You are on page 1of 12

European Journal of Operational Research 262 (2017) 1052–1063

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Production, Manufacturing and Logistics

Combining Wang–Landau sampling algorithm and heuristics for


solving the unequal-area dynamic facility layout problem
Jingfa Liu a,b, Dawen Wang a,b,∗, Kun He c, Yu Xue a,b
a
School of Computer and Software, Nanjing University of Information Science and Technology, Nanjing 210044, China
b
Jiangsu Engineering Center of Network Monitoring, Nanjing University of Information Science and Technology, Nanjing 210044, China
c
School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan 430074, China

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

Article history: The dynamic facility layout problem (DFLP) is the problem of placing facilities in a certain plant floor for
Received 8 April 2016 multiple stages so that facilities do not overlap and the sum of the material handling and rearrangement
Accepted 2 April 2017
costs are minimized. We describe a model, where the facilities have unequal-areas and the layout for
Available online 6 April 2017
each stage is produced on the continuous plant floor. The most difficulty of solving this problem consists
Keywords: in the lack of a powerful optimization method. Wang–Landau (WL) sampling algorithm is an improved
Global optimization Monte Carlo method, and has been successfully applied to solve many optimization problems. In this
Dynamic facility layout paper, we combine the WL sampling algorithm and some heuristic strategies to solve the unequal-area
Unequal-area DFLP. In the WL sampling algorithm, a vacant point strategy is applied to update layout at one stage. To
Wang–Landau sampling algorithm prevent overlapping of facilities and reduce the empty space among facilities, a pushing strategy and a
Group decision-making pressuring strategy are applied. We have tested the proposed algorithm on four groups of cases and the
Heuristic strategies
computational results show that the proposed algorithm is effective in solving the unequal-area DFLP.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction Liu and Meller (2007) researched the continuous-representation-


based SFLP, and proposed a genetic-algorithm-based heuristic
Because of the important significance of saving the actual which combined the sequence-pair representation with the MIP
production costs, the facility layout problem (FLP) is very early model. Anjos and Vannelli (2006) presented a mathematical-
to be studied. In the early time, the objective of the FLP is to programming framework based on the combination of two
minimize the material handling costs by arranging facilities in new mathematical-programming models. Komarudin and Wong
a static production environment. As material flows among fa- (2010) proposed an ant system which used slicing tree repre-
cilities will not change, the layout problem is called the static sentation to easily represent the SFLP without too restricting
facility layout problem (SFLP) which has been studied by many the solution space. Ulutas and Kulturel-Konak (2012) proposed a
scholars. Mir and Imam (2001) employed a hybrid optimization clonal selection algorithm which had a new encoding and a novel
approach which contained a simulated annealing method and procedure to copy with dummy facilities to fill the empty space in
an analytical method. Bozer and Meller (1997) researched the the plant floor. Gonçalves and Resende (2015) developed a hybrid
distance-based facility layout problem, and proposed an alterna- approach which combined a biased random-key genetic algorithm,
tive distance measure basing on the expected distance between a novel placement strategy and a linear programming model.
two facilities. Meller, Narayanan, and Vance (1999) proposed Singh and Sharma (2006) and Drira, Pierreval, and Hajri-Gabouj
some general classes of valid inequalities for a mixed-integer (2007) presented a literature review of the SFLP.
programming (MIP) model of SFLP. They used these inequalities In today’s market, with increasing global competition and short
in a branch-and-bound algorithm and moderately increased the life cycle of production, material flows among facilities change
range of solvable problems. Motivated by the work of Meller et al. during the planning horizon, so that the problem becomes the
(1999), Sherali, Fraticelli, and Meller (2003) presented an improved dynamic facility layout problem (DFLP). The dynamic facility layout
MIP model and several effective solution strategies for the SFLP. problem generally includes two categories: equal-area dynamic
facility layout problem and unequal-area dynamic facility layout

Corresponding author at: School of Computer and Software, Nanjing University
problem, which correspond to the facilities with equal-area and
of Information Science and Technology, Nanjing 210044, China. facilities with unequal-area, respectively. Rosenblatt (1986) is the
E-mail address: wangdw1993@163.com (D. Wang). first one to study the equal-area DFLP and proposed a heuristic

http://dx.doi.org/10.1016/j.ejor.2017.04.002
0377-2217/© 2017 Elsevier B.V. All rights reserved.
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1053

