Oper Res Int J

DOI 10.1007/s12351-017-0334-5

ORIGINAL PAPER

Solving the multiple level warehouse layout problem

using ant colony optimization

Jean-Paul Arnaout1 • Caline ElKhoury2 • Gamze Karayaz3

Received: 29 August 2016 / Revised: 29 May 2017 / Accepted: 30 June 2017

 Springer-Verlag GmbH Germany 2017

Abstract This paper addresses the multiple level warehouse layout problem, which
involves assigning items to cells and levels with the objective of minimizing
transportation costs. A monthly demand and an inventory requirement are associ-
ated with every item type along with vertical and horizontal unit transportation
costs. The warehouse has one port to transport items vertically from ground floor to
the other levels, where each item must be assigned to exactly one cell on the
assigned level. An ant colony optimization (ACO) algorithm is adapted to this NP-
complete problem and its performance is evaluated by comparing its solutions to the
ones obtained using genetic algorithms (GA) as well as the optimal solutions for
small problems. The computational results reflected the superiority of ACO in large-
size problem instances, with a marginally better performance than GA in smaller
ones, while solving the tested instances within a reasonable computational time.
Furthermore, ACO was able to attain most of the known optimal solutions for small-
size problem instances.

Keywords Multiple level warehouse layout problem  Metaheuristic  Ant

colony optimization

& Jean-Paul Arnaout

arnaout.j@gust.edu.kw
1
Gulf University for Science and Technology, Mishref, Kuwait
2
Industrial Engineering Department, Lebanese American University, Beirut, Lebanon
3
Department of Management, Işik University, Istanbul, Turkey

123
 J.-P. Arnaout et al.

1 Introduction

In this paper, we adapt an ant colony optimization algorithm for the MLWLP, where
items are assigned to cells and levels with an objective of minimizing the total
vertical and horizontal transportation costs. In particular, the warehouse has only
one I/O port with sufficient capacity to transport items vertically from ground floor
to every other level at all times. Each item must be assigned to exactly one cell, and
a cell may store more than one item type; no fractional distribution of products is
allowed. Furthermore, the capacity of each cell is the same on all levels, and the
levels could have a different number of cells. A monthly demand and an inventory
requirement are associated with every item type along with a vertical unit
transportation cost (cost to move one unit from ground to another level) and
horizontal unit transportation cost (move one unit of the item one meter
horizontally). Lai et al. (2002) proved that the single-floor layout problem is an
NP-hard one; in particular, the authors transferred an arbitrary instance of the Bin-
packing problem into an instance of the reduced single-floor layout problem where
each cell cannot store more than one item type, and showed that the reduced
problem’s optimal solution gives also the optimal solution to the Bin-packing
problem. Since the single-floor is a special case of the multi-floor layout problem,
then the latter is also NP-complete. Accordingly, heuristic algorithms are more
appropriate than exact approaches, where with the latter it becomes time consuming
to find optimal solutions.
One of the most popular and earlier suboptimal methods to solve this problem was
CRAFT of Armour and Buffa (1963) and Buffa et al. (1964). CRAFT considers an
initial layout and attempts on improving it by exchanging facilities. The starting layout
could be an existing one or an arbitrary one for a new facility. Scriabin and Vergin
(1975) compared the performances of several computer algorithms (including
CRAFT) for the plant layout problem with selected human visual methods. While
their study concluded that the latter outperformed the computer programs, it was soon
critiqued by Block (1977) who concluded that the superiority of one approach over the
other is highly dependent on the flow dominance. Johnson (1982) proposed an
algorithm called SPACECRAFT which transforms the multi-floor problem to a single-
floor one by affixing the floors to a single floor. Bozer et al. (1994) introduced the
MULTIPLE method that uses space-filling curves to identify block layouts. Similar to
CRAFT, an initial layout is suggested and improved by swapping. Kochhar and
Heragu (1998) presented a genetic algorithm based heuristic called MULTI-HOPE for
the problem under study, compared it to existing simulated annealing heuristics, and
showed its superiority. Abdinnour-Helm and Hadley (2000) presented a pair of two-
stage heuristics, where in the first stage a minimal inter-floor flow is obtained, and in
the second stage Tabu search is used to refine the layout. Zhang et al. (2002) proposed
an IP model and showed that it is NP-hard. The authors also developed and showed the
effectiveness of several genetic algorithm heuristics for the problem. Lee et al. (2005)
proposed an improved genetic algorithm for layouts with inner structure walls and
passages. They showed that it outperformed an existing simulated annealing algorithm
by Meller and Bozer (1996). Yang and Feng (2006) implemented a fuzzy layout

123
Solving the multiple level warehouse layout problem using…

optimization algorithm to deal with a multi-level warehouse layout problem under

fuzzy environment, in which different types of items need to be placed in a warehouse
and the monthly demand of each item type and horizontal distance are treated as fuzzy
variables. Önüt et al. (2008) introduced a particle swarm optimization algorithm for the
problem and compared it to branch and bound. Zhang and Lai (2010) addressed the
same variant of the MLWLP, with the addition of adjacent cells. In particular, when the
inventory requirement of an item type exceeds the cell capacity, the adjacency
constraint allows to split the type for storing in multiple cells. The authors proposed an
integer programming model to formulate the problem, and developed a cube-per-order
index policy based heuristic, the standard tabu search (TS), greedy TS, and dynamic
neighborhood based TS to solve the problem. Gabor and van Ommeren (2010)
proposed a new integer programming formulation for the multilevel facility location
problem, where facilities are organized on different and demand points have to be
assigned to a path of open facilities passing each level, with an objective to minimize
the total cost (of opening facilities and transportation). Matić et al. (2011) introduced a
genetic algorithm for the same variant of the MLWLP addressed in this research and
through extensive computational tests, proved its superiority over branch and bound.
Arnaout (2013) addressed the Euclidean location-allocation problem with an unknown
number of facilities, and an objective of minimizing the fixed and transportation costs.
A three stage ant colony optimization (ACO) algorithm was introduced for the problem
and its solutions compared to genetic algorithms. This study was later extended in
Arnaout et al. (2016) where an optimization–simulation–optimization approach was
proposed to solve the stochastic version of the problem. Having said that, both studies’
ACO differ considerably from the one proposed in this work. In particular, as the actual
studied problems are different, the greedy rules must be different as they are problem
dependent, as well as the ACO implementation; e.g. Arnaout (2013)’s ACO was
modeled in three stages while the one here in only two, leading to a reduced
computational time. Obviously, also the ACO parameters’ values are problem
dependent.
Matić et al. (2014) considered the multilevel uncapacitated facility location
problem, proposed a new efficient integer programming formulation, and designed a
parallel memetic algorithm (MA) with a new strategy for applying the local search
improvement within the MA frame. Guerriero et al. (2015) presented a rollout-based
heuristic for the product allocation problem, where products are allocated to storage
positions in a multi-layer warehouse, with the objective of optimizing the
warehouse space utilization and handling costs. Ortiz-Astorquiza et al. (2015)
modeled the multi-level uncapacitated facility location problem (MUFLP) as two
different combinatorial optimization problems, and corrected a previous conclusion
stating that the MUFLP is not submodular.
To the best of the authors’ knowledge, there does not exist published research
that developed ant colony optimization algorithm for the MLWLP. Having said this,
in this study, a newly designed ant colony optimization (ACO) algorithm is
introduced for the problem and the results are compared to those obtained using a
genetic algorithm (GA) by Matić et al. (2011) as well as an exact algorithm (B&B).
Some preliminary promising results for this problem have been reported by El
Khoury and Arnaout (2012).

