Scientia Iranica E (2017) 24(1), 364{377

Sharif University of Technology

Scientia Iranica
Transactions E: Industrial Engineering
www.scientiairanica.com

An ANN-based optimization model for facility layout

problem using simulation technique
P. Azimi and P. Soo
Faculty of Industrial and Mechanical Engineering, Qazvin Branch, Islamic Azad University, Qazvin, Iran.
Received 31 May 2015; received in revised form 12 September 2015; accepted 4 April 2016

KEYWORDS Abstract. A real manufacturing system faces lots of real-world situations, such as
Facility layout; stochastic behaviors; the lack of attention to this issue is noticeable in the previous research.
Articial neural The aim of this paper is to nd the optimum layout and the most appropriate handling
network; transporters for the problem by a novel solving algorithm. The new model contains two
Discrete-event objective functions including the Material Handling Costs (MHC) and the complication
simulation; time of jobs (makespan). Real-world situations such as stochastic processing times, random
Non-dominated breakdowns, and cross tracs among transporters are considered in this paper. Several
sorting genetic experiment designs have been produced using DOE technique in simulation software and an
algorithm. Articial Neural Network (ANN) as a meta-model is used to estimate the objective functions
in the metaheuristic algorithms. A hybrid non-dominated sorting genetic algorithm (H-
NSGA-II) is applied for the optimization task. The proposed methodology is evaluated
through a real case study. First, simulation model is validated by comparing it with a real
data set. Then, the prediction performance of ANN is investigated. Finally, the ability
of H-NSGA-II in searching the solution space is compared with the traditional NSGA-II.
The results show that the proposed approach, combing simulation, ANN, and H-NSGA-II,
provides promising solutions for practical applications.
© 2017 Sharif University of Technology. All rights reserved.

1. Introduction vehicles. Material handling vehicles such as forklifts,

trucks, and Automated Guided Vehicles (AGVs) are
A facility layout problem is concerned with determining used to transport materials between various points.
the arrangement of machines, departments, or cells The type of these transporters in uences the layout of
on the shop oor. The most important performance the machines and vice versa. In traditional systems,
measure to evaluate the eciency of a layout is the the decisions related to material handling systems
Material Handling Costs (MHC) [1]. Tompkins [2] are made after nalizing the layout or vice versa.
claimed that 20 to 50 percent of the total operating According to Meller and Gau [3], there exists a lack of
expenses in manufacturing were attributed to MHC parallel engineering in selecting the material handling
and eective facility layout could reduce these costs by system with respect to the facility layout. The lack
10 to 30 percent. The ow of materials and the distance of concurrent engineering results in a high degree of
between machines are important determinants of MHC. disharmony between the facility layout and material
Also, MHC depends on the employed material handling handling system. Therefore, the decisions on deter-
mining the type of material handling equipment and
*. Corresponding author. Tel: 021 22425773;
the place of machines should be made simultaneously.
Fax: 021 22425846 While MHC remains the critical index of layout
E-mail addresses: [email protected] (P. Azimi); eciency, shorter cycle times have become much im-
parham soo@yahoo.com (P. Soo) portant in today's manufacturing systems [4]. Today's
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 365

consumer market demands that manufacturers must be Simulation is known as a powerful tool to evaluate
competitive. This requires ecient operation of man- various alternatives of facility layout. Computer simu-
ufacturing plants and their ability to satisfy customer lation has been applied in dierent facility layout prob-
demand as quick as possible. On-time delivery and lems (for example, see [4,7-11]). Gupta [12] employed
short manufacturing cycle times, as practical issues, a simulation model to select the best layout from a set
should be considered during the layout design process. of candidate layouts. He rstly generated a number
This paper considers both MHC and completion of material ows based on a predened probability
time of jobs (makespan) as optimality criteria of distribution and found the optimum layouts of each
the layout. To be more practical, this paper takes ow. He dened the best solution as the layout with
the stochastic nature of transporters handling time each department pair being separated by the average
and transporters failure into account when calculating distance of the generated layouts. Then, the exibility
the makespan. Also, this paper considers eect of of the layout was measured by evaluating its deviation
work ow interference as a major concern, which has from the ideal distances.
previously been considered very poorly, in estimating According to Grajo [13], layout optimization and
the makespan. The cross trac of transporters can simulation are two tasks that are crucial to any facility
result in considerable delays [5]. These delays lead planning and layout study. This is because simulation
to an increase in cycle times of production system. models can re ect all of the attributes of real systems
Consequently, by taking into account the above factors, that are dicult to consider using analytical models for
we estimate the makespan with the highest possible the layout optimization problems. Azadivar and Tomp-
precision. Because of the described complexity of the kins [14] suggested a simulation model with a GA-based
manufacturing system, a closed-form analytical expres- optimization algorithm. In their method, simulation
sion to calculate makespan does not exist. Therefore, models were used to evaluate the objective functions
we utilize an Articial Neural Network (ANN) to and GA-based algorithm was used to search the op-
estimate the makespan. More specically, we rst timum solution. Azadivar and Wang [4] presented a
build a series of random layout designs and then facility layout optimization technique that considered
discrete-event simulation model is used to evaluate the the dynamic features and operational constraints of the
makespan of these designs. The obtained makespan system as a whole. In their proposed approach, the
is applied to structure an ANN. The ANN works as performance measures of system, such as cycle time
makespan estimator during the search process. In and productivity, were evaluated by simulation. Pagell
order to search the solution space, a hybrid algorithm and Melnyk [9] investigated three layouts consisting
based on non-dominated sorting genetic algorithm of the existing worker-paced assembly line, a modied
(NSGA-II) and an adaptive local search are devel- assembly line, and service cells to improve the overall
oped. The remainder of this paper is organized as operation of a service process. They used computer
follows. simulation to stimulate a critical analysis of the pro-
Section 2 reviews the existing literature. Section 3 cess.
presents the mathematical formulation and properties Kulturel-Konak et al. [15] presented a case in
of the problem. Section 4 explains the proposed which the demand rate was considered as a stochastic
methodology. To show performance of the suggested parameter. They also allowed routing exibility for the
method, computational experiments are done in Sec- products so that they were permitted to follow dierent
tion 5. Finally, the possible extensions of this research routes in the facility. They used a simulation approach
are listed along with the conclusions on the proposed to model the uncertainty. In their simulation, the
approach. mean, variance, and covariance of interdepartmental
ows were estimated; then, these estimations were
2. Literature review used in the design process. Tabu search metaheuristic
was employed to solve the problem. Jithavech and
There is no research similar to the work accomplished Krishnan [11] presented a simulation-based method to
in this paper as we review a series of approximately evaluate the uncertainty associated with the layout.
correlated studies. Rosenblatt [6] was the rst who They validated their simulation model against analyt-
introduced the concept of multi-objective approach to ical methods. Results from case studies showed that
facility layout problems. He considered both qualita- the simulation-based procedure resulted in reduction
tive and quantitative objectives together to come up of risk as high as 80%. Zhou et al. [10] introduced
with a multi-objective formulation. Rosenblatt also a method where simulation was combined with Ge-
proposed a graphical method to generate a series of netic Algorithm (GA) to optimize the layout. They
ecient solutions based on the con icting objectives of tested the optimized site layout within a simulation
minimizing the ow cost and maximizing the closeness environment. Altuntas and Selim [16] proposed four
rating. dierent weighted association rule-based data mining
366 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