optimization scheme based on dynamic programming. Afterwards, & Landau, 2001) is an improved Monte Carlo algorithm. Unlike
some scholars have also studied it and put forward all kinds of conventional Monte Carlo simulations that generate a probability
heuristic methods. Urban (1993) proposed a pair-wise interchange distribution at a given temperature, the WL sampling method can
procedure which made use of ‘forecast windows’ to find different estimate the density of states accurately via a random walk, which
sets of good layout plans for the planning horizon. Baykasoglu produces a flat histogram in energy space. There have been many
and Gindy (2001) proposed a simulated annealing algorithm with improvements and applications on the WL sampling algorithm.
an effective data structure and neighborhood generation mecha- For example, Schulz, Binder, and Müller (2003) proposed a simple
nism. Şahin and Türkbey (2009) presented a new hybrid heuristic modification of the WL sampling algorithm. This modification
based on the simulated annealing approach supplemented with a removed the systematic error that occurred at the boundary of the
tabu list. Pourvaziri and Naderi (2014) presented a novel hybrid range of energy over which the random walk took place. Seaton,
multi-population genetic algorithm. Unlike the available previous Wüst, and Landau (2010) used the WL sampling algorithm to de-
genetic operators, they designed operators to search only the scribe the thermodynamic behavior of a continuous homopolymer.
feasible space. Hosseini-Nasab and Emami (2013) proposed a Liu (2014) proposed an improved WL sampling method which
hybrid particle swarm optimization algorithm which combined incorporated the generation of initial conformation based on the
a simple and fast simulated annealing process. Their algorithm greedy strategy and the neighborhood strategy based on pull-
further explored the continuous solution space by using a coding moves into the WL sampling method to predict the protein struc-
technique. Mckendall and Liu (2012) presented three tabu search tures on the face-centered-cube (FCC) hydrophobic-hydrophilic
heuristics. The first is a simple tabu search heuristic, the second (HP) lattice model. The most challenge of solving the DFLP is that
adds diversification and intensification strategies to the first, and the function to be optimized is characterized by a multitude of
the third is a probabilistic tabu search heuristic. Hosseini, Khaled, local minima separated by high-energy barriers. The WL sampling
and Vadlamani (2014) proposed a robust and simply structured method can visit the all accessible states of the system, which
hybrid technique based on three heuristics: imperialist com- means it can jump out of these high-energy barriers. Therefore,
petitive algorithm, variable neighborhood search, and simulated the WL sampling method is an ideal global search algorithm for
annealing. Bozorgi, Abedzadeh, and Zeinali (2015) applied a data solving the DFLP. In this paper, by combining the WL sampling al-
envelopment analysis method and a tabu search-based algorithm gorithm and some heuristics, we put forward a heuristic WL (HWL)
which used a dynamic tabu list and a diversification strategy to sampling algorithm. The numerical results show that the proposed
find the most efficient layout. Ulutas and Islier (2015) studied the algorithm is an effective method for the unequal-area DFLP.
equal-area DFLP in footwear industry and proposed a clonal selec- The structure of this paper is as follows. In Section 2, we
tion algorithm based on affinity maturation and receptor editing introduce the mathematical model of the DFLP. In Section 3, we
processes. Kouvelis, Kuawarwala, and Gutierrez (1992) addressed introduce the WL sampling algorithm and the heuristic strate-
this problem by using the concept of robust layouts, and proposed gies. In Section 4, we list the computational results and make a
a branch and bound method to generate the robust layouts for the comparison. Section 5 is the conclusion of this paper.
manufacturing systems. Li, Li, Ma, and Tang (2015) considered the
remanufacturing equal-area DFLP with uncertainties and proposed
a simulated annealing algorithm. 2. Problem formulation
For the unequal-area DFLP, Montreuil and Venkatadri
(1991) first presented a proactive strategic methodology sustained In this section, we provide a mathematical model of the
by a linear programming model. Yang and Peters (1998) put for- unequal-area DFLP. First, some notations for describing the math-
ward a heuristic procedure based on the construction type layout ematical model are presented.
design algorithm. Dunker, Radons, and Westkamper (2005) pro- Indexes
posed a hybrid algorithm which combined dynamic programming i, j Facility index, i, j = 1, …, N, where N = the number of
and a genetic algorithm. For each stage the genetic algorithm facilities
evolved a population of layouts while the dynamic programming t Stage index, t = 1, …, T, where T = the number of stages
provided the evaluation of the fitness of the layouts. McKendall
and Hakobyan (2010) developed a boundary search heuristic which Parameters
placed facilities along the boundaries of already placed facilities, Ctij Material handling cost per unit distance between facility
and used a tabu search method to improve the obtained solutions. i and facility j in stage t
Jolai, Tavakkoli-Moghaddam, and Taghipour (2012) presented a Rti Rearrangement cost of facility i in the beginning of stage
multi-objective particle swarm optimization method which found t
the optimal scheme using a particle best position and a swarm Ftij Material flow between facility i and facility j in stage t
best position. Dong, Wu, and Hou (2009) considered the unequal- Wti Shorter edge length of facility i in stage t
area DFLP under dynamic business environment and proposed a Lti Longer edge length of facility i in stage t
simulated annealing algorithm based on a shortest path model. L Length of plant floor
Mazinani, Abedzadeh, and Mohebali (2013) considered a new kind W Width of plant floor
of DFLP with flexible bay structure, in which facilities are assigned Ii Interior of facility i
to parallel bays in a plant floor, and proposed a genetic algorithm. Variables
Ma, Cai, Zhang, and Li (2015) studied the unequal-area DFLP with xti , yti Centroid coordinate of facility i in stage t
uncertain product demands and proposed an improved genetic lti , wti Length and width of facility i in stage t
algorithm which combined triangular fuzzy number operation dti Orientation of facility i in stage t (0: vertical, 1: horizon-
and adaptive local search with genetic algorithm. For a review of tal)
solution methodologies for the DFLP, one can see Balakrishnan and 
1 If facility i is rearranged in stage t (xt−1,i
Cheng (1998) and Kulturel-Konak (2007). rti = xti or yt−1,i = yti or dt−1,i = dti )
Although some approaches mentioned above have been applied 0 Otherwise
to solve the unequal-area DFLP, their efficiency still needs to
be improved. Wang–Landau (WL) sampling algorithm (Landau, Next, a mathematical model of the equal-area DFLP is pre-
Tsai, & Exler, 2004; Liu, Hao, Li, Xue, Liu, & Huang, 2016; Wang sented.
1054 J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063

Objective: [10 0 0, 10 0 0+). The energy interval [ E


, E ) is denoted by [E],
where E
rounds E down to its nearest integer, E rounds E up
Minimize : total cost E (X )
to its nearest integer. For example, E = 704.36 falls into energy

T 
N 
N 
T 
N
interval [704, 705), which is denoted by [704.36].
= Cti j · Fti j (|xti − xt j | + |yti − yt j |) + Rti rti . (1)
The basic process of applying the WL sampling algorithm to
t=1 i=1 j>i t=2 i=1
the DFLP is as follows. Since at the beginning of the algorithm
Constraints: all possible energies and the density of states g(E(X)) are un-
known, we set respectively the density of states g(E(X)) and the
Ii ∩ I j = ∅ ∀i, j, (2)
corresponding histogram to 1 when accessing to a new energy.
In the simulations on the DFLP, the new configurations can be
xti − 0.5lti ≥ 0 ∀t, i, (3) generated by a heuristic configuration updating procedure (see
Section 3.2), but not all the new configurations can be accepted.
xti + 0.5lti ≤ L ∀t, i, (4) The acceptance probability of a new configuration is proportional
to the reciprocal of the density of states g(E(X)) associated with
yti − 0.5wti ≥ 0 ∀t, i, (5) the new configuration. Assume that the original configuration is
X1 , and the new configuration is X2 . The acceptance probability P
from X1 to X2 is as follows:
yti + 0.5wti ≤ W ∀t, i, (6)

xti , yti , lti , wti ≥ 0 ∀t, i, (7)  