123
 J.-P. Arnaout et al.

The rest of this paper is organized as follows. In Sect. 2, we discuss the ILP
formulation of the problem. In Sect. 3, the ant colony optimization algorithm is
presented in general then adapted for this problem in particular. The computational
tests are presented in Sect. 4 and finally we conclude this research in Sect. 5.

2 Mathematical formulation

The ILP formulation below was taken from Zhang et al. (2002) and repeated here
for the reader’s convenience. It will be solved using a branch and bound (B&B) to
obtain optimal solutions for small-size instances and near-optimal/feasible solutions
for large-size instances, as well as generate the running times for this exact method
(B&B) to be compared with the ACO times; this is explained in detail in Sect. 4.
Table 1 highlights the needed notations.
The problem is formulated as follows with xjlk representing the binary decision
variables, where xjlk ¼ 1 if item j is assigned to cell k on level l; 0 otherwise.
X
J X
L X
Kl  
Min Qj Dlk Cjh þ Cjlv xjlk ð1Þ
j¼1 l¼1 k¼1

subject to
X
L X
Kl
ð2Þ
xjlk ¼ 1; for j ¼ 1; . . .; J:
l¼1 k¼1

X
J
Sj xjlk  A; for l ¼ 1; . . .; L; k ¼ 1; . . .; Kl : ð3Þ
j¼1

xjlk 2 f0; 1g; 8j; l; k: ð4Þ

Objective (1) minimizes the total horizontal and vertical transportation costs. The
constraints (in the given order) ensure that each item type is assigned to one cell

Table 1 MLWLP’s notation

J Number of item types, with index j 2 f1; 2; . . .; J g
L Number of levels, with index l 2 f1; 2; . . .; Lg
Kl Number of available cells in level l, with index k 2 f1; 2; . . .; Kl g
Qj Monthly demand of item type j
Sj Inventory requirement of item type j
Cjh Horizontal unit transportation cost of item type j
Cjlv Vertical unit transportation cost of item type j to level l
A Storage capacity of a cell
Dlk Horizontal distance from cell k on level l to the I/O port or elevator

123
Solving the multiple level warehouse layout problem using…

Table 2 Example’s input data v v

Item j Qj Sj Cjh Cj1 Cj2

1 97 4 12.98699 17.82904 14.05882

2 68 4 13.0068 13.98738 1.878259
3 87 4 12.95581 1.656288 6.410947
4 116 14 11.50614 17.35655 7.830464
5 22 13 13.21278 12.1254 22.82416
6 40 3 12.1108 13.9456 1.09018

Table 3 Horizontal distances to the I/O port

Cell 1 Cell 2 Cell 3 Cell 4

Level 1 2 3 2 3
Level 2 1 2 4 3

Table 4 Optimal assignment

Cell 1 Cell 2 Cell 3 Cell 4

Level 1 5 3
Level 2 1, 2 4 6

only, that the cells’ capacity is not exceeded, and that the decision variables are
binary.
To better understand MLWLP, an example is presented as follows. A particular
warehouse has two levels, each constituting of four cells with a maximum
individual capacity of 14. The items’ monthly demand, inventory requirement,
horizontal costs and vertical costs (for each level) are shown respectively in
Table 2, and the horizontal distances from each cell to the I/O port are shown in
Table 3.
The optimal assignment that led to the lowest costs of 10,987.95 is shown in
Table 4. In particular, item types 1 and 2 are assigned to the first cell on the
second level, as their combined inventory requirement is eight (less than the
cell’s capacity). Cells 2 and 4 and cell 3 on levels 1 and 2 respectively are left
empty as their horizontal distances to the port are higher than the other cells.
Item types 5 and 3 are assigned to cells 1 and 3 respectively on the first level,
and item types 4 and 6 are assigned to cells 2 and 4 respectively on the second
level.

3 Ant colony optimization (ACO) for MLWLP

Following the natural ability of ants to find the shortest path between their nest and
food places, scholars were inspired to simulate this behavior to solve combinatorial
optimization problems. ACO was firstly proposed by Dorigo (1992), with the first

