You are on page 1of 20

A genetic algorithm approach to cellular manufacturing systems

G.C. Onwubolu
a,
*
, M. Mutingi
b
a
Department of Technology, The University of the South Pacic, P.O. Box 1168, Suva, Fiji
b
Olivine Industries Ltd., Birmingham/Plymouth Road, Southerton, P.O. Box 779, Harare, Zimbabwe
Abstract
A genetic algorithm (GA) metaheuristic-based cell formation procedure is presented in this paper. The cell
formation problem solved here is to simultaneously group machines and part-families into cells so that inter-
cellular movements are minimized. An option for considering the minimization of cell load variation is included
and another, which combines minimization of intercellular movements and cell load-variation, exists. The algo-
rithm solves this problem through improving a cell conguration using the GA metaheuristic. The designer is
allowed to specify the number of cells required a priori and impose lower and upper bounds on cell size. This
makes the GA scheme exible for solving the cell formation problems. The solution procedure was found to
perform well on tested large-scale problems and published data sets. Moreover, the proposed procedure compares
very favorably to a well-known algorithm, and another TSP-based heuristic available in the literature. The results
of computational tests presented are very encouraging. q 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Group technology; Cellular manufacturing; Metaheuristics; Genetic algorithm
1. Introduction
The production planning process involves transforming inputs using analytical and logical models
into expected outputs. These inputs typically include the number of units of each product to be produced
by each of the available alternative processes, target inventory levels for each products, quantities of
materials to be transported within and among cells, workforce levels, overtime. Others, include addi-
tional shifts and unused capacity, subcontracting plans, purchased material requirements, order release,
job schedule and sequence. The manufacturing system conguration has a direct impact on the produc-
tion planning activity for each planning period. For example, the ability of a manufacturing rm to meet
its target of the number of units of each product to be produced as well as considering available
alternative processes, is directly affected by the manufacturing system in place. Cellular manufacturing,
Computers & Industrial Engineering 39 (2001) 125144
0360-8352/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved.
PII: S0360- 8352( 00) 00074- 7
www.elsevier.com/locate/dsw
* Corresponding author.
which relies on grouping of machines that can produce families of similar parts in cells, increases
production effectiveness through performing similar activities together, and standardizing similar
tasks (Hyer & Wemmerlov, 1984). A number of benets arise from adopting cellular manufacturing,
such as: reduced inventory, reduced capacity, reduced labor and overtime costs, shorter manufacturing
lead times, faster responsiveness to internal and external changes such as machine and other equipment
failures, product mix and demand changes (Wemmerlov & Hyer, 1989). Production planning is there-
fore a core activity that realizes the benets of manufacturing cells. There are three different GT cell
formation models, namely standard, generalized and metaheuristic model.
Standard cell formation models ignore many manufacturing factors and only consider machining
operations of parts, so that a manufacturing system is presented by a binary machine-part incidence
matrix A, with
a
mn