g(E (X1 ))
p(X1 → X2 ) = min ,1
Here X represents a configuration and X = (x11 , y11 , d11 , …, x1 N , g(E (X2 ))
y1 N , d1 N; …; xt 1 , yt 1 , dt 1 , …, xtN , ytN , dtN; …; xT 1 , yT 1 , dT 1 , …, xTN ,
= min(eln(g(E (X1 )))−ln(g(E (X2 ))) , 1 ). (10)
yTN , dTN ), where (xti , yti ) is the pick-up/drop-off location of facility
i in stage t and (xt 1 , yt 1 , dt 1 , …, xtN , ytN , dtN ) is a layout in stage
t, 1 ≤ t ≤ T, 1 ≤ i ≤ N. Objective (1) is to minimize the sum of
the material handling and rearrangement costs. Constraint (2) in- If X2 is accepted, then g(E(X2 )) will be multiplied by a modifi-
dicates that there is no overlapping between any two facilities. cation factor fi , and the histogram function H([E(X2 )]) will add 1,
Constraints (3)–(6) make sure that the facilities are not placed that is, g(E(X2 )) = fi ∗ g(E(X2 )) (ln(g(E(X2 ))) = lnfi + ln(g(E(X2 )))),
out of the plant floor. Constraint (7) is the constraint of variables’ H([E(X2 )]) = H([E(X2 )]) +1 (the initial value of i is 0); otherwise,
range, and lti and wti are calculated as follows: g(E(X1 )) and H([E(X1 )]) will be updated as follows: g(E(X1 )) = fi ∗
g(E(X1 )) (ln(g(E(X1 ))) = lnfi + ln(g(E(X1 )))), H([E(X1 )]) = H([E(X1 )])
lti = Lti dti + Wti (1 − dti ) ∀t, i, (8) +1. Because the value of the density of states g(E(X)) will become
so large that it will lead to the data overflow, we update g(E(X))
wti = Lti (1 − dti ) + Wti dti ∀t, i. (9) with the formula that contains logarithm. The initial value of the
modification factor fi is set differently in different model. If the
3. Solution procedure for the DFLP value of f0 is too small, it will take a long time to find all the pos-
sible energies; conversely, if the value of f0 is too large, there will
In this section, we propose a heuristic Wang–Landau (HWL) be a statistical error. A value of e (2.7181...) is chosen for f0 in this
sampling algorithm for solving the DFLP with unequal-area facil- paper. The flatness of the histogram determines the convergence
ities. First, we introduce Wang–Landau sampling algorithm. Then, of the WL sampling algorithm. In the simulations, we check the
some heuristic configuration updating strategies are put forward flatness of the histogram every 103 MC steps, where a MC step
to update the current configuration. Finally, the framework of the represents a new configuration generated by the algorithm. While
proposed HWL is presented. the histogram is flat, the density of states g(E(X)) will converge to
its true value by proportioning to the accuracy of the modification
3.1. Wang–Landau sampling algorithm factor ln(fi ). After this, we
 reduce the modification factor fi by a
modified formula: fi+1 = fi , and reset the values of all entries of
Wang–Landau (WL) sampling algorithm is a novel Monte Carlo H([E(X)]) to 0. Then we start a new round of iteration. However, in
(MC) algorithm and was proposed by Wang and Landau (2001). reality, it is hard to get the absolute flat histogram. The so-called
It gets a flat histogram of energy by a random walk in the whole “flat histogram” in the WL sampling algorithm refers to the values
energy space following a certain acceptance criterion, so as to es- of all entries of H([E(X)]) are greater than the value that the
timate the density of states g(E(X)), i.e., the number of all possible average of all entries of histogram <H([E(X)])> multiplies k (0 <
states (configurations) for an energy level E(X) of the system. The k < 1), where k is based on the complexity of the simulation as
total cost E(X) represents the energy in this paper. The algorithm well as the accuracy of g(E(X)) to determine. In this paper, we set
samples the energies of the whole energy space comprehensively the value of k to 0.7. After the modification factor fi is less than a
while obtaining the density of states g(E(X)) for the range of threshold ffinal , the algorithm stops. At this time, g(E(X)) will also
possible energies. Because of the comprehensive sampling, we can converge to its true value by proportioning to the accuracy of the
find the approximate global optimal value of the energies, and this modification factor ln(ffinal ). ffinal is the control parameter of the ac-
is the theoretical principle we apply the WL sampling algorithm curacy of g(E(X)) and also determines the number of MC iterations
to solve the DFLP. in the whole simulation process. If the value of ffinal is too small,
Before introducing the WL sampling algorithm to the DFLP, we it will take a long time to terminate the simulation; conversely, if
first give a partition for the energies. Considering that the values the value of ffinal is too large, the density of states g(E(X)) will not
of energies in range are positive real numbers in this paper, we converge to its real value. In this paper, we set ffinal = 1.0 0 01.
divide all possible energies of the configurations into numerable In order to find the configuration with the lowest energy Emin ,
intervals. For example, we divide [0, 10 0 0] into 10 0 0 energy we record the energy and the corresponding configuration each
intervals [0, 1), [1, 2), …, [999, 10 0 0), and the numbers that are time we find a configuration with lower energy in the process of
bigger than 10 0 0 will be divided into the single energy interval the WL simulations.
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1055

Fig. 1. The envelope rectangle of a layout.

3.2. Heuristic configuration updating procedure

In the WL sampling algorithm, each iterative step must update


the current configuration. To update the configuration, for the
layout in each stage, we use the combination of the following
strategies:

• vacant point strategy;


• pushing strategy;
• pressuring strategy.

where are described in the next sections.


Fig. 2. The overlapping distance between facility i and facility j.

3.2.1. Vacant point strategy


Before introducing the vacant point strategy, the concepts of
stage t. Suppose we need to move facility j. The movement rules
the envelope rectangle and the vacant point are given first. The
are as follows.
envelope rectangle is the rectangle that could envelop all facilities
If facilities i and j overlap each other as Fig. 2(a), we first check
exactly. The vacant point is the point that is inside envelope
whether Atij is smaller than Btij or vice versa. If Atij is smaller,
rectangle, but not inside any facility. As shown in Fig. 1, the dotted
facility j will be moved to the right by Atij value. If Btij is smaller,
rectangle represents an envelope rectangle, and points in the
facility j will be moved to the up by Btij value. If facilities i and j
shaded region of the envelope rectangle can be selected as the
overlap as Fig. 2(b), and if Atij is smaller, facility j will be moved
vacant points.
to the right by Atij value; otherwise, facility j will be moved to the
The vacant point strategy of updating the layout is descripted
up by Btij value. If facilities i and j overlap as Fig. 2(c) or (d) and
as follows. First, generate randomly 100 vacant points in the
Atij is smaller, facility j will be moved to the right by Atij value;
envelope rectangle. Then, choose the facility with the largest unit
otherwise, facility j will be moved to the up by Btij value.
material handling cost (UMHC) and place temporarily its centroid
The process of the pushing strategy is as follows. First, choose
at every vacant point with vertical and horizontal directions of fa-
the facility with the largest overlapping area. Then, move the
cility. Here, the formula for calculating the unit material handling
chosen facility according to the movement rules. If the chosen
cost of facility i in stage t is as follows:
facility overlaps with more than one facility, the facility which
N
j=1 Cti j · Fti j (|xti − xt j | + |yti − yt j |) has the minimum index is the reference of the chosen facility. If
UMHCti = N there are other facilities that have the overlapping, this procedure
j=1 Fti j is repeated.
f or i = 1, 2, . . . , N, t = 1, 2, . . . , T . (11) We use an example to describe the process of the pushing
strategy. Suppose we have a layout as Fig. 3(a) in stage t. Facility
Next, calculate the unit material handling cost of the chosen
2 overlaps with facilities 1 and 3 and has the largest overlapping
facility associating with each vacant point, and place formally the
area, so we choose facility 2 to move. Facility 1 has the minimum
centroid of the chosen facility at the vacant point where the unit
index between facilities 1 and 3, and the overlapping between
material handling cost of the chosen facility is the smallest. With
facilities 2 and 1 is considered. Because Bt 12 is smaller than At 12 ,
the positions of other facilities unchanged, finally we gain two
facility 2 is moved to the up. The result of this movement is
new layouts due to the two placing directions.
shown in Fig. 3(b). As you can see, this movement leads to the
overlapping between facilities 2 and 4. We continue this pushing
3.2.2. Pushing strategy procedure and solve the overlapping between facilities 2 and 4.
It is inadvisable that facilities overlap with each other after The final layout is shown in Fig. 3(c).
executing the vacant point strategy. To make sure that constraint
(2) is satisfied, the pushing strategy is applied. In order to avoid
repetitive movement, a principle of to-up or to-right is used in the 3.2.3. Pressuring strategy
pushing strategy, that is, the facility can only move to the up or to After using the pushing strategy to prevent overlapping among
the right. Suppose facilities i and j overlap each other. As shown in facilities, we may gain a layout where facilities are placed out
Fig. 2, Atij labeled in Fig. 2(a)–(d) represents the overlapping of the plant floor. To make the layout compact, we apply the
distance in the direction of the X-axis between facilities i and j following pressuring strategy.
in stage t. Btij labeled in Fig. 2(a)–(d) represents the overlapping First, calculate the material handling cost (MHC) of each facility
distance in the direction of the Y-axis between facilities i and j in in each stage by the formula:
1056 J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063