123
 J.-P. Arnaout et al.

algorithm aiming to search for an optimal path in a graph. Since then, many studies
have emerged on the implementation of ACO to solve a wider range of numerical
problems. For more on ACO literature, the reader can refer to Dorigo and Stützle
(2004). To solve an optimization problem using ACO, the problem can be
represented as a connected graph (nodes and edges). Initially, a certain amount of
pheromone is deposited in each edge in the graph. After the ants move from a node
to another according to the State Transition Rule; which determines the probability
of an ant to move from one node to another. The probability is determined by two
factors: pheromone amount (s) and visibility (g). The probability of moving from
saij gbij
node i to j for ant k can be calculated as follows: Pkij ¼ P ; where W represents
sa g b
l2W il il
the set of non-visited nodes and (g) is usually determined by a greedy rule
(heuristic). Two important parameters to direct the search are a and b; which are the
exponents in the probability function that determine the importance of the
pheromone amount over the visibility amount. After all ants construct their
solutions, pheromone will be updated locally and globally according to the quality
of the constructed objective function value and evaporation rate (q) as shown in (5)
and (6). The first term accounts for the pheromone reduction (local update) due to
evaporation, and the second term (global update) represents the additional deposit of
pheromone in the links (arcs) according to the produced tour length when traveled
by all ants.
X
m
sij ð1  qÞsij þ Dskij 8ði; jÞ ð5Þ
k¼1
 k
1=Cmax if arc ði; jÞ is used by ant k
Dskij ¼ ð6Þ
0 Otherwise
k
where Cmax represents the objective function value of the solution constructed by
ant k.
As previously noted, MLWLP’s computational complexity is NP-complete;
subsequently, heuristic algorithms such as ACO are more appropriate than exact
approaches, where with the latter it becomes time consuming to find optimal
solutions.
In order to be able to model and solve the MLWLP using ACO, the latter must be
divided into two parts/stages. In the first stage, items J are assigned to levels L,
without considering the cells. Subsequently, in Stage 2, items are assigned to the
cells in their selected level (which was determined in Stage 1).
Stage 1 is represented by a 2D Matrix (M1) of size J  L, where cell M1ðj; lÞ ¼ 1
if item j is assigned to level l, 0 otherwise. The network representation of this stage
is shown in Fig. 1.
Initially, items are sorted in the decreasing order of their monthly demand ðQj Þ.
Next, the greedy rule/visibility amount gjl for assigning item j to level l is shown in
Eq. 7. The latter gives more priority of assignment to the level with the lower
vertical transportation cost Cjlv . An item j can be assigned to level l as long as the
latter’s space can fit the item’s inventory requirement.

123
Solving the multiple level warehouse layout problem using…

Fig. 1 Stage 1 network Items Levels

representation

j=1 {0,1} l=1

{0,1}

j=2 {0,1} l=2

{0,1}
. .
. {0,1} .
. .
{{0,1}
,1}
,1}

j=J {0,1} l=L

( .
1 v; if Sj  Resl ;
gjl ¼ Cjl ð7Þ
0 otherwise
where Resl is the remaining (residual) space on level l following each item
j assignment.
Next, ant t assigns item j to level l according to the probability in Eq. 8.
 a  
b
sIjl  gjl
t;I
Pjl ¼ P  a   ð8Þ
I b
hW s jh  g jh

Table 5 shows the output from Stage 1 for the example that was introduced in
Tables 2 and 3, where item types 3 and 5 are assigned to level 1, and item types 1, 2,
4, and 6 to level 2.

Table 5 Stage 1 output’s

Items Level 1 Level 2
example (M1)
1 0 1
2 0 1
3 1 0
4 0 1
5 1 0
6 0 1

123
 J.-P. Arnaout et al.

The pseudocode of Stage 1 is described as follows:

1. Sort items j in decreasing order of Qj
2. For j ¼ 1; . . .; L:
2:1. Generate gjl and Pkjl according to (7) and (8) respectively.
2:2. Generate a cumulative distribution function (CDF) according to Pt;I
jl .
2:3. Generate a random variable rv ¼ UNIFORM ð0; 1Þ.
2:4. Assign item j to level l according to rv and CDF.

Following Stage 1, items are already assigned to levels; thus, in Stage 2, items
will be assigned to the cells in their selected level. In particular, cells with the
smallest horizontal distances will be first saturated before moving to farther cells.
Initially, cells are sorted in the increasing order of their horizontal distances; then,
for every cell, the visibility is calculated following Eq. 9 for every item j. The
rationale of gjk is to locate closer the items with higher horizontal cost, monthly
demand and inventory requirement, if there is enough space in this particular cell.
Next, ant t assigns item j to cell k haccording to the probability in Eq. 10.
Cj Qj Sj ;; if Sj  Resk ;
gjk ¼ ð9Þ
0;; otherwise
where Resk refers to the residual space in cell k.
 a  
b
sIIjk  gjk
t;II
Pjk ¼ P  a   ð10Þ
II b
hW sjh  gjh

Fig. 2 Stage 2 network Items Cells

representation

j=1 {0,1} k=1

{0,1}

j=2 {0,1} k=2

{0,1}
. .
. {0,1} .
. .
{0,1}

j=J {0,1} k = Kl

123
Solving the multiple level warehouse layout problem using…

Stage 2’s output is also represented by a 2D Matrix (M2) of size J  Kl for every
level l, where M2ðj; kÞ ¼ 1 if item j is assigned to cell k, 0 otherwise. The network
representation of this stage is shown in Fig. 2.
Stage 2 output is shown in Table 6, where extending on Table 5, we have the
following assignments:
• On level 1, item type 5 is assigned to cell 1, item type 3 to cell 3.
• On level 2, item types 1 and 2 are assigned to cell 1, item types 4 and 6 are
assigned to cells 2 and 4, respectively.

The total costs associated with this optimal assignment were 10,987.95.
The pseudocode of Stage 2 is described as follows:
1. Sort cells k in increasing order of Dlk
2. For j ¼ 1; . . .; J:
2:1. Generate gjk and Pt;II
jk according to (9) and (10) respectively.
2:2. Generate a cumulative distribution function (CDF) according to Pt;II
jk .
2:3. Generate a random variable rv ¼ UNIFORM ð0; 1Þ.
2:4. Assign item j to level k according to rv and CDF.

After all ants finish their paths’ traversals, the pheromone amounts in each edge
are reduced locally by evaporation. The path that results in the best objective
function value; i.e., the minimum transportation costs has the pheromone amounts
increased on it globally.
sIjl ð1  qÞ  sIjl þ u  DsI;Best
jl ð11Þ

sIIjk ð1  qÞ  sIIjk þ u  DsII;Best

jk ð12Þ

where:
(
1
DsI;Best ¼ ;; if arc ðj; lÞ is used by best ant
jl Total Cost
0;; Otherwise

Table 6 Stage 2 output’s

Items Cell 1 Cell 2 Cell 3 Cell 4
example (M2)
1 1 0 0 0
2 1 0 0 0
3 0 0 1 0
4 0 1 0 0
5 1 1 0 0
6 0 0 0 1

123
 J.-P. Arnaout et al.

(
1
DsII;Best ¼ if arc ðj; kÞ is used by best ant
;;
jk Total Cost
0;; Otherwise
P P P l  
Note that Total Cost is defined as Jj¼1 Ll¼1 Kk¼1 Qj Dlk Cjh þ Cjlv xjlk .
The pseudo code for our ACO algorithm is summarized below:
 
