Received August 26, 2019, accepted September 8, 2019, date of publication September 13, 2019,

date of current version September 27, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2941442

A New Two-Stage Approach for a Bi-Objective

Facility Layout Problem Considering Input/
Output Points Under Fuzzy Environment
ARASH MOHAMADI1 , SADOULLAH EBRAHIMNEJAD 2, ROYA SOLTANI3 ,
AND MOHAMMAD KHALILZADEH1
1 Department of Industrial Engineering, Islamic Azad University, Science and Research Branch, Tehran 1477893855, Iran
2 Department of Industrial Engineering, Islamic Azad University, Karaj Branch, Karaj 3148635731, Iran
3 Department of Industrial Engineering, KHATAM University, Tehran 19941633356, Iran
Corresponding author: Sadoullah Ebrahimnejad (ibrahimnejad@kiau.ac.ir)

ABSTRACT Facility layout problem is one of the most important problems in a huge range of industries and
services organizations. Simultaneous study of some qualitative and quantitative parameters like closeness
relationship between facilities, physical constraints such as input/output points and how to arrange facilities
can play a key role to determine the facility layout. Considering these parameters can lead to reduce
production costs, increase production capacity, and remove additional displacements. A two-stage approach
is proposed to achieve these goals. In the first stage, a goal programming model is proposed to determine
weights of the attributes based on the experts’ opinions with different preference representation structures.
Afterwards, the closeness ratings of the facilities are calculated using weights of attributes. In the second
stage, an efficient layout is designed by determining facilities placement sequence, location of the next
facility adjacent to the previous one, location of input/output points, and rectilinear feasible shortest
path between facilities. Two meta-heuristic algorithms including particle swarm optimization and genetic
algorithm are designed due to computational complexity. The objective function is to minimize the sum
of products distance between facilities and closeness rating and also to minimize the dead space. A case
study of an Auto Body Parts company is demonstrated to verify the efficiency of the proposed two-stage
approach. Furthermore, in order to assess the performance of the proposed approach, a comparison is drawn
between the proposed approach and five existing approaches in the literature to solve different problems by
using the two meta-heuristic algorithms.

INDEX TERMS Facility layout problem, different preference representation structures, goal programming,
genetic algorithm, particle swarm optimization algorithm, fuzzy sets.

I. INTRODUCTION layout causes increased work-in-process and manufacturing

The facility layout problem (FLP) is to find an efficient
arrangement of most important assets of an organization,
such as machines, cells, and departments on a planar site.
Since 20-50% of the total operating cost of manufacturing
company and 15-70% of the total cost of manufacturing of
a product are attributed to MHC [1], the most significant
factor to determine the efficiency of a layout is the material
handling cost (MHC) [2]. If facilities are arranged efficiently,
organizations can reduce these costs by at least 10-30% [3]
and enhance their productivity. In return, an inferior facility
The associate editor coordinating the review of this manuscript and
approving it for publication was Seyedali Mirjalili.
lead time.
There are a variety of FLP representation methods, but
most of them fall into two main categories: discrete repre-
sentation and continuous representation. By representing the
FLP in a discrete fashion, the FLP is simplified but at the
penalty of eliminating many solutions from consideration [4].
In the simplest form of this representation, it can be formu-
lated as quadratic assignment problem (QAP) [2]. In the QAP
formulation, all facilities have the same area and the layout
is divided into N equal-sized locations, where each facility is
assigned to exactly one location, and vice versa [5]. However,
discrete representation cannot model the exact location of
facilities and therefore it is not suitable to consider details
like input/output (I/O) points and path design [6]. Continuous

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 134083

representation is more accurate and realistic than discrete
representation and thus it is capable of finding the real opti-
mal solution for FLP. However, the continuous representation
increases the complexity of the problem. In this category,
in order to find an optimal solution for FLP, various math-
ematical models are presented that most of them fall into
mixed integer programming (MIP). Generally, optimizations
methods are useful tools to find an optimal solution for FLP
but the computational time required to solve problem increase
exponentially with the problem size [7]. Therefore, there is a
need for approximated algorithms that provide good subopti-
mal solutions.
Most of the relevant studies in the literature focused only
on MHC as an objective function and neglected other impor-
tant objectives conflicting with it, such as total closeness
rating (TCR), dead space, distance between facilities. In most
real situations in order to avoid establishing a poor layout and
have an efficient layout, it is necessary to consider different
objectives which are in conflict with each other. When the
FLP involves more than one objective, the FLP is treated as a
multi-objective optimization problem (MOOP).
In order to solve a MOOP by using approximated algo-
rithms, two methods are available. The first method is inte-
grated formulation with a weighted sum of some conflicting
objectives and the second method is integrated formulation
with Pareto-based solution methods. The premise of the first
method consists in generating an objective function by aggre-
gating all the sub-functions objective with consideration of
their weights [8]. The second method is to find a set of Pareto-
optimal solutions. In this method, the objectives are defined
separately and by making comparison of the solutions on the
non-dominance criteria, a set of Pareto–optimal solutions are
obtained [9].
Systems development and the complexity of the relation-
ship between them, the different behaviors or insights of
individuals (such as optimism and pessimism, risk-taking and
risk-avoidance), different levels of their knowledge and expe-
rience due to environmental or geographical conditions, indi-
vidual creativity and synergy of all-encompassing view of the
group from different angles, and the perceived coping with
the challenges by using individual judgments and decisions
and their bias effects make it possible to use group decisions
instead of individual ones in order to minimize the impacts of
these biased judgments. Thus this judgment method for indi-
viduals can be defined by different preference representation
structures. Flexible structures representing these preferences
make it possible for all experts to express their own opinions
in the most convenient form which indicates the depth of a
true understanding of the experts in a pragmatic way.
Multi-attribute group decision-making method finds a way
to output the solution which reflects the experts’ opinions as
much as possible [10].
In order to achieve the goals of an organization, the impor-
tance of the effective attributes of the facility layout needs
to be determined. The method proposed in this research
uses the experts’ opinions based on different preference
134084
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
representation structures and determines the weight of each
attribute and the closeness rating.
In order to obtain an efficient layout, it is necessary to
consider the closeness ratings between facilities. The close-
ness ratings show the importance of adjacency of facilities to
each other. The value of closeness rating depends on various
factors such as material flow, environmental communication,
and supervision of personnel. Determining the crisp values
of these factors is almost impossible because a facility layout
plan is drawn up usually before the implementation. Due
to the ambiguity and uncertainty, using Fuzzy Set Theory
makes it possible to determine the closeness rating values
satisfactorily by combining quantitative and qualitative data.
Three important factors affecting the quality of facil-
ity layout are arranging facilities, locating I/O points, and
designing material flow network. In order to simplify the
problem solution, many previous studies have determined
the distance between the two facilities by considering the
distance between the centers; this results in a layout far
away from reality. As a result, determining the facility layout
requires a method that determines I/O points locations and
material flow design. In the present study, a novel approach
is presented to determine the arrangement of facilities, I/O
points locations, and rectilinear feasible shortest path (RFSP).
Two meta-heuristic algorithms are designed and their perfor-
mances are compared with each other in order to implement
this approach.
The type of FLP considered in this research is unequal
area FLP that falls into open field category in terms of layout
configuration. This study makes four main contributions to
the extant literature.
1- A new method is presented to determine closeness
ratings between facilities by using a set of experts’ opinions
based on different preference representation structures.
2- A new method is presented to concurrently deter-
mine facility layout, I/O points of facilities, and material
flow design.
3- Two meta-heuristic algorithms, a genetic algorithm
(GA) and a particle swarm optimization (PSO), are designed
to determine an efficient layout with respect to two objectives;
the sum of products distance between facilities and the
closeness rating, and also dead space.
4- It is verified that the layout obtained from proposed
method is more effective in comparison with some proposed
methods in the literature.
The rest of this study is organized as follows. Section II
reviews the literature. A new two-stage approach is presented
in section III. A case study is presented in section IV to inves-
tigate the applicability of the proposed approach. In section V,
computational experiments are conducted to compare the
performance of the proposed approach with the performances
of existing approaches in the literature. In section VI, anal-
ysis of the results is performed for better understanding of
effectiveness of each of presented methods on the overall
performance of the proposed approach. Finally, section VII
presents conclusions as well as directions for future work.
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
II. LITERATURE REVIEW
Appropriate facility layout is one of the most important
aspects of production, which has a major impact on orga-
nizations’ productivity [11]. Every production process has
its own characteristics and specifications that reflect on the
designing process and the layout pattern. When considering
a facility layout problem, it is desirable to consider multiple
attributes. Methods to address a facility layout problem with
multiple quantitative and qualitative attributes are presented
by [12], [13].
There have been numerous studies adopting
mathematical modelling approaches to find optimal solutions
for FLP. Haddou Amar et al. [8] presented a mathematical
model, aimed at minimizing the material flow and CO2 emis-
sions. Then, they used a rule-based algorithm to determine
the arrangement of facilities. Salmani et al. [14] considered
dimensions of departments dynamic and uncertain in order
to increase the effectiveness of facility layout. Accordingly,
they developed the mixed-integer programming model and
two new objective functions, one of which attempted to
minimize the area of the departments, and the other tried to
maximize the number of adjacencies. Furthermore, in their
study, for each of the objective functions, the upper and
lower bounds were calculated. Gai and Ji [10] presented
an integrated model to solve the FLP. First, their proposed
approach generated a set of layouts using a mathematical
programming model, aimed at minimizing the total trans-
portation cost among all facilities. Then, using experts’
opinions, the score of each alternative was determined by
considering the quantitative and qualitative criteria. It seems
that the generation of some initial layouts by considering
only the transportation cost, and then the ranking of obtained
layouts based on both quantitative and qualitative criteria
can be led to eliminating layouts that do not have the least
scores in terms of transportation cost, but have the highest
scores with respect to other criteria, such as information
flow, equipment flow, personnel flow, supervision links, etc.
Since the FLP is known to be NP-hard optimization prob-
lem [15], a considerable amount of researches focuses on
the improvement of tightness of the MIP. Montreuil [16]
presented a MIP model that hardly solved the facility layout
problem even with five facilities. Noticeable improvements
have been achieved by redefining binary variables, apply-
ing valid inequalities, and branching priorities [17]. Hence
Konak et al. [5] were able to optimally solve problems up to
fourteen facilities.
The existence of a large number of binary variables in
order to avoid overlapping of facilities makes the MIP model
be complicated to solve the FLP. Therefore, exact solution
methods such as mathematical modelling approaches are not
applicable to obtain optimal solutions for large-scale prob-
lems. Hence heuristic and meta-heuristic algorithms based on
MIP models have been designed to obtain efficient solutions.
Liu and Meller [4] proposed a genetic algorithm based heuris-
tic that combined the sequence-pair representation with the
MIP model.
VOLUME 7, 2019
Aiello et al. [18] developed a multi-objective genetic algo-
rithm for FLP with unequal areas. The proposed genetic
algorithm was used to determine the facility layout via
the slicing structure. In their study, four objectives were
the cost of transportation, the horizontal-to-vertical ratio,
the closeness rating and the distance between the facilities.
Paes et al. [19] proposed the two new algorithms, namely
QGA and DRQGA, with regard to unequal areas in an open
space. The first algorithm, by dividing the space into four
quadrants coordinates system, places each facility in such a
way that it is located only in one quarter and do not have
an intersection with horizontal and vertical axes. The second
algorithm acts like the first algorithm, with the difference
that a deconstruction and renewal steps are added to the first
algorithm. The obtained results had better quality 6 to 7 per-
cent on average than the previous ones. Sahin [20] considered
a two-objective facility layout problem (FLP) with prede-
termined weights. These objectives consisted of minimiz-
ing the cost of transporting materials and maximizing TCR.
He used a simulated annealing algorithm to solve the prob-
lem. Xiao et al. [21] proposed a two-stage model. In the
first stage, a zoning algorithm was used to determine the
relative location of the facilities. In the second stage, a linear
programming model was developed to identify the exact
location of the facilities along with their dimensions. They
also designed a simulated annealing algorithm to improve the
placing sequence. Abotaleb et al. [22] presented an algorithm
in which the facilities are not limited to rectangular or square
forms and can have freeform geometric representations.
In order to calculate the distance between facilities, their
proposed algorithm calculated the distance travelled by the
material instead of using the Euclidean or rectilinear distance.
Mohamadi et al. [23] proposed an approach to determine
facilities placement sequence and arrange them adjacent to
each other on the FLP.
In reality, in order to obtain an efficient layout there is
a need to address the other aspects of FLP. One of these
aspects is to consider the I/O points for facilities. Most
approaches assume that I/O points are located at the cen-
troids of facilities and few approaches consider the location
of I/O points on the boundaries of facilities [24].
Deb and Bhattacharyya [25] considered fixed and variable
I/O points on the boundary of each facility. In their study,
genetic algorithm, simulated annealing, and combined algo-
rithms were employed to find the best layout and the results
were compared. Another aspect is material flow design that
should be determined along with I/O points. Most proposed
approaches are based on a sequential approach compris-
ing three steps [26]: determining the arrangement of facil-
ities, the locations of material handling points of facilities,
and design of the underlying material handling system.
Kim and Kim [27] and Rajagopalan et al. [28] proposed
exact and heuristic methods for setting or improving the
I/O locations and for solving material flow design assuming
an already given layout. Few approaches concurrently con-
sider these three aspects [24], [29].
134085

