CHAPTER 3 IBTRODUCTION - Shodhgangaietd.inflibnet.ac.in/jspui/bitstream/10603/908/15/15_chapter...
Transcript of CHAPTER 3 IBTRODUCTION - Shodhgangaietd.inflibnet.ac.in/jspui/bitstream/10603/908/15/15_chapter...
CHAPTER 3
GRAFICS AND IT8 COMPARISON WITH OTHER
WELL-KNOWN ALGORITHMS*
3.1 IBTRODUCTION
The survey of machine-component cell formation methods
(Chapter 2) indicates that most of them deal with 0-1 matrix and
block-diagonalization of the matrix to obtain machine cells and part
families. The survey also indicates that there is scope to do further
research in the area of nonhierarchical clustering. In this chapter,
initially introduction to nonhierarchical clustering technique is given.
Then, two nonhierarchical clustering algorithms namely ZODIAC
(Chandrasekharan and Rajagopalan 1987) and GRAFICS (Srinivasan
and Narendran 1991) are explained in detail by giving suitable
examples. Then, by sequentially reviewing the published GT literature,
it is shown that GRAFICS can perform better than many well-known
algorithms. Then, the similarities and dissimilarities of the algorithms
namely ZODIAC and GRAFlCS are reported. Finally, the need for
improving GRAFICS algorithm is highlighted.
A paper entitled, 'Nonhierarchfcal clusturlng of machine-
component c u l t luing see& from an m c i e n t Seed Genuraffon
Algorithm (ESGA)' based on this part of the research work was
published in the Indwtrial Engineering Journal, Vol.XXN, No. 1 I Ck
No. 12, 1995.
is the machine-component incidence matrix entry for the i&
machine and the k" component,
is the machine-component incidence matrix entry for the jth
machine and the k'h component,
number of voids (zeros) within the diagonal blocks of the
machine-component incidence matrix in a n iteration (Bo
represents the number of voids in the previous iteration)
is the component vector corresponding to component type j of
the machine-component incidence matrix
is the dissimilarity (or distance) between two machines namely
machine i and machine j
number of exceptional elements in the machine-component
incidence matrix in an iteration (E,, represents the number of
exceptional elements in the previous iteration)
is the component index,
i~ machine group
i'h machine seed
is the similarity between two machines i and j of the machine-
component incidence matrix,
a variable which is equal to 1 if the machine i is grouped with
the machine j; which is equal to 0, otherwise
127
3.3 NO-CHICAL CLUSTERING
The identification of machines and part groups is similar to the
identification of "clustersw in a scattered data space (scatter of data
points). Researchers have applied cluster analysis in its varied form to
the problem of forming machine cells and component families.
Cluster analysis seeks to group data into clusters such that the
elements within a cluster are closely related while the clusters
themselves have little or no relationship amongst them. The major
classes of c lus te r ana lys i s a re hierarchical c luster ing and
nonhierarchical clustering.
A hierarchical clustering method first computes the similarity or
dissimilarity between each pair of parts or niach~nes. Some methods
use agglomerate philosophy while others use divisive philosophy for
clustering hierarchically. The hierarchical clustering algorithms
generate a hierarchy of feasible solutions each with a particular value
of a performance measure. The analyst chooses the best feasible
solution corresponding to the best value of the performance measure.
Nonhierarchical clustering algorithms start with an initial set of
machine seeds and results in a set of machine-component cells with
optimum or near optimum value of performance measure.
3.3.1 Advantage of Nonhlerarchical Clustering Methods over
Hierarchical Clustering Methods
The main drawback of hierarchical methods (Anderberg 1973) is
that when two points (row vectors or column vectors) are grouped
together a t some stage of the algorithm there is no way to retrace the
step even if it leads to suboptimal (or unnatural) clustering a t the end.
At every stage of clustering those points which have formed some sort
of groups face the rest of the data with a fait accompli that severely
limits further possibilities. In nonhierarchical clustering, the choice is
rather free, and the natural clusters emerge from the given data without
permanently binding any data unit due to the linking done in the
initial stages of execution.
Nonhierarchical clustering is capable of identifying the natural
groups in a data-set. Only three algorithms are available for formation
of machine groups and part families using the nonhierarchical
clustering technique. The three nonhierarchical clustering algorithms
are given below:
a. Ideal Seed Nonhierarchical Clustering Algorithm
(Chandrasekharan and Rajagopalan 1986a).
b. ZODIAC (Chandrasekharan and Rajagopalan 1987).
c. GRAFICS (Srinivasan and Narendran 1991).
Hence, the development of a lgori thms based on
nonhierarchical clustering methods needs to be explored
further.
Two nonhierarchical clustering algorithms namely ZODIAC and
GRAFICS are explained in the following sections.
3.3.2 Introduction to ZODIAC (Chandrmekhalm and Rajagoplan
1987)
The algorithm namely Z9DIAC (Zero-One Data-Ideal seed
Algorithm for Clustering) was proposed by Chandrasekharan and
Rajagopalan (1987). This is a nonhierarchical clustering algorithm.