Fig. 3. An example of pushing strategy.

touches facility 2. The next facility nearest to the basic facility is


facility 1. This facility can be first vertically moved until its up
edge touches the basic facility, and then be horizontally moved
with no obstruction until its centroid is aligned with the centroid
of the basic facility (see Fig. 4(c)). Facility 2 is vertically moved
until its down edge touches the basic facility. The final layout is
shown in Fig. 4(d).

3.2.4. Computation of rearrangement costs


We define the facility group as the combination of some
facilities. If the centroids of all facilities in the facility group are
unchanged relative to each other in two successive stages, we con-
sider the relative positions among these facilities are unchanged.
As long as the relative positions among facilities are unchanged
in two successive stages, we think the all facilities which are con-
tained in the facility group do not generate rearrangement costs.
We use an example to describe this computational process
of rearrangement costs. As shown in Fig. 5(a) and (b), because
the positions of facilities 1–4 change in two successive stages,
there are four rearrangement costs by expression (1). In fact, the
relative positions between facilities 3 and 4 are unchanged. We
take facilities 3 and 4 as a facility group. If we move the entire
layout upward shown in Fig. 5(b) until facilities 3 and 4 reach
the positions as same as the positions shown in Fig. 5(a), we can
obtain the layout shown in Fig. 5(c). Obviously, Fig. 5(b) and (c)
are isomorphic. If the layout in Fig. 5(c) is still a feasible layout
Fig. 4. An example of pressuring strategy.
which satisfies constraints (2)–(9), Fig. 5(c) will lead to only two
rearrangement costs compared with Fig. 5(a). In the computational
process of rearrangement costs, we always choose the maximum

N
MHCti = Cti j · Fti j (|xti − xt j | + |yti − yt j |) facility group that contains more facilities. This will reduce the
j=1 rearrangement costs of the layout.
f or i = 1, 2, . . . , N, t = 1, 2, . . . , T . (12)
3.2.5. Configuration updating process
Then, choose the facility with the largest material handling cost We update a configuration by updating the layouts at different
and name it the basic facility. Move the candidate facility towards stages in the configuration. For the first stage, we apply the vacant
the basic facility. Here, the candidate facility is the facility which is point strategy on the original layout to get two temporary layouts
nearest to the basic facility. There are two ways of the movement: where the picked facility has two different placing directions.
first vertically move and then horizontally move, and first hori- Hereafter, we apply the pushing strategy on these two temporary
zontally move and then vertically move. We choose the way by layouts to eliminate the overlapping among facilities, and use the
which we could get more improvement in the material handling pressuring strategy to further get two compact layouts. Finally,
costs of the layout. In each way of the movement, we move the we calculate respectively the material handling costs of these two
candidate facility until it touches another facility or its centroid is compact layouts, and choose the layout with the smaller material
aligned with the centroid of the basic facility. If these two cases handling cost as the updated layout.
may happen, we choose the former in priority. If there are still For other stages, first we copy the updated layout of the pre-
other facilities that can be moved, this procedure is repeated. vious stage as the original layout of this stage. Then we apply the
We use an example to describe the process of the pressuring vacant point strategy on the original layout to get two temporary
strategy. Suppose we have a layout as Fig. 4(a), and facility 4 is layouts, and apply the pushing strategy and the pressuring strategy
the basic facility. It is obvious that facility 3 is nearest to facility 4, on these two temporary layouts to get two compact layouts with-
so we move facility 3 towards facility 4. No matter which way we out overlapping. At this time, we calculate the material handling
choose for the movement, the movement will lead to a layout as costs and the rearrangement costs of these two compact layouts,
Fig. 4(b). The termination condition of the movement is that the respectively. Here, the rearrangement costs are generated by
down edge of facility 3 touches the basic facility, and its right edge comparing the newly produced compact layouts with the original
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1057

Fig. 5. An example of computation of rearrangement costs in two successive stages.

Table 1
Comparison of computational results by five algorithms for case P6-6.

HGA (2005) TS/BSH (2010) MOPSO (2012) MPSO (2015) HWL (in this paper)

ReCost Cost Time (minutes) Cost Time (minutes) Cost Time (minutes) Cost Time (minutes) Cost Time (minutes)

19 6569.0 29.4 6648.3 27.7 – – 6763.7 39.2 6523.5 4.6


50 – – – – 6659.4 8.4 – – 6597.0 5.2

layouts and computed by the strategy of Section 3.2.4. The sums of 4. Computational results and discussion
the material handling costs and rearrangement costs are the total
costs for these two compact layouts. We choose the smaller total In order to evaluate the performance of the HWL sampling
cost and compare it with the material handling cost of the original algorithm, we apply the algorithm on four groups of cases. The
layout. If the smaller total cost is less than the material handling first group contains two classical dynamic facility layout cases:
cost of the original layout, we set the compact layout with the P6-6 and P12-4 (Yang & Peters, 1998). The second group contains
smaller total cost as the updated layout. Otherwise, repeat this two optimality comparison test cases created in this paper. The
updating process starting from the vacant point strategy. If this third group contains a dynamic facility layout case with practical
process is repeated more than 103 times and at every time we application and the fourth group contains three single-stage facil-
do not gain an accepted layout, we set the original layout as the ity layout cases. The algorithm is compiled by using JAVA language
updated layout. and all experiments are carried out on a PC with an Intel Core 2
Duo, 2.94 gigahertz CPU and 2.0 gigabytes RAM. For each case, we
run the HWL sampling algorithm ten times independently.

