Int J Adv Manuf Technol (2013) 68:2215–2228

DOI 10.1007/s00170-013-4820-5

ORIGINAL ARTICLE

Managing a new multi-objective model for the dynamic

facility layout problem
Saeed Emami & Ali S. Nookabadi

Received: 3 February 2011 / Accepted: 28 January 2013 / Published online: 21 February 2013
# Springer-Verlag London 2013

Abstract In today’s manufacturing environment, the facil- in manufacturing systems. Facility layout planning plays an
ity layout needs to be adaptable to changes. This situation important role in the modern manufacturing systems which
requires the solution of the dynamic layout problem. But in seriously impact a company’s competitiveness [1]. The most
previous studies of dynamic facility layout optimization, the significant indicator for evaluating the efficiency of a layout is
main objective is to minimize the sum of the re-arrangement the material handling cost. Since 20–50 % of a manufacturing
and material handling costs. To be more realistic, each of company’s total operating costs is attributed to material han-
these cost terms in objective function might be of different dling, companies can reduce these costs and improve their
importance to decision makers. In this research, the objec- productivity if their facilities are arranged effectively [2].
tive function has been considered as two distinct functions. The current volatile changes in market will demand that
This formulation enables decision makers to apply their own layout to be flexible enough to adapt any dynamic changes in
views. On the other hand, in the proposed model the product design, process design, and schedule design [1].
adjacency-based objective aims at maximizing adjacency Actually, due to fluctuations in product volume and mix, the
scores between the facilities in a facility layout has also materials flow does occasionally change. For example, con-
been used. The proposed multi-objective model was defined stantly changing product volumes and mix models of mobile
as a complex combinatorial optimization problem. It has phones and short life cycles of industrial products makes man-
also been the objective of the present study to evaluate some ufacturing companies to frequently update the facility layout
of the known methods that have been proposed to solve the [3]. In addition, the innovations in manufacturing technologies
multi-objective problem. The results for test problems also make it necessary to modify facilities as needed [4].
showed that the population based metaheuristic methods Depending on the nature of the material flow data, the
are useful tools in solving proposed model. layout problem is classified into static and dynamic layout
problems. When the flow of materials between the depart-
Keyword Multi-objective optimization . MODFLP . ments does not change over time, this problem is known as
MADM . Nondominated solution set the static facility layout problem (SFLP) and in the simplest
form can be formulated as a quadratic assignment problem
(QAP) [5]. When material flows between departments
1 Introduction change during the planning horizon, the SFLP problem
becomes one of dynamic facility layout (DFLP). The
Facility layout decisions involve designing the arrangement DFLP depends on the material flow changes that are antic-
of elements (departments, machinery, human resources, etc.) ipated to occur in the future. The future is divided into a
number of time periods. The flow data for each period are
predicted and it is presumed that the flow data throughout
S. Emami (*) : A. S. Nookabadi
each period remain constant [6].
Department of Industrial & Systems Engineering,
Isfahan University of Technology, Isfahan 84156-83111, Iran The DFLP is the problem of arranging facilities on the
e-mail: [email protected] plant layout for multiple periods. The solution for the SFLP
A. S. Nookabadi is a single layout, and the solution for the DFLP is a layout
e-mail: [email protected] plan. A layout plan for the DFLP can be represented as a series
2216 Int J Adv Manuf Technol (2013) 68:2215–2228

of layouts each of which is associated with a period. The studies included above share a common assumption
Therefore, the layout plan in the planning horizon is deter- that all the departments are equal size. There are some other
mined that the sum of the material handling cost and the cost studies which do not make this assumption. Three recent
of re-arranging departments in consecutive periods is mini- examples of such studies are Dunker et al. [21], Mckendall
mized. Re-arrangement cost is incurred when the layout is and Hakobyan [22], and Jolai et al. [23].
rearranged after consecutive periods [7]. In a manufacturing In real-world scenarios, the quantitative analysis of ma-
environment, when facilities (machines) are moved from one terial handling and re-arrangement costs for the facility
location to another, re-arrangement cost is incurred. Re- layout may not be sufficient; the qualitative factors also
arrangement may also cause other costs such as production need to be considered. This paper seeks ways to obtain a
loss and it may need specialized equipment and labor [6]. more comprehensive outlook toward DFLP and to develop a
Compared to the SFLP, the DFLP is very recent. Like the classic formalization of the problem. Therefore, we will
SFLP, the DFLP is a combinatorial problem and computa- combine the qualitative or adjacency-based objective with
tionally intractable for which the optimal solution can be the quantitative-based objectives of facility layout to model
found only for very small problems. a multi-objective dynamic facility layout problem
Balakrishnan and Cheng [8] and Culturel-Konak [9] (MODFLP). It is also the objective of the present study to
reviewed the DFLP literature. They divided the solution evaluate some of the known approaches that have been
approaches for the DFLP into exact and heuristic algorithms. proposed to solve the problem.
They found that most of the DFLP approaches in the literature There has been only one study on DFLP with the multi-
use heuristics and the discrete representation of the layout and objective. Chen and Rogers [24] proposed a dynamic multi-
that most of the DFLP formulations are extensions of the QAP objective facility layout model to explore several aspects of
used for the SFLP. The difference between the QAP for the facility layout planning including time, distance-based ob-
DFLP and SFLP lies in the fact that the QAP for DFLP jective as well as the adjacency-based objective. They ap-
considers not only material handling cost, but also the costs plied ant colony optimization to solve the dynamic multi-
of re-arrangement between consecutive periods. objective facility layout problem. Although our model and
Rosenblatt [7] was the first to address the DFLP. The author Chen and Rogers’s model are both multi-objective, Chen
developed a dynamic programming model and a solution and Rogers have used the Urban [25] method to develop the
procedure based on the branch-and-bound scheme for deter- model, and finally, a single aggregate objective function has
mining an optimal layout for each of the several prespecified been introduced. In this study, the graph theory has been
future planning periods. Lacksonen and Enscore [10] pre- employed to incorporate both material handling costs and
sented five modified exact/heuristic algorithms to solve the adjacency scores between the facilities in a facility layout.
dynamic layout problems that include CRAFT, cutting planes, On the other hand, in decision making process on choosing
branch and bound, dynamic programming, and cut trees. the optimal layout, usually different decision makers with
Urban [11] proposed a steepest-descent pair wise interchange different opinions may be involved; so, to enable decision
procedure combined with the concept of forecast windows makers to apply their own views in the proposed MODFLP,
similar to CRAFT to solve the DFLP. both material handling and re-arranging costs have been
Conway and Venkataramanan [12] presented a solution expressed as two distinct but related functions.
technique based on genetic algorithm (GA) to solve the Since it is difficult to choose a single solution for a multi-
DFLP. Kaku and Mazzola [13] applied the tabu search algo- objective optimization problem, the general approach is to
rithm to solve DFLP. McKendall and Liu [14] developed three show a set of Pareto optimal solutions to decision makers.
tabu search heuristics to solve this problem. Rodriguez et al. Therefore, the decision maker selects one of the Pareto sol-
[15] presented a hybrid metaheuristic algorithm based on the utions based on his/her priorities. In these lines, a two-step
genetic algorithm and tabu search for the DFLP. Baykosaglu methodology is applied. In step 1, we evaluate two classic
and Gindy [16] were the first to apply simulated annealing methods namely weighted sum and ε-constraint methods and
(SA) algorithm to the DFLP. The approach taken by them is a also consider three metaheuristic methods known as elitist
straightforward implementation of SA to solve the DFLP. nondominated sorting genetic algorithm (NSGA-II), differen-
Şahin et al. [5] proposed a simulated annealing heuristic for tial evolution (DE), and Pareto-simulated annealing (PSA).
DFLP with budget costraint. McKendall and Shang [17] de- Except DE algorithm, the others have already been proposed
veloped three hybrid ant systems to solve the DFLP. Erel et al. for solving multi-objective optimization problems. In this
[18] presented a new three-phase heuristic scheme to solve the study, the DE algorithm is also modified to solve the proposed
DFLP. Krishnan et al. [19] proposed a new tool, Dynamic multi-objective model. Considering different comparison cri-
From Between Chart, to solve the DFLP. Mazinani et al. [20] teria and using the technique for order performance by simi-
studied a new kind of DFLP using flexible bay structure and larity to ideal solution (TOPSIS), in this step, the best
applied a genetic algorithm to solve it. algorithm for solving each problem is selected.
Int J Adv Manuf Technol (2013) 68:2215–2228 2217