Here each machine vector (row) is treated a s a point in a higher
dimensional zero-one space and clustered around some fured seed
points, which may be among these points themselves. In this section
the natural seed clustering version of the algorithm followed by
t h e ideal seed c lus te r ing algori thm a r e explained. ZODIAC
algorithm i s given in Appendix 1 along with the corresponding
flowchart (Figure I A. 1 ) .
The natural seed algorithm attempts to find the most natural
machine cells from the data. It uses Jaccard (Sukal and Sneath 1963)
similarity matrix to create initial seed points around which the
machines are to be grouped. The seed points should be a s far away
from each other, that is, a s dissimilar a s possible so that the points
clustered around them are similar and the clusters (groups) themselves
are dissimilar.
The Jaccard similarity coefficient is given by
where
S,, is the similarity between two machines i and j of the machine-
component incidence matrix,
a,, is the machine-component incidence matrix entry for the i'h
machine and the kth component,
a,, is the machine-component incidence matrix entry for the jth
machine and the kth component,
k is the component index,
1 ; if a, = a,, = 1 (or) if a, + a,,
0; if a, = a,, = 0
numbers of components which visit both machine i and
machine j
S!, = (Jaccard (number of components which visit both machine i and coefficient)
machine j) + (number of components which visit one or
other of the machines)
ZODIAC algorithm is explained using a sample problem. The
process sequences of the above problem is given in Table 3.1. The
corresponding initial 0- 1 machine-component incidence matrix of the
above problem is shown in Figure 3.1. The Jaccard similarity matrix
corresponding to Figure 3.1 is shown in Table 3.2. The average of the
Jaccard similarity coefficient for machines from Tables 3.2 is found to
be 0.1078. Since the matrix is symmetric, only the upper triangular
values are considered. The matrix is scanned to obtain a pair of
machines whose similarity is less than the average. Machine 1 and
Machine 4 have a similarity value of zero and hence these machines
are chosen a s the first two machine seed points. Machine 6 is chosen
a s the third seed point since its similarity with the existing seeds
namely machine 1 and machine 4 is less than the average similarity
value of 0.1078. It is to be noted that any machine that has a Jaccard
similarity value more than the average Jaccard similarity value even
with one existing seed (already chosen seeds) does not qualify to be a
seed.
Table 3.1 Process sequences of components of the sample problem
used to demonstrate ZODIAC algorithm
Note: C1, C2. ..., C6 represent component codes corresponding to
six different types of components.
Component code
1
Cl
I C2
I C3
I C4
1 C5 1 c6 L
M 1 ,M2,.... , M6 represent machine codes corresponding to
six different types of machines.
J
Total number of operations
2
2
2
2
2
2
Process sequence
M5- M1
M1 - M3 M6 - M5 M5 - bib M2 - M1 M4 - M2
--
Component k
Machine i
Figure 3.1. Init ial machine-component incidence matr ix used to
demonstrate ZODIAC algorithm
The Jaccard similarity matrix corresponding to th i s Figure 3.1
(calculated) is shown in Table 3.2
Table 3.2 Jaccard similarity matrix of machines corresponding to
Figure 3.1 used to demonstrate ZODIAC algorithm
Machine j
Machine i
The machines are clustered around these seed points. The three
seed points are given below:
1. S e e d l : ( 1 1 0 0 1 0 ) .....( machine 1).
2. Seed2: [ 0 0 0 0 0 1 ] .....( machine 41.
3. Seed3: [ O O l l O O ) .....( machine 6).
Each machine vector (machine row) is assigned to the machine
seed with which the distance (dissimilarity measure is given below) is
minimum. The dissimilarity (or distance) between two machines i and
j is given by the following formula:
do = ' / - I k
(3.3)
where
a,, is the machine-component incidence matrix entry for the ith
machine and the kth component,
aJk is the machine-component incidence matrix entry for the jth
machine and kLh component,
d,) is the dissimilarity (or distance) between two machines
namely machine i and machine j
and
k is the component index.
The machine groups thus obtained are given below:
1 . Machine group 1 : (1 ). .. . singleton cluster
2. Machine group 2 : (4,2,3 ) .
3. Machine group 3 : (6,s ) .
The above machine groups obtained using natural seeds (namely
machine 1, machine 4 and machine 6) will be used to cnate sood
points for ideal mad clustering hter. A similar procedure is performed
for the components using the transpose of the matrix in Figure 3.1.
The transpose of Figure 3.1 is given below in Figure 3.2.
