A Coevolutionary Algorithm For A Facility Layout Problem

This article was downloaded by: [Monash University
Library] On: 02 January 2015, At: 01:22

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954

Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of
Production Research
Publication details, including instructions for authors
and subscription information:
http://www.tandfonline.com/loi/tprs20
A coevolutionary algorithm for a

facility layout problem
T Dunker
, G Radons
& E Westkmper
a Fraunhofer Institute of Manufacturing Engineering and

Automation Nobelstrasse 12 70569 Stuttgart Germany
b TU Chemnitz Institut fr Physik Reichenhainer Strasse

70 09126 Chemnitz Germany
Published online: 06 Aug 2010.
To cite this article: T Dunker , G Radons & E Westkmper (2003) A coevolutionary algorithm
for a facility layout problem, International Journal of Production Research, 41:15, 34793500, DOI: 10.1080/0020754031000118125
To link to this article: http://dx.doi.org/10.1080/0020754031000118125
PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the
Content) contained in the publications on our platform. However, Taylor & Francis, our
agents, and our licensors make no representations or warranties whatsoever
as to the accuracy, completeness, or suitability for any purpose of the Content. Any
opinions and views expressed in this publication are the opinions and views of the
authors, and are not the views of or endorsed by Taylor & Francis. The accuracy
of the Content should not be relied upon and should be independently verified with
primary sources of information. Taylor and Francis shall not be liable for any
losses, actions, claims, proceedings, demands, costs, expenses, damages, and

other liabilities whatsoever or howsoever caused arising directly or indirectly in
connection with, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any
substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,
systematic supply, or distribution in any form to anyone is expressly forbidden.
Terms & Conditions of access and use can be found at
http://www.tandfonline.com/ page/terms-and-conditions
int. j. prod. res., 2003, vol. 41, no. 15, 34793500
A coevolutionary algorithm for a facility layout problem
Downloadedby[MonashUniversityLibrary]at01:2202January2015
T. DUNKER{*, G. RADONS{ and E. WESTKAMPER{
This paper presents a coevolutionary approach to the numerical

optimization of large facility layouts. Our work is based on a mixed
integer model for the layout constraints and objectives, which
improves formulations found in the literature. Nevertheless, layouts
with more than seven departments are dicult to solve. One way
out is to apply genetic algorithmssearching systematically for
solu-tions but without guarantee of finding an optimum. In this
paper we suggest some improved mutation and cross-over
operators. Yet, with increasing number of departments also genetic
algorithms take very long. In this case we propose to use additional
structures given by qualitative or quantitative reasoning. Clustering
the departments into groups we allow each group (species) to
evolve (genetic algorithm) in a separate area while position and size
of these areas (environment) undergo an evolution, too. Numerical
experiments verify this coevolutionary approach.
1. Introduction
One subproblem in factory planning consists in determining good
locations of a set of departments (or manufacturing cells) on a planar
site. This task is called the facility layout problem. The objectives
briefly described by the word good are manifold. In addition, many of
them are of a qualitative nature and it is not straight-forward to
formulate them as measurable quantities. One objective is certainly to
minimize the material handling cost. Yet, manpower requirements,
work-in-process inventory, flow of information, etc. play an important
role, too.
In all cases a basic assumption is that the proximity of certain
departments is more favourable than other configurations. The aim is to
arrange the departments in such a way that the desired proximity
relations are satisfied.
There exist many simple examples. If all departments require only

proximity with a predecessor and a successor a line layout is an
adequate solution. A star-like layout might be a good design in a
situation where few departments preprocess parts which are
assembled in a final department. But, as soon as the number of
departments increases and the exchange relations between the
departments are more complex, one cannot fulfil all desired proximity
relations and one has to decide between dierent possibilities. In such
a case it is natural to apply computer programs for finding solutions.
For several decades there has been research on this subject. Meller and Gau (1996)
present a detailed review of the dierent formulations of the facility layout
Revision received March 2003.

{ Fraunhofer Institute of Manufacturing Engineering and Automation,
Nobelstrasse 12, 70569, Stuttgart, Germany.
{ TU Chemnitz, Institut fur Physik, Reichenhainer Strasse 70, 09126 Chemnitz, Germany. * To whom
correspondence should be addressed. email: tmd@ipa.fhg.de
International Journal of Production Research ISSN 00207543 print/ISSN 1366588X online # 2003 Taylor &
Francis Ltd http://www.tandf.co.uk/journals
DOI: 10.1080/0020754031000118125
3480
T. Dunker et al.
problem and the variety of algorithms. Some additional recent references can
be found in the introduction of Chiang (2001).
One can roughly distinguish between three types of problem

formulation. Suppose there are a finite number of locations (e.g. on a
lattice) and a set of depart-ments or machines. In a first approach the
task consists in assigning the departments to the dierent locations
minimizing some cost function. This yields a quadratic assignment
problem. A second approach starts with the available floor area and
divides it successively into smaller subareas with respect to certain
constraints which finally yields the locations for the departments. In a
third formulation the necessary sizes and shapes of the departments
are given and the departments can be placed arbitrarily within the
available floor area. This last formulation, which results in a mixed
integer programming (MIP) problem, forms the basis of our considerations in this paper.
All these approaches lead to combinatorial optimization problems
which share the diculty of a computational complexity growing very
rapidly with the number of departments. Hence many suggested
algorithms for treating these problems search heuristically for good
solutions instead of aiming at globally optimal solutions. Evolutionary
methods like genetic algorithms (GA) have successfully been applied
in this context, see e.g. Chan and Tansri (1994), Rajasekharan et al.
(1998), Azadivar and Wang (2000) or Tavares et al. (2000).
In this work we present a new evolutionary approach in order to
attack such large-scale problems. Making use of additional structural
features, which might be given by qualitative or quantitative
reasoning, we apply principles of coevolution. Given a grouping of the
set of departments we consider each group as a dierent species and
let them pass through an evolutionary process in a common environment.
The remainder of this paper is organized as follows. In Section 2 we
describe the mixed integer formulation of the facility layout problem
chosen for our work. We present some improvements which reduce the
number of binary variables and accel-erate the solution. Section 3
provides the definition of the genetic operators and the description of
the coevolutionary algorithm. We introduce new operators and
demonstrate their excellent performance with some examples. Finally,
in Section 4, we discuss the numerical results of our coevolutionary
algorithm.
2. Mathematical formulation
2.1. Spatial considerations
2.1.1. Position and orientation
In order to make real facility layout problems manageable for

computational algorithms a certain abstraction is necessary. The
approximation of departments, floor areas, and others by rectangles is
a popular method, see e.g. Rajasekharan et al. (1998) or Chiang
(2001). In addition, we assume that their sides are parallel to the axis
of our coordinate system in the plane. Furthermore, we suppose that
the size and the shape of each department or facility is given in
2
advance. Denoting a rectangle by A _ R we introduce the following

notation:
I f1; . . . ; ng index set of all rectangles,
i; j 2 I
indices for the rectangles,
Coevolutionary algorithm for a facility layout problem
3481
Xi; Yi 2 R coordinates of the centre point of the ith rectangle:

Oi 2 f0; 1g orientation of the ith rectangle:
Oi 1 long side parallel to the x-axis,
Oi 0 long side parallel to the y-axis,
X
M
iY
2 f0; 1g flip up/down (symmetry axis parallel to x),
2 f0; 1g flip left/right (symmetry axis parallel to y).

