013) Assume that each
of the components of a feasible solution in the best model of the ILP 3.5. TSM and ITSM methods
model are strictly nonnegative. Then, the OSS of the ILP model is de-
termined as the subscription of the feasible zone of the best model and The sub-problems of the TSM method are similar to mod-
the feasible zone created by reversing the signs of the worst model els (5) and (6) Huang et al. (1995). In the ITSM method intro-
constraints. duced in Wang and Huang (2014), the first sub-problem is model
(5). To formulate sub-problem 2 of the ITSM, the following addi-
Allahdadi and Mishmast Nehi (2013) provides additional infor- tional constraints are added to the second sub-problem of the TSM
mation about this theorem. method:

n
a+ x ≥ b−
ij j i
, xj
3.3. ILP method j=1
⎧ +
⎪xj , if sign(a± ) = sign(c±j ), f or j ∈ A1 ,
Define the index sets A1 = { j : c±j ≥ 0} and A2 = { j : c±j ≤ 0}. In ⎪ ij
⎪x− ,
⎨ if sign(ai j ) = sign(c±j ),
±
f or j ∈ A1 ,
the ILP method presented in Huang and Moore (1993), two sub- jopt

problems are formulated as follows: = x− , if sign(a± ) = sign(c±j ), f or j ∈ A2 , (8)

⎪
⎪ +
j ij
sign(ai j ) = sign(c±j ),
±
Sub-model 1: ⎪
⎩x jopt , if f or j ∈ A2 .
42 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49

The additional constraints (7) and (8) are equivalent. The method as the set of indices of the binding constraints at x̄ from the system
to identify the additional constraints (8) is easier than that for the Ao x ≥ bo ; similarly, the set J = { j : x̄ j = 0} is defined corresponding
additional constraints (7). to the set x ≥ 0.

3.6. IILP method Theorem 4.1. Bazaraa, Jarvis, and Sherali (1977) A feasible point x∗
is an optimal solution to the LP model (2) iff
This subsection reviews the ILP improvement (IILP) method  o 
proposed in Allahdadi et al. (2016). In this method, two sub-
λoi ai + v j o e j = co
i∈I j∈J
models are introduced to solve the ILP model (1). The first sub-
model is as follows: λoi ≥ 0, i ∈ I, v j o ≥ 0, j ∈ J. (11)
 
Minimization Z = +
c+j x+j + c+j x−j In other words, x∗
is an optimal solution to the LP model (2) if and
j∈A1 j∈A2 only if the objective gradient co belongs to the cone generated by the
subject
 to:  gradients of the binding constraints at x∗ .
|a±i j |− sign(a±i j )x+j + |a±i j |+ sign(a±i j )x−j ≥ b+i , i = 1, . . . , m, (9)
j∈A1 j∈A2 In Eq. (11), if λoi > 0, i ∈ I, v j o > 0, j ∈ J, then x∗ is a unique op-
x+j ≥ 0, f or j ∈ A1 , timal solution to model (2).
x−j ≥ 0, f or j ∈ A2 . Suppose that the set of indices of the binding constraints at the
extreme point xo ∗ of the characteristic model (2) is I ∪ J. In model
Using the optimal solutions to model (9), the second sub-problem (2), if the gradient co belongs to the cone generated by the gradi-
is written as follows: ents of the constraints of model (2) whose indices lie in I ∪ J (i.e.,
  the system of Eq. (11) is feasible), then xo ∗ is an optimal solu-
Minimization Z − = c−j x−j + c−j x+j
tion to model (2). Obviously, the ILP model (1) is the union of all
j∈A1 j∈A2
subject to : characteristic models as in model (2). Therefore, if the index set
  of the binding constraints of all characteristic models is equal to
|a±i j |+ sign(a±i j )x−j + |a±i j |− sign(a±i j )x+j ≥ b−i , i = t1, . . . , m,
I ∪ J, then the union of the optimal solutions to these characteristic
j∈A1 j∈A2 (10)
models will be the OSS of the ILP model. This OSS is the same as
0 ≤ x−j ≤ x+jopt , f or j ∈ A1 ,
an extreme zone of the feasible region of the ILP problem; that
x+j ≥ x−jopt , f or j ∈ A2 , ±
  is, the subscription of the interval hyperplanes ai x± = b± i
, i∈I
− ( |a± δj | + −
x j
− |a±δ j |− x+jopt ) + (|a±δ j |− x+j − |a±δ j |+ x−jopt ) ≤ 0, and x±j = 0, j ∈ J. In fact, the constraints ai x± ≥ b± , i ∈ I and
±
i
j∈B2 j∈B4
x±j≥ 0, j ∈ J are binding in this extreme zone.
where A j , j = 1, 2 and B j , j = 2, 4 are defined in Sections 3.3 and Because each of the characteristic models of the ILP model is
3.4, respectively. The last constraint of model (10) is an additional an LP model, an LES (11) can be written for each of them. The
constraint to ensure that all of the solutions obtained by the IILP union of these systems forms an ILES. Let the index set of binding
method are optimal. In this additional constraint, δ is the index of constraints in all characteristic models be I ∪ J. Then, this ILES is:
the binding constraints in model (9). By solving these sub-models,  
±
the results are x±j = [x−j , x+j ] and Zopt
opt opt
± −
= [Zopt
opt
+
, Zopt ]. λ±i ai + v j ± e j = c±
Note that the sub-models of the ILP and TSM methods are simi- i∈I j∈J