Machine k
Camponrnt i
Figure 3.2 Transpose of Figure 3.1 (component-machine incidence
matrix)
The Jaccard similarity matrix corresponding to Figure 3.2 is
shown in Table 3.3. The average of the Jaccard similarity coefficient
for components from Table 3.3 is found to be 0.1656. Since the matrix
[Table 3.3) i s symmetric, only the upper triangular values are
considered. The matrix (Table 3.3) i s scanned to obtain a pair of
components whose similarity is less than the average. Component 1
and component 6 have a similarity value of zero and hence these
components are chosen a s the first two component seed points. Since
the remaining components' (namely components 2,3,4,5) Jaccard
Table 3.3 Jaccard similarity matrix of components arrived from
Figure 3.2
Component j
Component i
similarity values with the existing component seeds (namely
component 1 and component 2) are more than the average Jaccard
similarity value of 0.1656 (of Table 3.3), they do not qualify to be a
seed. The components are clustered around these seed points. The
two component seed points are given below:
1. S e e d l : [ 1 0 0 0 1 0 ) .....( component 1).
2. Seed 2: [ 0 1 0 1 0 0 ] .....( component 61.
Each component vector (component row) is assigned to the
component seed with which the distance (dissimilarity measure given
in equation 3.3) i s minimum. The component groups thus obtained
are given below:
1. Component group 1 : ( 1,2,3,4 ) .
2. Component group 2 : ( 6,5 ).
Intermediate solutions from the ZODIAC algorithm may yield unequal
numbers of machine groups and part groups. In such cases the
solutions are not evaluated. The number of seed points for machines
and parts are made equal by eliminating (discarding) a few seed points
(corresponding to small or singleton groups) and then the iterations
are con tinu ed .
In this case the number of machine groups and part groups
obtained using natural seed clustering are unequal. Hence this solution
is not evaluated. The number of seed points for machines and parts
are made equal by eliminating (discarding) singleton machine group 1
with machine 1 a s its constituent.
Hence the following machine groups and part groups are
considered a s output of natural seed clustering and the iterations
continued.
Machine groups:
1. Machine group 1 : ( 4 ,2 ,3 ) .
2. Machinegroup 2 : (6,s).
. . .(machine 1 discarded)
Component groups:
1. Component group 1 : ( 1,2,3,4 ) .
2. Component group 2 : ( 6,s ).
This solution i s improved using the ideal seed algorithm. Two
Component seeds are obtained from the two component groups. They
are given below:
1. Componentseedl: [ I 1 1 1001.
2. Component seed 2: [ 0 0 0 0 1 1 1. component seed 1 has been created from part family 1 (component
group 1) and has 1's in locations (1,2,3 and 4) representing the parts
grouped in part family 1. The second component seed is also created
similarly. Each component seed is a vector with 6 elements. The
machine vectors (machine rows from Figure 3.1) are clustered using
these component seed points and using the distance measure given
in equation 3.3. The machine groups thus clustered are given below:
1. Machine group 1 : ( 5,6 1.
2. Machine group 2 : ( 1,2,3,4 ).
From machine-component groups obtained using natural seeds
earlier, machine seed points are created to cluster components. The
machine group (4,2,3) yields machine seed 1 with a 1 in positions 4,2
and 3. The machine seeds thus obtained from two machine groups are
given below:
1. Machineseed 1: [ 0 1 1 l o o ] . 2. Machine seed 2: 10 0 0 0 1 1 1.
Each component is attached to a machine seed with which its
distance is minimum. The component vectors (component rows from
Figure 3.2) are clustered using these machine seed points and using
the distance measure given in equation 3.3. The component groups
thus clustered are given below:
1 . Component group 1 : ( 2,5,6 1.
2. Component group 2 : ( 1,3,4 1.
Finally, part groups are assigned to machine groups using a
procedure called diagondization. Each family can be assigned to one
machine cell. The number of 1's is computed for each of the 4 possible
pairs of machine cells and part families. The combinations with the
maximum number of 1's are paired together. This procedure is
continued until each machine cell has a part family assigned to it. In
the present example, the assignments are given below:
1. Machine group I assigned to part family 2 (component
group 2).
2 . Machine group 2 assigned to part family 1 (component
group 1 ) .
The results are summarised in Table 3.4. The block diagonalised
machine-component incidence matrix is given in Figure 3.3.
Table 3.4 Intermediate machine-component cells
The solution shown in Figure 3.3 h a s 1 intercell move. The
goodness of the solution is given by a measure called grouping efficiency
(Chandrasekharan and Rajagopalan 1987). Therefore grouping
efficiency corresponding to the present solution shown in Figure 3.3
is equal to 77.77%. Since our objective is to find a solution with
maximum grouping efficiency, we continue the ideal seeding procedure.
r--- - Machine-component cell no
1
2
It is found out whether the solution can be improved using
ideal seed algorithm. Two component seeds are obtained from the two
Mach~ne group
5,6
],2,3,4 --
- - I
Part family
1,3,4
2,5,6
Component k
Machine i
Figure 3.3 Block-diagonalised machine-component incidence matrix
component groups shown in Table 3.4. They are given below:
1. Componentseedl: { I 0 I 1 0 0 1 .
2. Componentseed2: [ O l 0 0 1 I ] .
Component seed 1 has been created from part family 1 (component
group 1) and has 1's in locations ( 1 , 3 and 4 ) representing the parts
grouped in part family 1. The second component seed is also created
similarly. Each component seed is a vector with 6 elements. The
machine vectors (machine rows from Figure 3.1) are clustered using
these component seed points and using the distance measure given
in equation 3.3. The machine groups thus clustered are given below:
1. Machine group 1 : ( 5,6 ).
2. Machine group 2 : ( 1,2,3,4 ) .
From the machine-component groups obtained using ideal seed
clustering earlier (shown in Table 3.41, machine seed points are created
to cluster components. The machine group ( 5,6 ) yields machine
seed 1 with a 1 in positions 5 and 6. The machine seeds thus obtained
from two machine groups are given below:
1. Machineseed 1: [ 0 0 0 0 1 1 1 .
2. Machine seed 2: [ 1 1 1 1 0 0 1.
Each component is attached to a machine seed with which its
distance is minimum. The component vectors (component rows from
Figure 3.2) are clustered using these machine seed points and using
the distance measure given in equation 3.3. The component groups
thus clustered are given below:
1. Component group l : ( 1,3,4 1.
2. Component group 2 : ( 2,5,6 1.
Finally, part groups are assigned to machine groups using the
procedure called diagonalisation, which was explained earlier. In the
present example (iteration), the assignments are given below:
1 . Machine group 1 assigned to part family 1 (component
group 1 ) .
2. Machine group 2 assigned to part family 2 (component
group 2).
The results are summarised in Table 3.5.
Table 3.5 Intermediate machine-component cells
.
- - - - - - - -- -
Part fam~ly
1,3,4
2,5,6
r-- Machine-component cell number
I I
I 1
I 2
1
- - -- - ,
Mach~ne group
5 6
1,2,3,4
This solution is same a s that of the solution obtained in the
previous iteration with one intercell move and a grouping efficiency of
77.77%. Finding that further iterations do not improve the grouping
efficiency, the ZODIAC algorithm stops here. intermediate solutions
from the ZODIAC algorithm may yield unequal numbers of machine
groups and component groups. In such cases the solutions are not
evaluated. In such situations the number of seed points for machines
and parts are made equal by eliminating (discarding) a few seed points
(corresponding to small or singleton groups) and the iterations
continued. The machine-component incidence matrix corresponding
to the final solution obtained using the ZODIAC algorithm is shown in
Figure 3.4. Since the objective is to maximize grouping efficiency,
intercell moves are not considered for choosing the best solution among
the solutions generated by ZODIAC.
Component k
Machine i
Figure 3.4 Final solution obtained using ZODIAC for the sample problem
(Grouping efficiency = 77.77% and
Number of exceptional elements ;e 1)
3.3.3 Introduction to (IRAFICS Algorithm
GRAFICS (Srinivasan and Narendran 1991) is a nonhierarchical
clustering algorithm in the area of cellular manufacturing systems.
GRAFICS i s a n acronym for 'Grouping using Assignment method
For Initial Cluster Seeds". GRAFlCS algorithm is given in Appendix I
along with the corresponding flowchart (Figure 1 A.2 ) .
GRAFICS h a s two phases. In the first phase, the machine similarity
matrix i s given a s input to assignment method (Hungarian method
(Budnick e t al. 1988)) to generate a n initial set of machine groups
such that the cumulative sum of the similarity values between the
machines within the initial machine groups is maximized. Subtours
(Bellmore and Nemhauser 1968) are identified from the assignment
solution and are used to determine the initial set of machine seeds to
cluster components.
SEED:
Consider the problem given in Figure 3.5. Let the solution
for :his problem using the assignment method [Hungarian
method (Budnick et al. 1988)l be a s follows: X,, = X,, =
X,, = X,, = X,, = X , , = 1 , where X,, is equal to 1 if the
machine i is grouped with the machine j ; it is equal to 0,
otherwise . From t h i s solut ion, we get two sub tour s .
1-2-5- 1 a n d 3 -4 -6 -3 . The machines in each subtour are
treated a s a machine group. The two machine groups are
a s follows: M , = (1,2,5) and M, = (3,4,6). These two
machine groups are used to generate the following two machine
seeds which are in turn used to cluster components in the next
stage: S, = (110010) and S,= (001101).Eachseedconsistsof
entries 1 or 0. In a given machine seed, the entry positions
represent machine numbers in ascending order. The e n t n
1 represents the presence of the corresponding machine
in that seed and the entry 0 represents the absence of
the corresponding machine in that seed. The formulation
of the above seeds namely S, and S, are explained in
Table 3.6 and Table 3 . 7 respectively.
(To improve readability, Figure 1.2 is reproduced below)
Component j
Machine vector corresponding to
machine i where i-1 : (1010)
0
Machine i 3 0 O 111
5 1 0 1 Componrnt vector
6 0 component J whrre 1-4 . (001101)
Figure 3.5 Machine-component incidence matrix (contains the d
shown in Figure 1.2)
The second phase of GRAFlCS is the formation of the machine-
component cells. In this phase, clustering is done based on maximum
density rule (Srinivaaan and Narendran 1991). The maximum density
rule is given below:
Maximum Density Rule:
Assign a component vector (machine vector) from the
machine-component incidence matrix to the machine seed
Table 3.6 Tabular explanations for the formation of machine seed,
8, : (1 10010)
Machine type, i
1
2
3
4
5
6
Whether a particular machine type is included in the machine group, M , (YES or NO)
YES
YES
NO
NO
YES
NO
The cornaponding machine seed, 8,
1
1
0
0
1
0
Table 3.7 Tabular explanation for the formation of machine seed,
5, : (001101)
r- - ---- I
Mach~ne type, I
I
1
1 2
3
4 1
I 5
6
L- -
- --- --
Wherhrr a part~cular mach~nr typc 1s ~ncludrd In thr machinr group. M, (YES or NO)
NO
NO
YES
YES
N 0
YES
-- -- --
----- _-
Thr corrrspand~ng machtnr srrd. 8,
0
0
1
1
0
1
- -- - - -- -
(component seed) with which it has the maximum common
1's and break ties by assigning to the machine seed
(component seed) which has the smallest number of 1's
among the contending machine seeds (component seeds)
[Component vector and machine vector are pictorially
represented in Figure 3.51.