In step 2, using an extension of the fuzzy TOPSIS meth- Ct Maximum adjacency rating in period t
od, the best solution (the best layout plan) out of the non- Xtij The 0, 1 variable for locating department i at location
dominated solutions obtained from the best algorithm is j in period t
selected. The rest of the paper is organized as follows: Ytijl The 0, 1 variable for shifting i from j to l in period t
Section 2 is allotted to the statement and modeling of the
multi-objective dynamic facility layout problem. Section 3 Therefore, a three-objective model of DFLP is developed
describes the methods used in this article to solve the as follows:
MODFLP. In Section 4, computational results are reported.
Finally, the last section presents the concluding remarks and
T X
X N X
N X
N X
N
Minimize Z1 ¼ ftik djl Xtij Xtkl ð1Þ
certain suggestions for future study. t¼1 i¼1 k¼1 j¼1 l¼1

T X
X N X
N X
N

2 Statement and modeling of MODFLP Minimize Z2 ¼ Atijl Ytijl ð2Þ

t¼2 i¼1 j¼1 l¼1

In the classical DFLP model, the objective function is T X

X N 1 X
N
expressed as the sum of different costs with an equal impor- Maximize Z3 ¼ btik ctik  Minimize Z3 ð3Þ
tance. On the other hand, in decision making process on t¼1 i¼1 k¼iþ1
choosing a facilities plan, usually different decision makers T X
X N 1 X
N

with different opinions are involved. It is necessary to prepare ¼ ðCt  btik ctik Þ
t¼1 i¼1 k¼iþ1
a model, enabling decision makers to apply their own views.
So, in the proposed MODFLP both material handling and re- Subject to
arranging costs are expressed as two distinct functions.
X
N
Important qualitative information that expresses adjacen- xtij ¼ 1 ; i ¼ 1; 2; . . . ; N and t ¼ 1; 2; . . . ; T ð4Þ
cy requirements is involved in the facility layout problems j¼1
that need to be duly considered. The relationship chart is
usually used to reflect “proximity ratios” between activities. X
N

In this study, the adjacency-based objective function (max- xtij ¼ 1 ; j ¼ 1; 2; . . . ; N and t ¼ 1; 2; . . . ; T ð5Þ
i¼1
imizing adjacency requirements) is also used for considering
the qualitative aspects of the problem. Ytijl ¼ Xðt1Þij Xtil ; i; j; l ¼ 1; 2; . . . ; N ; t ¼ 1; 2; . . . ; T
The assumptions for the MODFLP are as follows: Xtij ¼ f0; 1g for all i; j; t
Ytijl ¼ f0; 1g for all i; j; l; t
& The layout configuration is given.
ð6Þ
& The distances between locations are determined a priori.
& Flow between departments is dynamic and deterministic. The objective functions (1) and (2) are used to minimize
& Departments and locations are of equal size. the material handling and re-arrangement costs, respective-
& Re-arrangement costs between departments are deter- ly. Objective function (3) is used to maximize adjacency
mined a priori. requirements. Equations (4) and (5) ensure that each depart-
& Information of adjacency values is determined. ment is located and each location is occupied in every
period. The constraint set (6) helps to consider the re-
The following variables and symbols are used to formu- arrangement costs if at least one department is rearranged
late MODFLP: in consecutive periods.
The adjacency factor btik represents the adjacency ratio
N The number of departments in the layout between facilities i and k in period and it is as follows [26].
T The number of periods in the planning horizon
djl The distance from location j to l btik ¼1 if 0 < djl  dmax =6
ftik The flow cost from department i to k in period t btik ¼ 0:8 if dmax =6 < djl  dmax =3
Atij The cost of shifting department i from location j to l in btik ¼ 0:6 if dmax =3 < djl  dmax =2
btik ¼ 0:4 if dmax =2 < djl  2dmax =3
period t
btik ¼ 0:2 if 2dmax =3 < djl  5dmax =6
dma Maximum distance between the facilities
btik ¼0 if 5dmax =6 < djl  dmax
btik Adjacency factor between the facilities i and k in
period t The adjacency rating ctik is not always quantifiable and
ctik Adjacency value (0–5) between the facilities i and k in difficult to define. In fact, the optimization result can vary
period t depending on this value. In this study, the adjacency rating
2218 Int J Adv Manuf Technol (2013) 68:2215–2228

ctik is converted to numerical values in the following In multi-objective optimization, it is necessary to find all
way [27]. possible tradeoffs among multiple-objective functions that are
usually conflicting. Since it is difficult to choose a single
ctik =0 It is undesirable for facilities i and k to be solution for a multi-objective optimization problem without
located close together (X) iterative interaction with the decision makers, the general ap-
ctik =1 It is unimportant for facilities i and k to be proach is to show a set of Pareto optimal solutions to decision
located close together (U) makers [29]. Therefore, the decision maker selects one of the
ctik =2 It is ordinary for facilities i and k to be located Pareto solutions based on his/her priorities. This is a multi-
close together (O) attribute decision making (MADM) problem. MADM
ctik =3 It is important for facilities i and k to be located addresses the problem of choosing an option from a set of
close together (I) alternatives that are characterized in terms of their
ctik =4 It is especially important for facilities i and k to attributes. A MADM problem can be concisely
be located close together (E) expressed in a matrix format, in which columns indicate
ctik =5 It is absolutely necessary for facilities i and k to attributes considered in a given problem and lists the
be located close together (A) competing alternatives [30].
The optimization approach considered in this article is as
Although the model proposed in this paper is easy to follows:
comprehend, it is computationally intractable. In fact, this
model is an extension to the classic dynamic facility layout Step 1 Finding nondominated solutions set
problem. DFLP is a complex combinatorial optimization
problem for which only small problems can be solved 1.1. The following classical and meta-heuristic
optimally in a reasonable computation time [16]. methods are used separately to solve
MODFLP