1. Populate the paths with specified pheromone amounts sIjl ; sIIjk
2. For Iteration = 1, …, Iter:
2:1. For Ant = 1, …, No_of_Ants:
2:1:1. Solve for Stage 1 (Assignment to Levels)—Find M1 according to
Eqs. (7) and (8)
2:1:2. Solve for Stage 2 (Assignment to Cells)—Find M2 according to
Eqs. (9) and (10)
2:1:3. Find Total Cost(Ant) that are associated with M1 and M2

2:2. Update pheromone amounts locally and globally according to Eqs. (11)
and (12)

4 Computational tests

The proposed ACO was implemented in Java and compiled with Netbeans 7.0.1,
with 4 GB of memory available for working storage on a personal computer Intel
(R) Core (TM) i3-370 M CPU @ 2.4 GHz. As mentioned earlier, ACO was
compared to genetic algorithms (GA) and branch and bound (B&B). The GA results
were obtained from Matić et al. (2011), where the algorithm was coded in C
programming language and tested on Intel Core 2 Quad Q9400 @2.66 GHz with
8 GB RAM. The ILP presented in Sect. 2 was solved using B&B; the latter’s solver
used was Lingo 11.0 from Lindo Systems, tested on Intel (R) Core (TM) i5-4430 S
CPU @ 2.7 GHz with 16 GB RAM.
The data was obtained from Matić et al. (2011) for benchmarking. In particular,
the authors generated instances based on the characteristics of real problems, with
the following major attributes:
• The number of items J vary from 10 to 40 and from 100 to 400 for small and
large instances, respectively.
• The number of levels L vary from 2 to 5.
• The number of cells Kl in each level is determined on the inventory
requirements, the levels and the cells capacity A, where the latter is set to 16.
• 70% of the item types have an inventory requirement Sj less than A=2, 20% with
h
Sj 2 A=2; AÞ, and 10% with Sj ¼ A:

123
Solving the multiple level warehouse layout problem using…

4.1 Design of experiments

Design of experiments (DoE) was utilized to determine the appropriate values for
the ACO parameters that will minimize the Total Cost. Numerous publications
provide a good review of DoE (e.g., Fisher 1960; Taguchi 1993; NIST/SEMATECH
2012). The factors considered in this experiment along with their levels of low,
medium and high respectively are as follows: No of Ants : ð5; 33; 60Þ; q :
ð0:001; 0:1505; 0:3Þ; / : ð0:001; 0:1505; 0:3Þ; b : ð0:5; 1; 1:5Þ;
sIjl : ð0:5; 5:25; 10Þ, sIIjk : ð0:5; 5:25; 10Þ, and Iter : ð500; 7750; 15;000Þ; where q
and / are respectively the pheromone evaporation and global update rates as
defined earlier, and a was fixed at 1 as the different ratio levels are determined by b.
The values of the parameter levels were selected based on many runs under different
settings. To reduce the number of runs but reach sound conclusions, D-Optimal
Design was utilized, which has been shown to be an effective design (NIST/
SEMATECH 2012). JMP 11.0 from SAS was used to generate a D-Optimal design,
with 40 experiments. The experiments were conducted on a selected instance from
the 40 items and 5 levels problem size, as the latter was more congruent with the
studied problem sizes. In addition, each experiment was run 5 times and the average
was reported for that particular experiment. The factors along with their interactions
were analyzed using regression, ANOVA, and factors’ effect tests. Three-factor
interactions and higher were not considered as they typically have weak effect (Ross
1996). Tables 7 and 8 show that the overall model fit and factor selection is
meaningful with high R2 value and low p value for the overall model.
In Table 9, we show the parameters’ estimates obtained through the regression
results including the significant factors (with p value B0.05) that are bolded. Based
on a 95% Confidence Interval, a relatively large t-Stat, and a small p value (less than
0.05), a prediction expression was generated and solved for the minimum Total Cost
while varying the factors’ values. As a result, the following parameter values were
determined to provide the best performance for the ACO: No of Ants ¼ 32; q ¼
0:001; / ¼ 0:2624; b ¼ 0:71267; sIjl ¼ 0:5, sIIjk ¼ 2:1898, and Iter ¼ 9580.

Table 7 Summary of fit of the

R2 0.983559
model
R2 Adj 0.839704
Root mean square error 1923.317
Mean of response 100193.7
Observations (or sum wgts) 40

Table 8 Analysis of variance

Source DF Sum of squares Mean square F ratio

Model 35 885,207,700 25,291,649 6.8372

Error 4 14,796,592 3,699,148 Prob [ F
C. total 39 900,004,291 0.0366

123
 J.-P. Arnaout et al.

Table 9 Parameter estimates

Term Estimate Std error t ratio Prob [ |t|

Intercept 1,334,108 2,013,626 0.66 0.5439

No of Ants 13,721.01 7035.054 1.95 0.1229
Iter 6913.043 6609.79 1.05 0.3547
q -2239.52 7571.828 -0.3 0.7821
/ 1000.492 6979.632 0.14 0.8929
sIjl -21,223.8 6977.848 -3.04 0.0383
sIIjk -15,555.9 6763.122 -2.3 0.0829
b 2,646,226 4,265,574 0.62 0.5686
b2 1,408,479 2,252,336 0.63 0.5657
No of Ants  Iter 723.0577 379.6774 1.9 0.1296
No of Ants  q -882.905 399.9681 -2.21 0.0919
No of Ants  / -132.633 415.4071 -0.32 0.7655
No of Ants  sIjl 192.7325 412.558 0.47 0.6647
No of Ants  sIIjk -1118.51 394.3244 -2.84 0.047
No of Ants  b 16,149.71 7390.985 2.19 0.0942
Iter  q 1592.12 384.1581 4.14 0.0143
Iter  / 1278.329 438.7617 2.91 0.0435
Iter  sIjl -542.521 422.3496 -1.28 0.2683
Iter  sIIjk -232.967 376.7145 -0.62 0.5698
Iter  b 9239.698 6994.317 1.32 0.257
q/ 1985.805 405.5378 4.9 0.0081
q  sIjl -518.519 402.5423 -1.29 0.2672
q sIIjk 96.92,193 380.0392 0.26 0.8113
qb -3188.81 7931.281 -0.4 0.7082
/  sIjl -327.435 421.273 -0.78 0.4804
/ sIIjk 273.402 383.8292 0.71 0.5156
/b 2106.284 7347.345 0.29 0.7886
sIjl  sIIjk -734.541 396.0697 -1.85 0.1373
sIjl b -22,967.1 7401.227 -3.1 0.0361
sIIjk  b -17,183.6 7171.243 -2.4 0.0747
No of Ants2 1303.551 1101.392 1.18 0.3021
Iter2 3597.456 1144.393 3.14 0.0347
q2 248.3606 1140.6 0.22 0.8383
/2 2244.788 1014.977 2.21 0.0915
sI2
jl
-2388.49 1186.774 -2.01 0.1145
sII2
jk
816.4277 1274.17 0.64 0.5565