3.3. Description of WL combining with heuristics 4.1. Computational results of two classical dynamic facility layout
cases
By incorporating the heuristic configuration updating procedure
into the WL sampling algorithm, a heuristic Wang–Landau (HWL) The first classical dynamic facility layout case is to place 6 fa-
layout algorithm for the unequal-area DFLP is presented. The cilities in 6 stages (P6-6), and the second is to place 12 facilities in
detailed computational procedure is outlined below. 4 stages (P12-4). These two cases were first presented in Yang and
Step 1: Generate an initial configuration X1 (where the layout Peters (1998), and were further calculated in Dunker et al. (2005),
for each stage is as same as that in Yang & Peters, 1998). Calculate Jolai et al. (2012), McKendall and Hakobyan (2010), and Asl and
E(X1 ). Let the optimal configuration Xmin = X1 , and the total cost Wong (2015). Jolai et al. (2012), used 50 units and 200 units for
of optimal configuration Emin = E(X1 ). Initialize the set of intervals rearrangement cost of each facility (the following brief write down
containing visited energies (i.e., total costs) S = {[E(X1 )]}, the for ReCost) on the first and second cases, respectively, as in Yang
state density function g(E(X1 )) = 1, and the histogram function and Peters (1998); but it was 19 and 50 units for the first and sec-
H([E(X1 )]) = 1. Let i = 0, l = 0, k = 0.7, f0 = e. ond cases in Dunker et al. (2005), Mckendall and Hakobyan (2010),
Step 2: Update current configuration X1 by the heuristic con- and Asl and Wong (2015). In this paper, in order to compare the
figuration updating procedure, and gain a new configuration X2 . results with the above calculations, we use not only 19 and 50
Compute E(X2 ). Let g(E(X2 )) = 1, H([E(X2 )]) = 1, and l = l + 1. units for P6-6, but also 50 and 200 units for P12-4. The 30 × 30
/ S, let S = S ∪ {[E (X2 )]}.
Step 3: If [E (X2 )] ∈ and 50 × 50 plant floors are chosen for P6-6 and P12-4, respec-
Step 4: If random (0, 1) <min(eln(g(E (X1 )))−ln(g(E (X2 ))) , 1 ), accept tively, and the material handling cost per unit distance is set to 1.
X2 , and let X1 = X2 , E(X1 ) = E(X2 ), ln(g(E(X2 ))) = lnfi + ln(g(E(X2 ))), The computational results by the HWL sampling algorithm
H([E(X2 )]) = H([E(X2 )]) + 1, go to step 5; otherwise, do not accept for case P6-6 are shown in Table 1. As shown in the table, the
X2 and let ln(g(E(X1 ))) = lnfi + ln(g(E(X1 ))), H([E(X1 )]) = H([E(X1 )]) best objectives and running times obtained by the HWL are
+1, go to step 6. uniformly better than results by other four algorithms for both
Step 5: If E(X2 ) < Emin , let Xmin = X2 , Emin = E(X2 ). 19 and 50 ReCost. In fact, for ReCost of 19, the best objective
Step 6: If l% 10 0 0 = 0, go to step 7; otherwise, go to step 2. obtained by the HWL is reduced by (6569.0 − 6523.5)/6569.0 =
Step 7: If H ([E (X )] )k < H ([E (X )] ) > for all visited energy 0.69% compared with that by the hybrid genetic algorithm (HGA)
intervals [E (X )] ∈ S, then
 go to step 8; otherwise go to step 2. in Dunker et al. (2005), reduced by (6648.3 − 6523.5)/6648.3
Step 8: Let fi+1 =, fi , i = i + 1. = 1.88% compared with that by the tabu search and boundary
Step 9: If fi < 1.0 0 01, then the algorithm stops, output Xmin and search heuristic (TS/BSH) in McKendall and Hakobyan (2010),
Emin ; otherwise, for each [E(X)] in S, let H([E(X)]) = 0, but g(E(X)) and reduced by (6763.7 – 6523.5)/6763.7 = 3.55% compared
remains unchanged, and go to step 2. with that by the modified particle swarm optimization (MPSO) in
1058 J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063

Fig. 6. P6-6 layouts by the HWL for stages 1–6 (from left to right, from top to Fig. 7. P6-6 layouts by the HWL for stages 1–6 (from left to right, from top to bot-
bottom) with objective 6523.5. tom) with objective 6597.0.

Table 2
and n is the running times. For ReCost of 19, the deviation
Statistics of results by the HWL sampling algorithm for case P6-6.
between the best and worst objectives is 96 and the standard
ReCost Best Worst Average Standard deviation is small. For ReCost of 50, the deviation between the
objective objective objective deviation
best and worst objectives is 86 and the standard deviation is small
19 6523.5 6619.5 6582.1 26.1 too. This means the objectives obtained by the HWL sampling
50 6597.0 6683.0 6648.3 29.6
algorithm are not scattered for both ReCosts of 19 and 50.
The computational results by the HWL sampling algorithm for
case P12-4 are shown in Table 3. From the table, the objective
Asl and Wong (2015). The layouts for objective 6523.5 are shown in obtained by the HWL is a little worse than that by TS/BSH for Re-
Fig. 6. Facility 2 is replaced in the beginning of stage 4, and others Cost of 50. But the objectives are better than those by HGA, MPSO
are consistent with the original, which leads to one rearrangement and MOPSO for ReCosts of 50 and 200, respectively. In fact, for
in the layouts. For ReCost of 50, the best objective obtained by the ReCost of 50, the best objective obtained by the HWL is increased
HWL is reduced by (6659.4 − 6597.0)/6659.4 = 0.94% compared by (26950.5 − 26845.5)/26950.5 = 0.39% compared with that
with that by the multi-objective particle swarm optimization by TS/BSH, but reduced by (27748.0 − 26950.5)/27748.0 = 2.88%
(MOPSO) in Jolai et al. (2012). The layouts for objective 6597.0 compared with that by HGA and (28826.6 – 26950.5)/28826.6
are shown in Fig. 7. These 6 layouts keep the same positions in = 6.51% compared with that by MPSO. In terms of computation
different stages, which means there are no rearrangement costs. time, the HWL overmatches other algorithms for ReCost of 50. The
The statistics of the computing results for case P6-6 are shown layouts for objective 26950.5 are shown in Fig. 8. Ten facilities are
in Table 2. Table 2 shows the best and worst objectives, the replaced in the beginning of stage 2 except facilities 10 and 11.
average objectives, and the standard deviations of these objectives Besides, positions of all facilities are kept invariant in other stages.
over 10 independent runs. The standard deviation is calculated by There are 10 rearrangements in all layouts. For ReCost of 200,
the following formula: the best objective obtained by the HWL is reduced by (28115.5 −
 27924.5)/28115.5 = 0.68% compared with that by MOPSO, while
n
i=1 (ai − ā )2 the computational time of the former is a litter longer than that
(13)
n of the latter. The layouts for objective 27924.5 are shown in Fig. 9.
Here, ai represents the objective by the HWL sampling algo- There is none rearrangement in all layouts, which is similar to the
rithm in the ith run; ā represents the average of these objectives layouts of objective 6597.0 for case P6-6.
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1059

Table 3
Comparison of computational results by five algorithms for case P12-4.