3 Methodology 1.1.1. classical methods

&weighted sum method
Optimization refers to finding one or more feasible solutions, &ε-constraint
which correspond to extreme values of one or more objec- 1.1.2. population-based metaheuristic
tives. The fundamental difference between single-objective methods
and multi-objective optimizations lies in the cardinality in
the optimal set. From the standpoint of a user, however, only & NSGA-II
one solution is needed, no matter whether the associated & DE
optimization problem is single- or multi-objective. Thus, in a & PSA
multi-objective optimization, the effort must be ideally made: 1.2. The methods are compared and evaluated on
the following criteria and then the best method
& In step 1, to find the set of trade-off optimal solutions by is selected
considering all the important objectives.
& In step 2, to choose one of the obtained solutions using & convergence to the Pareto optimal front
higher-level information. & variety in solutions
& run time
Selection of the best method is a MADM
Definition 1: A solution x(1) is said to dominate the other problem, therefore, the TOPSIS technique is
solution x(2), if both conditions 1 and 2 are true: used.
1. The solution x(1) is no worse than x(2) in all 1.3. the solutions of the best method are consid-
objectives. ering as final nondominated solutions set
2. The solution x(1) is strictly better than x(2) in
Step 2 choosing one of the obtained solutions
at least one objective function.
Various criteria such as material handling costs, re-
arrangement costs, adjacency requirements of facili-
Definition 2 (nondominated set) Among a set of solutions P, ties, budget, etc. can affect on facilities layout deci-
the nondominated set of solutions P′ are those that are not sions. This is an MADM problem, therefore to select
dominated by any member of the set P. When the set P is the the best solution out of the nondominated solutions set
entire search space, or P=S, the resulting nondominated set, obtained in step 1, the decision makers opinions are
P′, is called the Pareto optimal set [28]. used and the TOPSIS technique also applied.
Int J Adv Manuf Technol (2013) 68:2215–2228 2219

Table 1 The importance weight of the attributes 3.3 NSGA-II algorithm

Handling cost Re-arrangement cost Adjacency
requirements Deb et al. [33] proposed an elitist nondominated sorting GA
which they termed NSGA-II. It sorts the combination of
DM 1 Very high Medium high High parents and children population with elitist strategy, intro-
DM 2 High High High duces the crowded comparison operator to improve diversi-
ty of solutions. As the literature shows [34–36], NSGA-II is
a well-known and a very popular multi-objective evolution-
3.1 Weighted sum method ary algorithm. This algorithm has been used to solve differ-
ent problems. Therefore, we were motivated to use this
The weighted sum method combines individual objective method for solving MODFLP.
functions into a single composite function (F(x)) by premulti-
plying each objective by a user-supplied weight. Initially, 3.4 NSGA-II for the MODFLP
Zadeh [31] applied this method with subsequent applications.
In the present article, the three objective functions to 3.4.1 Representing a solution
change into a single composite function by using weights
w1, w2, w3. The solution representation (chromosome) L ¼ ðl 1 ; l 2 ; . . . ; l T Þ
was used in employing NSGA-II algorithm for solving
T X
X N X
N X
N X
N
Minimize F ¼ w1  ftik dtjl Xtij Xtkl MODFLP, in which L signifies facility layout plan, lt shows
t¼1 i¼1 k¼1 j¼1 l¼1 facility layout in period t (l t ¼ ðl t ð1Þ; l t ð2Þ; . . . ; l t ðN ÞÞ), and
T X
X N X
N X
N l t ðiÞ is the department allocated to location i in period t.
þ w2  Atijl Ytijl ð8Þ
t¼2 i¼1 j¼1 l¼1 3.4.2 Initial population
T X
X N 1 X
N
þw3  ðCt  btik ctik Þ; The random method was used for generating an initial
t¼1 i¼1 k¼iþ1
population. The method involves the generation of a re-
placement of numbers 1 to N (number of departments) for
3.2 The ε-constraint method the intended number of periods (T) in order to generate each
chromosome in the population. In fact, by placing T replace-
Haimes et al. [32] suggested reformulating the multiple ments next to each other, a chromosome of the length N×T
objective programming by just keeping one of the objectives is generated. This process will be repeated to generate a
and restricting the rest of the objectives within user- population of a predefined size.
specified values. In this article, we consider mathematical
formula following form: 3.4.3 Crossover
P
T P
N P
N P
N P
N
Minimize Z1 ¼ ftik dtjl Xtij Xtkl In order to perform the crossover performance aimed at
t¼1 i¼1 k¼1 j¼1 l¼1 generating two offspring from two parents, the single point
subject to operator is used. Since selecting a favorable point on a
P
T P
N P
N P
N
ð9Þ
Atijl Ytijl  "1 number of solutions for crossover performance may lead
t¼2 i¼1 j¼1 l¼1 to a nonfeasible solution, points are selected on a chromo-
PT NP 1 P N
some, which contain period changes, that is, points along
ðCt  btik ctik Þ  "2
t¼1 i¼1 k¼iþ1 the boundary between two consecutive periods.