Given the set of machine seeds, component families are identified
such that each component is assigned to a machine seed which has
the maximum number of machines required by that component. Then
the component seeds are constructed from the component families,
Again, the machine cells are identified such that each machine is
assigned to the component seed which requires ~t the most. I f we start
the above process with a good choice of initial set of machine seeds
and alternatively form component seeds and machine seeds, it will
finally give a feasible set of machine-component cells. Singleton cluster
is eliminated by assigning it to the seed which has a t least one member
(machine or component) clustered around it. After updating this
solution a s the latest feasible solution, the procedure is continued
until 1) the number of exceptional elements reaches zero (E = 0), or 2)
the number of exceptional elements ceases to decrease (E > Eo), or 3)
two consecutive feasible solutions are identical (E, = E and Bo = B).
3.3.3.1 Rumsrid Exunple for QRAFICS
In this section, GRAFICS algorithm is demonstrated using a
problem [Vohra et al. 1990) from the GT literature. The initial 0-1
machine-component incidence matrix of this problem is shown in
Figure 3.6 (The process sequences of the above problem are not
reported by the authors). The machine similarity matrix of this
problem is shown in Table 3.8. The similarity values shown in
Table 3 . 8 a re computed using the method followed by Kusiak
(1987). Using the machine similarity matrix in Table 3.8 a s input ,
assignment method [Hungarian method (Budnick et al. (1988)) 1
h a s genera ted 2 machine cells: ( 1 , 2 . 4 . 6 , 7 ) a n d ( 3 , 5 ) . The
corresponding initial machine seeds to cluster components based on
these machine cells are shown in Table 3.9.
Component j
Figure 3.6 Initial machine-component incidence matrix of the example
problem (Vohra et al. 1990)
Table 3.8 Machine similarity matrix of the example problem (Vohra
et al . 1990)
Machine j
Machine i
Note: The similarity values shown in Table 3 . 8 are computed using the method followed by Kusiak (1987). The formula h a s been already given in page 37 of Chapter 1.
Table 3.9 Initial machine seeds to cluster components of example
problem (Vohra et al. 1990) [output of Hungarian method)
Mach~ne seed nurnbcr
Machine
The next step of GRAFlCS uses the data in Table 3.9 (initial set of
machine seeds) and cluster components based on maximum density rule.
Application of lYIxfmum Density Rule:
In this paragraph the concept of maximum density rule is
demonstrated using a component vector's assignment to a
machine seed. Let u s consider a component vector, C,
corresponding to component type 1 from Figure 3.6 (initial
machine-component incidence matrix):
Component vector, C,: ( 1 1 0 0 0 1 0 )
This component vector, C, has to be assigned to any one of the two
machine seeds shown in Table 3.9:
Machine seed 1 : ( 1 1 0 1 0 1 1 )
Machine seed 2 : ( 0 0 1 0 1 0 0 )
Common 1's between C, and machine seed 1 = 3
Common 1's between C, and machine seed 2 = 0
Hence cornponent vector C, is assigned to machine seed 1 . i.e.
component of type, 1 is assigned to machine group 1 : ( 1 , 2 , 4,6,
7 ). Similarly all the remaining components with identification 2 ,
3, 4 , 5, 6 and 7 are assigned to the initial machine groups based
on maximum density rule.
These component families are shown in Table 3.10. Since the number
of machine cells generated by Hungarian method and the number of
Component families shown in Table 3.10 are equal, a feasible set of machine
component cells is formed. The corresponding grouping efficiency and
grouping efficacy are calculated. The value of grouping efficacy is equal to
53.33%. The value of grouping efficiency is equal to 75.09%. The
corresponding machine-component cells are shown in Table 3.1 1.
Table 3.10 Component families for the example problem (Vohra et
al. 1990) using GRAFlCS in Iteration 1
Table 3.11 Intermediate machine-component cells for example
problem (Vohra et al. 1990) using GRAFlCS
- Component family number
1
2 I
- Components
1,2,3.4.7 1 5,6 I
I 2
Itemtion 2:
The component seeds corresponding to the component families
in Table 3.10 are shown in Table 3.12. These component seeds are
used to generate new machine cells based on maximum density rule.
The new machine cells are shown in Table 3.13.