approaches to solve facility layout problem. They ˆ The distances between machines are determined a
constructed a simulation model and compared the priori;
layouts obtained by the proposed approaches in terms ˆ Flow between machines is deterministic;
of ve performance measures, namely machine utiliza-
tion, total amount of products produced, cycle time, ˆ Machines and locations are of equal size;
transfer time, and waiting time in queue. Dombrowskia ˆ The initial allocation cost of a machine in a location
and Ernst [17] presented a scenario-based simulation is ignored;
approach that used scenario technique, morphological ˆ The transporters' failure may occur, which leads to
analysis, and discrete event simulation to nd out increase in their processing time.
factory layout. Karpe et al. [18] presented a state
of the art review of simulation methods for facility 3.1. Model formulation
layout problems. Azadeh et al. [19] presented an In this section, the nonlinear integer programming
integrated computer simulation-stochastic data envel- formulations of the problem are presented. Before
opment analysis approach to deal with the facility proceeding to the mathematical model, we introduce
layout problem. In their research, computer simulation the indices, parameters, and decision variables:
network was used for performance modeling of each
layout design. The outputs of simulation were average Indices and parameters
time-in-system, average queue length, and average i; j Index of machines
machine utilization. By comparing their study with k; l Index of locations
some of the relevant studies and methodologies in the tr Index of transporter
literature, they revealed the high ability of the method M The number of machines or the
to handle complex layout problems in manufacturing locations
systems. Fi;j Amount of material ow among
This is the rst study that introduces an inte- machines i and j
grated computer simulation, ANN, and H-NSGA-II as
an optimization approach for handling imprecision and Dk;l The distance between the locations k
non-linearity of layout problems in a special case of and l
manufacturing system. Ci;j Unit material handling cost between
machines i and j
3. Problem description
The manufacturing system addressed here consists of
m machines in which n types of parts, each requiring
a set of operations, are to be processed. During
the manufacturing process, the transporters move the
materials from one machine to another until all the
processes are completed. The transporters' handling
time is stochastic with known probability distribution.
A desired design for the system requires an ar-
rangement of m machines in m predened positions
and assignment of transporters to each pair of machines
such that both MHC and makespan are minimized.
Since minimization of MHC does not match minimiza-
tion of makespan, the problem falls into the class of
multi-objective optimization problems. To explain the
con ict between two objective functions, adapted from
Chiang et al. [20], an eight-machine example with the
work ow matrix has been shown in Figure 1. The
solution to this problem using MHC-based layout is
shown in Figure 1(a). While this layout planning will
minimize the MHC, it is clear that there are numerous
points at which transporters interference occurs. An
alternative layout, taking work ow interference into
account, can be providing a work ow in which there Figure 1. Solution to the eight-facility example: (a)
are any con icting work ows (Figure 1(b)). Minimizing MHC, and (b) minimizing work ow
Other assumptions considered are as follows: interference [20].
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 367

F C tr Fixed cost of establishing the the variable MHC and the second term of it is related
transporter type tr to xed cost of the transporters. The rst objective
mtr Maximum available transporter type tr function of the problem is computed by analytical
relationships and it focuses on minimizing the handling
captr Capacity of the transporter type tr costs. Objective function (2), which is evaluated by
ANNs, focuses on minimizing the whole processing
Decision variables completion time. Constraint set (3) states that each
8
<1;
> if the machine i machine is assigned to a location and Constraint set
Xi;k is assigned to the location k (4) guarantees that each location is occupied by only
>
:
0; otherwise one machine. Constraint set (5) allocates a trans-
porter for movement between two special machines
8 only when material ow exists. Constraint set (8)
<1;
> if the transporter tr is selected to transfer controls the maximum available number of each type
tr
Yi;j parts from the machine i to the machine j
> :
of transporters. The decision variables are kept either
0; otherwise at 1 or at 0 by Constraint set (9).
 M M M M TR  
XXXX X Fi;j 4. Proposed ANN-based optimization
MinZ1 =
cap tr  Ci;j
i=1 j =1 k=1 l=1 tr=1 Because of complexity and uncertainty of many real-
 life problems, it is very dicult to create a precise
 Dk;l  Xi;k  Xj;l  Yi;jtr analytical model. In such complicated situations,
simulation is proven as a powerful computer-based
M X
M X
TR tool that can be used instead of the analytical models
X to study the behavior of complex real systems [21].
+ tr  F C tr ;
Yi;j (1)
i=1 j =1 tr=1
Even though simulation models are skillful in capturing
complex system behaviors, simulation is essentially a
MinZ2 = Makespan; (2) test approach and the way to get the optimum solutions
is not clear in it. Therefore, as Fu [21] pointed
M
X out, there is a need to develop algorithms that take
Xi;k = 1; 8i = 1; 2; :::; M; (3) advantage of the optimization technique, while being
k=1 as accurate as simulation. The principle of simulation
M optimization is that during the optimization process,
X objective function and constraints are evaluated by
Xi;k = 1; 8k = 1; 2; :::; M; (4)
simulation model.
i=1
Another weakness of simulation model comes
TR
X from its requirement of a number of replications, which
tr = sgn ;
Yi;j i;j 8i = 1; 2; :::; M 1; usually make simulation a very time consuming pro-
tr=1 cess. In order to take a step towards lling this gap, this
paper proposes a novel simulation-based optimization
8k = i + 1; 2; :::; M ; (5) framework, which integrates the simulation modeling,
8 articial neural network, and metaheuristic optimiza-
<1;
> if Fi;j > 0 tion algorithm. The framework is shown in Figure 2. A
sgni;j
>
8i; j = 1; 2; :::; M series of dierent scenarios are generated. Then, based
:
0; otherwise (6) on these scenarios, discrete event simulation model is
run. This input-output data is used to train an ANN
tr = Y tr ;
Yj;i i;j 8i; j = 1; 2; :::; M ; 8tr = 1; 2; :::; T R; to approximate the objective function. ANN acts just
(7) like a tremendous intelligent brain, which is trained
by simulation data and has the capability to estimate
X1 X
M M the makespan as fast as analytical relationships and as
tr  mtr ;
Yi;j 8tr = 1; 2; :::; T R; (8) accurate as simulation models.
i=1 j =i+1 To search the solution space, a multi-objective
tr 2 f0; 1g; i; j; k = 1; 2; ::; M ;
Xi;k ; Yi;j optimization algorithm, called NSGA-II, has been com-
bined with an adaptive local search. The NSGA-II
8tr = 1; 2; :::; T R: (9) showed the capacity to robustly solve large complicated
multi-objective problems [22]. In our hybrid NSGA-
The rst term of objective function (1) is related to II (H-NSGA-II), ANN is considered as a chromosomes
368 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