Table 2 Nondominated solutions for the N4T3 problem for each method used

method NSGA-II DE PSA GAMS

Objective solution f1 f2 f3 f1 f2 f3 f1 f2 f3 f1 f2 f3

1 26,149 0 63 26,149 0 63 26,149 0 63 26,149 0 63

2 26,119 2,330 66.6 26,119 2,330 66.6 26,119 2,330 66.6 26,119 2,330 66.6
3 26,026 5,321 73.8
2220 Int J Adv Manuf Technol (2013) 68:2215–2228

Fig. 1 Nondominated solutions

for the N4T3 problem

3.4.4 Mutation Where, r1, r2, r3 are random  mutually

 different integer
numbers taken over the range 1; Np , and are not equal to i.
For mutation to take place, substitution of two departments on a F is a real constant scaling factor in the region (0,2) [37].
chromosome is used. However, as the random selection of two Each mutated vector is subjected to parameter mixing
departments along a chromosome may lead to a nonfeasible using the crossover operation and the resultant vector is
solution, a period is randomly selected before the random called the trial vector. The trial vector, UiGþ1 ¼
selection of the two departments intended for the substitution. u1;iGþ1 ; u2;iGþ1 ; . . . ; uD;iGþ1 , where D is the dimension of
the problem defined by the number of variables, is obtained
3.5 DE algorithm from the mutated vector according to

Storn and Price [37] defined an optimization and evolution- vk;iGþ1 if random  pc ;
uk;iGþ1 ¼ ð11Þ
ary search method, namely, the DE algorithm, which is used xk;iG otherwise
k2½1;D;i2½1;Np 
for solving nonlinear programming problems associated
with continuous variables. As the literature shows [38–41], Where, random is a function that returns a uniform float-
this algorithm has been used to solve different problems. ing point number in the range [0, 1]. Pc is the crossover
Since DE algorithm has a simple and standard structure for probability (a user-defined parameter).
crossover and mutation operators, we were motivated to use If the trial vector yields a lower cost function value than
this method for solving MODFLP. the target vector for a minimization problem, then the trial
In this algorithm, each vector (chromosome) of the pop- vector replaces the target vector in the following generation.
ulation with size Np, in each generation becomes a target This constitutes the selection operation.
vector. Crossover and mutant operators are used to produce
a trial vector. Let XiG ; i ¼ 1; 2; . . . ; Np , be the target vector; 3.5.1 DE for the MODFLP
where i denotes the ith member of the population, and G the
generation to which the population belongs, a mutant vector In this research, for development of an algorithm
is generated according to intending solving multi-objective problems, the idea
dominating on NSGA-II algorithm is used. In other
ViGþ1 ¼ Xr1 G þ F  ðXr2 G  Xr3 G Þ ð10Þ

Table 4 Metrics relevant to all the methods used for the N4T3 problem
Table 3 Convergence metric (C) for N4T3 problem
Metric Method
C NSGA-II DE PSA GAMS Total
NSGA-II DE PSA GAMS
NSGA-II – 0.6667 1 1 2.6667
DE 1 – 1 1 3 C 2.6667 3 2.6667 2.6667
PSA 1 0.6667 – 1 2.6667 S 1,114 343 1,114 1,114
GAMS 1 0.6667 1 – 2.6667 RT 15 25 21 134
Int J Adv Manuf Technol (2013) 68:2215–2228 2221

Table 5 Comparison results of the methods used for the N4T3 prob- 3.5.4 Crossover
lem using TOPSIS

Method Closeness coefficient Ranking order In order to achieve the trial vector, the crossover operator
was used as in Relation (11), where Pc is considered to be a
NSGA-II 0.7009 2 controlling parameter whose appropriate value is obtained
DE 0.8243 1 through running the algorithm for different values of Pc and
PSA 0.6144 3 by comparing the results thus obtained.
GAMS 0.4211 4
3.6 PSA algorithm
This shows that the DE algorithm outperforms the others in solving the
N4T3 problem
Czyzak and Jaszkiewicz [42] modified the SA algorithm
for multi-objective optimization problems and proposed
words, instead of common mutation and crossover oper- a technique called PSA. Instead of using just one can-
ators in genetic algorithm that used in NSGA-II algo- didate for the final solution, PSA uses a set of interact-
rithm, DE mutation and crossover operators are used ing solutions called the generating set S at each iteration
accordingly. to propagate new solutions. SA has an advantage over
In order to make the DE algorithm compatible with the other metaheuristic algorithms in terms of the ease
discrete variables, the value for each variable is rounded to of implementation and gives reasonably good solutions
the closest integer. Considering the MODFLP, the two var- for many combinatorial problems. This algorithm has
iables xtij and Ytijl are binary (0 or 1) ones; thus, the values been used to solve different problems [43, 44].
for these variables are rounded to 0 or 1. Therefore, we were motivated to use PSA method, that
is modified SA algorithm, for solving MODFLP.
3.5.2 Representing a solution
3.6.1 PSA for the MODFLP
For showing a layout plan, real values of xtij (0 or 1) are used.
Accordingly, each chromosome in the DE algorithm is repre- In order to apply the PSA algorithm for solving MODFLP,
sented by L ¼ ðl 1 ; l 2 ; . . . ; l T Þ, in which L expresses facility the approach applied in NSGA-II representing a solution
layout
 plan, lt denotes facility layout in period t ( l t ¼ and generating an initial population are used. The other
xt1j ; xt2j ; . . . ; xtnj ), and xtij is a 1×N vector with one on steps are as follows:
the (1,j)th element and zero elsewhere.
& To generate the neighbor solution (y) from the cur-
3.5.3 Mutation rent solution (x ∈ S), the random pair wise exchange
is used. With random selection of the period and
The DE algorithm was developed for continuous varia- substitution of two departments in it, the neighbor
bles, but the problem under study has both continuous solution is generated.
and discrete variables. Hence, the following steps are & To determine the closest solution to x and also its non-
taken to make mutant performance possible for certain dominated one, all the solutions available in the gener-
target vectors: ating set (S) are first evaluated to see if they dominate x.
From among the solutions not dominating x, the solution
& For each target vector (XiG), select three random vectors with the minimum distance from x is selected and
corresponding to r1, r2, r3 which are random mutually termed x′. Euclidean distance is used to determine the
different integer numbers taken over the range )1; Np 2 distance between solutions.
and calculate ViGþ1 ¼ Xr1 G þ F:ðXr2 G  Xr3 G Þ; F>0 is a & Regarding the annealing schedule, the current tempera-
real constant scaling factor in the range (0,2). ture is determined by Tc ¼ T0 ar1 ; r ¼ 1; 2; . . . ; R equa-
& Round the real numbers available in ViG+1 to the closest tion. T0 is the initial temperature, a is called the cooling
integer numbers and then change integer numbers greater ratio, and R is the number of temperature reductions.
than one to one and those smaller than zero to zero. & The stopping criterion Tc <Tmin is used to stop the PSA
& If, in a specific period, there are more than one case of algorithm.
numbers equal to one in a row or if the sum of all the
numbers in a row are equal to zero, then such a row 3.7 Comparison of multi-objective optimization algorithms
should be deleted and the corresponding row in the
target vector be replaced. The same procedure will be Since convergence to the Pareto optimal front and the main-
repeated for each column. tenance of a diverse set of solutions are two distinct objectives
2222 Int J Adv Manuf Technol (2013) 68:2215–2228