lar, and the sub-models of the MILP and ITSM methods are similar.
The solution methods presented above will be addressed in the
final section of this paper using a numerical example.
4. Solving the ILP problem: a new algorithm
Here, the exact OSS of the ILP model is determined with at-
tention to the optimality condition of the LP problem. The objec-
tive gradient of the optimal solution to an LP problem belongs to
the cone generated by the gradients of the binding constraints; i.e.,
based on the set of indices of the binding constraints of the op-
timal solution to the LP model, an LES that is (strongly) feasible
can be formed. This system is written similar to the dual prob-
lem constraints. Because an ILP model is the union of numerous LP
models, an ILES can be formed, which is the union of the conven-
tional systems presented above. A new algorithm is introduced to
determine the exact OSS of the ILP model. First, an arbitrary char-
λ±i ≥ 0, i ∈ I, v j ± ≥ 0, j ∈ J.
Note that formulating all of the characteristic models of the ILP
model is NP-hard. In this paper, an arbitrary characteristic model
of the ILP model was solved, and the index set of its binding con-
straints was determined. Using this index set, the ILES presented
above was formed and solved. If this system is strongly feasible,
the OSS of the ILP model is a unique extreme zone of the feasible
region of the ILP model.
Definition 4.1. Rudin (1963) Suppose V ⊂ Rn . A point x ∈ Rn is an
inner point of V if there is a neighbourhood of x, Nr (x), such that
Nr (x) ⊂ V. The inner of the set V is the union of all of its inner
points and is denoted by Vo .
Definition 4.2. Let the set of indices of the binding constraints of
the worst model (i.e., model (4)) be I ∪ J. The set U is defined as
follows:
acteristic model is solved, and its binding constraints are obtained.  
Then, by using the results, an ILES is formed and solved. If all of 
n

the components of the interval solutions to this system are strictly U= x| a− x ≥ b+

ij j i
, i∈
/ I, x j ≥ 0, j ∈
/J .
non-negative, the OSS of the ILP model is determined by the bind- j=1

ing constraints of the best model and the binding constraints of

The inner of U is denoted by Uo .
the worst model with converse signs. Several essential definitions
and theorems are introduced to obtain the OSS of the ILP problem. Definition 4.3. Based on the ILP model (1), the index set I ∪ J, T
A stepwise solution algorithm is then studied to obtain the OSS. and S were defined as follows:
Consider a characteristic model (2). Let ai o be the ith row of  
Ao , 1 ≤ i ≤ m, and let ej be a vector in which its jth component is 1 
n
o T = x: a+ x ≥ b− , ∀i ∈ I, x j ≥ 0, ∀ j ∈ J ,
and its other components are 0. The set I = {i : ai x̄ = boi } is defined ij j i
j=1
 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49 43

 

n and
S= x: a− x ≤ b+ , ∀i ∈ I, x j ≤ 0, ∀ j ∈ J . 
ij j i λoi aoi j = coj , j ∈/ J
j=1
i∈I

Remark 4.1. In the following theorem, an arbitrary characteristic λoi > 0, i ∈ I.

model of model (1) was solved, and the index set of its binding
Therefore, vector co ∈ c ± exists such that
constraints was determined to be equal to I ∪ J. An ILES whose so-  o 
lutions specify the OSS of the ILP model was then formed. λoi ai + v j o e j = co
i∈I j∈J
Lemma 4.1. Allahdadi and Mishmast Nehi (2013) Consider the sets S
λoi > 0, i ∈ I, v j o > 0, j ∈ J. (15)
and T defined in Definition 4.3. If x∗ ∈ S ∩ T, then there exist aoi j ∈ a±
ij
and boi ∈ b± such that Hence, the gradient vector of the objective function co ∈ c±
is in-
i
troduced as a strictly non-negative linear composition of the gradi-

n
ent vector of the binding constraints at x∗ (i.e., ai o , i ∈ I, and ej , j ∈ J).
aoi j x∗j = boi , ∀i ∈ I, Thus, by the optimality conditions expressed by Karush, Kuhn, and
j=1
Tucker, x∗ is the optimal solution to the characteristic problem
and for j ∈ J, x∗j = 0. (13).
Conversely, suppose that x∗ is the optimal solution to an ar-
Theorem 4.2. Consider an arbitrary characteristic model of the ILP bitrary characteristic model (2). Because the ILES (12) is strongly
model with the index set of its binding constraints as I ∪ J, where |I| + feasible, system (15) is feasible. Therefore,
|J| = n. In addition, let S ∩ T ⊂ Uo . If the solution to the following ILES 
n
 aoi j x∗j = boi , i ∈ I
λ ± ±
a
i ij
+v = ±
j
c±j , j ∈ J,
j=1
i∈I
 x∗j = 0, j ∈ J. (16)
λ±i ai j ± = c±j , j ∈/ J (12)
i∈I Hence,