Note that we use small letters for given constants, while potential variables
position and orientationare denoted by capital letters. For a simple rectangle
we need only two orientations but as soon as the internal structure of such an
area is not symmetric with respect to its central axes we have to consider further
transformations. Figure 1 shows all possible orientations of a rectangle with an
internal structure.
Next, we wish to model nested areas. This is formalized below. Let i* be the index
of the total available floor area. We define a map P : I ! I, i.e. i !7 Pi, with the
following meaning. For each i 2 I the number Pi gives the index of a rectangle in
which Ai shall be contained. For i* we define Pi* i*. In this way we define a
hierarchical structure of embedding relations. Most of the time one will have simply
Pi i*. Furthermore we will distinguish two types of mrectanglesmovablef or

fixed. For this purpose we define two index sets I _ I
and I _ I with
m
f
m
f
I \ I 1 andf I [ I f I, where m stands for movable
and f for fixed.
Obviously, i* 2 I and i 2 I will mean Ai is fixed to APi.
m
f
and I one can
Depending on the containment structure P and the sets I
determine whether the above-defined quantities will be variables or constants.
Let us use the following notation:
k
P i P_ _ _ Pi _ _ _ :
|{z}
k times
Then we can distinguish the following three cases.
(1) If i 2 Im then certainly

Xi, Yi and
Oi are variables. Whether there is a need
X
Y
for the variables M

and Mi depends on the internal structure of Ai, i.e.
ki
on the Aj s with P j i for some k.
Mi
=0
M iY = 0
Mi
=1
M iY = 0
Mi
=0
Mi =1
M iY = 1
M iY = 1
Oi =0
Oi =1
Figure 1. All dierent orientations of a rectangle with an internal
structure.
3482
T. Dunker et al.
f
(2) If i 2 I Yand P i 2 I for all k then all quantities are constants and there is no need for M i
and Mi .
f
k
m
k
m
k
(3) If i 2 I and P i 2 I for some k then set k* minfk : P i 2 I g and j P _
X
i. In this case Xi, Yi and Oi are variables depending on the Xj , Yj , Oj , Mj and

Y
Mj . This will be investigated in the following paragraph.
In order to describe the transformation of a fixed rectangle Ai attached to a

movable one Aj we need constants for modeling this dependency:
j
Then we obtain the centre point coordinates from the relative coordinates by the
linear transformation
j
Yj
Yi
M yj
!;
_ _
_
_
i
with an orthonormal 2 _ 2 matrix M which can be decomposed in the following
way
j X
1 _ Mj
1 _ Mj
Vj
1 0
1 !B
11 :
_0 C
@ _
A
This is more advantageous since the variables U j and Vj take values in f0; 1g
only. One can verify that the constraints
A1
y 43
x
y
34
xi ; yi 2 R relative coordinates of Ai with respect to Aj (long side corresponds

to x-axis), see figure 2,
j
o i 2 f0; 1g relative orientation of Ai with respect to Aj (1 corresponds to
parallel long sides), see figure 2.
2
s
A4
A2
Figure 2. Illustration of the relative coordinates and orientations for nested rectangles. The
3
relative orientation o 4 equals 1 as the long sides of A4 and A3 are parallel.

Y
0 _ Oj Mj _ Uj
0 _ Oj _
Y
Mj
Uj
0 _ _Oj Mj Uj
Y
Oj Mj Uj _ 2
X
Oj Mj _ Vj _ 1
X
Oj _ Mj Vj _ 1
_Oj
3483
X
Mj
Vj _ 1
8
X
1 _ Oj Mj Vj
ensure the correct assignment of entries in the matrices M. The orientation of Ai
depends only on the orientation of the Aj
Oi
and its relative orientation
O
1j
for oj
Oj
_
2.1.2. Geometric containment constraints
10
0:
for oj
If Pi j and i 2 I we can express the relation Ai _ Aj by the following

four inequalitiesfor the left side
sj
li
si
1
Xj _ 2 Oj _ 2 _ Oj _ Xi _ 2 Oi _ 2 1 _ Oi;
lj
the right side
l
Xi
2 Oi
s
i
lj
2 Oj
1
_ Oi _ Xj
sj
11
1 _ Oj ;
12
1 _ Oi
13
1 _ Oj :
14
the bottom
sj
Yj _
and the top
2 Oj
s
Yi
lj
_
Oi
1 _ Oj _ Yi _
si
2 Oi
l
i
2 Oj
1
_ Oi _ Yj
li
_
Observe that as si, li, sj and lj are given constants the inequalities (11)(14)
are linear with respect to all other quantities.
2.1.3. Non-overlapping constraints

Now we turn to the part of the model which makes the numerical
treatment so hard. The non-overlapping of two rectangles Ai and Aj
means that Ai \ Aj 1. In the literature there exist dierent suggestions
for representing this placement restric-tion. The first mixed integer
formulation of the facility layout problem which is due to Montreuil
(1990) proposes four binary variables for each non-overlapping
relation. In Meller et al. (1999) this approach is optimized in order to
improve the performance of the branch and bound algorithm (b&b).
Das (1993) and Rajasekharan et al. (1998) use a formulation which
needs three binary variables.
3484
T. Dunker et al.
We will propose a model using two variables only. Let us start with a
short motiva-tion by computational results.
We have compared our formulation to that used in Das (1993) and
Rajasekharan et al. (1998) using the CPLEX MIP-solver (ILOG, Inc.) with
two dierent branching strategies. The second problem from Das
(1993) which consists of six departments served us as a test example.
Table 1 shows that the number of nodes and the solution time (on a
Pentium II 400 MHz) CPLEX needed for our formulation were considerably smaller.
Two non-overlapping rectangles Ai and Aj must be separated by a vertical or
a horizontal line. We introduce variables
Sij 2 f0; 1g line direction (vertical/horizontal)

O
Sij 2 f0; 1g order of Ai and Aj .

D
The setting Sij ; Sij 0; 0 stands for Ai left of Aj , 0; 1 for Ai right of Aj , 1; 0 for Ai
below Aj and 1; 1 for Ai above Aj , respectively. In order to avoid redundant
variables we require i > j. Let lmax maxi li. Using the four linear expressions
Eij
lmaxSij Sij
D
Sij
15
Eij
lmax1
Eij
lmax1 _ Sij Sij
Eij
lmax2 _ Sij _ Sij
_ Sij
16
17
18
D
where each takes the value zero for exactly one setting of Sij ; Sij
we can derive the necessary inequalities
li
si
lj
sj
Xi
Oi 2 1_ Oi _ Xj _
lj
sj
Oj _
li
1 _ Oj Eij
si
19
20
21
22
Xj
Oj
2
s
2 1_
l
Oj _ Xi _
Oi _
sj
1_ Oi Eij
lj
Yi
Oi 2 1_ Oi _ Yj _
sj
lj
2
s
Oj _
2
l
1 _ Oj Eij
4
Yj
2 Oj
2 1_
Oj _ Yi _
Branching strategy
2 Oi
1_ Oi Eij
Automatic
Number of binaries
3
Number of nodes (_10 )

Time in s
Strong
Two
Three
Two
Three
11:5
25.5
47:9
187.6
5
46.3
31
512
e
1.
C
o
m
pa
ris
on
of
th
e
co
m
pu
tat
io
na
l
co
m
pl
ex
ity
(n
od
es
of
th
e
b&
b
tre
e
an
d
ti
m
e)
of
ou
r
m
od
el
(t
w
o)
wit
h
th
e
m
od
el
(th
re
e)
w
hi
ch
ca
n
be
fo
un
d,
e.
g.
in
D
as
(1993) or Rajasekharan et al. (1998). For the data of the test example with six
departments see Das (1993).
3485