1 if part n is processed on machine m
0 otherwise
_
Many methods of cell formation have been developed and published (for an extensive review of these,
see Wemmerlov and Hyer (1986)). Matrix formulation is one of the popular approaches which has been
employed for example, by Chan and Milner (1982), King (1980), King and Nakornchai (1982) and
McCormick, Schweitzer and White (1972). These methods, however ignore major practical issues such
as the sequence of machine operations for each part type and the size of the cells. Mathematical
programming models have been developed for cell formation by Askin and Subramaniam (1987) and
Shafer and Rogers (1991). Most of these models can incorporate more complicated features of the
problem but often become computationally intractable for large problems. In spite of their shortcomings,
mathematical programming models for cell formation allows researchers to compare the solution quality
of heuristic methods, see for example, Wei and Gaither (1990). Kumar, Kusiak and Vannelli (1986),
Rajagopalan and Betra (1975) and Vannelli and Kumar (1986) have employed graph-decomposition
procedures for the machine-part cell formation problem. All these methods are not suitable because they
cannot cater for practical problems which have large number of parts to be classied into families, and
require exibility of specifying the number of cells and cell size limitations a priori (Hindi & Hamam,
1994).
The cell formation problem is a combinatorial optimization problem that is NP-hard. The optimiza-
tion algorithms yield a globally optimal solution in a possibly prohibitive computation time. Most of the
approaches presented in the previous sections are heuristic-based tailored algorithms for solving specic
part family and machine group problems. None of these approaches guarantee near optimal solutions.
Metaheuristics have emerged to solve combinatorial optimization problems with global or near-global
optimal solution in a reasonable computation time. Metaheuristics have been recently used to solve
combinatorial optimization problems, of which the cell formation problem is typical. Since the cell
formation problem is a difcult combinatorial optimization problem, recent search metaheuristics,
which are capable of solving practical problems have been developed such as simulated annealing
(Boctor, 1991; Harhalakis, Nagi & Proth, 1990), and neural networks (Kaparthi & Suresh, 1992;
Moon, 1990).
Venugopal and Narendran (1992) modeled the cell formation problem based on minimization of the
total cell load variation using genetic algorithms (GAs). Cheng, Gupta, Lee and Wong (1998) formulated
the cell formation problem as a traveling salesman problem (TSP) and a solution method based on the
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 126
GA software, genitor developed by Whitley (1989) was proposed to solve the TSP-cell formation
problem.
In practice, it is useful to consider multi-objectives when considering the cell formation problem.
There is limited literature in this area. In their study, Hsu and Su (1998) proposed a GA-based procedure
to solve the machine cell formation problem. More specically, the study minimizes: (i) total cost,
which includes intercell and intracell part transportation costs and machines investment costs; (ii)
intracell machine loading imbalances; and (iii) intercell machine loading imbalance under many realistic
considerations. More recently, Onwubolu and Songore (2000) presented a tabu search based procedure
for solving the cell formation problem, which is modeled with three options of minimizing objective
functions: (i) intercellular movements; (ii) cell load variation; and (iii) a combination of cell load
variation and intercellular movements. An interesting article that critically reviews the methods of
cell formation problem to include multi-objectives cluster analysis method, is that by Sarker and Xu
(1998). Up to six multi-objectives were presented in their article, which may be useful to consider when
dealing with multi-objectives problems.
This paper presents another metaheuristic called GA for the cell formation problem. Different objec-
tive functions are used depending on the requirement of the user. In our GA implementation, three
options were included: (i) minimization of the number of intercellular moves due to exceptional parts;
(ii) minimizing intracell work load imbalances; and (iii) multi-objective function based on the two
previous options. The algorithm takes into account lower and upper bounds on the number of machines
in a cell, and user-specied number of cells. The performance of the GA-procedure is measured by
comparing its results with those published for small problem sizes. Large-scale size problems are solved
very efciently in terms of reasonable computation time and performance measures such as grouping
efciency, clustering measure and bond energy measure. For this class of problems, non-metaheuristics
are normally confounded because they may spend prohibitive computation time to nd solutions, which
have very low quality when compared with the optimum solutions using performance measures already
mentioned. Most published literature on GA does not highlight the practical problems encountered
during implementation. This paper lls the gap.
2. Problem formulation
Cell formation is an assignment-type problem and can be summarized as follows: given n items and m
resources, the problem is to determine an assignment of the items to the resources optimizing an
objective function and satisfying a stated set of additional side constraints. An objective function
provides the basis for evaluating the machine groupings arrived at by a searching method (in this
case the GA method). In the present work, three problem formulation (objective function) models
were implemented. The rst is, based on the minimization of cell load variation. The second objective
function is based on the minimization of exceptional elements. The third is a combination of the previous
two.
2.1. Minimization of cell load variation model (model 1)
This model is based on the model proposed by Venugopal and Narendran (1992). In this model we
dene m as the number of machines, k as the number of cell, and n as the number of parts.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 127
W w
ij
is an m p n machine component incidence matrix where w
ij
is the workload on machine i
induced by part j
X x
il
is an (m p k) cell membership matrix
where
x
il