151
? Component 1
family
1,2,3,4,7
5,6 I
- - - - - - Machine-component
I cell number
1
1
2
, E = 1
B = 13
Grouping efficacy = 53 33'10
Grouplng efficiency = 75 09% ~ -
~
-
Machine cell
1,2,4,6,7
3 ,5
Table 3.12 Component seeds to cluster machines of example problem
(Vohra et al . 1990)
Component seed nurnbrr
Componrnt J
Table 3.13 Machine cells for example problem (Vohra et al. 1990)
using GRAFICS in Iteration 2
Since the number of machine cells in Table 3.13 and the numbrr
of component families in Table 3.10 are equal, a feasible set of machine-
component cells is formed. The corresponding grouping efficiency
and grouping efficacy are calculated. The value of grouping efficacy is
equal to 53.33%. The value of grouping efficiency is equal to 75.07%.
Finally, GRAFICS procedure stops since the number of exceptional
elements in the current feasible solution is same a s the number of
exceptional elements in the previous feasible solution and the number
. - -
I Machlne cell I number I
I I
1 I
I 2 -
---
Mach~nes
I
1,2,4.6,7 1
of voids in the current feasible solution is same a s the number of voids
in the previous feasible solution. The final machine-component cells
and the corresponding block diagonalised machine-component
incidence matrix are shown in Table 3.14 and Figure 3.7 respectively.
Table 3.14 Final machine-component cells for example problem
(Vohra et al. 1990) using GRAFlCS
Component j
1 --Machme-component 1 cell number
1 I 2
Machine i
Figure 3.7 Final block diagonal form of example problem (Vohra et al. 1990) using GRAFICS
- - --
Machlne cell
1,2,4,6,7
( E = 1, B = 13, Grouping efficacy = 53.33% and Grouping efficiency = 75.09% )
- -
Component family
1.2,3,4,7
C 3,5 5 6
1 E = l B = 13 1 Group~ng efficiency = 75 09%
1 Grouping efficacy = 53 33% 1 ---
3.3.4 Comp8riron of QRAFIC8 with other well-knom algorithm.
In th i s section, it is shown that GRAFICS (Srinivasan and
Narendran 1991) is better than many well-known algorithms by
sequential review of the GT literature. Miltenburg and Zhang (1991)
have compared nine well-known algorithms. The details of these
algorithms are given in Table 3.15.
They have concluded that the performance of Ideal Seed Non-
Hierarchical Cluster ing Algorithm (ISNC) developed by
Chandrasekharan and Rajagopalan (1986a) is relatively better than
the rest of the eight algorithms. But, ZODIAC (Chandrasekharan and
Rajagopalan 1987) is an improved version of ISNC.
Kandiller (1994) selected a subset of six well-known cell formation
techniques for a detailed analysis. The techniques selected for analysis
and comparison are:
1. Lattice-theoretic combinatorial grouping (COMBGR)
developed by Purcheck ( 1974).
2. Modified rank order clustering (MODROC) developed by
Chandrasekharan and Rajagopalan ( 1986b).
3. Machine-component cell formation (MACE) developed by
Waghodekar and Sahu (1984).
4. within-cell utilization based clustering (WUBC) developed
by Ballakur and Steudel (1987).
5. Cost analysis algorithm (CAA) developed by Kusiak and
Chow (1987).
6. Zero-one data: ideal seed algorithm for clustering
(ZODIAC) developed by Chandrasekharan a n d
Rajagopalan (1987).
Table 3.15 Details of well-known algorithms
S.No. Underlying Procrdurr Algorithm' Rrmarks
I Rank Order Clustering ROC/ROC Machines a s well a s components are clustrrcd uslna ROC.
2 Similarity Corfficirnt SC/ROC Machines a r r clustcrrd using SC and componcnts a r r clustcrrd using ROC.
3 Similarity Coemcicnt SC/SC Both machines and components arc clustrrrd using SC.
4 Modified Similaritv MSCIHOC Machines a r r clustrrrd using MSC and components arc clustrrrd using ROC.
5 Modified Similarity MSC/MSC Ih th machinrs and componrnts wrr cluslrrrd using MSC.
0 Mod~hrd Rank Order MROC Clustrrlng
7 Srrd Clust r r~ng ISNC
X Srrd Clust r r~ng SC-Srrd
9 Ilond E n r r u REA
'ROC - Rank Order Clustering Algor~thm (King 1980).
SC - Similarity Coefficient Algnrlthm (Slnglr L~nkagr Clustrring Algorithm (McAulry 1972)).
MSC - Modifird Similarity Coemcicnt Algorithm (Avrragr I.inkagr Clustrrlng Algor~thm (Andrrberg 19731)
MROC - Modified Rank Order Clust r r~ng Algorithm (Chandraarkhari~n and Rajagopalan 19AObl.
1SNC - Ideal Seed Non-Hierarchical Clustrring Algorithm (Chendraarkharan and Rajagopalan 1986a).
SC-Seed - Modified ISNC (Milttnhurg and Zhang 1991)
BEA - Bond E n e r ~ y Clustering Algorithm (McCormick rt .al 1972)
Kandiller (1994) camed out an extensive study of the six prominent
algorithms. Kandiller (1994) reported that ZODIAC (Chandrasekharan
and Rajagopalan 1987) is one of the best well-known algorithms.