D
There are relative positions which allow two settings of Sij , e.g. if Ai is to
the right of and above Aj so that they could be separated by a vertical as
well as by a horizontal line.
In order to break this symmetry we wish to
D
require that if possible Sij 1 is chosen. This could be achieved by
sj
lj
si
li
Yj _
Oj _
21
_ Oj < Yi
and
si
2 Oi
li
2 Oi
21
_ O l
i
max
23
ij
Yi _
_ Oj
_ Oi < Yj
2 Oj
1 lmaxSij
24
Unfortunately, these are strict inequalities which cannot be treated by

a linear program. Using them with a _ in our model we would miss
only cases where Ai and Aj touch the same horizontal line from
opposite sides.
Note that in our model for the genetic algorithm we will not make use of (23) and (24). On
the contrary, we prefer to use the more flexible setting for the SijDs and SijOs.
If the problem is invariant with respect to rotation and reflection then we can
D
break these symmetries by fixing, e.g. S21
1 (rotation), S21
1 (reflection
symme-try along x) and by adding the constraint 0 _ X2 _ X1 (reflection symmetry

along y).
Finally, we denote by I
no
_ fi; j : i > j; i; j 2 Ig the set of the non-overlapping
relations which are necessary for the model. The aim is to keep I
no
as small as
possible. For example consider the following situation: A2 _ A1 Ai_ , A3 _ A1,

A4 _ A2, A5 _ A2, A6 _ A3 and A7 _ A3. In order to have the no overlap between A2, A3 and
no
A4; . . . ; A7 it suces to set I f3; 2; 5; 4; 7; 6g. The relations 6; 4; 6; 5;
7; 4 and 7; 5 are automatically true because of the containment
relation and 3; 2.
2.2. Distance based objective function
The objective is to minimize the distances between dierent points.
In most cases these points will represent the drop o and pick up
points of a department. However, we can also imagine other points
which possess a certain importance, e.g. for the production process,
the internal communication process, or the security. In the following
we will call these points IO-points. They are attached to some floor
area or a department, i.e. a rectangle may possess no, one, two or
more IO-points. Hence, whether they are fixed or how they can be
moved, depends on the corresponding rectangles. We use the
following notation:
E
E
I f1; . . . ; n g set of the indices of the IO-points,
_;
_i
2 Ii
E
xE
_;
indices for the IO-points,
y_ 2
point
relative coordinates of the _-th IO2
to the centre of the ith rectangle, i 2 I
I*, with
respect
_2
m
,
2
_X
R
w
__
3486
T. Dunker et al.
Given an existing layout it might be necessary to consider the cost of changing

m
the layout at the same time. For this purpose we introduce for all i 2 I :
0
0
0
Xi ; Yi ; Oi
position and orientation values of the original layout,
Mi 2 f0; 1g
this variable isX1 if Xi,YYi or Oi have changed from the original
layout or if Mi
or Mi is equal to 1,
w m
i i2I
weights for moving a rectangle representing e.g. the cost for
moving a department.
With this notation the objective is to minimize the following expression
min
w__ _X
__
_Y
__
i I
2
a;_2I
wM
_>_
25
X
X
where the following constraints have to be satisfied
X
X
Y
_X
_ _X
_X
_ _X
_Y
_ _Y
_Y
_ _Y
__
26
__
27
__
28
__
29
_ lmaxMi
30
_ Xi _ lmaxMi
31
for all _; _ 2 I with _ > _ and w__ 6 0 and

0
Xi _ Xi
0
Xi
Yi _ Yi
0
Yi
_ lmaxMi
32
_ Yi _ lmaxMi
33
Oi _ O i
0
Oi
_ Mi
34
_ Oi _ Mi
35
Mi _ Mi
36
Y
Mi
37
_ Mi
for all i 2 I .
If all objects attached to a rectangle Ai are located on the same symmetry
axis then it is not necessary to employ (1) with all the constraints (2)(9). In this
case we need only two binary variables U i; Vi 2 f0; 1g implementing the rotation
and a single reflection. Consider, e.g. a IO-point _ attached to the rectangle Ai.
We can use the following model with inequality constraints
_2Oi _ Ui _ Vi _ 2Oi
Oi _ Ui Vi _ 2 _ Oi
and equations
E
X _ Xi x _
38
39
Ei
Ui Vi _ 1
40

for the symmetry axis parallel to the long side or, alternatively,
i
E
E
Ui _ Vi
X _ Xj y_
i
E
E
Y _ Yj y_ Ui Vi _ 1
3487
42
43
for the symmetry axis parallel to the short side of the rectangle. Thus our
mixed integer programming model is completely described by the
equations (1)(43). All equalities, inequalities and the objective are linear.
The model contains binary
variables Oi, MiX , MiY , Ui, Vi, Mi, SijD and SijO. While the number of SijDs and SijOs increases
quadratically, the number of the others increases linearly with the number
of rectangles.
3. Coevolutionary algorithm
The goal of finding an optimal solution of the mathematical model
introduced in Section 2 and to prove its optimality can be achieved
only for a small number of departments (up to _ 7). This is due to the
quadratic increase in the number of binary variables. That is why
various heuristic methods have been developed to find systematically
at least suboptimal solutions.
One such approach uses genetic algorithms, see e.g. Chan and
Tansri (1994), Conway and Venkataramanan (1994), Gero and Kazakov
(1998), Kochhar and Heragu (1998), Kochar and Heragu (1998) (all
quadratic assignment problems), Shnecke and Vornberger (1997),
Azadivar and Wang (2000) (both slicing tree repre-sentations),
Rajasekharan et al. (1998), Tavares et al. (2000) (both mixed integer
models), Gau and Meller (1999) (slicing tree and mixed integer). In our
case the information about the setting of the binary variables is
translated into a genetic code. Then a population of individuals
carrying this genetic code undergoes an evolution-ary process which
creates improved generations of this population by selection, crossover, and mutation.
In the following we introduce a coevolutionary approach which goes
beyond the standard genetic algorithms. The philosophy of coevolution
can be described as follows. Let us suppose that a large problem can
be decomposed into smaller ones which are linked to each other. Then
one can assign to each such subproblem a population of individuals
representing possible solutions. Dierent subproblems form dierent
species which undergo an evolution. Observe that there is no
exchange of genetic material between dierent species. However, the
fitness of an individual from one population now depends also on the
other populations.
One interesting field of research is to obtain the dierent species
themselves by an evolutionary process of specialization. In our case
we generate the problem decom-position by ourselves. We form
groups of departments. For each group we reserve a separate area.
Inside each such area group layouts are evolved by genetic
algorithms. The fitness of one group layout depends in addition on the
best layouts of the other groups. This is the coevolutionary part of our
algorithm. A second genetic algorithm changes size and position of the
areas. This is done for two purposes. First this allows further
improvement. Secondly, by changing the size we can control the
evolution of a groupmore space allows more variation while
tightening up stops evolution.
The remainder of this section is organized as follows. First, we
describe the genetic operators which are adapted to the problem.
Then the coevolutionary algorithm is described in detail.
3488
T. Dunker et al.
3.1. An adapted genetic algorithm