A review of the studies demonstrated that there is a need for
a comprehensive framework to consider all attributes affect-
ing the layout effectiveness. In other words, a framework
that can be used to receive the experts’ opinions depending
on their different behaviors or insights, and reflects experts’
opinions as much as possible. So far, only a few attempts have
been made to provide this framework.
Khilwani et al. [30] considered a multi-criteria facility
layout problem with three different preference representation
structures. In order to integrate different preference structures
into one preference structure, they used transformation func-
tions for each of the preferences and then a psychoclonal
algorithm was proposed to solve a layout design problem.
In general, when different preference structures integrate
into one preference structure, the part of the main informa-
tion of the primary structure may be removed or changed.
Therefore, there is a need for an approach that does not
require the integration of different preference structures into
one preference structure.
Anjos and Vieira [31] comprehensively reviewed the stud-
ies that have been conducted in the field of facility layout.
In their study, three categories of layout problems, including
row layout, unequal-areas facility layout, and multi-floor lay-
out have addressed. Recently, a review study was conducted
by Hoseini-Nasab et al. [32] to classify FLPs. They classified
FLPs in terms of four aspects as follows: layout evolution,
workshop characteristics, problem formulation, and resolu-
tion approaches.
Figure 1. indicates the address of this study in whole
FLPs proposed by Hoseini-Nasab et al. [32].
In order to better show the research gaps, we have com-
pared the recently published papers with each other as shown
in Table 1.
In this paper, there are two novelties: First, a new method
is developed to determine the closeness rating between facil-
ities by considering different attributes and accordingly, each
expert can choose one of the four preference structures so as
to compare the attributes. The main advantage of the proposed
method is to determine the weights of attributes without
integrating different preference structures into one preference
structure. Second, a method is developed for facility layout
with I/O points and also RFSP is determined from output
point of one facility to input point of another facility.
III. THE PROPOSED TWO-STAGE APPROACH
Two key factors to achieve an efficient layout are the deter-
mination of the closeness rating between facilities by consid-
ering different attributes, and a method to arrange facilities.
The proposed two-stage approach presents a method in the
first stage to determine the closeness rating, and a method in
the second stage to determine the facility layout.
A. STAGE ONE: DETERMINING CLOSENESS RATING
In this stage the attributes affecting the facilities layout are
determined and then the experts’ opinions are collected in
order to calculate their weights.
134086
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
Preference information on attributes provided by the
experts is represented in four different formats which are
defined as follows:
(1) As a Preference ordering of the attributes. In this case
the attributes are arranged in order of importance from best
to worst, without providing any additional information.
(2) As a utility function. In this case the utility value of
each attribute is determined.
(3) As a fuzzy preference relation. In this case when
an expert supplies a fuzzy binary relation over the set of
attributes, reflecting the degree to which an attribute is pre-
ferred to another. This degree is between 0 and 1.
(4) As a multiplicative preference relation. In this case
the preferences are represented as the pairwise comparison
matrix with the same ratio scale of the AHP.
In most previous studies [33], [34], the weights of attributes
are determined by integrating the four different preference
structures into one preference structure, leading to miss some
of the information in the process, but in the present study,
because of presenting a goal programming model there is no
need to integrate structures, and as a result, information is
preserved. The proposed model determines the weights of the
attributes in a such way that they have the lowest deviations
from the experts’ opinions. This model is also designed in
such way that can consider different importance for different
experts. Steps to determine the closeness rating are as follows:
Step 1: Forming the decision matrix
A decision matrix consists of the attributes and alternatives
for a decision problem. In this study, the attributes are factors
affecting the closeness rating, and each alternative shows the
relationship between each two facilities. For example, if there
are n
facilities; hence the number of alternatives is equal
to 2n combinations.
Step 2: Comparing the attributes
After forming the decision matrix, the attributes are
compared. These comparisons are based on experts’ opin-
ions in four different formats, i.e., multiplicative preference
relations, fuzzy preference relations, utility functions, and
preferences orderings.
Step 3: Determining weights of the attribute
Let X = {x1 , x2 , . . . , xn } (n ≥ 2) be a finite set of attributes
and W = (w1 ,w2 , . . . , wnP )T be final collective weight vector
of the attributes, where ni=1 wi = 1, wi ≥ 0, i = 1, . . . , n.
Let E = {e1 , e2 , . . . , em } (m ≥ 2) be a finite set of experts
and C P= (c1 , c2 , . . . , cm )T be the weight vector of the experts
where m h=1 ch = 1, ch ≥ 0, h = 1, . . . , m . In the following,
the problem is how to determine final collective weights of
attributes based on the four formats of preference relations.
When the exper’s opinions are expressed a a multiplicative
preference relation or a fuzzy preference relation, if the matri-
ces of these comparisons are fully consistent, the following
relations are true [35], [36].
wi
a∗ij = ∀i, j ∈ {1, . . . , n} (1)
wj
wi
p∗ij = ∀i, j ∈ {1, . . . , n (2)
wi + wj
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

FIGURE 1. The address of this study in whole FLPs.