Figure 4. An example to input in ANN model.

outputs (Y ) and desired outputs. Finally, the trained
neural network can be used to immediately predict the
simulation results of new congurations.
There are many kinds of neural network models.
Figure 2. Framework of the proposed method. Multi-Layered Perceptron Neural Networks (MLPNNs)
with nonlinear transfer functions have been considered
tness function evaluator. The combination of sim- in the present study. They are purely empirical models
ulation, ANN, and optimization technique provides that can theoretically mimic any relationship to any
an eective means for the high complex optimiza- degree of precision ([26,27]). They consist of three
tion problems. For more information on simulation- layers, including one input layer, one or more hidden
optimization approaches, advances, and applications, layers, and one output layer. The input layer consists of
we refer readers to recent reviews by Swisher et al. [23] the decision variables associated with machines' loca-
and Fu et al. [24]. tion (Xi;k ) and transporters' allocation considerations
tr ); and the output layer gives the outcome of the
(Yi;j
4.1. Articial neural networks process or the makespan. An example of the inputs is
Articial Neural Networks (ANNs) are known as ef- given in Figure 4. In this conguration, machine 1 is
fective techniques for approximating non-linear model assigned to location 3, machine 2 is assigned to location
functions [25]. Considering the highly non-linear re- 1, etc. Also, materials between machines 1 and 2 are
lation between the makespan and the selected trans- moved by transporter 1, materials between machines 1
porters, the ANNs can be eectively applied to nd and 3 are moved by transporter 2, etc.
this indenite relation. ANNs consist of dierent
interconnected processing elements that aim to solve 4.2. Hybrid non-dominated sorting genetic
a specic problem. These small computing elements algorithm
are called neurons (Figure 3). The neuron takes inputs, NSGA-II is one of the contemporary multi-objective
processes them, and transfers the outputs. Neurons are evolutionary algorithms that exhibits high performance
connected to each other by links known as synapses, and has been widely applied in various disciplines. The
and associated with each synapse there is a weight algorithm makes use of a fast non-dominating sorting
factor. First, all the input signals, design conguration approach to discriminate solutions, which is based on
(Xi ), transferred by synapses, have to be multiplied the concept of Pareto dominance and optimality. The
by their own weighting (Wi ). Then, a special value concept of Pareto dominance for minimization problem
bias (b) is added to the signals to generate a value (u). can be expressed as follows:
Finally, an activation function transfers the value (u) to Consider a multi-objective model with a set of
the output (Y ). The bias and activation function form con ict objectives, f (~x) = (f1 (~x); f2 (~x); :::; fn (~x)),
a node. The output can be the input of other nodes. subject to g(x) = (g1 (~x); g2 (~x); :::; gm (~x))  0, where
Through a learning algorithm, all the weights (Wi ) are ~x 2 X . ~x is the decision vector and X is the feasible
iteratively modied to minimize the dierence between solution space. f (~x) is the vector-valued function and
g(x) is a vector of constraints. We say solution ~a
dominates solution ~b if fi (~a)  fi (~b) 8i = 1; 2; :::; n
and 9i : fi (~a) < fi (~b).
The NSGA-II starts with random generation of
population. The binary tournament selection selects
the parents based on the rank and crowding distance.
Then, genetic operations such as crossover and muta-
tion are used to generate the child populations. The
detail of the complete method can be found in Deb's
Figure 3. An example of a neuron for the interested paper [28]. Also, in order to perform a careful search
problem. around the most promising area, an adaptive local
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 369

search is combined with NSGA-II. The local search calculated as follows:

operator helps to intensify the search in various areas PN 1 PN
pointed by the genetic mechanisms that in return can SC = a+1 SC ab ;
a=1 b= (12)
improve convergence towards real Pareto front. The N
local search is applied in a heuristic manner so that 2
it is only applied over some special generations. The
main components of the algorithm and the concepts of in which N is the number of chromosomes in pop-
adaptive local search are explained in the next sections. ulation. Finally, considering a pre-dened threshold
similarity coecient (') and the obtained average
4.2.1. Non-dominated sorting similarity coecient, the local search scheme will be
Before selection is performed, every individual (chro- automatically incorporated into the NSGA-II loop as
mosome) in the population is assigned a rank based on follows:
non-domination. First, the non-dominated solutions 8
are assigned rank 1. Then, the individuals of rank 1 are >apply local search
>
>
>
eliminated and non-dominated solutions are assigned < scheme to NSGA-II
> loop if SC < '
rank 2. This process is repeated until all individuals
>
are classied. The crowding distance metric proposed >
>
> do not use local search
>
by [28] is utilized, where the crowding distance of an :
in NSGA-II loop otherwise
individual is the perimeter of the rectangle with its
nearest neighbors at diagonally opposite corners. Thus, 4.2.3. Chromosome structure
if two individuals have the same rank, the one with a The proposed chromosome consists of two matrices,
larger crowding distance is better. each representing a special area of decision making.
4.2.2. Adaptive local search scheme The rst part shows how machines are placed in
The H-NSGA-II presented in this article uses adaptive locations; the second part represents the allocation
Simulated Annealing (SA) algorithm as the local search of transporters to each pair of machines. Figure 5
because of its good convergence rate. In order to adapt shows an example chromosome in which machine 1 is
SA to optimize multiple objectives simultaneously, the placed in location 1, machine 3 is placed in location 2,
Pareto dominance concept is utilized; this means that machine 2 is placed in location 3, and machine 4 is
a dominated or non-dominated neighbor is treated like placed in location 4. The materials between machine 1
a worse neighbor and moving towards it is done with and machine 2 are moved by transporter 1, materials
a certain probability. The rest of the algorithm is just between machine 1 and machine 3 are moved by
like a typical SA procedure. Since the SA is applied to transporter 4, etc.
all the Pareto front solutions and this may lead to high 4.2.4. Crossover operator
computational eorts, local search scheme is applied The crossover operator combines two chromosomes to
only in some generations. We develop a heuristic produce a new chromosome. We apply two crossover
index, called Similarity Coecient (SC), based on the types that only generate feasible solutions. The rst
executed local search. First, we compute the SC for proposed crossover operator has the following steps;
each pair of chromosomes by Eq. (10):
1. Strings related to machine layout are selected;
SCab = 2. A cross point is randomly selected (cross points
PM n PM PM o 1; :::; M );
i=1 @ (Xia ; Xib ) + i=1 j =1 f@ (Yija ; Yijb )g
; 3. The machine numbers before the cross point of
M (10) parent 1 are copied in the ospring. The remaining
machine numbers are put into empty positions
where Xia and Xib are the locations of machine `i' in according to their relative locations in parent 2.
the chromosomes `a' and `b' and Yija and Yijb are the
selected transporters to handle pares between machine
`i' and machine `j ' in the chromosomes a and b. M 0 is
the number of genes which are not empty. @ (;  ) is the
similarity between two especial genes and is expressed
by Eq. (11);
(
1 if  = 
@ (;  ) = (11)
0 otherwise
The average similarity coecient of the population is Figure 5. An example of chromosome representation.
370 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

Figure 6. An example of crossover type 1.

Figure 9. An example of neighborhood operator.
Figure 6 shows the crossover for a problem with
4 machines. a sub-matrix randomly. Then, two dierent cells are
In order to cross the sub-matrices related to selected randomly and the numbers in these cells are
transporters, a uniform crossover operator is used in exchanged. Figure 9 shows a neighborhood in which
which, rst, a Boolean matrix is generated; then, in sub-matrix 1 changes in genes 2 and 4.
cells in which Boolean matrix is equal to one, the data 4.2.7. Cooling schedule
are lled similar to the ospring from parent 1 and The performance of this algorithm also depends on the
in cells in which Boolean matrix is equal to zero, the cooling schedule, which is relevant to the temperature
ospring is lled similar to that from parent 2. updating function. In the proportional decrement
In order to apply the crossover of type 2, a 1  2 scheme, temperature decreases at steps k and k + 1
reference vector is rst generated. Then, if the rst of the outer loop by:
call of reference vector is 1, the machine layout matrix
of ospring will be copied from parent 1; if the rst Tk+1 = Tk ; (13)
call of reference vector is 0, the machine layout matrix
of ospring will be copied from parent 2. This way is where  is the cooling rate and is obtained by some
repeated for transporters matrix (Figure 7). experiments.
4.2.5. Mutation 4.2.8. Stopping criterion
In this paper, the mutation operation is performed only To limit the number of replications of both NSGA-II
on the machine layout matrix. First, a chromosome and SA algorithms, some convergence experiments are
is randomly selected. Then, to create a hard change, performed and the best criterion is applied as follows.
the genes of the selected chromosome are arranged NSGA-II will be stopped in the case when total
inversely (Figure 8). number of iterations reaches a predened number that
is set according to the result of experimental design.
4.2.6. The neighbourhood structure For stopping SA in a temperature level, rst, we dene
2-change neighborhood is used to dene the neighbor- the set of m iterations as a round. If the mean change
hood in the local search algorithm. First, we select between two successive rounds remains xed within
0.95% condence interval, we reduce the temperature.
For the outer loop of SA, a certain number of iterations
is set as the stopping criterion, and this value is
determined according to the result of experiment.

5. Computational experiments
To demonstrate and validate the simulation-based op-
timization framework proposed in this paper, a real-
life production system is studied. This case study
Figure 7. An example of crossover type 2. involves 6 machines, named M 1, M 2, M 3, M 4, M 5,
and M 6, and 6 products. The demand of product
is known (Table 1) and so the material ow between
machines is clear. The distance between locations and
capacity and speed of the available transporters are,
respectively, shown in Tables 2 and 3. The MHC is
Figure 8. An example of mutation operator. xed at 1 $/meter. The model is coded in Enterprise
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 371