Table 6 Nondominated solutions for the N6T5 problem for each method used

Method DE NSGA-II PSA GAMS

Objective solution f1 f2 f3 f1 f2 f3 f1 f2 f3 f1 f2 f3

1 102,640 16,155 250.2 112,990 10,530 231.8 106,510 11,533 248 106,320 2,814 243
2 105,160 9,378 253.4 106,090 3,867 255 104,120 5,077 256 106,070 2,330 249
3 102,040 12,969 252.6 103,060 8,467 255 104,330 4,250 256 110,140 6,681 240
4 111,150 0 233 114,370 8,495 235.8 102,560 11,014 254 108,260 0 241
5 108,240 16,947 237 105,640 18,756 240.6 102,040 8,615 259 106,420 0 261
6 110,520 2,187 237.8 105,430 4,405 255.4 102,120 12,440 259
7 112,650 5,760 232.6 107,100 12,463 241.4 102,740 9,383 253
8 105,780 8,063 241 106,670 15,526 239 104,870 4,339 254
9 113,680 3,778 232.6 109,550 5,681 249.8 106,390 5,867 249
10 114,640 3,253 232.6 105,020 6,952 250.6 105,570 7,473 251
11 108,450 2,436 237.4 106,060 18,018 239.4 103,790 9,205 255
12 115,180 6,031 232.2 110,950 9,903 235.8 104,680 5,517 253
13 108,180 7,503 237.8 112,950 9,215 237 101,960 14,844 257
14 115,350 6,470 231.8 106,250 16,802 240.2 105,570 12,450 248
15 102,480 16,942 251.8 118,450 9,378 232.6 103,460 8,117 259
16 103,320 103,320 249 93,811 15,441 256.6 102,640 12,081 252
17 104,340 10,466 244.6 94,733 9,117 263 105,070 12,631 246
18 106,860 7,315 239 116,890 4,516 235.4 102,940 9,805 252
19 109,460 5,777 237 103,420 7,867 243.8 104,170 11,618 252
20 107,820 9,817 238.6 96,783 13,378 251.4 103,920 13,529 249
21 107,230 6,752 247 94,832 14,632 253.8 103,570 12,275 251
22 104,680 6,831 256.2 109,080 9,657 239 110,070 4,228 253
23 101,940 13,694 254.2 113,240 6,965 236.2 105,230 11,814 250
24 115,330 10,360 231.8 100,780 12,303 248.6 106,550 2,964 264
25 107,710 3,667 251.4 112,200 12,631 235.4 108,680 8,004 244
26 106,090 4,832 250.6 110,010 10,555 236.6 109,020 4,928 246
27 102,440 11,431 254.2 107,630 11,477 239 104,060 9,566 252
28 103,540 10,409 250.2 99,040 10,752 251.8 105,290 7,974 252
29 103,850 9,432 248.6 106,820 14,694 240.2 113,060 7,902 246
30 108,230 5,182 249.8 117,840 0 243 105,620 4,151 262
31 105,740 15,569 243 96,515 11,631 253.4 103,860 6,521 253
32 102,490 13,466 251 118,610 1,365 240.6 102,220 9,452 256
33 110,720 9,289 232.6 103,170 6,044 261 103,980 5,809 255
34 103,740 11,781 246.6 112,510 16,868 233.4 106,720 10,289 248
35 107,940 6,577 238.6 116,980 2,142 246.2 105,930 7,340 251
36 110,130 10,027 234.6 107,630 4,245 251.4 103,850 7,277 261
37 101,000 12,169 247.4
38 109,240 9,033 238.2
39 110,230 7,019 239
40 108,890 14,605 236.2
41 112,330 5,709 238.6
42 113,950 4,633 236.6
43 117,000 1,165 242.2

in multi-objective optimization, no single metric can be used set coverage and spacing metrics have been used to assess
to evaluate the performance of an algorithm and there is a clear convergence and diversity. In addition to these, the runtime of
need for using at least two performance metrics. Here, the two algorithms has been used for their comparison.
Int J Adv Manuf Technol (2013) 68:2215–2228 2223

Fig. 2 Nondominated solutions

for the N6T5 problem

P  i 
Where, di ¼ mink2Q^k6¼i M  k
3.7.1 Set coverage metric m¼1 fm  fm P and d is the mean
jQj
value of the above distance measure d ¼ i¼1 di =jQj . The
This metric was suggested by Zitzler and Thiele [45]. above metric measures the standard deviations of different di
The set coverage metric C(A,B) calculates the propor- values. When the solutions are nearly uniformly spaced, the
tion of solutions in B which are weakly dominated by corresponding distance measure will be small. Thus, an algo-
solutions in A: rithm having a smaller spacing (S) is preferred. Along these
lines, an algorithm with a greater C metric, a smaller S, and a
shorter runtime is preferred in our study over other algorithms.
j f b 2 Bj9 a 2 A : a  b g j
C ðA; BÞ ¼
j Bj 3.8 Selecting the most appropriate MADM method

The metric value C(A,B) = 1 means that all members In the present study, simplicity of performance, infinite number
of B are weakly dominated by A. On the other hand, of alternatives (set of nondominated solutions), ease of coding,
C(A,B) = 0 means that no member of B is weakly shorter runtime, lower quantity of information required, ability
dominated by A. to handle qualitative criteria, ability to handle group decision
making, and less sensitivity to attribute weights warranted the
3.7.2 Spacing metric use of the TOPSIS method for selecting the best solution out of
the nondominated set. A basic principle of TOPSIS requires
Schott [46] suggested a metric that calculates the relative the selected alternative to have the shortest distance from the
distance between consecutive solutions in the non- positive ideal solution (PIS) and the farthest distance from the
dominated set (Q) obtained as follows: negative ideal solution (NIS). In principle, the best alternative
should be closest to the PIS and farthest from the NIS. In this
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi algorithm, a closeness coefficient is defined to determine the
u
u 1 X jQj
 2 ranking order of all alternatives. Therefore, we can determine
S¼t di  d ð13Þ the ranking order of all the alternatives according to closeness
jQj i¼1
coefficient. The decision matrix and weight of attributes

Table 7 Convergence metric (C) for N6T5 problem Table 8 Metrics relevant to all the methods used for the N6T5 problem

C NSGA-II DE PSA GAMS Total Method

Metric NSGA-II DE PSA GAMS

NSGA-II – 0.3721 0.4167 0.6 1.3888
DE 0.1944 – 0.3333 0 0.5277 C 1.3888 0.5277 0.1854 0.2643
PSA 0.1389 0.0465 – 0 0.1854 S 856 993 832 2,605
GAMS 0.0556 0.0698 0.1389 – 0.2643 RT 665 693 775 17,530
2224 Int J Adv Manuf Technol (2013) 68:2215–2228