(λ± , λ± , λ± ± ± ± 
n 
n
is equal to w± = 1 2
,··· |I| , v1 , v2 , · · · , v|J| ) such that for all
a+ x∗ ≥ aoi j x∗j = boi ≥ b− , i∈I
ij j i
i ∈ I, λ±
i
> 0, and for all j ∈ J, v±j > 0, then x∗ ∈ S ∩ T
iff x∗ is an j=1 j=1
optimal solution to a characteristic model of the ILP model.
Additionally, because of the feasibility of x∗ , the following is given:
Proof. Suppose that x∗ ∈ S ∩ T. Thus, by Lemma 4.1, there exist aoi j ∈ for all j, x∗j ≥ 0. Therefore, x∗ ∈ T. Similarly,
a±
ij
, and boi ∈ b±
i
such that 
n 
n
a− x∗ ≤
ij j
aoi j x∗j = boi ≤ b+
i
, i ∈ I,

n
j=1 j=1
aoi j x∗j = boi , ∀i ∈ I,
j=1 and because for all j ∈ J, x∗j = 0 is equivalent to x∗j ≤ 0 and x∗j ≥
0, then x∗j satisfies the constraints x∗j ≤ 0, ∀ j ∈ J. Therefore, x∗ ∈ S.
and for j ∈ J, x∗j = 0.
Accordingly,
The following characteristic model is considered:
x∗ ∈ S ∩ T .

n
Minimization coj x j 
j=1

n Note that, by considering an arbitrary characteristic model with
subject to: aoi j x j ≥ boi , i ∈ I, binding constraints I ∪ J, if the corresponding ILES (12) is strongly
(13) feasible, then any characteristic model (including the best and the
j=1

n
worst models) has same set of binding constraints at optimality,
a+ x ≥ b−
ij j i
, i∈
/ I, which, in this case, is referred to as I ∪ J. Hence, there exists a
j=1
unique extreme zone, and the complete OSS is given by S ∩ T.
x j ≥ 0, j = 1, · · · , n.
Remark 4.2. Based on the proof of Theorem 4.2, because x∗ ∈ S ∩ T

n
is an arbitrary solution, if system (12) is strongly feasible, the OSS
Because for all i ∈ I, aoi j x∗j = boi , and because S ∩ T ⊂ Uo , for all
of the ILP model is equal to S ∩ T. In addition, if system (12) is
j=1
i ∈ I, the following is true: strongly feasible, then each characteristic system through system
(12) is feasible, and the gradient vector of the objective function

n 
n 
n
of each characteristic model belongs to the cone generated by
a+ x∗ ≥ aoi j x∗j ≥ a− x∗ > b+ ≥ boi ≥ b− , (14) o
ij j ij j i i the gradients of the constraints ai x ≥ boi , i ∈ I and xj ≥ 0, j ∈ J (i.e.,
j=1 j=1 j=1
ai o , i ∈ I and ej , j ∈ J). Therefore, the largest cone generated by the
and ∀ j, x∗j ≥ 0, so x∗ is a feasible solution of the model (13). In gradients of the interval objective function lies in the cone that
addition, ∀ j ∈ J, x∗j = 0 because x∗ ∈ S. By assuming |I| + |J| = n and is equal to the subscription of all of the cones created by n vec-
by the activity of the n constraints at x∗ , the point x∗ is a basic tors ai o ∈ ai ± , i ∈ I and ej , j ∈ J. Therefore, from the viewpoint of geo-
solution of the model (13). metric interpretation, the extreme zone S ∩ T of the feasible region
By assumption, the solution to ILES (12) is λ± i
> 0, ∀i ∈ I, and of the ILP model, which is the subscription of n linearly indepen-
±
v j > 0, ∀ j ∈ J. Thus, there exists coj ∈ c±j such that
± dent interval hyperplanes ai x± = b± i
, i ∈ I and x±j = 0, j ∈ J, is the
 same as the OSS (optimal extreme zone) of the ILP model.
λoi aoi j + v j o = coj , j ∈ J
Solution algorithm 4.1: The solution algorithm to obtain the
i∈I
OSS of the ILP model is introduced as follows. Consider the ILP
λoi > 0, i ∈ I, voj > 0, j ∈ J, model (1).
44 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49