Table 1. Demand rate of products (per day). Table 3. Travel path distance between locations (meter).
Demand Sequence From/to 1 2 3 4 5 6
Product 1 300 1 !3 !6 1 0 10 20 15 20 30
Product 2 200 1 !2 !1 !3 !5 !6 2 10 0 10 20 15 20
Product 3 150 1 !3 !5 !2 !6 3 20 10 0 25 20 15
Product 4 220 1 !2 !3 !6 4 15 20 25 0 10 20
Product 5 150 1 !2 !4 !6 5 20 15 20 10 0 10
Product 6 90 1 !2 !6 6 30 20 15 20 10 0

data. Then, a neural network is trained using this data

set. Finally, some validation tests are conducted and
the ability of designed ANN to predict the makespan
is investigated. In this case, there are 16 factors,
including 6 factors related to machines' location, each
in 6 levels, and 10 factors related to transporters
selection, each in 4 levels. The total number of possible
experiments is 66  104 , which is so computationally
Figure 10. Representation of the real production system intractable. To overcome this problem, the cong-
in ED simulation software.
uration of all experiments is generated by Uniform
Dynamics 8 developer (ED). A graphical representation Design (UD). UD was proposed by Fang [29]. Its
of the system is shown in Figure 10. most important feature is that it decreases the number
The computational experiments are accomplished of experimental congurations, especially when the
in three phases. In the rst phase, we validate the experimental region has many factors and multiple
simulation model through real data analysis. In the levels. According to the uniform design table of
second phase, an ANN is developed and its accuracy the form Un (616 ), the number of experiments can be
in predicting the makespan is investigated. In the last in the range of 17 to 30 (n is the desired number
phase, a comparison between H-NSGA optimization of experiments). Because n should be the common
based on articial neural networks and normal NSGA- multiple of all the levels of factors, it can only be
 optimization based on articial neural networks is 24. Therefore, the uniform design table U24 , shown in
presented. The metaheuristic optimization algorithms Table A.1 in Appendix A, is selected. In Appendix A,
are coded by MATLAB and are implemented in a desk- for factors 1-6 (machines location) the numbers from
top with a 3.20-GHz CPU running Windows 7 (64 bit). 1 to 4 signal location 1, the numbers from 5 to 8
signal location 2, etc. For factors 7-16 (transporters
5.1. Simulation model validation selection), the numbers from 1 to 6 signal transporters
In this section, we investigate whether the simulation type 1, the numbers from 7 to 12 signal transporters
model behaves in accordance with the actual system or type 2, etc. The simulation results of all 24 experiments
not. To assess this, simulation results of 30 days are are presented in Appendix A.
compared with the actual measurement data in terms The present problem had 16 control factors as the
of makespan for each day. The results are presented in input neurons, and the makespan as the single output.
Table 4. As presented in this table, there is an error Thus, our ANN includes 16 input nodes and one output
less than 2%, showing high accuracy of the designed node. The number of nodes in the hidden layer can be
simulation model. estimated by Eq. (14) proposed by Chen and Yang [30]:
5.2. Developing an articial neural network i+o p
h= + N; (14)
In this section, a set of proper scenarios are selected 2
for the proposed simulation model to generate training where i is the number of input nodes, o is the number
Table 2. Characteristics of the transporters.
Capacity Speed (m/s) Fixed No of available
cost ($) transporters
Transporter 1 10 Normal (3,0.5) 500 4
Transporter 2 15 Normal (7,2) 700 3
Transporter 3 10 Normal (3,1) 450 5
Transporter 4 10 Normal (8,2.5) 600 2
372 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

Table 4. Results of real system and simulation model. After determining the structure of ANN, back-
Obtained makespan propagation algorithm is carried out to train the
Days network. The back-propagation algorithm has power-
(minute) % Error
no. ful approximation capacity and is applicable to both
Real system Simulation binary and continuous inputs. The type of transfer
1 387 389 0.52 function employed in this work is a sigmoid function
2 340 347 2.06 (Eq. (15)) at hidden layer and a linear transfer function
3 390 401 2.82 at output layer. Neural Network Toolbox V4.0 of MAT-
LAB mathematical software was used for makespan
4 395 386 2.28 prediction:
5 384 394 2.60
6 364 383 5.22 1
f (x) = : (15)
1+e x
7 400 397 0.75
8 400 407 1.75 For inter-comparisons between the simulated and mea-
9 386 393 1.81 sured makespans using the ANN model, two perfor-
mance measures, i.e. the Root Mean Squared Error
10 342 348 1.75 (RMSE) and coecient of determination (R2 ), are used
11 399 390 2.26 as follows [31]:
12 358 355 0.84 v
u
M  2
13 352 348 1.14 1 uX Yi Yi
RMSE = t (16)
14 396 391 1.26 m i=1 Yi
15 388 394 1.55
16 382 378 1.05 where m is the number of samples, Yi is the actual
response of sample i, and Y^i is the predicted response
17 340 351 3.24 of sample i. According to RMSE = 0.01545 and R2 =
18 383 391 2.09 0:9771, the ANN model has been properly trained and
19 375 379 1.07 has good quality predictions.
20 342 347 1.46 Usually, it is necessary to check the tted model
21 391 380 2.81
to ensure that it provides an adequate approximation
to the new input data. To this aim, the ANN results
22 363 372 2.48 are compared with respect to their deviations from
23 388 393 1.29 the simulation results for 15 new trials, which do
24 356 370 3.93 not belong to the training data set. We propose the
25 367 371 1.09 condence intervals to evaluate the overall performance
of the neural network, because interval estimates are
26 399 411 3.01 much more useful than point estimates for decision-
27 368 361 1.90 making. Each trial is simulated for 10 replications
28 372 374 0.54 and E (Ysim ) and V AR(Ysim ) are combined to form
29 385 382 0.78 condence intervals for each trial (Eq. (17)):
30 378 387 2.38
Interval = E (Ysim )  t 2 ;r
p
V AR(Ysim ): (17)
r
Average 375 379 1.92
As show in Figure 11, in each experiment trail, the
Table 5. The optimal ANN parameters. predicted result by ANN (red points) falls within
the interval obtained by the simulation, and so the
Parameter value capabilities of structured ANN will be proven.
Hidden layers 1
Nodes in hidden layer 14
5.3. Evaluation of the proposed metaheuristic
algorithm
Learning rate 0.015 We tested the optimization framework on numerous
random data sets that dier with respect to input
of output nodes, N is the number experiments, and h is parameters such as products demand, transporters
the numberpof nodes in the hidden layer. Consequently, capacity, material handling costs, etc. The main
2 + 24  14. The optimal ANN conguration,
h = 16+1 challenge in comparing the two multi-objective algo-
which was found experimentally, is summarized in rithms is that they do not try to nd one optimal
Table 5. solution, but a set of Pareto solutions. A good
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 373

Table 6. The parameters of the algorithms and their levels.