In Rajasekharan et al. (1998) a (01) sequence for setting the binary variables is
used as the genetic code. Standard cross-over and mutation operators are applied.
Obviously, some of the outcomes are infeasible settings. If an infeasible gene occurs
it is eliminated from the population immediately. There are for example, binary
variables which represent the order of the x-coordinate of two centre points. If we
take six centre points then we have 15 such variables giving 32768 possible genes,
but there are just 6! 720 possible arrangements, e.g. many of the settings are
infeasible as they do not satisfy the transitivity of relative positions (if Ai is left of Aj
and Aj is left of Ak then Ai cannot be right of Ak, it has to be left of Ak).
In order to avoid producing too many infeasible genes we introduce a coding

and operators which do not leave the space of feasible settings.
3.1.1. Coding
We assume that the problem to solve involves the rectangles with indices from the
g
where the letter g stands for group and k 1 2

is the index of
index set Ik _ I
; ;...
g
g
rectangles.
Denote by
the group.p Let Nk jIk j be the number of relevant
_ 1; . . . ;
n
nk
k individuals. Let
the index of an individual within a population of
1; 2; . . . be the index of the generation. The _th individual of the kth group in
generation will be represented by
p
i ;...;i
k;_
;i ;...;i
; b
ij
i; j2Ik
i>j
The bij 2 f0; 1g, with i; j 2 Ik and i > j, represent values of the binary variable Sij , i.e.
whether there is ax vertical or horizontal
separating line between the rectangles Ai
y
and A . The two vectors i ; . . . ; i g and i ; . . . ; i g are permutations of the elements
1
nk
1
nk
g j
of Ik and they represent the order of the x- and y-coordinates of the midpoints of
the rectangles.
Next we describe the translation of this genetic code to the variables of the model
O
g
with i j
I and i
D
described in section 2. Given an I
and
;
> j in
k;_
way. For each pair of indices 1
we set the Sij
<
ij
n g
2k
we check the following
_x 1 x 2 _ k
the following
x
x x
x
alternatives. Set i maxij1 ; ij2 and j minij1 ; ij2 . If bixjx 0 holds, i.e. there is a
vertical separating line between Aix and Ajx , then set
D
Sixjx 0
0 if
Hence the final order
j1
Sixjx
Note that if bixjx 1 the order
44
j1
<
8
1 if
x:
i :
j2
45
, ij2 does not enter in the problem formulation.
of the x-coordinates of
and
Aix
j1
Aix
j2
is not necessarily
y
Analogously, for the y-direction we set i maxij1 ; ij2 and

y
y
y
1 is true, i.e. there is a horizontal separating line between
j minij1 ; ij2 . If bi j
Aiy and Ajy , then we assign
X
ij1
_X
ij2
yy
3489

D
Siyjy 1
O
Siyjy
3.1.2.
y
0
if i
if iy
46
y
j1
47
iy :
j2
Genetic operators
Chan and Tansri (1994) introduced three dierent cross-over operators for per-mutations
partially matched, order and cycle cross-over. For our genetic algorithm
we use a version of the order cross-over. After selecting two parent genes Ik;_ and Ik;_
let us consider the parts of the genes representing the x- and the yorder. Take, for example,
1;_
; . . . ; inkg ; _1 and
;
1;_2 ;
. . . ; inkg _2
where we add the number of the individual as a second subindex. We select

g
randomly two cut positions c1; c2 2 f1; . . . ; n kg with c1 _ c2. Then we construct two
new genes from the two selected genes. First we fill the position from c1 to c2 with
the original parts of the sequence
x
. . . ; ic
x
ic 2;c2
1;_1
; . . . ; ic
2;_1
; . . . and . . . ; ic
1;c2
;...;
; . . .:
Then the positions to the left of c1 and to the right of c2 are filled with
the numbers from the other parent which are not contained in the
already filled part. While filling we keep the order given by the parent
we take the elements from. In terms of the notation above this means,
e.g. for the first ospring gene
x
x
x
x
x
x
i ; . . . ; i
;i ;...;i ;i
;...;i g
f 1;_2
with
fi
f c1_1;_2 c1;_1
;...;i
f 1;_2
c2;_1 f c1;_2
g \ fi
f nk_c2c1_1;_2
f nk_c2c1_1;_2
;...;i
c1;_1
g1
c2;_1
and the mapping j 7! f j is strictly increasing. In the same way the

cross-over is defined for the y-direction. The motivation for this
definition is the idea that we wish to keep the part that is located
between the cuts, hoping that it contributes to a good solution and
arranging the remaining elements in the order given by the other
parent.
For mutation a single parent is randomly selected.
Then
the mutation operator
x
x
exchanges two randomly chosen elements in
just y
in i1
; . . . ; ink
and independently
also
; . . . ; inkg .
The more complicated part is the modification of fbij gi;

or horizontal separating line. As we did not find a
j2Ikg;i>j
which codes the decision of having a vertical
method which could be geometrically motivated we decided to use the

standard cross-over with two cut positions and the standard mutation.
However, in addition we apply an improvement strategy. First, we fix the
D
variables Sij and Sij
for the given individual I
k ;_
according to (44)(47)
a
n
d
so
lv
e
th
e
re
m
ai
ni
n
g
m
ix
e
d
in
te
g
er
pr
o
bl
e
m
(1
)
(4
3)
:
evaluate (I
k;_
)
g
for all i; jD2 Ik with

i>j
O
fix Sij and Sij
endfor
solve remaining mixed integer

problem return solution
3490
T. Dunker et al.
x
Next, we update i1
; . . . ; inkg
and i1
y
; . . . ; inkg by sorting the centre point coordinates
whether they can be changed
of the obtained solution. Then we check for all SD
ij
without violating a constraint in the current solution.
change_is_possible (Sij )
D
if Sij 0 (x-direction)
if jYi _ Yj j _ siOi=2 li1 _ Oi=2 sj Oj =2 lj 1 _ Oj =2
return true
else
return
false endif
else (y-direction)
if jXi _ Xj j _ liOi=2 si1 _ Oi=2 lj Oj =2 sj 1 _ Oj =2

return true
else
return
false endif
endif
If it is possible we change the variable Sij . These actions are repeated

until the objective value does not decrease further. Summarizing, we
have sketched our improvement strategy below:
improve (I
k ;_ )
while the objective value decreases
evaluate (I
x k;_
obtain i1 ; . . . ; inkg
from the
and i1 ; . . . ; inkg solution
for all i; j 2 Ik with i > j
if change_is_possible (Sij )
if Sij 0
b
ij
ij
else
b
endif
endif
endfor
endwhile
g
return Ik
;_
with smallest objective value
We have tested these operators with a standard genetic algorithm with a

population of 50 dierent individuals. Each new generation is created by
cr
mu
co
n 20 cross-over operations, n 5 mutations and copying the n 5
best individuals. The selec-tion for the cross-over and the mutation
accepts individuals with objective value above average of the population
ac
mu
with a probability of p 20%. The mutation rate is chosen to be r
10%. We terminate if the average of the objective values has not changed
ch
more than m 0:01%, or the average of the best values has not
nc
ge
changed for the last n 10 generations, or a maximal number m
1000 of generations has been exceeded.
3491
1 cr
for 1; . . . ; n
_1
select parents from generation
cross-over generates two new individuals for generation
endfor
mu
for 1; . . . ; n
select parent from generation _ 1
mutation generates a new individual for
generation endfor
co
copy the n best individuals from generation _ 1 to generation
compute the average of the objective values and determine
whether the best individual has changed
while the change of the average is larger than m
individual has changed during the last n
nc
ch
and the best

ge
generations and _ m
Applying this to the third example from Das (1993) with eight departments we obtained
satisfactory results. Running the deterministic algorithm (31 h 30 min on a Pentium III 866
MHz with a memory use of approx. 1.5 GB) it has been proved that the optimal objective
value is 8778:3. Running our genetic algorithm 13 times, the optimal objective value was
reached three times. In the worst case the objective value was 9106:6 which lies just
3:7% above the optimum. Figures 3 and 4 show the convergence and distribution of the
solution values, respectively. In all cases com-putations took less than 10 minutes (on a
Pentium II 400 MHz). Also for all other examples our approach shows very good results.
Table 2 compares our best results to
10000
9800
best objective value
genetic_algorithm
initialize population and compute the average of the objective
values do
9600
9400
9200
9000
8800
0
100
200
300
400
elapsed time in seconds