Srinivasan and Narendran (1991) have S ~ O H ~ that the performance
of GRAFICS is better than ZODIAC. Hence, in thls research work,
GRAFICS is considered for further improvement.
3.3.5 Comparison of ZODUC and GEUFICS
In this section, ZODIAC [Chandrasekharan and Rajagopalan 1987)
and GRAFICS (Srinivasan and Narendran 1991) are compared and their
similarities and dissimilarities are reported in Table 3.16.
Table 3.16 Comparison of ZODIAC (Chandrasekharan and Rajagopalan 1987) and GRAFICS (Srinivasan and Narendran 1991)
- - - - - ---
ZODIAC
1 A nonhlerarchlcal clusterlng
algorithm
2 Input da t a 1s 0-1 machlne-
component ~nc~dence matnx
3 Jaccard s ~ m ~ l a r ~ t y measure IS
used to create inlt~al machlnt.
seed polnts around which the
machlnes are to be grouped
4 lnlt lal machine groups a re
formed uslng the ~ n ~ t ~ a l machlne
seed polnts and a dtsslmllanty
I measure
CRAFICS
1 A nonhlerarchlcal clusterlng
algonthm I
2 lnput data I S 0 -1 m a c h ~ n c -
component lncldence matnx I
3 Slmllarlty measure used by
Kuslak (1987) IS used to
generate ~ n ~ t i a l set of machlne
groups
4 In~bal set of machlne groupsare
formed by glvlng machlne
slm~larlty matrlx a s Input to
assignment method [Hunganan
method (Budnick et a1 1988) ]
Table 3.16 Comparison of ZODIAC (Chandrazrkhamn and Rajagopalan 1987)
and GRAFICS (Srinivasan and Narmdran 1991) (continued)
used to create initial component I formed using the initial set of
-
seed points around which the ( machine seeds generated
ZODIAC
5. Jaccard similarity measure is
components are to be grouped.
- GRAFICS
5. lnitial component groups are
us ing ass ignment method
[Hungarian method (Budnick
et al. 1988)].
component seed points and "a
dissimilarity measure".
6. Initial component groups are
formed us ing the initial
7. No objective function is used in
the formation of either initial
machine groups or initial
component groups which are
formed separately.
6 , Initial component groups are
formed using initial se t of
machine seeds and "maximum
density rule".
7, lnitial set of machine groups
are formed by giving machine
similarity matrix a s input to
assignment method
[Hungarian method (Budnick
et al. 1988)] with the objective
of maximizing the cumulative
sum of the similarity value8
between the machines within
the init ial se t of machine
groups.
8. Machine groups and component
groups are formed separately.
Subsequently, each part family
(component groups) is assigned
to a machine group based on a
procedure called diagonalisation.
8 . Machine groups and
component groups are formed
concurrently by using
"maximum density rule".
Table 3.16 Comparison of ZODlAC (ChandraseWlaran and Rajagopalan 1987)
and GRAFICS (Srinivasan and Narmdran 1991) (continued)
7 -
ZODIAC I
9 , The initial solution obtained
using natural seed clustering is I
improved us ing ideal seed
clustering. In ideal seed I
clustering machines are
clustered using component seed
points a n d also using a
I dissimilarity measure (a distance
measure). Similarly components
; are clustered using machine seed
points and a lso using a
; dissimilarity measure (a distance
measure). I 1
10.The goodness of the solution
(machine-component cell
formation) is evaluated using a
performance measure namely
grouping efficiency
(Chandrasekha ran and
Rajagopalan 1987).
!
!
i
--
GRAFICS
9. Initial set of machine groups
obtained us ing assignment
method is used subsequently to
form component groups using
"maximum density rule". Then
machine groups and component
groups are formed alternatively
using "maximum density rule".
I
10.The goodness of the solution
(machine-component cell I formation) is evaluated primarily
using a performance measure 1 namely grouping efficacy
(Kumar and Chandrasekharan I 1
1990) and secondarily using a 1
performance measure namely
grouping efficiency
(Chandrasekharan and 1 Rajagopalan 1987).
Table 3.16 Cornpariaon of ZODIAC (Chandrasekharan and Rajagopalan 1987) and GRAFlCS (Srinivasan and Narendran 1991) (continued)
h
I ZODIAC
11 .When further iterations do not
improve grouping efficiency,
ZODIAC algorithm stops.
I
I ~ I 1
! I
i 12.The choice of the initial number
of seeds and seed points needs
careful consideration. I
Sometimes, natural seeds which
) are generated based on the
/ characteristics of the data set
can result in fewer seeds than
required, thereby increasing the
i b lanks within the diagcnal
blocks considerably.
I I
I
GRAFICS
I I .GRAF!CS algorithm stops while
any one of the following
conditions occun:
(a) the number of exceptional
elements reaches zero (E=O)
(or1
(b) the number of exceptional
elements ceases to decrease
(E'E,)
(or)
(c) two consecutive feasible
solutions are identical (E,=E and
B,=B).