Algorithm Parameter Description Low High Optimum
level level level
Pc Percent of cross over 0.6 0.9 0.85
Pm Percent of mutation 0.05 0.15 0.08
N pop Initial size of pop 50 150 130
H-NSGA-II N iter Number of iteration 50 300 230
 Temperature decrement rate 0.96 0.99 0.985
m Number of iterations inside each inner loop 5 30 20
Out L Number of outer loop iterations 50 200 127

Pc Percent of cross over 0.6 0.9 0.80

Pm Percent of mutation 0.05 0.15 0.14
NSGA-II
N pop Initial size of pop 50 150 140
N iter Number of iterations 50 300 260

Figure 12. Performance of the algorithms in terms of

MHC.

spread over the objectives pace considering all the

Figure 11. Comparison of the results obtained by objectives.
simulation and ANN model. Before evaluating the algorithms, both algorithms
are tuned by response surface methodology. The
Pareto front is characterized by (i) converging to explored bound and optimum level of each parameter
the real Pareto-optimal front, and (ii) maximizing are presented in Table 6. Then, the parameters of the
the diversity of the Pareto solutions. Consequently, algorithms are xed to their optimum level and test
the quality of multi-objective optimization algorithms problems are solved by the algorithms. Figures 12
is often dicult to dene precisely by any single and 13 present the results of the proposed H-NSGA-
performance metric. A list of indicators has been II and normal NSGA-II in the context of MHC and
introduced over the past few decades. To analyze makespan. As shown in Figure 12, in terms of the best
the performance of the H-NSGA-II, we compare its MHC index, H-NSGA-II outperforms NSGA-II in all
Pareto fronts with those obtained by common NSGA- the test problems, except for P6. Also, in terms of the
II in terms of metrics Pareto Ratio (PR), Spacing (S), best makespan, H-NSGA-II nds better results than
Overall Pareto Spread (OPS), and Computational time NSGA-II in all problems, except for P5 (Figure 13).
(CPU time). PR is the ability of a multi-objective Furthermore, the average values for both objectives
optimization method to produce non-dominated so- considerably improve with the introduction of the
lutions. S measures the standard deviation of the adaptive scheme. We also analyze the performances
distances among solutions of the Pareto front. The of the algorithms using statistical tests.
smaller the value of the spacing metric, the better are Appendix B highlights the values of PR, S, OPS,
the solutions spread along the obtained front. OPS and CPU time performance measures. As discussed
quanties how widely the non-dominated solutions above, these scales are used to measure convergence
374 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

some generations of NSGA-II may lead to increase in

computational time, large increase in solution quality
and speed of convergence can compensate it.

6. Conclusion and remarks

This paper presents a new model and a novel solving
approach to solve facility layout optimization problems
for manufacturing systems with dynamic character-
istics. The proposed approach integrates computer
simulation, ANN, and H-NSGA-II techniques to over-
come the limitations of traditional layout optimization
Figure 13. Performance of the algorithms in terms of methods. The main motivation behind choosing this
makespan. problem is the necessity for integrating decisions for the
layout of machines and material handling vehicles. The
and diversity of Pareto-front for handling multiple results show that the application of an ANN model can
objectives. As presented in Appendix B, the H-NSGA- predict the makespan in a complicated manufacturing
II approach with adaptive local search shows better system, eciently. Despite the previous models, this
behavior than the approach without local search. In paper considers real aspects of material handling such
order to statistically analyze the results shown in as random breakdowns, random processing times, and
Appendix B, we use Student's t-test. The hypothesis waiting times during the handling process. Since
is: these aspects are stochastic variables and theoretically
H0 : 1 2 = 0; dicult to obtain, the ANN model, which has been
trained by the simulation model results, computes the
H1 :  1 2 6= 0; problem makespan.
In the experimental results, there were 66 alterna-
in which 1 is the obtained average value by H-NSGA- tives for the layout of machines and 104 alternatives for
II and 2 is the obtained value by NSGA-II algorithm. selecting the transporters; therefore, there were a total
The results are shown in Table 7. According to Table 7, number of 66  104 designs. However, in the experiment,
it can be seen that in terms of PR and CPU time index, only 24 simulation congurations were run to train
there is no meaningful dierence between algorithms. the data set needed for ANN model. A case study
But, H-NSGA-II outperforms NSGA-II with respect to was presented and the simulation results were given
spacing and OPS metrics. The last column of Table B.1 for validation of the ANN results. In all experiments,
in Appendix B shows the computational times of the the ANN precision was 95%. Finally, a hybrid non-
algorithms for each test problem. Although there was dominated sorting genetic algorithm was proposed to
no signicant CPU time dierence between algorithms, search the solution space. The performance of the
according to \Mean" statistics, it can be claimed that proposed algorithm was compared with that of the
H-NSGA-II is relatively faster than NSGA-II. This normal non-dominated sorting genetic algorithm. In
statement implies that although embedding the SA in terms of PR and CPU time criteria, both algorithms
Table 7. Results of two-sample T -test.
Source No Mean St. Dev 95% CI for dierence T -Value P -Value
H-NSGA-II 30 0.676 0.206
PR (-0.0697, 0.1217) 0.54 0.588
NSGA-II 30 0.650 0.168