123
Solving the multiple level warehouse layout problem using…

Table 10 CPU performance comparison

GA B&B ACO
Intel Core2 Quad Q9400 @ Intel Core i5-4430S @ Intel Core i3-370 M @
2.66 GHz 2.70 GHz 2.40 GHz

CPU mark 3385 5886 2025

Normalization 1 1.739 0.598
factor

Table 11 ACO, GA, and B&B (Opt) results on small instances with a = 0.2, 0.4, 0.5

Time (sec) Time (sec)

It L B&B / Opt GA ACO AvgDev B&B GA ACO It L B&B / Opt GA ACO AvgDev B&B GA ACO

10 2 21062.3 opt opt 0 0.2 0.391 0.2228 25 4 40814.64 opt opt 0 2.1527 0.725 0.483
10 3 22324.21 opt opt 0 0.2939 0.412 0.2448 25 5 57930.11 opt opt 0 0.3026 0.678 0.403
10 4 22389.98 opt opt 0 0.2069 0.4025 0.2442 30 2 58000.91 opt opt 0 0.3286 0.867 0.588
10 5 22473.23 opt opt 0 0.2017 0.413 0.2831 30 3 51355.36 opt opt 0 0.2034 0.852 0.553
15 2 31144.13 opt opt 0 0.2 0.5245 0.3773 30 4 47189.87 opt opt 0 0.8746 0.836 0.506
15 3 28124.82 opt opt 0 0.2 0.534 0.412 30 5 49518 opt opt 0 14.693 0.848 0.455
α = 0.2

15 4 29878.76 opt opt 0 0.2034 0.5115 0.3615 35 2 81913.01 opt opt 0 322.94 1.143 0.589
15 5 35417.05 opt opt 0 0.1774 0.4565 0.2984 35 3 59878.91 opt opt 0 0.2991 0.929 0.619
20 2 35156.61 opt opt 0 0.2069 0.6085 0.3976 35 4 57463.36 opt opt 0 0.8051 0.91 0.522
20 3 35682.76 opt opt 0 0.1635 0.569 0.4207 35 5 60809.13 opt opt 0 41.532 1.042 0.563
20 4 35470.8 opt opt 0 0.353 0.5595 0.4306 40 2 69045.67 69241.59 69133.7152 0.256 976.38 2.156 1.311
20 5 49135.65 opt opt 0 0.2782 0.5735 0.4249 40 3 67769.64 opt opt 0 1.5284 1.069 0.574
25 2 43962.44 opt opt 0 0.3582 0.7345 0.4846 40 4 62302.34 opt opt 0 121.83 1.039 0.679
25 3 41779.07 opt opt 0 0.1756 0.733 0.4639 40 5 64254.56 opt opt 0 1.8067 1.004 0.609

10 2 19557.82 opt opt 0 0.16 0.42 0.3191 25 4 51442.91 opt opt 0 3.3229 0.871 0.582
10 3 21909.18 opt opt 0 0.16 0.4595 0.3556 25 5 41413.67 opt opt 0 0.1861 0.697 0.521
10 4 22817.54 opt opt 0 0.1565 0.4075 0.3904 30 2 64887.96 opt opt 0 1474 0.994 0.63
10 5 22659.07 opt opt 0 0.2052 0.422 0.3182 30 3 76866.71 opt opt 0 5.7191 1.55 1.051
15 2 25757.23 opt opt 0.010 0.1478 0.528 0.4569 30 4 59766.96 opt opt 0 0.9685 1.025 0.755
15 3 23617.06 opt opt 0 0.1652 0.5325 0.3603 30 5 56562.64 opt opt 0 2.1336 0.787 0.51
α = 0.4

15 4 38256.88 opt opt 0 0.1461 0.5055 0.4049 35 2 174532.68 181193.5 174884.341 1.516 210.07 2.14 1.595
15 5 35376.11 opt opt 0 0.1391 0.4735 0.4154 35 3 111466.96 113420.3 111521.9 0.952 360.07 1.852 1.353
20 2 55555.54 opt opt 0 2.1892 0.808 0.6121 35 4 104516.72 105969.8 104612.611 0.864 21.544 1.68 1.044
20 3 60265.26 opt opt 0 0.6156 0.6205 0.4836 35 5 74214.03 opt opt 0 4.554 1.114 0.853
20 4 46134.17 opt opt 0 1.085 0.6165 0.4642 40 2 219995.72 232370.3 opt 0.011 20.737 2.47 1.689
20 5 44392.36 opt opt 0 0.9425 0.653 0.417 40 3 93736.42 95508.09 94121.148 1.312 3575.2 2.347 1.286
25 2 71392.76 opt opt 0 10.165 1.1365 0.8004 40 4 84640.98 opt opt 0 91.319 1.188 0.835
25 3 41933.35 opt opt 0 16.839 1.054 0.606 40 5 95826.2 97222.24 96470.0699 1.234 8.5847 2.095 1.443

10 2 22858.06 opt opt 0 0.1495 0.3965 0.1749 25 4 87364.52 opt opt 0 0.2347 0.929 0.564
10 3 22699.26 opt opt 0 0.1513 0.4245 0.2487 25 5 56509.16 opt opt 0 0.626 0.668 0.407
10 4 20987.27 opt opt 0 0.1461 0.3885 0.1805 30 2 52880.94 opt opt 0 125.38 1.16 0.571
10 5 23005.11 opt opt 0 0.153 0.4055 0.2427 30 3 76567.91 opt opt 0 4.6636 0.98 0.432
15 2 31823.6 opt opt 0 0.16 0.5235 0.2674 30 4 47069.24 opt opt 0 0.7929 0.817 0.508
15 3 40242.19 opt opt 0 0.1948 0.5425 0.3449 30 5 46630.49 opt opt 0 0.819 0.837 0.435
α = 0.5