Table 9 Comparison results of the methods used for the N6T5 prob- The following three assumptions were used in imple-
lem using TOPSIS
menting second step of the methodology:
Method Closeness coefficient Ranking order
& The three attributes of material handling cost, re-
NSGA-II 0.9376 1 arrangement cost, and adjacency requirements are used
DE 0.6978 2 for evaluating solutions.
PSA 0.6358 3 & There are two decision makers (the number of decision
GAMS 0.1766 4 makers can, of course, be expected to be more).
& The importance weight vector of the attributes is as in
This shows that the NSGA-II algorithm outperforms the others in
solving the N6T5 problem the following:

For executing the two classic multi-objective methods

comprise the inputs to the algorithm. In order to effect group (weighted sum and ε-constraint methods), the GAMS soft-
views of decision makers and to properly deal with the existing ware was used. In this article, two SBB and BARON solvers
uncertainties in their views, extensions of the TOPSIS has been were used for solving the proposed model. Therefore, for each
used in this study which were applied for group decision weight vector and upper bound vector (ε), a local optimal
making under the fuzzy environment presented by Chen [47]. solution was obtained through the SBB solver. Then, the
solution was improved using the BARON solver. Since the
runtime was tremendously large even for small-sized prob-
4 Computational results lems, a runtime limit was set for both solvers. In other words,
if the two solvers reached a global optimal solution within the
This section is intended to study and evaluate the methods time limit, the optimization process was stopped; otherwise,
investigated in this paper for solving MODFLP. The dataset the solution obtained was regarded as the nondominated
for comparing these methods are taken from the literature. solution achieved through either the weighted sum or the
Compared with the classic dynamic layout model, the pro- ε-constraint method at the end of the time limit. Being time-
posed multi-objective model involves more parameters. The consuming and needing a large computer memory for solving
data required for the new parameters have been randomly large-sized problems, the classic methods were solely used for
generated based on a uniform distribution. Effectiveness of solving the problems N4T3, N6T5, and N6T10.
the methods investigated has been studied using N4T3 (four MATLAB programming language was used for executing
departments and three periods), N6T5, N6T10, N15T5, the meta-heuristic algorithms (NSGA-II, DE, PSA). Running
N15T10, and N30T5 problems. these algorithms is based on the assumption that the initial
Based on the first step of the methodology, each of these population is generated randomly. Therefore, each test problem
problems was solved using the multi-objective methods pro- was solved five times using the heuristics methods. It should be
posed in this paper. These methods were analyzed based on the mentioned that all the values used for algorithm control param-
outputs obtained and the metrics introduced here. The fuzzy eters were selected through experimentation and that the set of
TOPSIS method was applied to determine the best method test problems were solved on a Pentium IV 3.2 GHz PC.
because the number of applied metrics intended for analyzing The results obtained from the classic weighted sum and
the methods was more than one and there was no single method ε-constraint methods were combined and reported as the
which could be the best with respect to all the metrics employed. output of the GAMS software. In what follows, a more
detailed description of the solution methodology is pre-
Table 10 Results from applying the proposed multi-objective optimi- sented by reviewing the solution steps for N4T3 and N6T5.
zation procedure Finally, a summary of the results obtained from solving
Problem The most suitable Objective function values of the best other problems is provided below (Table 1).
algorithm non-dominated solution
4.1 N4T3 problem
f1 f2 f3

N4T3 DE 26,149 0 63
4.1.1 Step 1 of methodology
N6T5 NSGA-II 111,150 0 233
N6T10 DE 229,720 13,512 486
A 2×2 layout has been considered for the N4T3 problem.
Table 2 and Fig. 1 show the nondominated solutions
N15T5 NSGA-II 536,280 23,830 1,655.4
obtained from the different methods employed.
N15T10 NSGA-II 1,050,200 72,006 3,301.8
Table 3 presents the values for the convergence metric
N30T5 NSGA-II 628,984 71,587 7,239
(C) between each two methods computed for the solutions
Int J Adv Manuf Technol (2013) 68:2215–2228 2225

in Table 2. The last column of this table presents the degree the solution obtained using the proposed methodology dom-
of the domination of each method over others. Table 4 inates the solution of Chen and Rogers’s method.
shows the values for convergence attribute (C), diversity
attribute (S), and runtime (RT).
Considering Table 4 as a decision matrix and assigning 5 Conclusion
very high importance to each of the criteria, the closeness
coefficient and the ranking order of the alternatives In this paper, a more comprehensive outlook was taken toward
(methods) will be as in Table 5 using the TOPSIS method. DFLP to develop a multi-objective model. The model allows
simultaneously for both quantitative and qualitative analyses
4.2 Step 2 of methodology of dynamic facility layout problems. In this study, the objec-
tive function related to the re-arrangement and material han-
The set of nondominated solutions obtained from the DE dling costs has been considered as two distinct but related
algorithm is used for selecting the best solution. Thus, the functions. Each of these cost terms might be of different
nondominated solution 1 in Table 2 form DE column is importance to decision makers. This formulation enables de-
selected as the best solution by applying the TOPSIS meth- cision makers to apply their own views. Like the classic
od and on the basis of the assumptions made above. models, the proposed model is both intractable and time-
consuming in terms of problem solution. In this study, we,
4.3 N6T5 problem therefore, investigated and evaluated several known methods.
In step 1, a given data set was used to study and evaluate two
4.3.1 Step 1 of methodology classic methods (weighted sum and ε-constraint methods) and
three metaheuristic methods (NSGA-II, DE, PSA). The three
A 2×3 layout has been considered for the N6T5 problem. metrics of convergence, diversity, and runtime were used for
Table 6 and Fig. 2 show the nondominated solutions obtained the comparison of these methods. The results showed that the
from the different methods employed (Table 7, 8 and 9). classic methods investigated had a poor efficiency while the
population-based metaheuristic methods were useful tools in
4.3.2 Step 2 of methodology solving the MODFLP. In addition, the two NSGA-II and DE
were found to have a better efficiency compared to the PSA.
The set of nondominated solutions obtained from the In step 2, to deal with decision making problem in an uncer-
NSGA-II algorithm is used for selecting the best solution. tain environment, an extension of the fuzzy TOPSIS method was
Thus, the nondominated solution 4 in Table 6 form NSGA- applied for group decision making under a fuzzy environment.
II column is selected as the best solution by applying the In the problem addressed here, we did not consider time
TOPSIS method and on the basis of the assumptions made dependability. Modeling and solving the MODFLP when re-
above. Table 10 presents the results from solving all the test arrangement cost depends on time periods, and developing
problems in the present study. the DFLP model for a multifloor layout are areas warranting
These results show that the population-based metaheuristic further study.
methods have considerable potential to cope with multi-
objective problems and that, in contrast, the classic multi-
objective optimization methods suffer from poor efficiency Appendix A
in solving MODFLP. Based on the assumptions made, the
NSGA-II algorithm proves to be more efficient compared to A dynamic facility layout consists of six departments
other algorithms investigated in this study. Generally speak- and three time periods is considered. The same
ing, the DE algorithm may be recommended for the solution
of small-sized problems while NSGA-II algorithm might have Table 11 Material flow matrix of the three time periods
a better performance in solving large-sized problems.
Time period 1 Time period 2 Time period 3
The computational study involving the proposed model
and solving methodology was also compared to the model 1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
of Chen and Rogers [24] using the data set from Chen [48].
The dataset consists of six departments and three time periods. 1 0 8 4 8 6 4 0 10 2 7 5 2 0 10 3 6 4 4
2 8 0 2 1 5 2 10 0 2 1 5 1 10 0 1 1 7 1
The rearrangement cost is 10. Appendix A shows the material
3 4 2 0 4 5 2 2 2 0 4 5 2 3 1 0 2 3 1
flow cost, distance matrix, and closeness ranking matrices for
4 8 1 4 0 2 4 7 1 4 0 3 6 6 1 2 0 2 6
each period. It also shows the computational results involving
5 6 5 5 2 0 4 5 5 5 3 0 6 4 7 3 2 0 8
the proposed model and solving methodology with compare 6 4 2 2 4 4 0 2 1 2 6 6 0 4 1 1 6 8 0
to the model of Chen and Rogers [24]. The results show that
2226 Int J Adv Manuf Technol (2013) 68:2215–2228