H-NSGA-II 30 0.557 0.177

OPS (0.0282, 0.1812) 2.75 0.008
NSGA-II 30 0.452 0.110

H-NSGA-II 30 6.02 5.33

Spacing (-9.29, -0.27) -2.14 0.038
NSGA-II 30 10.8 11.0

H-NSGA-II 30 1116 511 -

CPU time (-405, 169) 0.82 0.413
NSGA-II 30 1234 579
 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 375

performed similar. But, in terms of OPS and spac- changes: a genetic algorithm approach", Eur. J. Oper.
ing metric criteria, the performance of the proposed Res., 113, pp. 169-182 (1999).
hybrid algorithm was statistically better. For future 15. Kulturel-Konak, S., Smith, A.E. and Norman, B.A.
research, one may extend unequal area constraint \Layout optimization considering production uncer-
of departments, which demonstrates a more realistic tainty and routing exibility", Int. J. Prod. Res.,
representation of real-world manufacturing facilities. 42(21), pp. 4475-4493 (2004).
Also, other factors aecting the makespan, such as 16. Altuntas, S. and Selim, H. \Facility layout using
human factors, can be investigated. weighted association rule-based data mining algo-
rithms: Evaluation with simulation", Expert Syst.
References Appl., 39, pp. 3-13 (2012).
1. Singh, S.P. and Sharma, R.R.K. \A review of dierent 17. Dombrowski, U. and Ernst, S. \Scenario-based simula-
approaches to the facility layout problems", Int. J. tion approach for layout planning", 8th CIRP Conf. on
Adv. Manuf. Technol., 30, pp. 425-433 (2006). Intell. Comput in Manuf. Eng., 12, pp. 354-359 (2013).
2. Tompkins, J.A., White, J.A., Bozer, Y.A. and Tan- 18. Karpe, A.B., Kulkarni, C.N. and Jahagirdar, R.S.
choco, J.M.A., Facilities Planning, Wiley, New York \Simulation methodology for facility layout problems",
(2003). Int. J. Res. Eng. Sci. Technol., 1, pp. 1-10 (2011).
3. Meller, R.D. and Gau, K.Y. \The facility layout prob- 19. Azadeh, A., Nazari, T. and Charkhand, H. \Opti-
lem: Recent and emerging trends and perspectives", mization of facility layout design problem with safety
J. Manuf. Syst., 15(5), pp. 351-366 (1996). and environmental factors by stochastic DEA and
simulation approach", Int. J. Prod. Res., 1(1), pp. 12-
4. Azadivar, F. and Wang, J. \Facility layout optimiza- 17 (2014).
tion using simulation and genetic algorithms", Int. J.
Prod. Res., 38(17), pp. 4369-4383 (2000). 20. Chiang, W., Kouvelis, P. and Urban, T. \Single and
5. Egbelu, P.J. and Tanchoco, J.M.A. \Potentials for bi- multiobjective facility layout with work ow interfer-
directional guide-path for automated guided vehicle ence considerations", Eur. J. Oper. Res., 174(3), pp.
based systems", Int. J. Prod. Res., 24(5), pp. 1075- 1414-1426 (2006).
1097 (1986). 21. Fu, M.C. \Optimization for simulation: Theory vs.
6. Rosenblatt, M.J. \The facilities layout problem: A practice", Inf J. Comput., 14, pp. 192-215 (2002).
multi-goal approach", Int. J. Prod. Res., 17(4), pp. 22. Wang, G., Yan, Y., Zhang, X., Shangguan, J. and
323-332 (1979). Xiao, Y. \A simulation optimization approach for
7. Savsar, M. \Flexible facility layout by simulation", facility layout problem", Proc. of the IEEE Int'l. Conf.
Comput. Ind. Eng., 20(1), pp. 155-165 (1991). on Ind. Eng. Eng. Manage., pp. 734-738 (2008).
8. Morris, J.S. and Tersine, R.J. \A simulation compari- 23. Swisher, J.R., Hyden, P.D., Jacobson, S.H. and
son of process and cellular lay-outs in a dual resource Schruben, L.W. \A survey of recent advances in
constrained environment", Comput. Ind. Eng., 26(4), discrete input parameter discrete-event simulation op-
pp. 733-41 (1994). timization", IIE Trans., 36, pp. 591-600 (2004).
9. Pagell, M. and Melnyk, S.A. \Assessing the impact of 24. Fu, M.C., Glover, F.W. and April, J. \Simulation
alternative manufacturing layouts in a service setting", optimization: a review, new developments, and appli-
J. Oper. Manage., 22(4), pp. 413-29 (2004). cations", In: Kuhl ME, Steiger NM, Joines JA, Eds.,
10. Zhou, F., AbouRizk, S.M. and AL-Battaineh, H. Proceedings of the 2005 Winter Simulation Conference
\Optimization of construction site layout using a hy- (2005).
brid simulation-based system", Simul. Modell. Pract. 25. Ashtiani, A., Mirzaei, P.A. and Haghighat, F. \Indoor
Theory, 17, pp. 348-363 (2009). thermal condition in urban heat island: Comparison
11. Jithavech, I. and Krishnan, K.K. \A simulation-based of the articial neural network and regression methods
approach for risk assessment of facility layout de- prediction", Energy Build., 76, pp. 597-604 (2014).
signs under stochastic product demands", Int. J. Adv. 26. Funahashi, K. \On the approximate realization of
Manuf. Technol., 49(1), pp. 27-40 (2009). continuous mappings by neural networks", Neural
12. Gupta, R.M. \Flexibility in layouts: A simulation Netw., 2(1), pp. 83-92 (1989).
approach", Material Flow., 3, pp. 243-250 (1986). 27. Hornik, K., Stinchcombe, M. and White, H. \Multi-
13. Grajo, E.S. \Strategic layout planning and simulation layer feed forward networks are universal approxima-
for lean manufacturing a layOpt tutorial", Proceedings tors", Neural Netw., 2, pp. 359-66 (1989).
of the 1996 Winter Simulation Conference, pp. 564-568 28. Deb, K., Pratap, A., Agarwal, S. and Meyarivan, T.
(1996). \A fast and elitist multiobjective genetic algorithm:
14. Azadivar, F. and Tompkins, G. \Simulation optimiza- NSGA-II", IEEE Trans. Evol. Comput., 6, pp. 182-
tion with quantitative variables and structural model 197 (2002).
376 P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377