Figure 3. Convergence of the best objective values.
500
3492
T. Dunker et al.
6
average: 8937.6
deviation: 102
frequency
0
8800
8900
9000
9100

Figure 4. Distribution of the objective values.
the best known from Das (1993) and Rajasekharan et al. (1998) showing a
consider-able improvement in all cases.
3.2. Description of the coevolutionary algorithm
If we attack large problems with the above-described simple genetic algorithm we
have to wait a very long time for good results. One way to obtain results more rapidly
could be stopping the algorithm if it exceeds a time limit. The result might be good.
However, the quality of this method would vary more or less randomly.
One may question whether large problems are really that large in practice.
Most of the time we have additional information which can reduce the complexity
con-siderably. By some quantitative or qualitative method we can form groups of
depart-ments which should be placed together. Then the problem reduces to two
smaller problems (in terms of the number of objects to be placed). One has to
provide an area for each group of departments and one has to determine the
layout for each group within these areas. Tam and Li (1991) suggested
approaching the facility layout problem in a hierarchical manner by a divide-andconquer strategy. They formed groups, computed the layout for each of them,
and placed the groups in a final step.
Number of
departments
8
10
12
Four-step method
described in
Das (1993)
Genetic algorithm
by Rajasekharan
et al. (1998)
10 777:1
15 878:3
41 267:5
9 174:8
19 777:3
45 353:5
Best results of our

genetic algorithm as
described above
8 778:3
15 694:5
37 396:1
Table 2. Minimal objective value for three example problems from Das (1993) which are
reported in Das (1993) and Rajasekharan et al. (1998) in comparison with our best results.
3493
We propose a coevolutionary method of iterative nature. In a first

step we find initial layouts for each group. Next, we fit a rectangle
around each group and enlarge each side by a factor 1 zk giving
more space to each group for possible further change. This allows, e.g.
that a group becomes more oblong during the subsequent
optimization. Next, we arrange these rectangles using again a genetic
algorithm. In the first run we approximate the IO-points by the central
points of the rectangles. Afterwards it is necessary to consider all
relative positions of the IO-points of the group. Experiments with
continued approximation by the central point did not show satisfactory
results.
Keeping the external IO-points for all groups constant, each group
undergoes a short evolution by a genetic algorithm. Observe that the
objective function does not only include the weighted distances between
the group members; the weighted dis-tances to the constant IO-points
outside the group contribute to the objective func-tion, too. It is obvious
that these genetic algorithms can be computed in parallel. For each group
we decide whether we change zk for the next iteration. For this purpose we
compare the new dimensions to the old ones. If the proportional increase
maxlnew _ lold;
0
l
old
max
_
maxsnew _ sold;
;
0
s
old
exceeds a certain percentage p of zk then zk is multiplied by a factor f > 1.

_
If it remains below p zk then zk is divided by the same factor f .

Consequently, if the shape of the needed area does not change the
provided group area becomes tighter. We stop the iteration when all group
areas are close to the needed area of the group. The algorithm is
summarized below.
coevolutionary_algorithm
for all groups k
genetic_algorithm find a good layout for the departments of

group k (open floor)
fit a group area around the group and enlarge it by zk (i.e. multiply the
length of each side by the factor 1 zk)
endfor
do
genetic_algorithm find a good layout for the group areas (consider each area as a
department with several IO-points)
for all groups k ? ? ? coevolution ? ? ?

genetic_algorithm find a good layout for the departments of group k
(the placement is restricted to the group area and all external IOpoints are fixed) fit a new group area around the group
if the ratio of the sides has changed

much increase zk
else
decrease zk
endif
enlarge the new group area

by zk endfor
while there is still a large zk for some k (or the average of the zks is
large)
L
et
us
ad
d
a
re
m
ar
k.
W
he
n
th
e
pr
ovi
de
d
flo
or
ar
ea
for
a
gr
ou
p
be
co
m
es
sm
all
er
an
d
sm
all
er
it
is
m
or
e
di
c
ult
for
th
e
ge
ne
tic
alg
ori
th
m
to
fin
d
ne
w
fe
asi
ble
lay
ou
ts
3494
T. Dunker et al.
D
for this group. An awkward choice of the Sij and Sij may prevent the
rectangles from fitting inside the prescribed area. However, as we
keep the genetic pool from the last iteration we can copy at least one
feasible individual to the population of the next generation.
4. Results and conclusions
For our numerical experiments we created a random example with
62 depart-ments (rectangles) of dierent shapes. For simplicity we
placed one IO-point in the centre of each department. The weights w__
were generated randomly, too. In this example, we do not start from
an initial layout. Hence, all the weights wi in (25) are zero. Table 3 in
the appendix provides all the necessary quantities. Let us call this
problem P62.
In order to find a grouping we implemented the heuristic grouping
algorithm of Harhalakis et al. (1990). A similar clustering algorithm was
applied by Tam and Li (1991). The aim is to arrange the departments
in groups minimizing the sum of the weights between departments
belonging to dierent groups. In addition, one limits the size of each
group. In De Lit et al. (2000) one can find grouping algorithms using
genetic algorithm. One can construct examples where the
deterministic heuristic by Harhalakis et al. (1990) stops far from the
optimum as it can handle only simple exchange operations. In these
cases algorithms like the one by De Lit et al. (2000) can improve the
grouping.
For our tests we generated groupings with four, six, eight and nine
groups with a maximal group size of 16, 11, eight and seven departments,
respectively. Minimizing the sum of weights wij between IO-points from
dierent groups is equivalent to maximizing the sum of weights
connecting departments of the same group. All the groupings we used are
summarized in table 4 in the appendix.
There are several parameters which influence the performance of the
proposed coevolutionary algorithm. For the
parameters
introduced
in
mu
ac
nc
section
3.1 and 3.2 we used the settings r
10%, p 20%, n 5,
ch
m 0:5%; the values in the following table:
First arrangement of the areas for each group

Adjusting the positions of the changed areas
Coevolution of the group populations
and initial enlargement factor zk