VOLUME 7, 2019 134087

 A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

TABLE 1. A summary of the features of the recently published papers.

where a∗ij , p∗ij are the elements of fully consistent matrices, as follows:
A∗ and P∗ , respectively. It is hoped that the final collec- wi
tive weights of attributes reflect every expert’s opinions. ahij ≈ i, j = 1, . . . , n; h = 1, . . . , mf (3)
wj
As a result, the relationship between each expert’s opinions w i
phij ≈ i, j = 1, . . . , n; h = mf +1 , . . . , ms (4)
and the obtained final collective weights can be defined wi + wj

134088 VOLUME 7, 2019

A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

m
where values ahij and phij are the elements of the decision matrix X  
+ ch wi − wj − vhij 
 
for each expert. Given (5) and (6), the difference between ahij h=ms+1
and w i h wi
wj as well as the difference between pij and wi +wj are m
X  
represented by the following equations: + ch wi − wj − ghij 
 
h=mr +1
qhij (w) = wi − ahij wj i, j = 1, . . . , n; h = 1, . . . , mf (5) i, j = 1, . . . , n (12)
rijh (w) = wi − phij i, j = 1, . . . , n;


wi + wj
The objective is to minimize the value of zij (w) to narrow the
h = mf +1 , . . . , ms (6)
gap between the collective opinion and each expert’s opin-
When the expert’s opinions are expressed as a utility func- ions. To solve the above problem, a linear goal programming
tion, first, the utility values (ui ) are normalized (7), and model is proposed as follows:
then the difference between normalized values are cal- n X
X n
culated (8). Finally, difference between final collective Minimize z = (dij+ + dij− )
weights and the values obtained from (8) is calculated i=1 j=1
by (9). j6 =i
mf  
X
uhi subject to ch wi − ahij wj
whi = Pn h
i, j = 1, . . . , n; h = ms+1 , . . . , mr (7)
h=1
i=1 ui
ms
vhij = whi − whj i, j = 1, . . . , n; h = ms+1 , . . . , mr
 
(8)
X
+ ch wi − phij (wi + wj )
lijh (w) = wi − wj − vhij i, j = 1, . . . , n; h = ms+1 , . . . , mr h=mf +1
m
(9) X  
+ ch wi − wj − vhij
where whi is the normalized value of attribute i. In fact, h=ms+1
whi is weight of attribute i for expert h. vhij represents the m
X  
difference between two weights of attributes i and j for the + ch wi − wj − ghij − dij+ + dij−
expert h. lijh (w) denotes the difference between (wi − wj ) h=mr +1
and vhij . = 0 i, j = 1, . . . , n; i 6= j
n
If expert’s opinions are presented as a preferences order- X
ing, in order to obtain final collective weights, each expert wi = 1
i=1
presents an optimism level (α). Using both this information
wi ≥ 0 i = 1, . . . , n
and the minimax disparity approach developed by Wang and
Parkan [37], weights of attributes for each expert are deter- dij+ ≥ 0, dij− ≥ 0 i, j = 1, . . . , n; i 6= j (13)
mined. Equation (10) represents the difference between the
weights of attributes for each expert, then difference between The proposed goal programming model not only combines
final collective weights and the values obtained from (10) is four different preference representation structures, but also
calculated by (11). gains the final collective weight of each attribute.
Step 4: Determining scores of the alternatives
ghij = whi − whj i, j = 1, . . . , n; h = mr+1 , . . . , m (10) The scores of the alternatives for this decision making
problem are the values of the final closeness rating between
tijh (w) = wi − wj − ghij i, j = 1, . . . , n; h = mr+1 , . . . , m
each two facilities, which are calculated using Fuzzy Simple
(11) Additive Weighting (FSAW) [38] by considering weights of
the attributes.
where ghij represents the difference between two weights of
attributes i and j for the expert h. tijh (w) denotes the difference B. STAGE TWO: DETERMINING FACILITY LAYOUT BY
between (wi − wj ) and ghij . META-HEURISTIC ALGORITHMS
Considering the important degree of each expert, the devi- The aim of this stage is to present a comprehensive approach
ation degrees, qhij (w), rijh (w), lijh (w), and tijh (w), denoted to select the facilities to enter the layout, determine the
by (5), (6), (9), and (11) are combined to form a new collective location of the next facility adjacent to the previous one,
deviation degree as follows: determine locations of I/O points, and determine RFSP from
mf
X   output point of one facility to input point of another facility.
zij (w) = ch wi − ahij wj  In this study, an appropriate placement sequence of facilities
 
h=1 is selected based on TCR and the maximum closeness rating.
ms
X   In the following, two meta-heuristic algorithms, PSO and
+ ch wi − phij (wi + wj ) GA, are proposed to determine facility layout and facilities
 
h=mf +1 I/O points.