Step 1: First, an arbitrary characteristic model of the ILP model Table 1

Solutions to methods for solving the ILP model (19).
(1) is solved. Without loss of generality, a characteristic model with
centre values can be solved as follows: Method x±
1opt
x±
2opt
±
Zopt

n
BWC [5.18,6.37] [3.34,4] [22.14,34.86]
Minimization ccj x j IILP [5.21,6.23] [3.83,4.03] [22.28,33.58]
j=1 ILP and TSM [5.21,6.34] [3.32,4.03] [22.28,34.69]

n
(17) MILP and ITSM [4.57,6.34] [3.32,3.49] [19.59,34.69]
subject to: aci j x j ≥ bci , i = 1, . . . , m,
j=1
x j ≥ 0, j = 1, · · · , n.
6. Numerical examples
Let x̄ be the optimal solution to model (17).
Step 2: After solving model (17), the set of indices of the bind- Several examples are introduced to illustrate the algorithm pre-
ing constraints at x̄ of model (17) is specified. Let this index set sented above. The following examples are solved using MATLAB
n n
j=1 ai j x j ≥ bi be I = {i : j=1 ai j x̄ j = bi } and that
from the set c c c c
R2012 with interval arithmetic performed using the interval tool-
from the set xj ≥ 0 be J = { j : x̄ j = 0}. box INTLAB v6 (Rump, 1999). In Examples 6.1 and 6.2, the optimal
Step 3: The interval variable λ± = (λ± 1
, λ±
2
, · · · , λ±
|I| ) is consid- solutions of the ILP model are obtained by the MCS method. Then,
ered, where for i ∈ I, λ± i
is the corresponding variable to the ith the solutions obtained by the MCS method and S ∩ T are compared
constraint of the ILP model (1). In addition, the interval vari- to demonstrate the robustness and applicability of the proposed
able v± = (v±
1
, v±
2
, · · · , v± ±
|J| ) is considered, where for j ∈ J, v j cor-
algorithm. The optimal solutions obtained by the MCS method are
responds to x±j ≥ 0. shown in Figs. 4, 6–8 part (a).
Step 4: The following ILES is formed: Example 6.1. Consider the following ILP model:

λ±i a±i j + v±j = c±j , j ∈ J, Maximization Z ± = [5.2, 6]x± − [1, 1.2]x±
1 2
i∈I
 subject to:
λ±i ai j ± = c±j , j ∈/ J. (18) [8, 10]x± 1
− [12, 14]x± 2
≤ [3.8, 4.2]
(19)
i∈I [1, 1.1]x± 1
+ [0 . 19 , 0 . 2] x±
2
≤ [6.5, 7]
± ±
Step 5: The solution set of system (18) is calculated [1, 1.1]x1 + [0.8, 1]x2 ≤ [15, 16]
by the enclosure method as λ± = (λ± , λ± , · · · , λ± x± , x± ≥ 0(−x± ≤ 0, −x± ≤ 0 ).
|I| ) and v =
±
1 2 1 2 1 2
( v±
1
, v±
2
, · · · , v±
|J | ). The feasible region of the ILP model (19), which is the same
Step 6: If for all i ∈ I, λ±
i
> 0 and for all j ∈ J, v±j > 0, then as the feasible region of the best model (i.e., the greatest feasible
based on Theorem 4.2, the OSS of the ILP model (1) is a unique region), the feasible region of the worst model (i.e., the smallest
extreme zone that is equal to S ∩ T. feasible region), and the extreme zones of the feasible region of
Note that S and T are defined in Definition 4.3. The solution the ILP model (19) are presented in Fig. 3.
process for the algorithm presented above is represented in Fig. 1. First, the standard form of the characteristic model with centre
values is solved. The optimal solution to this characteristic model is
5. Monte Carlo simulation Z c ∗ = 28.14 and (x1 , x2 , s1 , s2 , s3 ) = (5.75, 3.67, 0, 0, 6.16 ). Because
the slack variables s1 and s2 are equal to zero, the first and second
Monte Carlo simulation (MCS) is applicable to many sciences. constraints are binding. Thus, I = {1, 2}, and J = ∅. Now, consider
Although its usage varies from science to science, they all have the variables λ± 1
and λ±2
, which correspond to the first and second
random variables as samples. MCS is studied in Carlin, Polson, and constraints of model (19), respectively. Then, the following ILES is
Stoffer (1992), Chib, Nardari, and Shephard (2002) and Tierney and formed:
Mira (1999). The simulation can be expressed based on two as- 
sumptions (Tierney & Mira, 1999): (i) All of the parameters of the [8, 10]λ±
1
+ [1, 1.1]λ±
2
= [5.2, 6]
(20)
ILP problem are random variables with unknown distribution func- −[12, 14]λ±
1
+ [0.19, 0.2]λ±
2
= −[1, 1.2]
tions, and (ii) The ranges of values of the random variables are
Using Theorem 2.2 and solving system (20) using the enclosure
specified by the interval parameters c±j , a± , and b± for the ILP
ij i method, we obtain λ± = (λ± 1
, λ±
2
) = ([0.12, 0.18], [3.29, 4.96] ). Be-
problem.
cause λ1 > 0, λ2 > 0, from Theorem 4.2, the OSS of the ILP model
± ±
The tool MATLAB R2012 is used to simulate the Monte Carlo
(19) is determined by S ∩ T as follows:
(MC) samples. Corresponding to each coefficient, a set of 10,0 0 0
MC samples is created and applied as input data for the ILP model. S ∩ T ={x = (x1 , x2 ) : 8x1 − 14x2 ≤ 4.2, x1 + 0.19x2 ≤ 7,
In fact, the number of LP models is chosen at random and solved. 10x1 − 12x2 ≥ 3.8, 1.1x1 + 0.2x2 ≥ 6.5}.
Obviously, the OSS of the ILP model is the set of the optimal solu-
tions of these LP models. The optimal solutions obtained through The set S ∩ T is shown in Fig. 4(b).
MCS method for an ILP model will be addressed in the next section Now, the ILP model (19) is solved using the solution methods
by a numerical example. introduced in Section 3. The solution areas of the TSM and the ILP,
Any arbitrary probability density function can be simulated for BWC, ITSM, and MILP methods are shown in Table 1. The worst ob-
each coefficient of the ILP model. In this paper, all of the coeffi- jective function obtained using the BWC method is equal to 22.14.
cients are random variables with uniform distribution. For exam- Obviously, the interval [22.14, 34.86] is the exact interval of ob-
ple, consider the coefficients c1± = [5.2, 6], a± 12
= [−14, −12], and ±
jective function values. The Zopt values corresponding to the BWC
b±3
= [15, 16]. The samples corresponding to c1 ∈ c1± , a12 ∈ a±
12
, and and MILP methods show that the interval of the objective func-
b3 ∈ b± 3
with uniform distribution are shown in Fig. 2; (a)–(c), re- tion of the MILP method is not a subset of the interval [22.14,
spectively. Fig. 2(a) shows that under a uniform distribution, the 34.86]. Therefore, the solution area of the MILP method contains
set of all created samples corresponding to c1 forms the interval non-optimal solutions.
c1± . Because the samples c1 are distributed uniformly in the interval The exact OSS obtained from the new algorithm and the solu-
c1± . Similarly, the other coefficients of an ILP model can be shown. tion area of the methods presented above are shown in Fig. 5.
 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49
Fig. 1. The solution process of the new algorithm.
Fig. 5 also shows that the solution areas of the BWC, ILP and
TSM methods contain infeasible and non-optimal solutions, and
some of the optimal solutions to the ILP model (19) are lost. How-
ever, the solution areas of the MILP and ITSM methods are com-
pletely feasible; they consist of non-optimal solutions, and several
of the optimal solutions to the ILP model (19) are lost. The solution
area of the IILP method is both completely feasible and optimal,
but several of the optimal solutions to the ILP model (19) are lost.
Remark 6.1. To compare the new algorithm with previous solution
methods, the following are given:
(1) The new algorithm is able to obtain the exact OSS of the ILP
model, but the previous methods only obtain some of the optimal
solutions. In addition, parts of the solution areas of the previous
solving methods may be infeasible and/or non-optimal, and some
of the optimal solutions may be lost (as shown in Fig. 5).
(2) In this algorithm, by solving an ILES, the exact OSS of the ILP
model can be obtained, but the solution process of the old meth-
ods is complex.
45
(3) The new algorithm is able to solve a general ILP model
(m ≤=≥ n) when the solution set of the related ILES is positive, but
the method expressed in Allahdadi and Mishmast Nehi (2013) can
solve the ILP model, where (i) m = n, and (ii) all components of
the feasible solutions to the best model are positive.
(4) The new algorithm can solve problems in which some of the
components of the optimal solutions to the worst and best mod-
els are zero, but the method introduced in Allahdadi and Mish-
mast Nehi (2013) cannot.
(5) In this algorithm, a necessary and sufficient condition for
determining the OSS of the ILP model is proved for the general
case (Theorem 4.2), while the old method was only showed to
work for the case m = n.
Now, the objective function in model (19) is changed.
By changing the objective function, the OSS of the ILP
model may be changed. Let the objective function be
Maximization Z ± = [3.8, 4]x±
1
+ [1.8, 2]x±2
. The solutions to
the characteristic model with centre values are Z c ∗ = 39.66 and
(x1 , x2 , s1 , s2 , s3 ) = (4.12, 12.41, 128.23, 0, 0 ). Because s2 = s3 = 0,
46 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49

constraints 2 and 3 are binding. Thus, I = {2, 3}, and J = ∅. Now,

consider the variables λ±1
and λ±2
, which correspond to constraints
2 and 3, respectively. Then, the following ILES will be formed:

[1, 1.1]λ±
1
+ [1, 1.1]λ±
2
= [3.8, 4]
(21)
[0.19, 0.2]λ±
1
+ [0.8, 1]λ±
2
= [1.8, 2]
Solving system (21), gives (λ± 1
, λ±
2
) = ([1.25, 2.97], [1.25, 2.24] ).
Because λ±1
> 0 , λ±
2
> 0 , using Theorem 4.2, the OSS of the ILP
model is specified by S ∩ T as follows:
S ∩ T ={x = (x1 , x2 ) : x1 + 0.19x2 ≤ 7, x1 + 0.8x2 ≤ 16,
1.1x1 + 0.2x2 ≥ 6.5, 1.1x1 + x2 ≥ 15}.
The set S ∩ T is shown in Fig. 6(b).
Now, the ILP model (19) will be solved with another objective
function: Maximization Z ± = [2, 4]x±1
+ [6, 8]x±
2
. The optimal solu-
tion to the standard form of the characteristic model with centre
values is as follows:
Z c = 120.56, (x1 , x2 , s1 , s2 , s3 ) = (0, 17.22, 227.89, 3.39, 0 ).
Fig. 2. Samples chosen within (a) c1± = [5.2, 6], (b) a± = [−14, −12], and (c) b± =
[15, 16] with uniform distributions.
12 3
Because s3 = 0, the third constraint is binding; thus, I = {3}, and
Because x1 = 0, J = {1}. The variable is considered to be λ ± , which
corresponds to the third constraint of model (19), and the variable
v ± corresponds to x1 ≥ 0 (i.e., −x1 ≤ 0). Now, the following ILES is
formed:

[1, 1.1]λ± − v± = [2, 4]
(22)
[0.8, 1]λ± = [6, 8]
Solving system (22) gives (λ± , v± ) = ([5.9, 10], [1.33, 9] ). Because
λ ± > 0, and v ± > 0, the OSS of the ILP model is parallel to S ∩ T as
follows:
S ∩ T ={x = (x1 , x2 ) : x1 + 0.8x2 ≤ 16, 1.1x1
+ x2 ≥ 15, x1 ≥ 0, x1 ≤ 0}
={x = (x1 , x2 ) : x1 = 0, 15 ≤ x2 ≤ 20}.
The set S ∩ T is shown in Fig. 7(b).

Example 6.2. Consider the following ILP model:

Minimization Z ± = [1, 5]x±1
+ [3, 6]x±
2
subject to:
± ± (23)
[2, 5]x1 + x2 ≥ [3, 7]
x±1
, x±
2
≥ 0.
First, the standard form of the characteristic model with the
centre values of model (23) is solved, and its optimal solution is
obtained as Z c = 4.29 and (x1 , x2 , s ) = (1.43, 0, 0 ). Because s = 0,
Fig. 3. The feasible region and extreme zones of the ILP model (19). the constraint of this characteristic model has been binding, and

Fig. 4. (a) Obtained solutions of the MCS method, and (b) the OSS obtained through the new algorithm (or S ∩ T) for ILP model (19).
 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49 47

Fig. 5. The exact OSS of the ILP model (19): (a) the solution areas of the BWC and IILP methods, and (b) the solution areas of the ILP, TSM, MILP, and ITSM methods.

Fig. 6. (a) Solutions obtained by the MCS method, and (b) the OSS obtained using the new algorithm (or S ∩ T).

Fig. 7. (a) The solutions obtained by the MCS method, and (b) the OSS obtained using the new algorithm (or S ∩ T).

I = {1}. In addition, x2 ≥ 0 is binding, so J = {2}. The variable is as follows:

considered to be λ ± , which corresponds to the constraint of the
model (23), and the variable v ± corresponds to x2 ≥ 0. Now, the
following system is formed and solved: S ∩ T ={x = (x1 , x2 ) : 5x1 + x2 ≥ 3, 2x1 +x2 ≤ 7, x2 ≥ 0, x2 ≤ 0}
 ={x = (x1 , x2 ) : 0.6 ≤ x1 ≤ 3.5, x2 = 0}.
[2, 5]λ± = [1, 5]
(24)
λ± + v± = [3, 6]
The set S ∩ T is shown in Fig. 8(b).
The solution set of system (24) is (λ± , v± ) = ([0.2, 2.5], [0.5, 6.79] ). The following example studies the ILP model with 270 con-
Because λ ± > 0, and v ± > 0, the OSS of model (23) is equal to S ∩ T straints and 270 variables.
48 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49

Fig. 8. (a) Solutions to the ILP model (23) obtained by the MCS method, and (b) the OSS obtained using the new algorithm (or S ∩ T).