29. Fang, K.T. \The uniform design: application of Biographies

number-theoretic methods in experimental design",
Acta Math. Appl. Sin., 3, pp. 363-372 (1980). Parham Azimi was born in 1974. He received his
30. Chen, M.C. and Yang, T. \Design of manufacturing BS, MS, and PhD degrees in Industrial Engineering
systems by a hybrid approach with neural network from Sharif University of Technology in 1996, 1998,
meta modelling and stochastic local search", Int. J. and 2005, respectively. Currently, he is the Assistant
Prod. Res., 40, pp. 71-92 (2002). Professor of Industrial Engineering at Islamic Azad
University (Qazvin Branch). His research interests
31. Geyikci, F., Kilic, E., Coruh, S. and Elevli, S. \Model- include optimization via simulation techniques, facility
ing of lead adsorption from industrial sludge leachate
on red mud by using RSM and ANN", Chem. Eng. J., layout problems, location allocation problems, graph
183, pp. 53-59 (2012). theory, and data mining techniques.
Parham Soo Parham Soo is a PhD candidate at
Appendix A the Faculty of Industrial and Mechanical engineering,
The uniform design of experiments is shown in Table Azad university, Qazvin branch. His research elds are
A.1. MCDM and Simulation optimization. He has been also
a project manager at the Iranian Fuel Conservation
Appendix B Organization for designing and production of 500,000
pcs. of CNG valves and also exporting of LPG valves
The multi objective performance measures obtained for to the European markets including Germany and the
the algorithms is shown in Table B.1. Netherlands.
Table A.1. Uniform design of experiments.
Exp. Factors Makespan
no. X1 X2 X3 X4 X5 X6 Y12 Y13 Y23 Y24 Y26 Y35 Y36 Y46 Y52 Y56 (min)
1 24 2 5 16 10 19 13 9 1 8 4 8 3 10 21 10 429
2 8 24 20 14 1 12 15 8 2 15 24 15 12 11 2 5 672
3 9 4 17 23 16 19 7 12 12 20 21 20 7 4 22 20 570
4 2 17 11 6 15 22 4 11 3 21 11 4 20 9 19 21 414
5 4 10 5 20 23 12 23 23 6 14 14 14 11 6 15 4 712
6 14 7 22 8 6 24 10 16 15 13 1 12 24 7 20 1 514
7 1 8 21 12 14 20 12 10 24 10 16 21 17 24 6 6 957
8 11 23 4 15 11 8 20 13 17 1 5 17 22 5 8 22 408
9 18 13 3 11 2 16 5 24 19 17 12 16 10 19 23 24 737
10 7 3 14 17 23 11 17 4 8 7 7 22 13 20 17 16 868
11 21 12 19 2 22 17 2 5 9 4 13 18 18 3 14 9 344
12 16 17 6 3 7 23 22 2 7 22 20 11 23 17 11 14 1003
13 23 18 12 3 9 7 21 14 22 19 18 24 6 12 16 3 1028
14 19 1 23 9 13 8 24 15 5 3 15 3 8 21 7 23 431
15 13 4 18 21 12 6 1 17 10 23 2 19 1 23 9 7 739
16 17 22 15 13 21 1 15 18 13 9 23 7 21 22 24 11 667
17 15 6 3 11 20 24 6 3 18 12 22 13 4 8 3 19 457
18 22 19 13 2 6 12 9 20 4 11 6 23 14 13 4 18 692
19 12 1 6 15 18 24 18 22 14 24 8 6 16 15 1 8 834
20 3 21 21 10 24 18 19 6 20 16 3 10 5 16 13 15 794
21 20 11 16 21 5 2 14 7 23 18 10 1 15 2 5 13 529
22 5 3 20 13 9 15 3 21 21 5 19 9 19 14 12 17 449
23 10 16 2 24 17 7 7 1 16 2 9 2 9 18 18 2 290
24 6 20 8 3 12 21 11 19 11 6 17 5 2 1 10 12 317
P. Azimi and P. Soo/Scientia Iranica, Transactions E: Industrial Engineering 24 (2017) 364{377 377

Table B.1. Multi-objective performance measures obtained for the algorithms.

Pareto Ratio Overall Pareto Spacing CPU time
(PR) Spread (OPS) (S) (sec)

H-NSGA-II

H-NSGA-II

H-NSGA-II

H-NSGA-II
No.

NSGA-II

NSGA-II

NSGA-II

NSGA-II
Problem

P1 0.808 0.593 0.520 0.399 2.57 8.2 571 610

P2 0.818 0.607 0.652 0.628 3.26 5.93 489 525
P3 0.898 0.714 0.531 0.576 1.66 2.17 505 544
P4 0.880 0.696 0.695 0.600 6.73 9.55 436 441
P5 0.870 1.352 0.236 0.435 3.8 11.51 642 652
P6 0.748 0.787 0.283 0.241 4.6 2.56 566 610
P7 0.499 0.737 0.598 0.512 5.33 8.99 664 685
P8 0.834 0.431 0.195 0.416 2.14 11.11 521 562
P9 0.981 0.806 0.398 0.458 8.38 0.92 609 638
P10 0.959 0.663 0.748 0.418 2.19 18.95 400 394
P11 0.662 0.717 0.842 0.531 5.05 10.83 1038 1182
P12 0.895 0.660 0.566 0.669 7.96 0.81 1113 1220
P13 0.520 0.635 0.471 0.418 2.55 8.55 975 1083
P14 0.980 0.735 0.780 0.478 6.15 2.08 964 1012
P15 0.942 0.729 0.613 0.361 0.77 1.99 913 954
P16 0.662 0.648 0.458 0.354 3.04 3.66 963 1038
P17 0.465 0.690 0.856 0.239 18.86 39.54 1226 1384
P18 0.566 0.727 0.457 0.456 10.84 19.36 1294 1369
P19 0.674 0.632 0.516 0.510 1.69 3.55 1096 1231
P20 0.426 0.640 0.748 0.400 4.93 1.71 1146 1264
P21 0.500 0.433 0.433 0.306 10.22 18.66 1996 2181
P22 0.556 0.606 0.489 0.394 0.9 1.3 1767 1910
P23 0.485 0.586 0.462 0.546 2.87 6.25 1791 2078
P24 0.660 0.480 0.438 0.507 1.66 2.68 1933 2235
P25 0.590 0.438 0.749 0.392 9.92 17.18 1612 1863
P26 0.680 0.674 0.408 0.643 10.75 17.29 1749 1993
P27 0.360 0.550 0.856 0.529 23.52 36.98 1568 1798
P28 0.318 0.472 0.547 0.445 2.33 5.36 1752 1968
P29 0.243 0.618 0.695 0.390 3.36 7.54 1648 1837
P30 0.793 0.432 0.471 0.319 12.58 38.73 1538 1769