HGA (2005) TS/BSH (2010) MOPSO (2012) MPSO (2015) HWL (in this paper)

ReCost Cost Time (minutes) Cost Time (minutes) Cost Time (minutes) Cost Time (minutes) Cost Time (minutes)

50 27748.0 160.0 26845.5 82.0 – – 28826.6 185.1 26950.5 57.8


200 – – – – 28115.5 50.5 – – 27924.5 63.4

Table 4
Statistics of results by the HWL sampling algorithm for case P12-4.

ReCost Best Worst Average Standard


objective objective objective deviation

50 26950.5 27581.0 27374.4 176.1


200 27924.5 28321.0 28057.5 93.2

The statistics of the computing results for case P12-4 are


shown in Table 4. Table 4 shows the best and worst objectives, the
average objectives, and the standard deviations of these objectives
over 10 independent runs. For ReCost of 50, the deviation between
the best and worst objectives is 630.5, and the standard deviation
is small. This means the objectives obtained by the HWL are not
scattered. For ReCost of 200, the deviation between the best and
worst objectives is 396.5, which is much smaller than that for
ReCost of 50. The standard deviation is small too, it means the
objectives obtained by the HWL are not scattered as the former.

4.2. Optimality comparison test

Two optimality comparison test cases are created to illustrate


the deviations between the results by the HWL and the optimal
results. The first case, which places 3 facilities in 3 stages, is
denoted as P3-3. The second case, which places 4 facilities in 2
Fig. 8. P12-4 layouts by the HWL for stages 1–4 (from left to right, from top to
stages, is denoted as P4-2. For P3-3, we define a 12 × 9 plant
bottom) with objective 26950.5.
floor, i.e., length of the plant floor along the X-axis is 12 and width
of the plant floor along the Y-axis is 9. The ReCost is 15 units and
the dimensions of facilities and the material flows among facilities
in each stage are presented in Appendix 1. For P4-2, a 10 × 11
plant floor is defined, and the ReCost is 9 units. The dimensions of
facilities and the material flows among facilities in each stage are
presented in Appendix 2.
For P3-3, we can construct an optimal configuration by refer-
ring Eqs. (1)–(9). The objective of this optimal configuration is
1510.5 and the corresponding layouts are shown in Fig. 10, where
two facilities are reset in the beginning of stage 2 except facility 3,
and positions of all facilities in stage 3 are consistent with those in
stage 2. This case is also solved by the proposed HWL sampling al-
gorithm. Over 10 independent runs, the best objective obtained by
the HWL is equal to that of the optimal configuration by referring
Eqs. (1)–(9), and the best layouts by the HWL are the same as lay-
outs of the optimal configuration, which means the HWL sampling
algorithm can find the optimal configuration for case P3-3.
For P4-2, we can also construct an optimal configuration by
referring Eqs. (1)–(9). The objective of the optimal configuration is
827.5 and the corresponding layouts are shown in Fig. 11. Further-
more, we use the proposed HWL sampling algorithm to solve case
P4-2. By generating a legitimate initial layout, the best layouts with
the objective 848.5 are obtained by the HWL over 10 independent
runs. The best layouts by the HWL are shown in Fig. 12. The
deviation between the best result by the HWL and the objective of
the optimal layouts is 21. We can see from Fig. 11, in the beginning
of stage 2 we reset facility 4 so that its centroid is aligned with
Fig. 9. P12-4 layouts by the HWL for stages 1–4 (from left to right, from top to the centroid of facility 1, and an optimal configuration (Fig. 11) is
bottom) with objective 27924.5.
obtained. However, it is hard for the HWL to search this optimal
configuration. Note that in the beginning of stage 2, as facility 3
1060 J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063

Fig. 10. P3-3 optimal layouts for stages 1–3 (from left to right) with objective 1510.5.

Table 5
Comparison of computational results by four algorithms for three single-stage cases.

Case FLOAT (1993) HOT (2001) MPSO (2015) HWL (in this paper)

P8 – – 208.7 193.1
P11 – – 1335.6 1280.1
P20 1333.8 1287.3 1264.2 1264.6

class-B batteries, 500 class-C batteries in the third stage. The


material flows among all regions and the matrix of material
handing cost per unit distance at each stage are summarized in
Appendix 3. It costs 500 dollars if one region is reset.
The sum of costs by the HWL for this case is 236821.5 dollars,
Fig. 11. P4-2 optimal layouts for stages 1–2 (from left to right) with objective 827.5.
and the output layouts are shown in Fig. 13, where positions of all
regions in the second stage are consistent with the original and
seven regions are reset except region 1 in the third stage, which
leads to 7 rearrangements in the layouts. To test the difference
of the effectiveness of the HWL on the DFLP relative to the SFLP,
we amount the material flows of three stages in the plant to the
material flows of single stage, that is, we convert the DFLP into a
SFLP which is a special DFLP with one stage. The sum of costs by
executing the HWL for the SFLP is 257622.0 dollars. The former
is reduced by (257622.0 – 236821.5)/257622.0 = 8.1% in product
cost compared with the latter, which shows the significance of the
DFLP. The optimal configuration by the HWL over ten independent
runs for the SFLP is shown in Fig. 14.

Fig. 12. P4-2 layouts by the HWL for stages 1–2 (from left to right) with objective 4.4. Computational results for three single-stage facility layout cases
848.5.
To check further the effectiveness of the HWL, the computa-
tional results for three single-stage facility layout (i.e., SFLP) cases
has the largest UMHC according to Eq. (11), the centroid of facility are presented in this section. These three cases which include 8
3 is reset within vacant region and thereafter the pushing strategy facilities, 11 facilities, and 20 facilities are denoted as P8, P11, P20,
and the pressuring strategy are executed. Obviously, it could not respectively (Asl & Wong, 2015). The material handling cost per
yield the optimal layout in stage 2 as shown in Fig. 11, so the unit distance is set to 1 for each of these cases. The dimensions
layout in stage 1 remains invariant according to the configuration of the plant floors are chosen as 12 × 12, 15 × 15, and 14 × 14,
updating process described in Section 3.2.5. But the configuration respectively.
by the HWL is sub-optimal, demonstrating the effectiveness of the The computational results by the HWL sampling algorithm for
HWL in solving hard case for the unequal-area DFLP. the three single-stage cases are shown in Table 5. The objectives
obtained by the HWL are better than the results by the modified
4.3. Computational results for a dynamic facility layout case with particle swarm optimization (MPSO) in Asl and Wong (2015) for
practical application both P8 and P11. In fact, it decreases (208.7 − 193.1)/208.7 = 7.5%
for P8 and (1335.6 − 1280.1)/1335.6 = 4.2% for P11 compared with
We give a dynamic facility layout case with practical applica- those by MPSO. The improvements obtained by the HWL for P20
tion. There is a plant which is mainly responsible for production are (1333.8 − 1264.6)/1333.8 = 5.2% and (1287.3 − 1264.6)/1287.3
of automotive battery start-up line. The dimension of the plant is = 1.8%, in contrast to FLOAT in Imam and Mir (1993) and HOT
24 × 22 square meters. In accordance with the principle of similar in Mir and Imam (2001), respectively. It only increases (1264.6 −
facilities placed together, this plant is divided into 8 regions (i.e., 1264.2)/1264.6 = 0.03% for P20 compared with that by MPSO. The
facilities). The names and dimensions of all regions are presented optimal layouts by the HWL over 10 independent runs for P8, P11,
in Appendix 3. The production is divided into three stages, and and P20 are shown in Fig. 15, Fig. 16, and Fig. 17, respectively.
plans of each stage are as follows: product 4300 class-A batteries, The statistics of the computing results are shown in Table 6.
10 0 0 class-B batteries, 2400 class-C batteries in the first stage; In the table, the best and worst objectives, the average objectives,
product 1600 class-A batteries, 0 class-B batteries, 5500 class-C and the standard deviations over 10 runs are presented. For all
batteries in the second stage; product 20 0 class-A batteries, 580 0 the three single-stage cases, the standard deviations are small
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1061