Example 6.3. Consider the following ILP model: λ±4+9k = [0.02, 0.49], λ±5+9k = [0.009, 0.66], λ±6+9k = [0.32, 0.74],

270 λ±7+9k = [0.20, 0.58], λ±8+9k = [0.20, 0.67], λ±9+9k = [0.008, 0.25].
Maximization Z ± = c±j x±j
j=1 Because for i = 1, . . . , 270 the variables λ±
i
> 0, the OSS of the ILP

270 (25) model (25) is equal to S ∩ T as follows:
subject to: a± x± ≤ b±
ij j i
, i = 1, . . . , 270,
j=1 
270 
270

x±j ≥ 0, j = 1, · · · , 270. S ∩ T = {(x1 , . . . , x270 )| a− x ≤ b+

ij j i
, a+ x ≥ b−
ij j i
, i = 1, . . . , 270}.
j=1 j=1
For k = 0, 1, 2, . . . , 29, the coefficients of the ILP model (25) are
assumed to be as follows: For k = 0, 1, 2, . . . , 29, consider the point x = (x1 , . . . , x270 ),
±
c1+9 = [2.3, 2.5], ±
c2+9 ±
= c5+9 = [−1.3, −1], ±
c3+9 = whose components are: x1+9k = 4.0759, x2+9k = 1.016, x3+9k =
k k k k
± ± ± 3.5759, x4+9k = 4.2933, x5+9k = 1.0243, x6+9k = 3.4356, x7+9k =
[1.6, 1.7], c4+9 = [2 . 8 , 2 . 9] , c = [1 . 5 , 1 . 9] , c =
±
k 6+9k
±
7+9k 10.6746, x8+9k = 2.6655, x9+9k = 4.4732. This point does not lie in
[0.7, 0.9], c8+9 = [1 . 2 , 1 . 5] , c = [0 . 4 , 0 . 5] , and
k 9+9k S ∩ T and is not optimal because it does not satisfy the constraint
a±
(1+9k )(1+9k )
= a±
(2+9k )(4+9k )
= a± (7+9k )(3+9k )
= a± (8+9k )(6+9k )
= 2x1 + 1.5x2 + 3x3 ≤ 20 and it is an infeasible point of the ILP
[2, 3], a±
(1+9k )(2+9k )
= a± (2+9k )(5+9k )
= a± (3+9k )(8+9k )
= model (25).
[1.5, 2], ± ±
a(1+9k )(3+9k ) = a(2+9k )(6+9k ) = [3, 3.4], ±
a(3+9k )(7+9k ) = The point x = (x1 , . . . , x270 ) with components x1+9k =
[0.2, 0.4], a± = [3.1, 3.3], a± = [3.75, 4.25], 1.2855, x2+9k = 0.7468, x3+9k = 2.6798, x4+9k = 1.2241, x5+9k =
(3+9k )(9+9k ) (4+9k )(1+9k )
0.7378, x6+9k = 2.7291, x7+9k = 3.435, x8+9k = 1.7207, x9+9k =
a±
(4+9k )(2+9k )
= a ±
(5+9k )(5+9k )
= [2 . 5 , 3 . 5] , a ±
(6+9k )(9+9k )
=
± ± 3.0862 is a feasible point of the ILP model (25), but it is not
[−1.7, −1.5], a(4+9k )(3+9k ) = a(5+9k )(6+9k ) = [−2, −1.5],
optimal, because it does not satisfy the optimality constraint
a±
(6+9k )(7+9k )
= [0.95, 1.05], a± (6+9k )(8+9k )
= [2.8, 3.5], a± (5+9k )(4+9k )
= 3x1 + 2x2 + 3.4x3 ≥ 15. In addition, it does not lie in S ∩ T.
± ±
[3.5, 4.5], a(7+9k )(1+9k ) = a(8+9k )(4+9k ) = [1, 1.6], a±(7+9k )(2+9k )
= The point x = (x1 , . . . , x270 ) with components x1+9k =
a± = [−7, −6], a± = [1.2, 1.6], a± = 2.52, x2+9k = 0.89, x3+9k = 3.01, x4+9k = 2.52, x5+9k =
(8+9k )(5+9k ) (9+9k )(7+9k ) (9+9k )(8+9k )
[−6.75, −6.25], a± = [2 . 5 , 3] , and the other a ±
= 0 , 0.89, x6+9k = 3.01, x7+9k = 6.5, x8+9k = 2.18, x9+9k = 3.67 be-
(9+9k )(9+9k ) ij
longs to S ∩ T, so it is optimal. In addition, this point is the optimal
and b±
1+9k
= b±
2+9k
= b ±
3+9k
= [15 , 20] , b ±
4+9k
= b ±
5+9k
= b ±
6+9k
=
solution of the characteristic model of the ILP model (25) with
[5, 10], b±7+9k
= b± 8+9k
= b±9+9k
= [4, 6]. centre values.
The standard form of the characteristic model with centre val-
ues is solved, and for k = 0, 1, 2, . . . , 29, the results are as follows:
7. Conclusions
x1+9k = 2.52, x2+9k = 0.89, x3+9k = 3.01, x4+9k = 2.52,
x5+9k = 0.89, x6+9k = 3.01, x7+9k = 6.5, x8+9k = 2.18, Obtaining the OSS of the ILP model is important for researchers.
x9+9k = 3.67, si = 0, i = 1, . . . , 270, In this paper, an arbitrary characteristic model was solved, and the
set of indices of its binding constraints was obtained. Then, using
and Z c ∗ = 931.887. Because si = 0 for i = 1, . . . , 270, the index this index set, an ILES was formed and solved. If all of the com-
set of the binding constraints is I = {i = 1, . . . , 270}. The variables ponents of the interval solutions to this system are strictly non-
λ±i , i = 1, . . . , 270 corresponding to constraints 1 to 270 of the ILP negative, the OSS of the ILP model can be determined as the sub-
model (25) are considered, and the following ILES is formed: scription of the zone created by reversing the signs of the bind-

