Professional Documents
Culture Documents
International Journal of
Production Research
Publication details, including instructions for authors
and subscription information:
http://www.tandfonline.com/loi/tprs20
, G Radons
& E Westkmper
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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.
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.
Downloadedby[MonashUniversityLibrary]at01:2202January2015
problem and the variety of algorithms. Some additional recent references can
be found in the introduction of Chiang (2001).
3481
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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
P i P_ _ _ Pi _ _ _ :
|{z}
k times
=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
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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.
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
3483
X
Mj
Vj _ 1
8
X
1 _ Oj Mj Vj
Oi
O
1j
for oj
Oj
_
2.1.2. Geometric containment constraints
10
0:
for oj
sj
li
si
1
Xj _ 2 Oj _ 2 _ Oj _ Xi _ 2 Oi _ 2 1 _ Oi;
lj
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.
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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
Eij
_ 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
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
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
2 Oi
21
_ O l
i
max
23
ij
Yi _
_ Oj
_ Oi < Yj
2 Oj
1 lmaxSij
24
If the problem is invariant with respect to rotation and reflection then we can
D
1 (rotation), S21
1 (reflection
Finally, we denote by I
no
relations which are necessary for the model. The aim is to keep I
no
as small as
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
_;
y_ 2
point
I*, with
respect
_2
m
,
2
_X
R
w
__
3486
T. Dunker et al.
w__ _X
__
_Y
__
i I
2
a;_2I
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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
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
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.
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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.
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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;_
<
ij
n g
2k
_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
j1
Sixjx
44
j1
<
8
1 if
x:
i :
j2
45
of the x-coordinates of
and
Aix
j1
Aix
j2
is not necessarily
y
_X
ij2
yy
3489
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,
Downloadedby[MonashUniversityLibrary]at01:2202January2015
1;_
; . . . ; inkg ; _1 and
;
1;_2 ;
. . . ; inkg _2
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
just y
in i1
; . . . ; ink
and independently
also
; . . . ; inkg .
j2Ikg;i>j
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
endfor
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
ij
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
return
false endif
else (y-direction)
else
return
false endif
endif
k ;_ )
evaluate (I
x k;_
obtain i1 ; . . . ; inkg
from the
if change_is_possible (Sij )
if Sij 0
b
ij
ij
else
b
endif
endif
endfor
endwhile
g
return Ik
;_
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
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
genetic_algorithm
initialize population and compute the average of the objective
values do
9600
9400
9200
9000
8800
0
100
200
300
400
500
3492
T. Dunker et al.
6
average: 8937.6
deviation: 102
Downloadedby[MonashUniversityLibrary]at01:2202January2015
frequency
0
8800
8900
9000
9100
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
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
maxlnew _ lold;
0
l
old
max
_
maxsnew _ sold;
;
0
s
old
endfor
do
genetic_algorithm find a good layout for the group areas (consider each area as a
department with several IO-points)
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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:
ncr
nmu
nco
mge
20
0
12
5
5
3
5
5
3
15
1
3
increase
decrea
limit p 75%, se
limit
3495
4
1
deviation: 51708,76107
best: 4,23898E6
worst: 4,44855E6
x1000000
Downloadedby[MonashUniversity
Library]at01:2202January2015
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
4,6
4,5
4,4
4,3
4,2
0,0
0,5
1,0
1,5
2,0
2,5
3,0
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
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
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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.
Downloadedby[MonashUniversityLibrary]at01:2202January2015
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
D23
D62
D45 D17
D23
D4 D41
D39
D17
D30
D50
D1
D31
D8
D16
D61
D34
D9
D22
D3
D60
D24
D32
D28
D56
D18
D43
D36
D57
D5
D10
D2
D14
D26
D19
D58
D54
D51
D9
D33
D49
D14
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
D59
D55
D29
D31
D52 D50
D26
D4 D46
D5
D60
D30
D6
D35
D20 D56
D38
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
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
D61 D59
D40
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
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
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.
Downloadedby[MonashUniversityLibrary]at01:2202January2015
Gero, J. S. and Kazakov, V. A., 1998, Evolving design genes in space layout planning problems. Artificial
Intelligence in Engineering, 12, 163176.
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., Narayanan, V. and Vance, P. H., 1999, Optimal facility layout design.
Tam, K. Y. and Li, S. L., 1991, A hierarchical approach to the facility layout problem.
International Journal of Production Research, 29, 16584.
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.