VOLUME 7, 2019 134089


FIGURE 2. Different sections for locating an incoming facility.
FIGURE 3. The proposed string solution for the facility layout.
1) GENETIC ALGORITHM
Determining the solution structure and selecting proper oper-
ators play a key role in the performance of GA. How to choose
chromosomes for the crossover and mutation operators and
select individuals for the next generation are effective factors
on the quality of generated solutions. The main elements of
the proposed genetic algorithm are described below.
a: SOLUTION REPRESENTATION
In this study, the approach provided by Khilwani et al. [30]
is developed and promoted to determine the length and width
of each facility, the facility layout, best I/O points, and also
RFSP. The factors that are considered in the proposed solution
representation include:
- Orientation: A facility can be horizontal or vertical and
therefore parallel to the coordinate axes.
- Proportion: Due to the facility’s flexibility, the proper
length and width for each facility should be considered in
such a way to make sure that the best layout is achieved
proportional to the determined objectives and this ratio
must be in the defined interval between maximum and
minimum aspect ratios.
- Direction: A facility can be placed on either side of the four
sides of the previous facility.
- Section: In order to design an effective layout, each facility
is located at a particular candidate point. Here, the incom-
ing facility can be placed in three positions adjacent to
another facility. For example, in the upper direction of the
facility located on the shop floor, incoming facility can be
placed in three ways, as shown in Fig. 2.
The string solution of the proposed layout problem with-
out I/O points consists of four substrings, i.e., orientation,
134090
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
proportion, direction, and section. The string solution for
L facilities in the proposed FLP is shown in Fig. 3.
Furthermore, in order to determine the locations of
I/O points, each side of a facility is divided into 3 parts,
then one part is selected and one point is chosen randomly
in the selected part as the I or O point. In this situation two
substrings are added to the mentioned string solution. These
two substrings are shown in Fig. 4 to determine the I/O points.
In order to better understanding of the presented structure,
an example is provided to determine layout of three facilities.
Facilities placement sequence and the facility adjacent to
each of them must be determined before determining the
string solution. In this study, facilities placement sequence
and adjacency of facilities are determined based on TCR
and the maximum closeness rating. For details, refer to
Mohamadi et al. [23]. Consider two sets A and B, which indi-
cates facilities placement sequence and the facility adjacent
to each of them, respectively.
A = {2, 3, 1}
B = {2,2}
In order to construct a layout, the first member of the
set A, the facility 2, is placed at the center of the shop floor
first. Then the second member of this set, the facility 3,
is placed adjacent to the first member of the set B, i.e., the
facility 2. In general, the i th facility in the set A is placed
adjacent to the (i-1)th facility in set B; i = 2, . . ., n. Finally,
the facility 1 should be placed adjacent to the facility 2.
Now, how to decode a string solution to a facility layout is
presented. Consider the string solution provided in Fig. 5.
As can be seen, the string solution has six parts as follows:
(1) The first part of the string solution is related to Ori-
entation of facilities in the order existing in set A (facili-
ties placement sequence). The first number, i.e., number 1
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
FIGURE 4. Determination of the I/O points in the string solution.
FIGURE 5. An example for the string solution with three facilities.
indicates that the facility 2 is placed horizontally, and the
facilities 3 and 1 are placed horizontally and vertically,
respectively on the shop floor.
(2) The second part shows the ratio between longer and
shorter sides (Proportion) of each facility. The proportion
value for facility 2, i.e., the first member of set A is equal
to 1.2. As shown in part 1, the facility 2 should be placed
horizontally, so its side length in the x-direction (horizontal
side) is 1.2 times longer than its side length in the y - direction
(vertical side). Likewise, the horizontal side of facility 3,
i.e., the second member of set A is 1.5 times longer than
the vertical side. Moreover, the vertical side of facility 1 is,
i.e., the third member of set A is 1.1 longer than the horizontal
side.
(3) The third part shows that the location of each facility
is on which side(Direction) of their adjacent facility located
in the set B. There is not any information for the facility 2,
i.e., the first member of set A in this part because this facility
is the first facility to enter the layout and it does not have any
adjacent facilities. The number 3 in this part indicates that
facility 3 should be placed above the facility 2 (see Part 3
in Fig. 3.). The number 1 indicates that the facility 1 should
be placed on the right-hand side of the facility 2.
(4) The fourth part of the string solution is related to
Section of facilities. According to the information provided
in Part 4 of Figure 3., if a facility is placed above or below
the another facility, there are three situations for determining
Section which include right, left, and random. These situa-
tions are shown in Fig. 2. Similarly, if a facility is placed
on the right-hand or left-hand side of another facility, there
are three situations for determining Section which include
up, down, and random. By using the information of part 3
VOLUME 7, 2019
(Direction), the facility 3 should be placed above the
facility 2. The number 1 in Part 4 indicates that the selected
Section is right. As a result, the facility 3 should be located
above and to right of the facility 2, as shown in Fig. 2(c).
Moreover, the number 3 in this part indicates that facillity 1
which was already detemined to be placed on the right-hand
side of the facility 2, is randomly located on the right side
(the right vertical side) of the facility 2.
(5) In this part, the coordinates of Input Points of facilities
are specified in the order existing in the set A. Given that the
information presented in Part 5 of the string solution, the input
point of the facility 2 is in the area 3 and also the input points
of facilities 3 and 1 are in the areas 7 and 11, respectively.
These areas are shown in Figure 4. After determining the area
of input point, the exact location of this point is randomly
determined in the selected area.
(6) In this part, the coordinates of Output Points of facilities
are specified in the order existing in the set A. Given that
the information presented in Part 6 of the string solution,
the output points of facilities 2,3, and 1 are in the areas 5,8,
and 9, respectively. Similar to the previous part, the exact
locations of output points are determined.
The Figure 6. shows the final facility layout obtained from
the string solution.
When the number of facilities in FLP increases, the proba-
bility of placing more than one facility on one side (Direction)
of a particular facility increases. Placing several facilities on
one side of a particular facility may not be possible due to
the limitation of the length or width of a facility. Hence in
this study a zone algorithm presented by Xiao et al. [39] is
used. After locating each facility, this algorithm specifies the
possible zones for placing next facility entering the layout.
134091

FIGURE 6. Facility layout obtained from the string solution.
Therefore, if for the incoming facility is not possible to
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
zone algorithm [39] to place incoming facilities next to their
adjacent facilities, the obtained layout is always feasible and
can be evaluated.
c: EVALUATION OF A SOLUTION
After generating a layout, the score of the layout should be
computed. In this study, in order to evaluate the generated
layouts, two criteria, including the sum of products distance
between facilities and closeness rating, and the dead space are
considered.
In order to calculate the rectilinear distance between two
facilities without considering the I/O points, following equa-
tion is used:
dmn = Xmc − Xnc  + Ymc − Ync 
be placed next to its adjacent facility at selected location,
the nearest possible zone is selected and the incoming facility
is placed there.
b: FEASIBILITY OF THE SOLUTION
Generally, after generating a solution, its feasibility should
be checked. When the facilities are placed on the shop floor
   
(15)
where Ymc and Xmc represent the coordinates of the center of
gravity of facility m. Ync and Xnc denote the coordinates of the
center of gravity of facility n.
If the I/O points are defined for each facility, the distance
traveled from facility m to facility n is computed by:
dmn = Xmo − Xmc  + Ymo − Ymc  + RFSPL mn
   
according to a string solution, if they do not overlap with h   i
each other, the string solution is considered as a feasible solu- + Xni − Xnc  + Yni − Ync  (16)
   
tion. The following equations are used to check the overlap
between facilities: where Ymo and X om represent the coordinates of the output
point of facility m. Yni and X in denote the coordinates of
Xi1 − Xj2 ≥ 0 i = 1, . . . , L − 1; j = i + 1, . . . , L the input point of facility n. RFSPL mn is rectilinear feasible
Xj1 − Xi2 ≥0 i = 1, . . . , L − 1; j = i + 1, . . . , L shortest path length from the output point of facility m to the
input point of facility n.
Yi1 − Yj2 ≥0 i = 1, . . . , L − 1; j = i + 1, . . . , L In this study, the actual distance (RFSPL mn ) traveled by
Yj1 − Yi2 ≥0 i = 1, . . . , L − 1; j = i + 1, . . . , L (14) materials is determined as follows. First, facility layout along
with I/O points of each facility is considered (see Fig. 7(a).)
where Xi2 and Xi1 represent coordinates on the right and left Given the generated layout, a graph is designed. The nodes
sides of facility i, respectively. Yi2 and Yi1 denote coordinates of the graph are the facilities vertices and I/O points. The
on the top and bottom sides of facility i, respectively. If one edges of the graph are the sides of the facilities (see Fig. 7(b).)
of the four above relations is verified between two facilities After designing the graph, the shortest path is determined
i and j, it is concluded that the string solution is feasible from the node corresponding to the output point of facility m
and can be evaluated. In this research, because of using to the node corresponding to input point of facility n by

FIGURE 7. Determining the actual distance traveled by materials from the output point of facility m to the input point of facility n.

134092 VOLUME 7, 2019

A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
using the Dijkstra’s algorithm [40] (see Fig. 7(c).) The value
obtained from RFSPmn is represented by (16).
In order to evaluate the generated layout, based on the
distance traveled by the materials and closeness rating,
an evaluation criterion is given as follows:
L X
L
Z1 =
X
ARmn dmn
m=1 n=1
n6=m
where ARmn represents the closeness rating between the two
facilities m and n.
In order to determine dead space in the generated layout,
another evaluation criterion is also given as follows:
    XL
Z2 = Xt2 − Xt1 × Yt2 − Yt1 − Am
m=1
where Am is the area of facility m. Xt2 is the maximum value of
right side of facilities among L facilities. Xt1 is the minimum
value of left side of facilities among L facilities. Yt2 is the
maximum value of the top side of facilities among L facilities.
Yt1 is the minimum value of the bottom side of facilities
among L facilities .Finally, final evaluation criterion is given
by:
Z = Z1 + CZ2
where C is a constant value, representing the preference rate
of one unit of dead space against one unit of the sum of
products distance between facilities and the closeness rating.
In this study, this value is set to 2.
d: CHROMOSOME SELECTION
Selecting parent chromosome for crossover can be done in a
variety of ways. To select the best selection method, differ-
ent methods including Tournament selection with different
tournament sizes, Random Selection, and Roulette Wheel
Selection are compared. The results show that the Random
Selection method leads to better solutions.
e: CROSSOVER OPERATOR
In order to select the best crossover operator, the probabilities
of three crossover operators, including single-point crossover,
two-point crossover and uniform crossover are compared
with each other. The results indicate that a combination of
the two-point and uniform crossover operators (crossover
probabilities 0.3 and 0.7, respectively) leads to the high-
quality solutions.
f: PARAMETER SETTING
The parameters affecting the efficiency of the GA are initial
population, crossover probability, mutation probability, and
the number of generations. Table 17 in the Appendix rep-
resents the factors and the corresponding levels. To set the
parameters of the GA, Central Composite Design (CCD) of
the response surface methodology (RSM) is used.
VOLUME 7, 2019
In order to select the most appropriate parameter values,
a FLP with 8 facilities is considered. The closeness ratings are
given in Table 18 in the Appendix. These values are randomly
generated from 0 to 10.
The data from Table 17 are entered into the Design Expert
software. The value of α is set to 1.3. Furthermore, the
(17) number of factorial and axial experiments is equal to 1 and
the number of the central experiments is 10. The num-
ber of experiments determined by this method is 34. Each
experiment is performed 10 times and its average value is
considered.
In order to get better results from RSM, it is better to get the
obtained results non-scaled. In this study, the obtained results
are non-scaled using the relative deviation index (RDI),
according to (20).
(18)
f∗ −f
RDI = (20)
f∗ −f−
where f ∗ and f − represent the best and worst values for all
experiments, respectively. Quadratic regression is the best
regression relationship by using RSM that can clearly indicate
the relationship between the values of the parameters and
obtained results.
Finally, in the situation which I/O points are not consid-
ered, the best values for the parameters Pop Size, PCrossover,
(19) PMutation and Generations are set to 496, 0.77, 0.3, and 149,
respectively. In the situation which I/O points are considered,
the best values for the parameters Pop Size, PCrossover,
PMutation and Generations are set to 600, 0.74, 0.26,
and 500, respectively. The value of utility of these parameters
is equal to 1.
Figure 8. illustrates the outline of the proposed two-stage
approach by using the GA in the second stage.
g: TIME COMPLEXITY
The time complexity is defined as a function of the input
size n using Big-O notation where n is the number of
facilities. In the presented algorithm, without considering
I/O points for facilities, the most complexity is related to
triple-nested loop. The outer loop counter is It, which counts
up from 1 to npop. In this loop, npop is population size. The
middle loop counter is i, which counts up from1 to n that n is
the numbers of facilities. Finally, the inner loop counter is j,
which counts up from1 to n. The total iterations in this triple-
nested loop is npop∗ n2 . Since npop is a constant number, the
complexity of the presented algorithm is O(n2 ).
The mentioned triple-nested loop is also in the presented
algorithm with considering the I/O points. The difference is
that the Dijkstra’s algorithm is used to determine the shortest
distance between output points and input points inside of the
middle loop and outside of the inner loop. The complexity
of the Dijkstra’s algorithm is O(V2 ) [41] where V is the
number of nodes of a graph. In this study, there are six
nodes for each facility including the facility vertices and I/O
points. As a result, the number of nodes of the designed graph
is 6n and complexity of the Dijkstra’s algorithm is O(36n2 ).
134093