270 ing constraints of the worst model and the binding constraints of
λ±i ai j ± = c±j , j = 1, . . . , 270. (26) the best model. Numerical examples were introduced to illustrate
i=1 the algorithm. Finally, the ILP model was solved using the MCS
method. The solutions obtained by the MCS method and the OSS
System (26) is solved, and its solutions for k = 0, 1, 2, . . . , 29 are as
obtained using the introduced algorithm were then compared to il-
follows:
lustrate the applicability of this algorithm. The applicability of the
λ±1+9k = [0.11, 0.59], λ±2+9k = [0.05, 0.70], λ±3+9k = [0.16, 0.43], proposed algorithm was studied by solving a large-scale ILP model.
 H.A. Ashayerinasab et al. / Expert Systems With Applications 93 (2018) 39–49
Acknowledgements
The authors would like to thank to the Editor-in-Chief and the
anonymous referees for their constructive comments and sugges-
tions that have helped to improve this paper.
Supplementary material
Supplementary material associated with this article can be
found, in the online version, at 10.1016/j.eswa.2017.10.020
References
Allahdadi, M., & Mishmast Nehi, H. (2013). The optimal solutions set of the interval
linear programming problems. Optimization Letters, 7, 893–1911.
Allahdadi, M., Mishmast Nehi, H., Ashayerinasab, H. A., & Javanmard, M. (2016). Im-
proving the modified interval linear programming method by new techniques.
Information Sciences, 339, 224–236.
Bazaraa, M. S., Jarvis, J. J., & Sherali, H. D. (1977). Linear programming and network
flows (2nd ed.). Newyork: John Wiley and Sons.
Carlin, B. P., Polson, N. G., & Stoffer, D. S. (1992). A monte carlo approach to non-
normal and nonlinear state-space modeling. Journal of the American Statistical
Association, 87, 493–500.
Chib, S., Nardari, F., & Shephard, N. (2002). Markov chain monte carlo methods for
stochastic volatility models. Journal of Econometrics, 108, 281–316.
Fan, Y. R., & Huang, G. H. (2012). A robust two-step method for solving interval
linear programming problems within an environmental management context.
Journal of Environmental Informatics, 19, 1–12.
49
Fiedler, M., Nedoma, J., Ramik, J., Rohn, J., & Zimmermann, K. (2006). Linear opti-
mization problems with inexact data (pp. 79–100). New York: Springer. Chapters
2, 3.
Hladik, M. (2016). Robust optimal solutions in interval linear programming
with forall-exists quantifiers. European Journal of Operational Research, 254(3),
705–714.
Huang, G. H., Baetz, B. W., & Patry, G. G. (1995). Grey integer programming: An
application to waste management planning under uncertainty. European Journal
of Operational Research, 83, 594–620.
Huang, G. H., & Moore, R. D. (1993). Grey linear programming, its solving approach,
and its application. International Journal of Systems Science, 24, 159–172.
Lu, H. W., Cao, M. F., Wang, Y., Fan, X., & He, L. (2014). Numerical solutions compar-
ison for interval linear programming problems based on coverage and validity
rates. Applied Mathematical Modelling, 38, 1092–1100.
Ren, A., Wang, Y., & Xue, X. (2017). A novel approach based on preference-based
index for interval bilevel linear programming problem. Journal of Inequalities and
Applications, 112. Open Access.
Rohn, J. (1993). Cheap and tight bound: The recent result by e. hansen can be made
more efficient. Interval Computations, 4, 13–21.
Rudin, W. (1963). Real and complex analysis (2nd ed.). New York.: McGraw - Hill
Book Company.
Rump, S. M. (1999). INTLAB- INTerval LABoratory. In T. Csendes (Ed.), Developments
in reliable computing (pp. 77–104). Dordrecht: Kluwer Academic Publishers.
Tierney, L., & Mira, A. (1999). Some adaptive monte carlo methods for bayesian in-
ference. Statistics in Medicine, 18, 2507–2515.
Tong, S. C. (1994). Interval number, fuzzy number linear programming. Fuzzy Sets
and Systems, 66, 301–306.
Wang, X., & Huang, G. (2014). Violation analysis on two-step method for interval
linear programming. Information Sciences, 281, 85–96.
Zhou, F., Huang, G. H., Chen, G., & Guo, H. (2009). Enhanced-interval linear pro-
gramming. European Journal of Operational Research, 199, 323–333.