15 4 30581.51 opt opt 0 0.1356 0.504 0.262 35 2 126942.68 opt opt 0 1.2589 1 0.46
15 5 29707.79 opt opt 0.05 0.1356 0.4985 0.2955 35 3 62448.6 62647.81 opt 0.256 567.74 1.66 1.056
20 2 45827.5 opt opt 0 0.4869 0.607 0.3064 35 4 59084.84 opt opt 0.102 0.3982 1.228 0.642
20 3 44979.5 opt opt 0 0.1478 0.5845 0.2928 35 5 59641.18 59641.47 opt 0 20.581 1.483 0.987
20 4 37847.18 opt opt 0 0.2313 0.5495 0.2454 40 2 132694.08 133420.5 133060.17 0.566 14.448 2.453 1.315
20 5 34516.11 opt opt 0 0.2765 0.6095 0.2608 40 3 120387.72 120963.6 120655.52 1.213 1.1372 2.016 1.391
25 2 75767.56 opt opt 0 6.9554 0.94 0.552 40 4 90783.44 91027.21 90883.34 0 19.668 1.895 1.067
25 3 104124.99 opt opt 0 0.266 0.6995 0.4024 40 5 64575.8 opt opt 0 4.8862 1.261 0.633

123
 Table 12 ACO, GA, and B&B (Opt) results on small instances with a = 0.6, 0.8

Time (sec) Time (sec)

It L Opt GA ACO AvgDev B&B GA ACO It L Opt GA ACO AvgDev B&B GA ACO

123
10 2 30918.31 opt opt 0 0.1582 0.396 0.2833 25 4 52059.42 opt opt 0 3.9837 0.87 0.506
10 3 22218.82 opt opt 0 0.1652 0.422 0.2757 25 5 51936.51 opt opt 0.001 0.2591 0.693 0.359
10 4 26122.86 opt opt 0 0.1461 0.374 0.2364 30 2 212119.73 212354.2 opt 0.856 0.5843 1.537 0.9
10 5 25758.65 opt opt 0 0.1426 0.3785 0.2319 30 3 76593.42 opt opt 0 4.754 1.161 0.596
15 2 33228.06 opt opt 0 0.153 0.4805 0.3256 30 4 60721.99 opt opt 0 476.73 0.852 0.564
15 3 34761.06 opt opt 0 0.1252 0.478 0.2528 30 5 84895.25 opt opt 0 0.28 1.56 0.867
15 4 33129.28 opt opt 0 0.1287 0.468 0.2929 35 2 154200.14 158622.7 155432.209 2.567 7.4127 2.172 1.399
15 5 36192.38 opt opt 0 0.1391 0.4785 0.2534 35 3 125140.12 126929.9 opt 0.263 58.397 1.807 0.912

α = 0.6
20 2 86095.67 opt opt 0 0.1913 0.639 0.4354 35 4 93518.43 93857.1 opt 0.855 0.7929 1.535 0.871
20 3 64981.42 opt opt 0 0.233 0.5655 0.3219 35 5 85915.12 opt opt 0.025 1.252 1.426 0.818
20 4 40920.64 opt opt 0 0.233 0.538 0.3051 40 2 132187.99 132959.2 132450.762 0.612 30.764 2.137 1.249
20 5 40293.5 opt opt 0 0.2156 0.6195 0.3509 40 3 75738.35 75915.98 opt 0.225 4.7123 2.131 1.051
25 2 71542.27 opt opt 0 0.2382 1.0255 0.646 40 4 74256.03 opt opt 0 20.908 1.966 1.124
25 3 81936.29 opt opt 0 2.3161 0.814 0.4224 40 5 130371.24 132187 131579.076 0.123 1.445 2.518 1.616

10 2 19224.56 opt opt 0 0.1339 0.3855 0.3555 25 4 63208.6 opt opt 0 0.2504 0.733 0.591
10 3 25336.8 opt opt 0 0.1617 0.3745 0.2869 25 5 39551.93 opt opt 0 0.1878 0.621 0.492
10 4 22321.65 opt opt 0 0.1565 0.394 0.3183 30 2 62563.54 opt opt 0 1.6554 1.511 1.145
10 5 23153.77 opt opt 0 0.153 0.378 0.3312 30 3 122539.98 122804.9 opt 0.215 3.3803 1.733 1.231
15 2 38487.7 opt opt 0 0.1287 0.4445 0.343 30 4 74068.19 opt opt 0 0.2973 1.3 0.886
15 3 27029.93 opt opt 0 0.1148 0.4505 0.3436 30 5 74408.72 opt opt 0.045 0.2765 1.204 0.892
15 4 24769.59 opt opt 0 0.1356 0.426 0.3068 35 2 85351.41 86283.69 85362.9268 0.256 8.9759 1.943 1.31
15 5 24279.72 opt opt 0 0.1095 0.437 0.361 35 3 154474.56 opt opt 0 0.7286 1.775 1.337

α = 0.8
20 2 39261.9 opt opt 0 2.3805 0.5875 0.3781 35 4 116058 116626.2 116519.261 2.324 0.546 1.99 1.286
20 3 59359.67 opt opt 0 0.16 0.5545 0.4558 35 5 85853.76 85855.08 85855.08 0.512 0.28 1.55 0.952
20 4 38962.22 opt opt 0 0.2469 0.576 0.39 40 2 338547.78 349660.2 339649.08 2.567 35.292 2.529 1.568
20 5 37114.59 opt opt 0 0.2226 0.576 0.3844 40 3 144358.86 148026.2 145165.371 1.234 4.5314 2.416 1.714
25 2 79158.02 79369.4 opt 0.112 0.6573 0.799 0.4953 40 4 196334.48 197342.4 opt 0 0.266 1.95 1.271
25 3 122439.5 opt opt 0 0.3043 1.038 0.6213 40 5 92246.28 92311.57 opt 0.213 1.3963 1.986 1.239
J.-P. Arnaout et al.
Solving the multiple level warehouse layout problem using…