FIGURE 8. Flowchart of two-stage approach by using GA in the second
stage.
Since 36 and npop are constant numbers, the complexity of
the presented algorithm is O(n 3 ).
2) PARTICLE SWARM OPTIMIZATION ALGORITHM
The solution representation, feasibility of the solution, eval-
uation of a solution, and time complexity in PSO algorithm
are the same as the ones in GA. Parameter setting for PSO is
one of the most important factors affecting the ability to find
good solutions, which is discussed the following.
The PSO has 5 parameters a, b1, b2, the number of parti-
cles, and the number of iterations. It should be noted that a
is the coefficient that is given to the current location of the
particle, b1 is the coefficient that is given to the best location
134094
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
TABLE 2. Facilities information.
of the particle, and b2 is the coefficient that is given to the best
location of the group. Table 19 in the Appendix represents the
factors and the corresponding levels.
The mentioned data from Table 19 are entered into Design
Expert software and the value of α is set 1.3. Moreover,
the number of factorial and axial experiments is equal to 1
and the number of the central experiments is equal to 6. The
number of experiments determined by this method is 48. Each
experiment is performed 10 times and its average value is
considered. In the next step, the obtained results are non-
scaled using the RDI.
In this study, the number of segments of each particle with-
out considering I/O points and with considering I/O points
is 4n-2 and 6n-2, respectively. where n represents the number
of facilities.
The parameter a represents the proportion of the segments
of each particle, where values should not be changed. The
parameter b1 represents the maximum proportion of seg-
ments where a particle can be similar to its best individual
position. The parameter b2 is the maximum proportion of the
segments where a particle can be similar to the best position of
all particles. In order to select the most appropriate parameter
values, a FLP with 8 facilities which was mentioned in the
previous subsection is considered.
A linear regression is the best regression relationship by
using RSM that can clearly show the relationship between
the values of the parameters and the obtained results.
Finally, in the situation which I/O points are not consid-
ered, the best values for the parameters a, b1, b2, number of
particles, and number of iterations are set to 0.41, 0.83, 0.5,
366, and 150, respectively. In the situation which I/O points
are considered, the best values for the parameters a, b1, b2,
number of particles, and number of iterations are set to 0.38,
0.79, 0.51, 656, and 390, respectively. The value of utility of
these parameters is equal to 1.
IV. CASE STUDY
An actual company is used in this study to investigate the
applicability of the proposed two-stage approach. The case
study is an Auto Body Parts Company named Chekad in
Safadasht, Iran. In the first stage, the weights of attributes are
determined by the proposed goal programming model. Then
FSAW method is used to determine the closeness rating.
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

TABLE 3. Fuzzy linguistic conversion scales of attributes.

In the second stage, the GA and PSO are used to determine to integrate experts’ opinions, triangular fuzzy numbers
facility layouts for two situations, with I/O points and without merge together by using the arithmetic mean. Table 4 shows
I/O points. Moreover, two obtained layouts are compared the integrated decision matrix.
with each other. Step 2: Comparing the attributes
Area, maximum, and minimum aspect ratios for facilities In ste 2, the opinions of the six experts to compare the
are shown in Table 2. In this case, facilities are company attributes are collected. Each expert can select one preference
departments. representation structure. Two experts use multiplicative pref-
erence structure (e1 , e2 ), one expert applies fuzzy preference
structure (e3 ), one expert employs the utility function (e4 );
A. STAGE ONE: DETERMINING CLOSENESS RATING
and two experts use the preferences ordering (e5 , e6 ). The
Step 1: Forming the decision matrix
opinions provided by six experts (e1 , e2 , e3 , e4 , e5 , and e6 )
In this case study, four effective attributes considered to
to compare the four attributes X = {x1 , x2 , x3 , x4 } are as
determine the facility layout are as follow:
follows:
(1) Flow of material (x1 ): the flow of parts, raw material,    
work-in-process and/or finished products between facilities. 1 5 2 7 1 4 3 5
(2) Supervision of personnel (x2 ): It is measured by the 1 1 1 1
   
 
total number of employees in both facilities that come under
 1 2  1 2
5 5  4 2 
a common supervision. e1 =  1
 e2 =  1 
5 1 5 2 1 3
  
(3) Environmental communication (x3 ): the conditions that 2 3
 
 
exist between facilities, such as noise, temperature, etc. 1 1 1  1 1 1 
1 1
(4) Information communication (x4 ): The communication 7 2 5 5 2 3
between facilities can be measured using a survey of involved (21)
 
personnel. The unit can be the number of communications per 0.5 0.8 0.6 0.7
time unit.  0.2 0.5 0.3 0.4 
e3 =   0.4 0.7 0.5 0.6  e4 = [0.8, 0.4, 0.5, 0.4]

To form a decision matrix, first, the value of each alterna-
tive is expressed in terms of each attribute using linguistic 0.3 0.6 0.4 0.5
variables. Then, these linguistic variables are transformed (22)
into triangular fuzzy numbers by using the fuzzy linguistic e5 = [1, 3, 2, 4] e6 = [1, 4, 3, 2] (23)
conversion scales, as shown in Table 3.
In this case study, the values of relationship between each Step 3: Determining weights of the attributes
two facilities (the values of each alternative) with respect to In step 3, the goal programming model (13) is used to
the mentioned attributes are obtained from the six experts in determine the weights of the attributes for the facility lay-
order to form the decision matrix. Subsequently, these lin- out. This model was coded in General Algebraic Modeling
guistic variables are transformed into triangular fuzzy num- System (GAMS) 25.1. Then GAMS software was applied to
bers by using the fuzzy linguistic conversion scales. In order solve and optimize the proposed goal programming model.

VOLUME 7, 2019 134095

 A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

TABLE 4. The integrated decision matrix of opinions.

In this study, the importance of experts is considered to be TABLE 5. Obtained closeness ratings.
equal.
The weights of Flow of material (x1 ), Supervision of per-
sonnel (x2 ), Environmental communication (x3 ), and Infor-
mation communication (x4 ) are 0.447, 0.147, 0.286 and
0.120, respectively.
Step 4: Determining scores of the alternatives
The final scores are obtained by using the FSAW method.
In fact, the final scores of alternatives are the closeness rat-
ings. These closeness ratings are shown in Table 5.

B. STAGE TWO: DETERMINING FACILITY LAYOUT BY

META-HEURISTIC ALGORITHMS
After determining the closeness ratings, the appropriate lay-
out configuration is determined by the PSO and GA with and
without considering I/O points. MATLAB programming lan-
guage was used for executing these meta-heuristic algorithms
on a core i5-2450M 2.5 GHz PC.