0:6,
ncr
nmu
nco
mge
20
0
12
5
5
3
5
5
3
15
1
3
increase
decrea
limit p 75%, se
limit
p 50% and change factor f 1:5.
There is always a conflict between good exploration of the search

space and fast computation. The first requires large populations which
again needs more time for computation. Certainly, these parameters
can still be optimized.
We applied the coevolutionary algorithm about 20 times to each

grouping of P62. Figure 5 shows the convergence behaviour of the
dierent runs and, in addition, displays the average, standard deviation,
and the best and worst case of both the objective values of the solutions
and the time of the computations. In figure 8, in the appendix, one can
find the best layouts obtained. We observe that both the average objective
value and the time of computation increase with the number of
3495
convergence of 20 runs with 6 groups
best objective value x1000000
4
1

average: 5390,635
deviation: 1160,37707
best: 3397,35
objective value
worst: 7170,13
average: 4,35529E6
elapsed time in seconds x1000

average: 5699,116
best: 3098,17
objective value
worst: 7687,68
average: 4,39307E6
deviation: 112706,55086
best: 4,18422E6
worst: 4,64994E6
best: 4,23898E6
worst: 4,44855E6
x1000000
Downloadedby[MonashUniversity
Library]at01:2202January2015

average: 4966,666
objective value
best: 3783,38
average: 4,27436E6
worst: 7294,18
best: 4,18614E6
worst: 4,44648E6

average: 4533,0181
best: 3979,55
worst: 5213,89
objective value
average: 4,06205E6
deviation: 131592,77082
best: 3,93936E6
worst: 4,44988E6
Figure 5. Computational results for problem P62 with four, six, eight and nine groups
(obtained on a PC, Pentium IV, 1.5 GHz).
groups. The first is due to a worse packing. Since the groups never
exactly fill the provided rectangular area there is more space lost when
the number of groups increases. Secondly, it turned out that the
computations treating the part where whole groups are moved are
very time consuming. Here not only the orientation also the dierent
reflection symmetries have to be considered. This is the reason why
the time of computation increases. Thus the clustering into four groups
appears to be the best choice for a facility layout problem of this size
the restrictions introduced by the grouping are the least yet the
computation is still fast.
In contrast to computations of less than two hours of the
coevolutionary algo-rithm the simple genetic algorithm needs on
average about two days and 17 hours for P62. The quality of the
solution (objective value) lies in the same range as the solutions
obtained by the coevolutionary algorithm. For comparison the corresponding results for 11 runs on a Pentium IV, 1.5 GHz are summarized
in figure 6. A trial with a MIP-solver, a good starting solution and lower
bounds did not yield any feasible solution after one week.
Of course, there are further related problems of interest, which we
will not treat here. For example, one may ask what happens when the
problem size is further increased. Is there a point from which on six
groups perform better than four? A second interesting question is
whether dividing groups into subgroups can be of advantage in some
cases. Here one introduces further restrictions so one would in general
expect worse objective values.
3496
T. Dunker et al.
convergence of 11 runs without grouping
4,6

average: 237530,45455
objective value
best: 103948
average: 4,26083E6
worst: 331484
best: 4,18105E6
worst: 4,38149E6
4,5
4,4
4,3
4,2
0,0
0,5
1,0
1,5
2,0
2,5
3,0

Figure 6. Summary of computational results of the genetic algorithm without grouping for
P62. While the values of the objective function are in the same range as the results
obtained with grouping, the computation time exceeds that of the coevolutionary
algo-rithm by an order of magnitude.
Summarizing, we conclude that the proposed coevolutionary algorithm opens