Fig. 13. Layouts by the HWL for stages 1–3 (from left to right) with objective 236821.5 in dynamic case with practical application.

Fig. 14. Layout by the HWL with objective 257622.0 in single-stage case with prac-
tical application.
Fig. 17. P20 layout by the HWL with objective 1264.6.

indicating that the objective values obtained by the HWL are not
scattered.

5. Conclusions

In this paper, we have discussed the dynamic facility layout


problem (DFLP) with unequal-area facilities which aims at mini-
mizing the sum of the material handling and rearrangement costs.
Firstly, we present a mathematical model to describe this problem.
Then we combine Wang–Landau (WL) sampling algorithm and
some heuristics to solve the problem. The WL sampling algorithm
obtains a flat energy histogram by walking in the energy space
Fig. 15. P8 layout by the HWL with objective 193.1.
randomly, which can estimate the density of states of the system
accurately. It will reach the near lowest energy when estimating
the density of states. Four groups of cases are used to test the pro-
posed algorithm, and the results indicate the effectiveness of the
algorithm. Furthermore, the proposed method will be beneficial
for the DFLP with variable shapes and areas of facilities through-
out the time horizon, and can be generalized to the research and
application on multi-objective group decision-making.

Acknowledgments

This work is supported by the National Natural Science Foun-


dations of China (Grant Nos. 61373016, 61472147, 61373064 and
61403206), the Major Program of the National Social Science Foun-
dation of China (Grant No. 16ZDA047) and the Natural Science
Fig. 16. P11 layout by the HWL with objective 1280.1. Foundation of Jiangsu Province (Grant No. BK20141005).

Table 6 Appendix 1. Parameters of optimality comparison test case


Statistics of results by the HWL sampling algorithm for three single-stage cases. P3-3
Case Best objective Worst objective Average objective Standard deviation

P8 193.1 198.0 195.9 1.3


Dimensions of facilities in P3-3.
P11 1280.1 1302.4 1288.6 6.1
Facility 1 2 3
P20 1264.6 1313.8 1289.7 16.9
Dimension (Li , Wi ) (6, 4) (6, 4) (9, 6)
1062 J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063

Direction of Transporting material/item


Material flows of stage 1 for P3-3. transporting
Stage 1 Stage 2 Stage 3
material
1–3 4300 1600 200
Facility 1 2 3 1–4 10 0 0 0 5800
1 0 50 1 2–4 2400 5500 500
2 50 0 40 3–4 4300 1600 200
3 1 40 0 3–5 0 0 700
3–7 3400 7100 6300
Material flows of stage 2 for P3-3. 4–5 7700 7100 6500
4–6 4300 1600 200
5–4 3400 5500 6300
Facility 1 2 3 5–6 4300 1600 200
6–3 3400 5500 6300
1 0 50 40 6–7 4300 1600 200
2 50 0 1 7–8 7700 7100 6500
3 40 1 0
Matrix of material handing cost per unit distance (Cij ).
Material flows of stage 3 for P3-3.
Region 1 2 3 4 5 6 7 8

Facility 1 2 3 1 0 0.5 0.2 0.3 0.3 0.5 0.3 0.4


2 0.5 0 0.6 0.5 0.5 0.4 0.3 0.4
1 0 20 50
3 0.2 0.6 0 0.4 0.3 0.2 0.6 0.2
2 20 0 1
4 0.3 0.5 0.4 0 0.5 0.4 0.2 0.3
3 50 1 0
5 0.3 0.5 0.3 0.5 0 0.5 0.5 0.6
6 0.5 0.4 0.2 0.4 0.5 0 0.5 0.3
7 0.3 0.3 0.6 0.2 0.2 0.5 0 0.5
8 0.4 0.4 0.2 0.3 0.6 0.3 0.5 0