1) WITHOUT CONSIDERING I/O POINTS

In order to determine the suitable layout, each algorithm is
run 5 times. The objective function value of the best solution
for each run is shown in Table 6.
As can be seen in the Table 6, the best solution belongs to
PSO, which is 0.42% better than the best solution from GA.
But, GA has a significant advantage over PSO in terms of run
time. Figure 9(a). shows the best generated layout.

2) WITH CONSIDERING I/O POINTS

In order to determine the suitable layout, each algorithm is
run 5 times. The objective function value of the best solution
for each run is shown in Table 7.
As can be seen the best solution obtained from GA is 6.7%
better than the best solution obtained from PSO. Furthermore,
the run time of the two algorithms is almost the same. The
best layout generated by GA is shown in Fig. 9(b). In this
figure, the I/O points are marked with ∗ and o, respectively.
The coordinates of these points are presented in Table 8.
The triangular points are also used to help pinpoint the
locations of I/O points on the boundary of each facility.
Finally, the final layout by considering aisle structure is
shown in Fig. 9(c).
As shown in Fig. 9 considering the I/O points changes the
location of the facilities, their lengths and widths, the flow
diagram, and dead space value. As can be seen in Fig. 9(a),
if materials are transported from facility 6 to facility 2,
they have to pass through the facility 5, which may not

134096 VOLUME 7, 2019

A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

FIGURE 9. The best layouts generated by the proposed approach for case study.

TABLE 6. Results of two meta- heuristics algorithms without considering TABLE 7. Results of two meta- heuristics algorithms with considering
I/O points for case study. I/O points for case study.

be physically possible. Furthermore, this path and the path layout without I/O points can simplify the problem and makes
of transporting materials from facility 1 to facility 4 inter- it easier to solve, but makes it away from reality, thus making
sect. This intersection causes delays and increases production errors in calculations and predictions. Moreover, designing
time; as a result, production costs increase. In fact, the facility the facility layout without I/O points causes deviation from

VOLUME 7, 2019 134097


TABLE 8. Coordinates of the I/O points obtained from the proposed method and random selection.
the actual value of the objective function. In this case
study, the deviation percentage is about 26% (3960.75 out
of 5350.04). This is a deviation percentage for a company
having six facilities. Now consider huge plants having a
large number of facilities, such as research and development
department, press shop, paint department, assembly depart-
ment, warehouses for raw materials, semi-finished products,
and final products, etc. For this kind of plant, if facility layout
is designed without considering the I/O points, it will incur
considerable expense.
So far it has been clarified how much considering
I/O points are important. Now, in order to determine
I/O points, a comparison is drawn between the proposed
method and random selection.
In random selection, two random numbers are generated
between zero and perimeter of each facility to determine
I/O points. Bottom left corner is defined as origin. Then,
as much as each random number moves from origin clock-
wise on the boundary of facility in order to determine coordi-
nates of I/O points. The coordinates obtained from the points
are shown in Table 8. Objective function value of layout
in Fig. 9(b) along with the I/O points determined by random
selection is 9683.85 whereas this value is 5350.04 with the
I/O points determined by proposed method. In this example,
objective function value increases by 81%. This result shows
that determining the I/O points by the proposed method play
a significant role to reduce transportation costs.
V. COMPUTATIONAL EXPERIMENTS
In this section, computational experiments are conducted to
evaluate the performance of proposed approach in two sub-
sections. In the first subsection, the performance of proposed
approach without the I/O points is compared with the perfor-
mances of the five approaches proposed by Harmonosky and
Tothero [12], Khilwani et al. [30], Xiao et al. [21], McKendall
and Hakobyan [42], and VIP-PLANOPT software [43].
In the second subsection, the performances of two meta-
heuristic algorithms to solve three facility layout problems
with I/O points are compared with each other.
134098
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
A. COMPARISON OF DIFFERENT APPROACHES
WITHOUT I/O POINTS
In this subsection, the results of the proposed approach are
compared with the results of five existing approaches in
the literature [12], [21], [30], [42], [43]. Nine datasets that
fall into three categories are considered for this purpose.
The first category includes datasets A6 (with six facilities),
A8 (with eight facilities) and A10 (with ten facilities) taken
from Mohamadi et al. [23]. The second category includes O7
(with seven facilities), O8 (with eight facilities), and O9 (with
nine facilities) from Bozer and Meller [44], and one dataset
with fairly large size SC30 (with thirty facilities) including
information about one real problem in industry taken from
Liu and Meller [4]. The third category includes datasets L20
(with twenty facilities) from Imam and Mir [45] and L50
(with fifty facilities) taken from VIP-PLANOPT18 commer-
cial software [43]. The rectilinear distance measure is used
to calculate distance between two facilities for all datasets
except L50. For L50 dataset the Euclidean distance measure
is used to.
The results obtained from Harmonosky and Tothero [12]
approach, Khilwani et al. [30] approach, and proposed
approach to solve the first category datasets by using
GA and PSO are shown in Table 9 and 10, respec-
tively. Each of these problems is run 10 times, and the
average value of runs for each problem is shown in
Tables 9 and 10.
As shown in Table 9, by using GA, objective function
value of the proposed approach is less than two other
objective function values of existing approaches. The devi-
ation Percentages are 2.39% and 0.97% on average for
Harmonosky and Tothero [12] and Khilwani et al. [30]
approaches, respectively.
As shown in Table 10, by using PSO, objective func-
tion value of the proposed approach is less than two other
objective function values of existing approaches. The devi-
ation Percentages are 2.28% and 1.33% on average for
Harmonosky and Tothero [12] and Khilwani et al. [30]
approaches, respectively.
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
TABLE 9. The results of the first category datasets using GA.
TABLE 10. The results of the first category datasets using PSO.
TABLE 11. Comparing GA and PSO algorithms without I/O points for the
first category datasets for proposed approach.
TABLE 12. The results of the second category datasets.
In Table 11, the results of GA and PSO for the proposed
approach are compared with each other in terms of deviation
percentage of obtained results from the best results.
Table 11 shows that the results obtained from the PSO
are better than the results obtained from GA. The deviation
percentage is 0.8% on average which is not that considerable.
Table 12 shows the results obtained from proposed
approach by using GA and PSO and also Xiao et al. [21]
approach to solve the second category datasets. Each of these
problems is run 10 times, and the best value of runs for each
problem is shown in this table.
As shown in Table 12, the results obtained from proposed
approach by using GA and PSO and the results obtained from
Xiao et al. [21] to solve O7, O8, and O9 are almost the same.
VOLUME 7, 2019
FIGURE 10. The best obtained layout for SC30 problem.
But results are different for SC30 problem. In this problem,
the results of PSO are better than GA for the proposed
approach. Moreover, the results of PSO are much better than
the results of Xiao et al. [21] approach. The best obtained
layout for SC30 problem is shown in Figure 10.
Table 13 shows the results obtained from proposed
approach by using PSO, McKendall and Hakobyan [42]
approach, and VIP-PLANOPT18 software [43] to solve the
third category datasets. The orientation and proportion parts
in string solution shown in Fig. 3 are not considered because
the length and width of the facilities have already been spec-
ified in this category.
As shown in Table 13, the results of proposed approach
are better than the results of the other two approaches for
134099