Table 12 Closeness rating matrix of the three time periods Table 15 Comparison results of the methods used for this problem
using TOPSIS
Time period 1 Time period 2 Time period 3
Method Closeness coefficient Ranking order
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6
NSGA-II 0.8239 1
1 – A I A E I – A U E I U – A O E I I DE 0.6671 2
2 A – U X I O A – U X I U A – X U E U PSA 0.5777 3
3 I U – I I U U U – I I U O X – O I U GAMS 0.4385 4
4 A X I – O I E X I – O E E U O – O E
5 E I I O – I I I I O – E I E I O – E
6 I O U I I – U U U E E – I U U E E – departments are to be arranged or rearranged across
three time periods. The re-arrangement cost is 10. The
material flow cost matrices (from–to charts), distance
Table 13 Distance matrix, and closeness rating matrices are shown as be-
matrix Locations
low (Tables 11, 12, and 13):
1 2 3 4 5 6
Step 1 of methodology:
1 0 1 2 1 2 3 Table 14 shows the nondominated solutions obtained
2 1 0 1 2 1 2 from the different methods employed (Table 15).
3 2 1 0 3 2 1 Step 2 of methodology:
4 1 2 3 0 1 2 For choosing the best solution, the set of nondomi-
5 2 1 2 1 0 1 nated solutions related to the NSGA-II algorithm
6 3 2 1 2 1 0 (Table 14) is considered. By applying the TOPSIS
method and based on the explained assumptions in

Table 14 Non-dominated solu-

tions for each method used Method NSGA-II DE PSA GAMS

Objective f1 f2 f3 f1 f2 f3 f1 f2 f3 f1 f2 f3
solution

1 520 20 152.2 520 20 152.2 548 50 158.6 520 20 154.2

2 516 60 154.2 518 50 152.6 536 110 155.8 520 40 153
3 516 70 153.4 520 0 155 534 80 156.6 524 0 153
4 524 0 153 518 40 154.2 570 60 159.4 518 80 153
5 518 50 152.6 518 20 155 598 20 161.4 516 100 154.2
6 520 20 152.2 516 60 154.2 566 80 158.2 518 50 153.4
7 518 20 155 438 40 169 570 30 160.2
8 518 30 154.2 480 30 167.4
9 514 50 158.2
10 508 20 163.4
11 460 40 163.8
12 504 40 162.2
13 496 60 160.6
14 484 30 166.2
15 516 80 153.4
16 518 50 152.6
17 516 60 154.2
18 524 0 153
19 520 20 152.2
20 520 30 153.8
21 520 20 154.2
Int J Adv Manuf Technol (2013) 68:2215–2228 2227

Table 16 The solutions provided by proposed methodology and Chen and Rogers’s methodology

Assignment of facilities to Assignment of facilities to Assignment of facilities to f1 f2 f3

the sites for time period 1 the sites for time period 2 the sites for time period 3

Proposed methodology 6 1 4 2 5 3 1 6 4 2 5 3 1 6 4 2 5 3 520 20 152.2

Chen and Rogers’s methodology 2 5 3 6 1 4 3 5 2 4 6 1 3 5 2 4 6 1 520 50 152.2

This shows that the NSGA-II algorithm outperforms the others in solving this problem