Appendix 2. Parameters of optimality comparison test case


P4-2 References

Anjos, M. F., & Vannelli, A. (2006). A new mathematical-programming framework


Dimensions of facilities in P4-2.
for facility-layout design. Informs Journal on Computing, 18(1), 111–118.
Asl, D. A., & Wong, K. Y. (2015). Solving unequal-area static and dynamic facility lay-
out problems using modified particle swarm optimization. Journal of Intelligent
Facility 1 2 3 4
Manufacturing, 26, 1–20.
Dimension (Li , Wi ) (8, 3) (11, 4) (6, 3) (7, 3) Balakrishnan, J., & Cheng, C. H. (1998). Dynamic layout algorithms: A
state-of-the-art survey. Omega, 26(4), 507–521.
Material flows of stage 1 for P4-2. Baykasoglu, A., & Gindy, N. N. Z. (2001). A simulated annealing algorithm for dy-
namic layout problem. Computers & Operations Research, 28, 1403–1426.
Bozer, Y. A., & Meller, R. D. (1997). A reexamination of the distance-based facility
Facility 1 2 3 4 layout problem. IIE Transactions, 29, 549–560.
Bozorgi, N., Abedzadeh, M., & Zeinali, M. (2015). Tabu search heuristic for efficiency
1 0 10 1 1 of dynamic facility layout problem. International Journal of Advanced Manufac-
2 10 0 10 10 turing Technology, 77, 689–703.
3 1 10 0 10 Dong, M., Wu, C., & Hou, F. (2009). Shortest path based simulated annealing algo-
4 1 10 10 0 rithm for dynamic facility layout problem under dynamic business environment.
Expert Systems with Applications, 36, 11221–11232.
Drira, A., Pierreval, H., & Hajri-Gabouj, S. (2007). Facility layout problems: A survey.
Material flows of stage 2 for P4-2.
Annual Reviews in Control, 31(2), 255–267.
Dunker, T., Radons, G., & Westkamper, E. (2005). Combining evolutionary computa-
tion and dynamic programming for solving a dynamic facility layout problem.
Facility 1 2 3 4
European Journal of Operational Research, 165, 55–69.
1 0 1 10 80 Gonçalves, J. F., & Resende, M. G. C. (2015). A biased random-key genetic algorithm
2 1 0 10 10 for the unequal area facility layout problem. European Journal of Operational Re-
search, 246(1), 86–107.
3 10 10 0 10
Hosseini-Nasab, H., & Emami, L. (2013). A hybrid particle swarm optimization for
4 80 10 10 0
dynamic facility problem. International Journal of Production Research, 51(14),
4325–4335.
Hosseini, S., Khaled, A. A., & Vadlamani, S. (2014). Hybrid imperialist competitive
algorithm, variable neighborhood search, and simulated annealing for dynamic
facility layout problem. Neural Computing & Applications, 25(7–8), 1871–1885.
Appendix 3. Parameters of the case with realistic Imam, M. H., & Mir, M. (1993). Automated layout of facilities of unequal areas. Com-
implementation puters and Industrial Engineering, 24(3), 355–366.
Jolai, F., Tavakkoli-Moghaddam, R., & Taghipour, M. (2012). A multi-objective particle
swarm optimisation algorithm for unequal sized dynamic facility layout prob-
Names and dimensions of machine regions. lem with pickup/drop-off locations. International Journal of Production Research,
50(15), 4279–4293.
Komarudin, & Wong, K. Y. (2010). Applying ant system for solving unequal area fa-
No. Name Dimension No. Name Dimension cility layout problems. European Journal of Operational Research, 202(3), 730–746.
(meter × (meter × Kouvelis, P., Kuawarwala, A. A., & Gutierrez, G. J. (1992). Algorithms for robust sin-
meter) meter) gle and multiple period layout planning for manufacturing systems. European
Journal of Operational Research, 63, 287–303.
1 Raw material 4.0 × 4.0 5 Tin solder 3.2 × 2.5 Kulturel-Konak, S. (2007). Approaches to uncertainties in facility layout problems:
region region Perspectives at the beginning of the 21st century. Journal of Intelligent Manufac-
2 Clip assemble 3.6 × 2.0 6 Punching 3.8 × 3.6 turing, 18(2), 273–284.
region region Landau, D. P., Tsai, S. H., & Exler, M. (2004). A new approach to Monte Carlo simula-
3 Coil wire region 3.0 × 2.5 7 Testing region 4.0 × 3.5 tions in statistical physics: Wang–Landau sampling. American Journal of Physics,
4 Clamping region 4.8 × 4.0 8 Packaging 4.0 × 3.6 72(10), 1294–1302.
region Li, L., Li, C., Ma, H., & Tang, Y. (2015). An optimization method for the remanufac-
turing dynamic facility layout problem with uncertainties. Discrete Dynamics in
Material flows among machine regions. Nature and Society, 1–11.
J. Liu et al. / European Journal of Operational Research 262 (2017) 1052–1063 1063

Liu, J. F., Hao, L., Li, G., Xue, Y., Liu, Z. X., & Huang, J. (2016). Multi-objective layout Rosenblatt, M. J. (1986). The dynamics of plant layout. Management Science, 32(1),
optimization of a satellite module using Wang–Landau sampling method with 76–86.
local search. Frotiers of Information Technology & Electronic Engineering, 17(6), Şahin, R., & Türkbey, O. (2009). A new hybrid tabu-simulated annealing heuristic for
527–542. the dynamic facility layout problem. International Journal of Production Research,
Liu, J. F., Song, B. B., Yao, Y. L., Xue, Y., Liu, W. J., & Liu, Z. X. (2014). Wang–Landau 47(24), 6855–6873.
sampling in face-centered-cube hydrophobic-hydrophilic lattice model proteins. Schulz, B. J., Binder, K., Müller, M., et al. (2003). Avoiding boundary effects in
Physical Review E, 90(3), 042715. Wang–Landau sampling. Physical Review E, 67(6), 067102.
Liu, Q., & Meller, R. D. (2007). A sequence-pair representation and MIP-model-based Seaton, D. T., Wüst, T., & Landau, D. P. (2010). Collapse transitions in a flexible ho-
heuristic for the facility layout problem with rectangular departments. IIE Trans- mopolymer chain: Application of the Wang–Landau algorithm. Physical Review
actions, 39, 377–394. E, 81(1), 011802.
Ma, S., Cai, H., Zhang, Y., & Li, A. (2015). Dynamic facility layout method under Sherali, H. D., Fraticelli, B. M. P., & Meller, R. D. (2003). Enhanced model formula-
uncertain product demands. Chinese Journal of Mechanical Engineering, 26(11), tions for optimal facility layout. Operations Research, 51(4), 629–644.
1494–1502. Singh, S. P., & Sharma, R. R. K. (2006). A review of different approaches to the facil-
Mazinani, M., Abedzadeh, M., & Mohebali, N. (2013). Dynamic facility layout prob- ity layout problems. International Journal of Advanced Manufacturing Technology,
lem based on flexible bay structure and solving by genetic algorithm. Interna- 30(5–6), 425–433.
tional Journal of Advanced Manufacturing Technology, 65, 929–943. Ulutas, B. H., & Kulturel-Konak, S. (2012). An artificial immune system based algo-
McKendall, A. R., Jr, & Hakobyan, A. (2010). Heuristics for the dynamic facility lay- rithm to solve unequal area facility layout problem. Expert Systems with Appli-
out problem with unequal-area departments. European Journal of Operational Re- cations, 39(5), 5384–5395.
search, 201(1), 171–182. Ulutas, B., & Islier, A. A. (2015). Dynamic facility layout problem in footwear indus-
Mckendall, A. R., Jr, & Liu, W. H. (2012). New Tabu search heuristics for the dy- try. Journal of Manufacturing Systems, 36, 55–61.
namic facility layout problem. International Journal of Production Research, 50(3), Urban, T. L. (1993). A heuristic for the dynamic facility layout problem. IIE Transac-
867–878. tions, 25(4), 57–63.
Meller, R. D., Narayanan, V., & Vance, P. H. (1999). Optimal facility layout design. Wang, F., & Landau, D. P. (2001). Efficient, multiple-range random walk algorithm to
Operations Research Letters, 23(3–5), 117–127. calculate the density of states. Physical Review Letters, 86(10), 2050–2053.
Mir, M., & Imam, M. H. (2001). A hybird optimization approach for layout design of Yang, T., & Peters, B. A. (1998). Flexible machine layout design for dynamic and
unequal-area facilities. Computers & Industrial Engineering, 39(1), 49–63. uncertain production environments. European Journal of Operational Research,
Montreuil, B., & Venkatadri, U. (1991). Strategic interpolative design of dynamic 108(1), 49–64.
manufacturing systems layouts. Management Science, 37(6), 682–694.
Pourvaziri, H., & Naderi, B. (2014). A hybird multi-population genetic algorithm for
the dynamic facility layout problem. Applied Soft Computing, 24, 457–469.

You might also like