TABLE 13. The results of the third category datasets.
L20 and L50 problems. The best obtained layout for L20 and
L50 problems are shown in Figures 13. and 14. in the
Appendix, respectively. The deviation percentages are 0.52%
and 5.64% on average for McKendall and Hakobyan [42] and
VIP-PLANOPT [43] approaches, respectively.
B. COMPARISON OF TWO META-HEURISTIC
ALGORITHMS WITH I/O POINTS
In this subsection, the performances of two meta-heuristic
algorithms for proposed approach with I/O points to solve
the three datasets of the previous subsection are compared.
The results are shown in Table 14. In contrast to the previous
subsection, the GA is better than the PSO in terms of objective
function value and run time. The deviation percentages are
9.1% and 71.67% on average for objective function value and
run time, respectively.
Comparing the best results (objective value) of these
datasets with I/O points with the best results of thesituation
without I/O points indicates that disregarding I/O points leads
to deviation from actual value of the objective function. The
deviation percentages for datasets A6, A8, and A10 are 26%,
28%, and 25%, respectively.
VI. ANALYSIS OF RESULTS
The approach presented in this research is the result of inte-
grating a set of different methods to solve FLP. These methods
are determining facilities placement sequence, how to place
them next to each other, locating the I/O points of each
facility, and RFSP. Analysis of results is performed for better
understanding of effectiveness of each of these methods on
the overall performance of the proposed approach. This anal-
ysis involves investigating three situations in the proposed
approach. In the first situation, facilities placement sequence
is selected randomly, in the second situation, the I/O points
are selected randomly and in the third situation material flow
path is not determined. Now, the effectiveness of each of
these situations is investigated to solve FLP. For this analysis,
three datasets A6, A8 and A10 are considered. Each of these
datasets is run 10 times, and the average value of runs for each
dataset is shown in Tables 15.
As can be seen, random selection of placement sequence
and the I/O points increases the value of the objective function
but disregarding path design causes the objective function
value to either remain unchanged or decrease. Decreasing
the value of the objective function in this situation does not
reflect any improvement in the results, in fact it shows that
134100
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
FIGURE 11. The rectilinear distance between the output point of
facility m and the input point of facility n.
FIGURE 12. Deviation percentage of the results of mentioned situations
from results of the presented approach.
the value of the objective function is moving away from
reality. As shown in Fig. 11, if path design is not determined,
the distance that the materials travel from the output point
of facility m to the input point of facility n is calculated as a
rectilinear distance between these two points.
The path determined in this situation is impossible in
reality; because the output point of facility m is the point at
which the materials should exit from the facility m, but in
this situation, the materials pass through the facility m again
to reach the input point of facility n.
Figure 12. shows the deviation percentage of the results
of mentioned situations from results of the presented
approach. As can be seen, the most deviations percentage are
related to the situation in which the I/O points are selected
randomly and the least deviations percentage are related to the
situation in which path design is not determined. These per-
centage of deviations are 8.71%, 57.85%, and 1.24% on aver-
age for random selection of facilities placement sequence,
random selection of I/O points, and disregarding path design,
respectively.
In the following, a sensitivity analysis is performed for C,
which is the coefficient of dead space, representing the prefer-
ence rate of one unit of dead space(Z2 ) against one unit of the
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP

TABLE 14. Comparing two meta-heuristic algorithms with I/O points.

TABLE 15. Comparing the results obtained from the proposed approach with the results of different situations.

TABLE 16. The best values obtained from the two objective functions in TABLE 17. Closeness rate among eight facilities.
terms of different values of C.

sum of products distance between facilities and the closeness

rating(Z1 ). For this analysis, the dataset A8 is considered. TABLE 18. Factors and selected levels for parameters of the genetic
Table 16 shows the values obtained from the two objective algorithm.

functions in terms of different values of C. As shown in this

table, as C increases, the value of Z2 decreases and the value
of Z1 increases slightly.

VII. CONCLUSION
In this study, a goal programming model is proposed to com-
bine the four different preference representation structures
and determine the weights of attributes based on the experts’
opinions. The main advantage of this model is to eliminate the
need for transformation functions to integrate several pref-
erence structures into uniform preference structure. Using
transformation functions lead to miss some of the information
of the primary structure. Thus the weights of attributes cannot In the second stage, an integrated approach is presented
accurately determine. Another benefit of this model is that it to determine the facility layout. This approach consists of
can consider the importance of the experts. determining facilities placement sequence, determining the
Moreover, in this research a two-stage approach is pro- location of the next facility adjacent to the previous one,
posed to solve facility layout problem. In the first stage, locating the I/O points of each facility, and designing flow
weights of attributes affecting the facility layout are deter- network.
mined by using the goal programming model and then close- An Auto Body Parts company is implemented as a
ness ratings between each two facilities are calculated. case study to demonstrate and validate the applicability