the possibility of finding good solutions for large facility layout problems within
some hours where global optimization algorithms fail. In addition there is still a
high potential for further computational acceleration by parallelization, if
appropriate hardware is available.
Acknowledgements
We would like to thank the referees for their valuable suggestions. This
research was supported by DFG research project SFB 467 Wandlungsfahige
Unternehmensstrukturen fur die variantenreiche Serienproduktion.
Appendix: Solutions and data
D2
D2
D4
D1
D9
D2
D3
D8
D4
D4
D7
D6
D1
D3
D8
D9
D8
D6
D5
D11
D3
D6
D7
D12
D5
8 Departments, Objective value: 8778.3
D7
D5
D10
D10
10 Departments, Objective value: 15694.5
D1
12 Departments, Objective Value: 37396.1
Figure 7. Best layouts for the last three example problems from Das (1993) found by the
above-described GA. For the example with eight departments it was possible to
prove optimality by solving the complete mixed integer problem.
Weights w__
si
li
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
15 14 51 93 60 42
13 10
44
88 22
16 14
33 35 6
20 13
92 90
16 13
83
21 14
19 17
19 14
21 21
20 17
13 11
14 12
19 18
21 17
21 20
21
7
20 19
16
9
17 11
16 15
14 10
19 16
20 15
18
9
18 14
14 14
16 11
12 12
17 13
13 10
14 13
17
8
21 19
21 10
15 10
12
9
21 17
16
9
18 14
15
9
17 12
17 12
11
9
20
8
21 17
20 13
19 10
20 14
18 10
13
8
18 11
16 10
20 10
19 19
21 11
20
7
11
7
17 11
21 11
13
7
21 10
17 16
13
32
8
11
13
6
16
45
56
31
42
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
45
63 96
25 37 12 46 23
13 14 71 94 17
11
18 18 98
59 61 53 54 18
72 20 63 84
83 60 41 62 94
18 65 8 7 37
61 91
42 75 77
23 99
14
32
43
99 10 86
92 54 56 67 62
83
56
94 59
46
35
11
62
44 52 97 86 42
98 17 62 90 54
88
23 75 11
52 21 96 12
30
24
82
17 38
37
40
29
87 81
52 45
18 23 16
63 29 93
81 23
56 19 65
76 43
42 83
38 36 25
42 50 29
28 17 8
55 58
65 90
62 96 9
82 52
39 91
17 56 92
26 46
90 97 70
27 13
91
94 55 30
74 93 8
55 85 84
84
27
51
6
26
24 82 45
49
52
76
48
31
11 32 75
76
61
6
32
17 72 42 82
21
71
49
7
71
27
22
52
40
84
98
59 90
98
71 14 36 95
98 76 98 25
66 86
79 49
77
59 72
7 57 69
73 37 24
5
55
23
32 29
26 14
82 7 36
61
69
32 99
25 52
65 28 83
9
70 43
98
87 53 72
44
10
16
96 51 97
15 41 18
77 37 19
31 27 79
67 82
53 35
17
35
29
30
9
61
65
87
91
12
67
36
22
94
95
96
56
87
27
52
29
32
11
14
72
46
61
20
11
37
32 51
41 21 37 51 25
85 7 52
54
91
65 12
42
30 76 48
47
65
17 26 85 79 79
39 82 48 71 48
75 9 54
23
65 40 64
50
55 69
30
61
97
74
15
36 83 54 36 77
25
95
94
78
78 71
14 12 91 16 27
60 60 10 52 48
45 51
69 59 35 62 72
19
64
63 47
28
29
19
50 55
10 10
84
34
68 91
31 21
45
77
34
43
57
85
86
75
85
76
90 34
56 43
15 30 44 10 45
42
60
24
29
90 71 59
74 49 91
42
65 33 50
46
23
25
36
52
84
43
9
82
68
20
48
78
9
57
6
29
62
67
75
17
31
41
27
75
32
20
68
68
88
56
71
63
41
8
54
30 40
85
40 41 84
26 77 23 93
82
84
83 16 70 76
21
16 16
17 18 33 60
14 55 23
76 27 26 97
84 53
18
85
15
19
61 34
65
68 88 74
56
66
17
93 41 42
44 7 36
80 51 18
93
33
69
21
33
53
62
78 68 78 73
19 61 83
9 10 75 56
54
90
85 93 35
68
84 26
8
89 48 90 97
70 77 63
25 6 22 21
53
94 27
70 73
72
49
86
20 64
12
54 16 30 76 96
31 57
43 8 24 26
38
22 42 92
10 35 80
71
9
65 72
79
24
92
36 58 84 94 91
80
9
39
23
73 22 12 7 11
47 42 24 84 57
33
39 90
6
24 89
77
27 45
62 70 94
29 78 85
52 71 20
91
79 30
72
97 98 70 22 93
19
21
11
31 26 19
65 42
47 62 42
72
82
13
34 58 51 19 80
71
46
82
41 75
47
6 18
8 45 28 16 99 32
5
12 43 84
96 17 75
77 14
47 91 92 85
41 48
22
80
80
22
85
22 54 30 88
23
6
40
70
28
47 40
59 58 38
68 12 38 82
78
41
88 81 95
58 24 36
68 89
13 44
13
93
80
25
52
28
71
74
84
97
51
51
75
45 59
67 69 61
65 84 34
54 96
62 93 7
73
94 29
13
32
52
64
45
30
41
83
29
17 85 10 85
51
37
80
66 39 70 68
95 9 99
71
27
8
54
64
37 25
53
83 29
82
88
67
54 95
50
98
90 28 66 56 19 75 69
11 40
73
89
55
28
22
80
78
97 10
40 92 59 22
60
42
59
51 69
23
42 62 35 38 58 79 54
17 37
63
43
98 94
30
24
17 32 7 84
74
69 48
88
15 78 22
95
35 56
89 90 29 64
46 85 86 7 99
59 94
62
73 5
84
33
79
53
99
58
42
5
47
69
9
74
49 72
91 50
55 50
21
35
60 11
64 91
95 51
18
74 68
79
24 57
74 13
24 71
79
92
87
89 62
74 41
72 61
58
73 32
45 40
41
79 55
87
40 41
27 31
76
67
22
85 58
96
74
79 86
39
94
67
13
86
7
43
88 83
26
50
72
41
91
61
74 51
53
65
19
18
13
60
65
88
10
26
9
48
76
25
7
52
68
5
83
69
9
69
65
44
59 68
68
31
76 33
88
27 29
17
97 19
72
44
47
99 70
8 45
67
60 54
33 51
98
94
44
15 22
21 67 63
34 73 85
32
9
85
57
24
16
13
75 93 58
67
83
6 90 71
7
6
87
23 70 83
59
46 72
14
70
54 98 31
56 35 65
32
49
36 22
17
74
77
55 21 86
39
46
79
44 72
11
9
21 33
11
8
70
83 99
55
89
28 47
84
81
24 95
18
27 54
10
55 12
11
33
88 31
37 22
62
6
24
90
79
25
37 41
74
25 95
11
66 18
50 91
82 53
95
39
56
54
55 18
87
54 24
74 8
40
51 43
25
24
88 23
54
33 78
11 89
41
36
51
24
27
36
50 13
98
62
83
59
53 55
8
8 20
60
87
79
73 91
59
72
68
11
8
11
10
65
57
35
36
37
16
16
63
40
48
10
46
54
60
91
5
69
95
7
36
14
76
11
10
68
44
87
31 80 25
76
82
Table 3. Side lengths and weight matrix for the randomly generated example of 62 departments (P62).
3498
T. Dunker et al.
Number
Group elements
1
2
3
4
Sum:
1
2
3
4
5
6
Sum:
1
2
3
4
5
6
7
8
Sum:
1
2
3
4
5
6
7
8
9
Sum:
f1; 4; 6; 8; 10; 12; 17; 23; 30; 33; 38; 39; 41; 45; 54; 58g
f3; 18; 19; 20; 21; 22; 24; 25; 26; 32; 35; 36; 51; 53; 56; 60g
f2; 5; 15; 27; 29; 37; 40; 42; 43; 44; 46; 47; 49; 52; 62g
f7; 9; 11; 13; 14; 16; 28; 31; 34; 48; 50; 55; 57; 59; 61g
f2; 3; 10; 17; 18; 24; 28; 37; 41; 55; 57g
f13; 19; 20; 22; 25; 32; 36; 37; 40; 49; 56g
f1; 7; 21; 23; 38; 43; 48; 53; 59; 61; 62g
f9; 11; 12; 26; 30; 31; 33; 34; 44; 50; 52g
f14; 27; 42; 47; 51; 54; 58g
f4; 5; 6; 8; 15; 16; 29; 35; 45; 46; 60g
f1; 12; 23; 33; 38; 45; 54; 58g
f18; 19; 20; 24; 25; 26; 32; 36g
f27; 37; 40; 42; 46; 52; 62g
f3; 21; 22; 35; 51; 53; 56; 60g
f6; 7; 10; 11; 13; 28; 48; 57g
f4; 8; 16; 30; 31; 39; 41; 50g
f9; 14; 17; 34; 55; 59; 61g
f2; 5; 15; 29; 43; 44; 47; 49g
f7; 21; 22; 35; 48; 53; 60g
f1; 23; 38; 43; 54; 61; 62g
f4; 6; 18; 24; 45; 55; 59g
f10; 13; 19; 28; 36; 42; 56g
f2; 3; 5; 17; 33; 41; 51g
f16; 20; 25; 32; 39; 40; 49g
f11; 15; 27; 29; 46; 47; 52g
f9; 12; 26; 30; 31; 50g
f8; 14; 34; 37; 44; 57; 58g
__
group
4819
4998
4944
4417
19 178
2245
2632
2879
2375
671
2649
13 451
1605
1222
946
1528
1244
1380
1165
1699
10 789
1067
1462
1176
1274
909
1263
1324
811
849
10 135
Table 4. Grouping all departments into four, six, eight and nine groups,
respectively.
D35
D17
D34
D24
D9
D14
D31
D48
D13
D7
D11
D28 D50
D57
D16
D52
D47
D15
D37 D29
D44
D49
D42 D40
Library]at01:2202January2015
D25 D56 D32

D12
D23
D62
D45 D17
D23
D4 D41
D39
D17
D30
D50
D1
D31
D8
D16
D61
D34
D9
D12 D58 D38
D22
D3
D60
D24
D32
D28
D56
D18
D43
D36
D57
D5
D10
D2
D14
D26
D19
D58
D54
D51
D9
D33
D20 D44 D47
D49
D14
Result, Iter. 13, Total Cost: 4186140.4 (6038 sec.)
D14
D48
D46
D52
D47
D29
D9
D8
D29
D56
D40
D11
D25
D31
D57
D13 D6
D13 D42
D27
D44
D15
D37
D25
D7
D58
D60
D10
D28
D36
D30
D61
D35
D38
D43
D45 D6
D24
D49
D59
D20
D12
D21
D19
D39
D16
D32
D26 D50
D23
D62
D18
D2
D5
D55
D3
D4
D17
D41
D53
Downloadedby[MonashUniversity
D47
D42
D7
D34
D11
D15
D16
D27
D11 D44 D34
D59
D55
D29
D31
D52 D50
D26
D4 D46
D5
D60
D30
D6
D35
D20 D56
D38
D37 D42 D46