4.2 Results and computational times

As highlighted earlier, ACO, GA, and B&B were coded in different languages
and run on different platforms; i.e., directly comparing their computational times
would not be a reasonable assessment. However, an indirect comparison of times
can be done by comparing the different computer platforms and normalizing the
CPU times accordingly. In particular, http://www.cpubenchmark.net/ was used to
compare the different processors, and the CPU marks and normalization factors
are shown in Table 10. The CPU mark refers to the performance of the pro-
cessor, where a higher number refers to better performance. The GA platform
was used as the base (factor of 1), and the running times of ACO and B&B are
multiplied by their respective factors in order to have a sense of how the three
algorithms measure up.
Tables 11 and 12 show the results for small-size problem instances, with a
referring to the percentage of cells with the same distance. In the case of the exact
solution using B&B, it was run until global optimal was reached; thus referred to in
the tables as Opt. The latter is needed to verify the ability of the tested
metaheuristics to reach the optimal solution. ACO was run five times for each
problem instance, and the best value was given in the column ACO. Subsequently,
AvgDev refers to the relative deviation in the five runs, and is calculated as
P5 TotalCosti BestTotalCosti 
100 i¼1 BestTotalCosti
AvgDev ¼ 5 . The normalized computational times are also
reported in the three columns that follow.
It is clear that both GA and ACO reached the optimal solution in the majority
of instances, with ACO attaining it more frequently. Furthermore, on average,
ACO performed better than GA, and the average normalized computational time
was lower in all problem instances. However, there were many cases where GA
and ACO performed the same; consequently, Paired t-tests over all the small
instances (140 instances) were conducted with 95% CI to determine the finest
rule, and the results are presented in Table 13. The latter indicates that the mean
difference is statistically significant as the 95% CI does not overlap with zero.
Table 14 shows the results for large-size problem instances, where B&B was run
for 30 min in order to provide an upper bound for comparison. The results clearly
indicate that ACO performed better than GA and B&B in all 56 instances, with a
lower normalized computational time than both algorithms. It was also interesting to
see that B&B, when it was able to attain a feasible solution within the allocated time
(30 min), generated better solutions than GA. However, in several large-size

Table 13 Paired t tests with 95% CI for small problems

Problem size Paired Mean Standard t stat Two-tailed p 95% CI on MD
t tests difference deviation
(MD)

Small (140 GA– 335.2667075 129.28703 2.593196753 0.010557123 [79.577,414.843]

instances) ACO
10 \ It \ 40

123
 Table 14 ACO, GA, and B&B results on large instances with a = 0.5, 0.8

Time (sec) Time (sec)

It L B&B GA ACO AvgDev GA ACO It L B&B GA ACO AvgDev GA ACO

123
100 2 863089.38 886029.67 781635.6387 0.417 7.310 4.192 250 4 4174870.45 4182619.77 3673299.269 0.501 22.128 13.052
100 3 572881.31 599825.17 481355.3198 0.503 6.531 3.719 250 5 2586253.91 2626602.19 2232341.06 0.466 22.626 13.319
100 4 536960.39 545607.15 459961.8721 0.260 6.743 3.811 300 2 12911500.00 13488441.57 11875779.97 0.458 31.022 18.370
100 5 512667.54 531832.02 493357.4137 0.202 6.992 3.996 300 3 5503853.31 5687478.2 4370366.624 0.153 29.413 17.398
150 2 2305247.18 2497848.77 2001682.704 0.402 11.764 6.840 300 4 3440208.44 3879476.68 2767290.67 0.562 31.050 18.383
150 3 1606305.63 1736344.65 1416940.014 0.326 11.450 6.667 300 5 N/A 5080521.97 4383188.787 0.397 29.009 17.133
150 4 1623774.14 1627720.06 1511036.533 0.556 11.328 6.586 350 2 N/A 11303227.28 8573801.583 0.260 42.345 25.147
150 5 948701.67 958259.04 869199.6062 0.361 10.897 6.336 350 3 N/A 7092958.06 5384304.015 0.218 38.839 23.014

α = 0.5
200 2 3247510.63 3407078.25 2782863.329 0.282 18.548 10.896 350 4 5430650.00 4571192.03 3873550.977 0.319 37.935 22.497
200 3 2884961.97 3044631.21 2609412.551 0.518 17.047 9.992 350 5 N/A 5908626.74 5105895.318 0.504 35.847 21.214
200 4 1427720.00 1736508.16 1394576.87 0.464 16.341 9.579 400 2 12656882.16 12951254.84 9833450.08 0.364 51.168 30.379
200 5 1591283.94 1649873.13 1426434.599 0.201 16.831 9.842 400 3 N/A 14815517.7 12837418.03 0.168 47.462 28.180
250 2 7993128.65 8007271.66 6912329.517 0.170 24.474 14.457 400 4 N/A 7055304.14 5269846.088 0.656 45.953 27.282
250 3 3212582.22 3561326.29 2733976.244 0.636 25.112 14.831 400 5 N/A 4741152.81 3621058.265 0.553 47.175 27.998

100 2 2069766.57 2121400.17 1786443.789 0.389 6.638 3.059 250 4 4099736.14 4191459.92 3389029.805 0.367 23.675 13.354
100 3 614389.80 635835.65 502419.7676 0.236 6.859 3.434 250 5 4236807.90 4494294.91 3925600.118 0.423 23.261 13.135
100 4 346379.86 360014.75 288417.3257 0.407 6.396 3.130 300 2 22675177.99 23368751.78 20207811.29 0.314 31.974 18.325
100 5 599700.63 627278.24 499943.4581 0.218 7.132 3.396 300 3 N/A 9532127.78 7904039.803 0.328 30.853 17.528
150 2 4029879.76 4324473.12 3566664.911 0.371 11.653 6.180 300 4 7389046.22 7966096.83 6806838.417 0.673 30.977 17.616
150 3 2260690.24 2330456.23 1797922.345 0.498 11.663 6.127 300 5 6675460.39 6924902.5 5963270.592 0.255 28.765 16.115
150 4 1704147.67 1803497.57 1479967.522 0.490 11.292 5.797 350 2 15843566.44 17019324.08 12374122.27 0.464 41.080 23.618
150 5 1441061.22 1495425.95 1325948.22 0.687 10.028 5.059 350 3 12486310.79 13459788.95 11262931.27 0.658 38.953 22.285

α = 0.8
200 2 6039473.20 6363820.94 5226417.67 0.412 17.820 9.989 350 4 N/A 9987551.66 8028276.603 0.207 37.978 22.009
200 3 2999400.82 3253969.22 2671910.697 0.417 17.788 9.722 350 5 8648125.33 9232550.61 7873573.185 0.340 36.754 21.122
200 4 3396466.13 3471624.5 2813565.174 0.222 16.360 8.919 400 2 N/A 30567646.34 25438658.05 0.310 51.654 29.854
200 5 3595443.00 3797900.27 3218695.31 0.602 17.259 9.602 400 3 N/A 20004110.91 17107844.15 0.486 49.479 28.900
250 2 12106008.22 12477844.33 10751265.87 0.427 25.663 14.459 400 4 N/A 11911918.6 9688214.966 0.276 44.790 26.085
250 3 5572386.08 6077753.05 4922813.89 0.356 24.764 13.835 400 5 9527649.61 10116167.3 8596674.721 0.560 42.687 24.782
J.-P. Arnaout et al.
Solving the multiple level warehouse layout problem using…

problem instances, the solution was reported as ‘‘N/A’’ as B&B was not able to
attain a solution within a feasible computational time (a max of 60 min was used for
these instances).