VOLUME 7, 2019 134101


TABLE 19. Factors and selected levels for parameters of PSO algorithm.
FIGURE 13. The best obtained layout for L20 problem.
of the proposed approach. Subsequently, a comparison is
made between the proposed approach and five existing
approaches in the literature. Computational experiments indi-
cate that the proposed approach has a reasonably good per-
formance. Moreover, our findings indicate that determining
the I/O points has a significant effect on the locations of
facilities, their lengths and widths, material flow design, and
finally objective function value. The obtained results show
that without considering I/O points, deviation percentage is
between 25% and 28% for different problem sizes.
A sensitivity analysis is performed to evaluate effective-
ness of each of methods used in proposed approach on
its overall performance. The results indicate that locating
I/O points, determining facilities placement sequence, and
designing flow network have the most impact on objective
function value, respectively.
The limitations of proposed approach in this research are
rectangular shapes of the departments and also the unchanged
closeness rating for all periods of time. In order to enhance
the efficiency of the layout derived from this approach,
a future study may consider the facility layout problem
without imposing a limit for shape of facilities. Furthermore,
134102
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
FIGURE 14. The best obtained layout for L50 problem.
closeness ratings can be considered dynamically for different
time periods. In this situation, a robust layout or several lay-
outs can be determined for different time periods depending
on the type of the problem.
APPENDIX
See Tables 17–19 and Figures 13 and 14.
REFERENCES
[1] A. Mohamadghasemi and A. Hadi-Vencheh, ‘‘An integrated synthetic
value of fuzzy judgments and nonlinear programming methodology for
ranking the facility layout patterns,’’ Comput. Ind. Eng., vol. 62, no. 1,
pp. 342–348, 2012.
[2] S. Emami and A. S. Nookabadi, ‘‘Managing a new multi-objective model
for the dynamic facility layout problem,’’ Int. J. Adv. Manuf. Technol.,
vol. 68, nos. 9–12, pp. 2215–2228, 2013.
[3] V. M. Pillai, I. B. Hunagund, and K. K. Krishnan, ‘‘Design of robust layout
for dynamic plant layout problem,’’ Comput. Ind. Eng., vol. 61, no. 3,
pp. 813–823, 2011.
[4] Q. Liu and R. D. Meller, ‘‘A sequence-pair representation and MIP-model-
based heuristic for the facility layout problem with rectangular depart-
ments,’’ IIE Trans., vol. 39, no. 4, pp. 377–394, 2007.
[5] A. Konak, S. Kulturel-Konak, B. A. Norman, and A. E. Smith, ‘‘A new
mixed integer programming formulation for facility layout design using
flexible bays,’’ Oper. Res. Lett., vol. 34, no. 6, pp. 660–672, 2006.
[6] A. Drira, H. Pierreval, and S. Hajri-Gabouj, ‘‘Facility layout problems:
A survey,’’ Annu. Rev. Control, vol. 31, no. 2, pp. 255–267, 2007.
[7] S. Vitayasak, P. Pongcharoen, and C. Hicks, ‘‘A tool for solving stochastic
dynamic facility layout problems with stochastic demand using either a
genetic algorithm or modified backtracking search algorithm,’’ Int. J. Prod.
Econ., vol. 190, pp. 146–157, Aug. 2017.
[8] S. H. Amar, A. Abouabdellah, and Y. El Ouazzani, ‘‘Interactive facility
layout problem: A bi-objective design,’’ Int. J. Interact. Des. Manuf.,
vol. 12, pp. 151–159, Feb. 2018.
[9] D. Singh and S. Ingole, ‘‘Multi-objective facility layout problems using
BBO, NSBBO and NSGA-II metaheuristic algorithms,’’ Int. J. Ind. Eng.
Comput., vol. 10, no. 2, pp. 239–262, 2019.
[10] L. Gai and J. Ji, ‘‘An integrated method to solve the healthcare facility
layout problem under area constraints,’’ J. Combinat. Optim., vol. 37,
pp. 95–113, Jan. 2019.
[11] S. Kamaruddin, S. Y. Khoo, Z. A. Khan, and A. N. Siddiquee, ‘‘The effect
of layout design on productivity: An empirical study,’’ Int. J. Prod. Qual.
Manag., vol. 7, no. 4, pp. 484–500, 2011.
VOLUME 7, 2019
A. Mohamadi et al.: New Two-Stage Approach for a Bi-Objective FLP
[12] C. M. Harmonosky and G. K. Tothero, ‘‘A multi-factor plant layout
methodology,’’ Int. J. Prod. Res., vol. 30, no. 8, pp. 1773–1789, 1992.
[13] Z. D. U. Durmusoglu, ‘‘A TOPSIS-based approach for sustainable layout
design: Activity relation chart evaluation,’’ Kybernetes, vol. 47, no. 10,
pp. 2012–2024, 2018.
[14] M. H. Salmani, K. Eshghi, and H. Neghabi, ‘‘A bi-objective MIP model
for facility layout problem in uncertain environment,’’ Int. J. Adv. Manuf.
Technol., vol. 81, nos. 9–12, pp. 1563–1575, 2015.
[15] A. Kusiak and S. S. Heragu, ‘‘The facility layout problem,’’ Eur. J. Oper.
Res., vol. 29, no. 3, pp. 229–251, 1987.
[16] B. Montreuil, ‘‘A modelling framework for integrating layout design and
flow network design,’’ in Material Handling (Progress in Material Han-
dling and Logistics). Hebron, Kentucky: Springer, 1991, pp. 95–115.
[17] H. D. Sherali, B. M. P. Fraticelli, and R. D. Meller, ‘‘Enhanced model
formulations for optimal facility layout,’’ Oper. Res., vol. 51, no. 4,
pp. 629–644, 2003.
[18] G. Aiello, G. La Scalia, and M. Enea, ‘‘A multi objective genetic algorithm
for the facility layout problem based upon slicing structure encoding,’’
Expert Syst. Appl., vol. 39, no. 12, pp. 10352–10358, Sep. 2012.
[19] F. G. Paes, A. A. Pessoa, and T. Vidal, ‘‘A hybrid genetic algorithm with
decomposition phases for the unequal area facility layout problem,’’ Eur.
J. Oper. Res., vol. 256, pp. 742–756, Feb. 2017.
[20] R. Şahin, ‘‘A simulated annealing algorithm for solving the bi-objective
facility layout problem,’’ Expert Syst. Appl., vol. 38, no. 4, pp. 4460–4465,
2011.
[21] Y. J. Xiao, Y. Zheng, L. M. Zhang, and Y. H. Kuo, ‘‘A combined zone-
LP and simulated annealing algorithm for unequal-area facility layout
problem,’’ Adv. Prod. Eng. Manag., vol. 11, no. 4, pp. 259–270, 2016.
[22] I. Abotaleb, K. Nassar, and O. Hosny, ‘‘Layout optimization of con-
struction site facilities with dynamic freeform geometric representations,’’
Automat. Construct., vol. 66, pp. 15–28, Jun. 2016.
[23] A. Mohamadi, S. Ebrahimnejad, and R. Tavakkoli-Moghaddam, ‘‘A novel
two-stage approach for solving a bi-objective facility layout problem,’’ Int.
J. Oper. Res., vol. 31, no. 1, pp. 49–87, 2018.
[24] A. Klausnitzer and R. Lasch, ‘‘Optimal facility layout and material
handling network design,’’ Comput. Oper. Res., vol. 103, pp. 237–251,
Mar. 2019.
[25] S. K. Deb and B. Bhattacharyya, ‘‘Solution of facility layout problems with
pickup/drop-off locations using random search techniques,’’ Int. J. Prod.
Res., vol. 43, no. 22, pp. 4787–4812, 2005.
[26] J.-G. Kim and M. Goetschalckx, ‘‘An integrated approach for the con-
current determination of the block layout and the input and output point
locations based on the contour distance,’’ Int. J. Prod. Res., vol. 43, no. 10,
pp. 2027–2047, 2005.
[27] J.-G. Kim and Y.-D. Kim, ‘‘A branch and bound algorithm for locating
input and output points of departments on the block layout,’’ J. Oper. Res.
Soc., vol. 50, no. 5, pp. 517–525, 1999.
[28] S. Rajagopalan, S. S. Heragu, and G. Taylor, ‘‘A Lagrangian relaxation
approach to solving the integrated pick-up/drop-off point and AGV flow-
path design problem,’’ Appl. Math. Model., vol. 28, no. 8, pp. 735–750,
2004.
[29] C. Friedrich, A. Klausnitzer, and R. Lasch, ‘‘Integrated slicing tree
approach for solving the facility layout problem with input and output
locations based on contour distance,’’ Eur. J. Oper. Res., vol. 270, no. 3,
pp. 837–851, Nov. 2018.
[30] N. Khilwani, R. Shankar, and K. Tiwari, ‘‘Facility layout problem:
An approach based on a group decision-making system and psychoclonal
algorithm,’’ Int. J. Prod. Res., vol. 46, no. 4, pp. 895–927, 2008.
[31] M. F. Anjos and M. V. C. Vieira, ‘‘Mathematical optimization approaches
for facility layout problems: The state-of-the-art and future research direc-
tions,’’ Eur. J. Oper. Res., vol. 261, no. 1, pp. 1–16, 2017.
[32] H. Hosseini-Nasab, S. Fereidouni, M. B. Fakhrzad, and S. M. T. F. Ghomi,
‘‘Classification of facility layout problems: A review study,’’ Int. J. Adv.
Manuf. Technol., vol. 94, nos. 1–4, pp. 957–977, 2018.
[33] F. Chiclana, F. Herrera, and E. Herrera-Viedma, ‘‘Integrating multiplica-
tive preference relations in a multipurpose decision-making model based
on fuzzy preference relations,’’ Fuzzy Sets Syst., vol. 122, pp. 277–291,
Sep. 2001.
[34] F. Herrera, E. Herrera-Viedma, and F. Chiclana, ‘‘Multiperson decision-
making based on multiplicative preference relations,’’ Eur. J. Oper. Res.,
vol. 129, pp. 372–385, Mar. 2001.
[35] E. Herrera-Viedma, F. Herrera, F. Chiclana, and M. Luque, ‘‘Some issues
on consistency of fuzzy preference relation,’’ Eur. J. Oper. Res., vol. 154,
no. 1, pp. 98–109, 2004.
VOLUME 7, 2019
[36] T. L. Saaty, The Analytic Hierarchy Process. New York, NY, USA:
McGraw-Hill, 1980.
[37] Y.-M. Wang and C. Parkan, ‘‘A minimax disparity approach for obtaining
OWA operator weights,’’ Inf. Sci., vol. 175, nos. 1–2, pp. 20–29, 2005.
[38] S. Y. Chou, Y. H. Chang, and C. Y. Shen, ‘‘A fuzzy simple additive weight-
ing system under group decision-making for facility location selection
with objective/subjective attributes,’’ Eur. J. Oper. Res., vol. 189, no. 1,
pp. 132–145, 2008.
[39] Y. Xiao, Y. Seo, and M. Seo, ‘‘A two-step heuristic algorithm for layout
design of unequal-sized facilities with input/output points,’’ Int. J. Prod.
Res., vol. 51, no. 14, pp. 4200–4222, 2013.
[40] E. W. Dijkstra, ‘‘A note on two problems in connexion with graphs,’’
Numer. Math., vol. 1, no. 1, pp. 269–271, Dec. 1959.
[41] P. M. Spira and A. Pan, ‘‘On finding and updating spanning trees and
shortest paths,’’ SIAM J. Comput., vol. 4, no. 3, pp. 375–380, 1975.
[42] A. R. McKendall, Jr., and A. Hakobyan, ‘‘Heuristics for the dynamic
facility layout problem with unequal-area departments,’’ Eur. J. Oper. Res.,
vol. 201, no. 1, pp. 171–182, 2010.
[43] VIP-PLANOPT 18 Engineering Optimization Software.
Accessed: Aug. 3, 2019. [Online]. Available: https://planopt.com
[44] Y. A. Bozer and R. D. Meller, ‘‘A reexamination of the distance-based
facility layout problem,’’ IIE Trans., vol. 29, no. 7, pp. 549–560, 1997.
[45] M. H. Imam and M. Mir, ‘‘Automated layout of facilities of unequal areas,’’
Comput. Ind. Eng., vol. 24, no. 3, pp. 355–366, 1993.
ARASH MOHAMADI received the B.S. and M.S.
degrees in industrial engineering from Islamic
Azad University (IAU), South Tehran Branch,
Tehran, Iran, in 2008 and 2011, respectively. He
is currently pursuing the Ph.D. degree in indus-
trial engineering with IAU, Science and Research
Branch, Tehran.
His research interests include facility layout,
fuzzy theory, mathematical modeling, MCDM,
and metaheuristics algorithms.
SADOULLAH EBRAHIMNEJAD received the
B.S. degree from the Iran University of Science
and Technology, Tehran, Iran, in 1986, the M.S.
degree from the Amirkabir University of Tech-
nology, Tehran, in 1993, and the Ph.D. degree
from Islamic Azad University (IAU), Science and
Research Branch, Tehran, in 2001, all in industrial
engineering. He is currently an Associate Profes-
sor with the Department of Industrial Engineering,
IAU, Karaj, Iran. His research interests include
operations research, risk management, supply chain management, operation
management, and fuzzy MADM.
ROYA SOLTANI received the Ph.D. degree from
the Department of Industrial Engineering, Iran
University of Science and Technology, Tehran,
Iran. She is currently an Assistant Professor
with the Department of Industrial Engineering,
KHATAM University. Her current research inter-
ests include operations research and management
science, optimization under uncertainty, robust
and stochastic optimization, reliability optimiza-
tion, logistics network design, and heuristic and
meta-heuristic applications.
MOHAMMAD KHALILZADEH received the
Ph.D. degree industrial engineering from the
Sharif University of Technology, Tehran, Iran.
He is currently an Assistant Professor with the
Department of Industrial Engineering, Islamic
Azad University (IAU), Tehran Science and
Research Branch, Tehran. His research interests
include project management, applied mathemat-
ics, industrial engineering, and human resource
management.
134103