Section 4, the nondominated solution 6 from NSGA-II 17. McKendall JAR, Shang J (2006) Hybrid ant systems for the
dynamic facility layout problem. Comput Oper Res 33:790–
column is selected as the best solution.
803
The solutions obtained using both proposed and 18. Erel E, Ghosh JB, Simon JT (2003) New heuristic for the dynamic
Chen and Rogers’s methodology are shown in layout problem. J Oper Res Soc 54:1275–1282
Table 16. As it can be seen, the solution of the proposed 19. Krishnan KK, Cheraghi SH, Nayak CN (2006) Dynamic from-
between chart: a new tool for solving dynamic facility layout
methodology dominates the solution of Chen and
problems. Int J Ind Syst Eng 1(1/2):182–200
Rogers’ method. 20. Mazinani M, Abedzadeh M, Mohebali N (2012) Dynamic facility
layout problem based on flexible bay structure and solving by
genetic algorithm. Int J Adv Manuf Technol. doi:10.1007/
s00170-012-4229-6
References 21. Dunker T, Radons G, Westkamper E (2005) Combining evo-
lutionary computation and dynamic programming for solving
a dynamic facility layout problem. Eur J Oper Res 165(1):55–
1. Tompkins J, White J, Bozer Y, Tanchoco J (2003) Facilities plan- 69
ning, 3rd edn. Wiley, New Jersey 22. McKendall ARJ, Hakobyan A (2010) Heuristics for the dynamic
2. Tompkins J, White J (1984) Facilities planning. Wiley, New Jersey facility layout problem with unequal-area departments. Eur J Oper
3. Franke C, Basdere B, Ciupek M, Seliger S (2006) Remanufactur- Res 201(1):171–182
ing of mobile phones—capacity, program and facility adaptation 23. Jolai F, Tavkkoli R, Taghipour M (2012) A multi-objective particle
planning. Omega 34:562–570 swarm optimisation algorithm for unequal sized dynamic facility
4. Benjaafar S, Heragu SS, Irani SA (2000) Next generation factory layout problem with pickup/drop-off locations. Int J Prod Res
layouts: research challenges and recent progress. I_FORMS 50:4279–4293
32(6):58–76 24. Chen G, Rogers J (2009) Managing dynamic facility layout with
5. Şahin R, Ertogral K, Türkbey O (2010) A simulated annealing multiple objectives. PICMET Proceeding, August 2–6, Portland,
heuristic for the dynamic facility layout problem with budget OR, USA
constraint. Comput Ind Eng 59:308–313 25. Urban TL (1987) A multiple criteria model for the facilities layout
6. McKendall JAR, Shang J, Kuppussamy S (2006) Simulated problem. Int J Prod Res 5(12):1805–1812
annealing heuristics for the dynamic facility layout problem. Com- 26. Lee KY, Roh MI, Jeong HS (2005) An improved genetic algorithm
put Oper Res 33:2431–2444 for multi-floor facility layout problems having inner structure walls
7. Rosenblatt MJ (1986) The dynamics of plant layout. Manag Sci and passages. Comput Oper Res 32:879–899
32(1):76–86 27. Lee HJ (1988) Heuristic graph-theoretic approach in facility layout
8. Balakrishnan J, Cheng CH (1998) Dynamic layout algorithms: a problem: the development of a decision support system. Disserta-
state-of-the art survey. Omega 26(4):507–521 tion, University of Texas, Arlington, USA.
9. Kulturel-Konak S (2007) Approaches to uncertainties in facility 28. Deb K (2001) Multi objective using evolutionary algorithms.
layout problems: perspectives at the beginning of the 21st century. Wiley, LTD
J Intell Manuf 18(2):273–284 29. Hisashi T, Hajime K, Shigenobu K (1996) Multi-objective
10. Lacksonen TA, Enscore EE (1993) Quadratic assignment algorithms optimisation by genetic algorithms: a review. Proceedings of
for the dynamic layout problem. Int J Prod Res 31(3):503–517 the 3rd International Conference on Evolutionary Computa-
11. Urban TL (1993) A heuristic for the dynamic facility layout tion. 517–522.
problem. IIE Trans 25(4):57–63 30. Yang T, Chen MC, Hung CC (2007) Multiple attribute decision-
12. Conway DG, Venkataramanan MA (1994) Genetic search and the making methods for the dynamic operator allocation problem.
dynamic facility layout problem. Comput Oper Res 21(8):955–960 Math Comput Simul 73(5):285–299
13. Kaku BK, Mazzola JB (1997) A tabu-search heuristic for the 31. Zadeh L (1963) Optimality and non-scalar-valued performance
dynamic plant layout problem. Informs J Comput 9(4):374–384 criteria. IEEE Trans Autom Control 8:59–60
14. McKendall ARJ, Liu WH (2012) New tabu search heuristics for 32. Haimes YY, Lasdon LS, Wismer DA (1971) On a bicriterion
the dynamic facility layout problem. Int J Prod Res 50:867–878 formulation of the problems of integrated system identification
15. Rodriguez JM, MacPhee FC, Bonham DJ, Bhavsar VC (2006) Solving and system optimization. IEEE Trans Syst Man Cyber 1(3):296–
the dynamic plant layout problem using a new hybrid meta-heuristic 297
algorithm. Int J High Performance Comput Netw 4(5/6):286–294 33. Deb K, Agarwal S, Meyarivan T (2002) A fast and elitist multi-
16. Baykasoglu A, Gindy NNZ (2001) A simulated annealing algorithm objective genetic algorithme: NSGA-II. IEEE Trans Evol Comput
for dynamic layout problem. Comput Oper Res 28:1403–1426 6(2):182–197
2228 Int J Adv Manuf Technol (2013) 68:2215–2228

34. Jemai J, Zekri M, Mellouli K (2012) An NSGA-II algorithm for the 41. Dexuan Z, Haikuan L, Liqun G, Steven L (2011) An improved
green vehicle routing problem. Evolut Comp Comb Opt 7245:37–48 differential evolution algorithm for the task assignment problem.
35. Bhattaacharya R, Bandyopadhyay S (2010) Solving conflicting Eng Appl Artif Intell 24:616–624
bi-objective facility location problem by NSGA II evolution- 42. Czyzak P, Jaszkiewicz A (1998) Pareto simulated annealing—a
ary algorithm. Int J Adv Manuf Technol 51:397–414 metaheuristic technique for multipleobjective combinatorial opti-
36. Wei Z, Feng YX, Tan JR, Wang JL, Li ZK (2009) Multi-objective mization. J Mul Cri Dec An 7(1):34–47
performance optimal design of large-scale injection molding ma- 43. Drexl A, Nikulin Y (2008) Multicriteria airport gate assignment
chine. Int J Adv Manuf Technol 41:242–249 and Pareto simulated annealing. IIE Trans 40:385–397
37. Storn R, Price K (1997) Differential evolution—a simple and 44. Amany MM, Hisham MA (2012) Optimal composition of virtual
efficient heuristic for global optimization over continuous spaces. enterprises with interval cost parameters. The 8th International
J Glob Optim 11:241–354 Conference on INFOrmatics and Systems, May 14–16.
38. Nearchou AC (2006) Meta-heuristics from nature for the loop 45. Zitzler E, Thiele L (1999) Multiobjective evolutionary algorithms:
layout design problem. Int J Prod Econ 101:312–328 a comparative case study and the strength Pareto approach. IEEE
39. Yanmei H, Minghao Y, Xiangtao L (2011) A novel objective Trans Evol Comput 3:257–271
function for job-shop scheduling problem with fuzzy processing 46. Schott JR (1995) Fault tolerant design using single and multi criteria
time and fuzzy due date using differential evolution algorithm. Int genetic algorithms. Master’s thesis, Boston, MA: Department of
J Adv Manuf Technol 56(9–12):1125–1138 Aeronautics and Astronautics Massachusetts Institute of technology.
40. Yang SH, Natarajan U, Sekar M, Palani S (2010) Prediction of 47. Chen CT (2000) Extensions of the TOPSIS for group decision-
surface roughness in turning operations by computer vision using making under fuzzy environment. Fuzzy Set Syst 114:1–9
neural network trained by differential evolution algorithm. Int J 48. Chen G (2007) Multi-objective evaluation of dynamic facility layout
Adv Manuf Technol 51:965–971 using ant colony optimization. Dissertation, University of Texas.