5 Conclusions and future research

In this paper, we have introduced an ant colony optimization algorithm (ACO) for
the MLWLP and compared its performance to Genetic Algorithms (GA). The
computational tests that were carried out between the rules proved the superiority of
ACO. In particular, for small-size problem instances, ACO and GA were able to
reach the global optimal solution for many instances, with ACO still performing
better, and the difference in performance was shown to be statistically significant
over 95% Confidence Interval. As the problem size increased, ACO’s performance
improved and the GA’s one deteriorated. In fact, for lagre-size problem instances,
ACO generated the lowest costs for all instances, followed by B&B, with GA
performing the worst. However, while the exact approach (B&B) was able to
generate optimal solutions for small-size problem instances when the time limit was
relaxed, it was not able to solve all the large-size problem instances within an
acceptable computational time, reaffirming the need for heuristics when solving NP-
hard problems.
As an extension to this research, it would be interesting to compare the results
obtained by ACO to algorithms that are more advanced than genetic algorithm, such
as particle swarm optimization or the newly developed worm optimization.

Acknowledgements The authors would like to express their sincere gratitude to Reviewer 1 for his
detailed and constructive comments, which aided in significantly improving the quality of this work.

References
Abdinnour-Helm S, Hadley SW (2000) Tabu search based heuristics for multi-floor facility layout. Int J
Prod Res 38(2):365–383
Armour GC, Buffa ES (1963) A heuristic algorithm and simulation approach to relative allocation of
facilities. Manag Sci 9:294–300
Arnaout J-P (2013) Ant colony optimization algorithm for the Euclidean location-allocation problem with
unknown number of facilities. J Intell Manuf 24:45–54
Arnaout J-P, Arnaout G, Khoury J (2016) Simulation and optimization of ant colony optimization
algorithm for the stochastic uncapacitated location-allocation problem. J Ind Manag Optim
12(4):1215–1225
Block TE (1977) Note—a note on ‘‘comparison of computer algorithms and visual based methods for
plant layout’’ by M. Scriabin and R. C. Vergin. Manag Sci 24(2):235–237
Bozer YA, Meller RD, Erlebacher SJ (1994) An improvement-type layout algorithm or single and multi-
floor facilities. Manag Sci 40:918–932
Buffa ES, Armour GC, Vollmann TE (1964) Allocating facilities with craft. Harv Bus Rev 42:136–158
Dorigo M (1992) Optimization, learning and natural algorithms. PhD thesis, Politecnico di Milano, Italie
Dorigo M, Stützle T (2004) Ant colony optimization. MIT Press, Cambridge
El-Khoury C, Arnaout J-P (2012) Metaheuristic for the multiple level warehouse layout problem. In:
Proceedings of the 2012 international conference on modeling and simulation, NY, USA.
pp 240–248
Fisher RA (1960) The design of experiments. Hafner Publishing Company, New York

123
 J.-P. Arnaout et al.

Gabor AF, van Ommeren JKC (2010) A new approximation algorithm for the multilevel facility location
problem. Discrete Appl Math 158(5):453–460
Guerriero F, Pisacane O, Rende F (2015) Comparing heuristics for the product allocation problem in
multi-level warehouses under compatibility constraints. Appl Math Model 39(23):7375–7389
Johnson RV (1982) SPACECRAFT for multi-floor layout planning. Manag Sci 30(5):648–649
Kochhar JS, Heragu SS (1998) MULTI-HOPE: a tool for multiple floor layout problems. Int J Prod Res
36(12):3421–3435
Lai KK, Zhang GQ, Xue J (2002) Layout design for a paper reel warehouse: a two-stage heuristic
approach. Int J Prod Res 40(3):731–744
Lee KY, Roh M-I, Jeong HS (2005) An improved genetic algorithm for multi-floor facility layout
problems having inner structure walls and passages. Comput Oper Res 32:879–899
Matić D, Filipovic V, Savic A, Stanimirovic Z (2011) A genetic algorithm for solving multiple warehouse
layout problem. Kragujev J Math 35(1):119–138
Matić M, Stanimirović Z, Djenić A, Stanojević P (2014) Memetic algorithm for solving the multilevel
uncapacitated facility location problem. Informatica 25(3):439–466
Meller R-D, Bozer Y-A (1996) A new simulated annealing algorithm for the facility layout problem. Int J
Prod Res 34(6):1675–1692
NIST/SEMATECH e-Handbook of Statistical Methods (2012). http://www.itl.nist.gov/div898/handbook/.
Accessed 28 June 2017
Önüt S, Tuzkaya UR, Dogac B (2008) A particle swarm optimization algorithm for the multiple-level
warehouse layout design problem. Comput Ind Eng 54(4):783–799
Ortiz-Astorquiza C, Contreras I, Laporte G (2015) Multi-level facility location as the maximization of a
submodular set function. Eur J Oper Res 247(3):1013–1016
Ross P (1996) Taguchi techniques for quality engineering. McGraw Hill, NewYork
Scriabin M, Vergin RC (1975) Comparison of computer and visual methods for plant layout-A rejoinder.
Manag Sci 23:172–181
Taguchi G (1993) Taguchi methods: design of experiments. American Supplier Institute Inc, Michigan
Yang L, Feng Y (2006) Fuzzy multi-level warehouse layout problem: new model and algorithm. J Syst
Sci Syst Eng 15(4):493–503
Zhang GQ, Lai KK (2010) Tabu search approaches for the multi-level warehouse layout problem with
adjacency constraints. Eng Optim 42(8):775–790
Zhang GQ, Xue J, Lai KK (2002) Layout design for a paper reel warehouse: a two-stage heuristic
approach. Int J Prod Econ 75(3):231–243

123