D7
D51 D35
D53
D12
D61
D15
D52
D21
D21
D36
D32
D22 D49
D48
D27
D45
D8
D19
D40
D18
D62
D40
D54
D39
D43
D6
D10
D33
D59
D53
D25
D28
D1
D39
D38
D54
D62
D2
D55 D57
D4
D41
D8
D3
D41 D24
D33 D23
D13
D37
D30
D1
D58
D2
D51
D22 D26 D19
D3
D21
D5
D46
D36
D18
D43
D27
D10
D60
D20
D61
D45
D53
D55
D59
3499
D48
D22
D54
D1
D33 D51
D25
D39
D45
D48
D7
D49
D1 D51
D34
D47
D10
D2
D42 D23
D44
D52
D27
D14
D62
D55
D33
D12
D20 D58 D38
D61 D59
D40
D21 D36 D32
D22 D24
D26
D9
D6
D35
D56
D18
D29
D16
D43
D11
D57
D28
D15
D54
D53
D3
D37
D60
D13 D31
D41
D50
D5 D30
D17
D4
D8
D46
D19
62 departments, objective function: 4181054.0
Figure 8. The best layouts for four, six, eight and nine groups obtained by our coevolutionary algorithm and the best result of the genetic algorithm without grouping.
R
ef
er
e
nc
es
iva
r,
F.
an
d
W
an
g,
J.,
20
00
,
Fa
cili
ty
la
yo
ut
op
ti
mi
za
tio
n
us
in
g
si
m
ul
ati
on
an
d
ge
ne
tic
al
go
rit
h
m
s.
Int
er
na
tio
na
l
Jo
ur
na
l
of
Production Research, 38, 436983.

Chan, K. C., and Tansri, H., 1994, A study of genetic crossover operations on the facilities
layout problem. Computers and Industrial Engineering, 26, 537550.
3500
Chiang, W. C., 2001, Visual facility layout design system. International Journal of Production
Research, 39, 181136.
Conway, D. G. and Venkataramanan, M. A., 1994, Genetic search and the dynamic facility layout problem.
Computers and Operations Research, 21, 95560.
Das, S. K., 1993, A facility layout method for flexible manufacturing systems. International Journal of Production
Research, 31, 279297.
De Lit, P., Falkenauer, E. and Dechambre, A., 2000, Grouping genetic algorithms:
an ecient method to solve the cell formation problem. Mathematics and
Computers in Simulation, 51, 257271.
Gau, K.-Y. and Meller, R. D., 1999, An iterative facility layout algorithm. International Journal of Production
Research, 37, 37393758.
Gero, J. S. and Kazakov, V. A., 1998, Evolving design genes in space layout planning problems. Artificial
Intelligence in Engineering, 12, 163176.
Harhalakis, G., Nagi, R. and Proth, J. M., 1990, An ecient heuristic in

manufacturing cell formation for group technology applications.
International Journal of Production Research, 28, 18598.
Kochhar, J. S. and Heragu, S. S., 1998, MULTI-HOPE: a tool for multiple floor layout problems. International Journal
of Production Research, 36, 342135.
Kochhar, J. S. and Heragu, S. S., 1999, Facility layout design in a changing environment.

Meller, R. D. and Gau, K.-Y., 1996, The facility layout problem: recent and emerging trends and perspectives.
Journal of Manufacturing Systems, 15, 351366.
Meller, R. D., Narayanan, V. and Vance, P. H., 1999, Optimal facility layout design.
Operations Research Letters, 23, 117127.

Montreuil, B., 1990, A modelling framework for integrating layout design and flow network design. Proceedings
of the Material Handling Research Colloquium, 90(2), pp. 4358.
Rajasekharan, M., Peters, B. A. and Yang, T., 1998, A genetic algorithm for facility layout design in flexible
manufacturing systems. International Journal of Production Research,
36, 95110.
Schnecke, V. and Vornberger, O., 1997, Hybrid genetic algorithms for constrained place-ment problems. IEEE
Transactions on Evolutionary Computation, 1, 266277.
Tam, K. Y. and Li, S. L., 1991, A hierarchical approach to the facility layout problem.
Tavares, J., Ramos, C. and Neves, J., 2000, Addressing the layout design problem
through genetic algorithms and constraint logic programming. In M. H.
Hamza (ed), Artificial Intelligence and Soft Computing. Proceedings of the
IASTED International Conference, IASTED/ACTA Press, pp. 6571.

A Coevolutionary Algorithm For A Facility Layout Problem

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Coevolutionary Algorithm For A Facility Layout Problem

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [Monash University

Library] On: 02 January 2015, At: 01:22

Informa Ltd Registered in England and Wales Registered Number: 1072954

A coevolutionary algorithm for a

a Fraunhofer Institute of Manufacturing Engineering and

b TU Chemnitz Institut fr Physik Reichenhainer Strasse

PLEASE SCROLL DOWN FOR ARTICLE

losses, actions, claims, proceedings, demands, costs, expenses, damages, and

int. j. prod. res., 2003, vol. 41, no. 15, 34793500

A coevolutionary algorithm for a facility layout problem

T. DUNKER{*, G. RADONS{ and E. WESTKAMPER{

This paper presents a coevolutionary approach to the numerical

There exist many simple examples. If all departments require only

Revision received March 2003.

One can roughly distinguish between three types of problem

In order to make real facility layout problems manageable for

advance. Denoting a rectangle by A _ R we introduce the following

indices for the rectangles,

Coevolutionary algorithm for a facility layout problem

Xi; Yi 2 R coordinates of the centre point of the ith rectangle:

2 f0; 1g flip left/right (symmetry axis parallel to y).

Pi i*. Furthermore we will distinguish two types of mrectanglesmovablef or

Then we can distinguish the following three cases.

(1) If i 2 Im then certainly

for the variables M

i. In this case Xi, Yi and Oi are variables depending on the Xj , Yj , Oj , Mj and

In order to describe the transformation of a fixed rectangle Ai attached to a

xi ; yi 2 R relative coordinates of Ai with respect to Aj (long side corresponds

Coevolutionary algorithm for a facility layout problem

ensure the correct assignment of entries in the matrices M. The orientation of Ai

depends only on the orientation of the Aj

and its relative orientation

If Pi j and i 2 I we can express the relation Ai _ Aj by the following

the right side

2.1.3. Non-overlapping constraints

Sij 2 f0; 1g line direction (vertical/horizontal)

Sij 2 f0; 1g order of Ai and Aj .

lmax1 _ Sij Sij

lmax2 _ Sij _ Sij

Number of nodes (_10 )

Coevolutionary algorithm for a facility layout problem

Unfortunately, these are strict inequalities which cannot be treated by

break these symmetries by fixing, e.g. S21

symme-try along x) and by adding the constraint 0 _ X2 _ X1 (reflection symmetry

_ fi; j : i > j; i; j 2 Ig the set of the non-overlapping

possible. For example consider the following situation: A2 _ A1 Ai_ , A3 _ A1,

indices for the IO-points,

relative coordinates of the _-th IO2

to the centre of the ith rectangle, i 2 I

Given an existing layout it might be necessary to consider the cost of changing

for all _; _ 2 I with _ > _ and w__ 6 0 and

Coevolutionary algorithm for a facility layout problem

3.1. An adapted genetic algorithm

and Aj is left of Ak then Ai cannot be right of Ak, it has to be left of Ak).

In order to avoid producing too many infeasible genes we introduce a coding

where the letter g stands for group and k 1 2

way. For each pair of indices 1

we set the Sij

we check the following

Hence the final order

Note that if bixjx 1 the order

, ij2 does not enter in the problem formulation.