I 12.If there exists multiple solutions
for a given assignment problem
[solved using Hungarian method
(Budnick et al. 1988) 1, though
there is a possibility of obtaining
reasonably large number of
subtours (i.e. reasonably large
number of machine seeds a s the
initial set of machine seeds)
which will lead to a meaningful
final solution (final machine-
component cells) , t he
assignment method (Hungarian
Table 3.16 omp par is on of ZODlAC (Chandrasekharan and Rajagopalan 1987)
and CX4FlCS (Srinivasan and Narmdran 1991) (continued)
1- ---
ZODIAC 4
1
I
I
13 The clustenng cntenon based on
" m ~ n i m u m value of d ~ s t a n c e
measure" does not reflect the
extent of processlng requrred bv
components Cons~der two seeds
S,=(llOOOOOOOO) and
S,=(0000011111) and a vector
V=(0000110000) The d~stance
1 measure between V and S, 1s 4
and between V and S2 1s 5 I f we
cluster based on the "rn~nrmurn
value of d~s tance measure", V
should be clustered to S, and not
a S, though they do not have a
' 1 ' In common The above
s l tuat lon often a r l se s whlle
clustenng ~llstructured matnces
This is a very important
h a c k of ZODIAC algorithm. I
-
GRAFICS
method) may glve a slngle
machrne seed a s the lnrhal set
ofmach~ne seeds or a very lrmrted
number of machrne seeds a s the
~ n ~ t ~ a l set of machrne seeds whlch
are undesirable
13 Thr clustenng cntenon based on
'max~rnurn dens~ty rule" does
reflect thr rxtent of processlng
requirrd by components
able 3.16 Comparison of ZODIAC ( C h a n d r a a e m and Rajagopalan 1987)
and GRARCS (Srin~asan and Narmdran 1991) (continued)
3.3.6 Meed for improving GRAFICS Algorithm
,-
ZODIAC
14. No constraint on machine group I
: size or part family size. I
i15.No cons t r a in t on number of ! I machine groups or number of
part families.
Limitation of Using Assignment Method for Generating Initial
Set of Machine Seeds:
GRAFICS
14. No constraint on machine group
size or part family size.
15.No const ra in t on number of
machine groups or number of
part families. ...~..
If there exists multiple solutions for a given assignment
problem, there i s a possibility of getting anv one of the
following three types of assignment solution:
1. A solution which consists of a single tour.
2. A solution which consists of a very limited number
of subtours.
3. A solution which consists of a reasonably large
number of subtours.
Among these alternate solutions, if the assignment method
gives any one of the first two types of solution (type 1 or
type 2), we will end up with a single machine aeed or a
very limited number of machine seeds. In the above
s i tuat ion, though there i s a posaibility of obta ining
reasonably large number of subtoura (type 3) which will
lead to a meaningful final so lu t ion (final mach ine -
component cells) in the next phase, the assignment method
may give a single machine seed or a very limited number
of machine seeds. This i s the major limitation of using the
assignment method to generate the initial set of machine
seeds.
To overcome the above limitation of the muignment method,
an Efficient Seed Generation Algorithm (ESGA) is proposed in this
research work.
Limitation of Phase 2 of GRAFICS:
Consider a s i tua t ion in which t h r r e r x i s t s a s ing l r
component with processing r equ i r rmrn t s on a set of
machines ( a machine group) and also that set of machines
i s not required by any other component. In GRAFICS, this
type of component (single component cluster or singleton
cluster) i s assigned to another machine seed (machine
group) with which that component h a s zero similarity.
Consider another situation in which a single machine
which will fully satisfy the processing requirements of a
set of components ( a component family) and also that
machine i s not required by any other component . In
GRAFICS, th i s type of single machine (single machine
cluster or singleton cluster) will be assigned to another
component seed (component family) with which it has zero
similarity. To overcome the above problem, singleton
clusters (i.e. single component cluster or single machine
cluster) should be allowed in the machine-component cells
which will improve the final solution. Hence, in th i s
r e s e a r c h work , c o r r e s p o n d i n g mod i f i ca t ions a r e
incorporated in GRAFICS to allow singleton clusters in
the solution if necessary.
To h a n d l e t h e above s i t u a t i o n s , two a lgo r i thms , namely
ALGORITHM 1 and ALGORITHM 2 are proposed in the first part of
this research work. The initial set of machine seeds obtained from the
seed generation algorithm namely ESGA is used in these algorithms.
Later, in the second part of this research work, an improved version of
the ALGORITHM 2 namely SA ALGORITHM is proposed.
3.4 SUMMARY
In th is chapter, detailed explanations have been given about two
n o n h i e r a r c h i c a l c l u s t e r i n g a lgo r i thms namely ZODIAC
(Chandrasekharan and Rajagopalan 1987) and GRAFlCS (Srinivasan
and Narendrarl 1991). The similarities and dissimilarities of these
algorithms have been reported. The drawbacks of GRAFlCS have been
highlighted. Also, it h a s been shown that GRAFlCS is better than many
well-known algorithms by sequential review of the published GT
literature.