1 if machine is in cell l
0 otherwise
_
M m
lj
is a (k p n) matrix of average cell load, where
m
jl

m
i1
x
il
w
lj

m
i1
x
il
The total load of cell l induced by part j is given as:

m
i1
x
il
w
lj
The number of machines in cell l is given as:

m
i1
x
il
The mathematical programming formulation of the grouping problem is as follows:
minimize z
1

m
i1

k
l1

n
j1
w
ij
2m
lj

2
1
subject to

k
l1
x
il
1 ;
i
2

m
i1
x
il
$ 1 ;
l
3
The expression

n
j1
w
ij
2m
lj

2
is the objective function. Eq. (1) gives the extent of variation of the
load on machine i in cell l (induced by all parts) from the mean load of cell l. The objective function z
1
adds this quantity for all the machines and cells. Hence this formulation requires a solution for which the
total cell load is minimized such that every machine belongs to exactly one cell and no cell is empty.
Eq. (2) ensures that for particular i, machine i is assigned to one cell only. Eq. (3) ensures that no
cell is empty.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 128
2.2. Model illustration
As an illustration, consider a 3 3 problem whose workload incidence matrix is given as:
w
ij

0:3 0:0 0:2
0:0 0:9 0:0
0:5 0:0 0:5

_
_
_

_
_
_
where rows represent machines and columns represent parts. Suppose the machine grouping arrived at
by the searching algorithm is:
Machine 1 2 3
Cell 1 2 1
This means that machines 1 and 3, respectively are in cell 1, while machine 2 is in cell 2. This implies:
x
il

1 0
0 1
1 0

_
_
_

_
_
_
Since
z x
11
w
1j
2m
1j
1x
12
w
1j
2m
1j
1x
21
w
2j
2m
2j
1x
22
w
2j
2m
2j
1x
31
w
3j
2m
3j

1x
32
w
3j
2m
3j

and
m
lj

m
i1
x
il
w
lj

m
i1
x
il
then
m
11
x
11
w
11
1x
31
w
31
=x
11
1x
31
0:3 10:5=2 0:4
m
12
x
11
x
12
1x
31
w
23
=x
11
1x
31
0 10:0=3 0:0
m
13
x
11
w
13
1x
31
w
33
=x
11
1x
31
0:2 10:5=2 0:35
m
21
x
22
w
21
=1 0:0
m
22
x
22
w
22
=1 0:9
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 129
m
23
x
22
w
23
=1 0:0
m
lj

0:4 0:0 0:35
0:0 0:9 0:0
_ _
The matrix, m
ij
, is then used to evaluate the objective function value.
2.3. Minimization of exceptional elements model: (model 2)
Minimize z
2

p
j1
z
jl
4
subject to

k
l1
x
il
1 ;
i
5

m
i1
x
il
$ 1 ;
l
6
where z
jl
y
ji
p x
il
, z
p
jl
i.e. number of exceptional elements; z
p
il
arg Maxz
jl
i.e. part is assigned to
cell where most of its operations are carried out
y
ji
w
ij

T
;
y
ij

1 if w
ij
. 0
0 otherwise
_
y
ji
is the `1's and `0's transpose matrix of w
ij
. Eq. (5) ensures that for particular i, machine i is assigned
to one cell only. Eq. (6) ensures that no cell is empty.
2.4. Model illustration
As an illustration, consider a 3 3 problem whose workload incidence matrix is given as:
w
ij

0:3 0:0 0:2
0:0 0:9 0:0
0:5 0:4 0:5

_
_
_

_
_
_
y
ji

1 0 1
0 1 1
1 0 1

_
_
_

_
_
_
and x
il

1 0
0 1
1 0

_
_
_

_
_
_
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 130
z
ij
y
ji
p x
il
1 0 1
0 1 1
1 0 1

_
_
_

_
_
_
p
1 0
0 1
1 0

_
_
_

_
_
_

2 0
1 1
2 0

_
_
_

_
_
_
Therefore we have: z
p
11
arg Max2; 0 2 implies that part 1 belongs to cell 1; z
p
21
arg Max1; 1
1 implies that part 2 belongs to cell 2; and z
p
31
arg Max2; 0 2 implies that part 3 belongs to cell 1.
The solution becomes:
Number of exceptional elements z 1; corresponding to part 3 processed by machine 2.
2.5. Model 3 (A combination of model 1 and model 2)
In this model, cell grouping is rst performed using the minimization of exceptional elements model.
The solution found then becomes the initial solution for model 2. The exceptional elements model
arrives at machine distribution that results in the minimum intercell movement. This solution may at
times have a very high cell load variation-value. Model 2 then searches in the neighborhood of the
solution from model 1 for a machine distribution to cells that result in reduced cell load variation. At
times the solution that results from this redistribution may have a higher number of exceptional elements
(higher intercell movement). This model is ideal for cellular manufacturing system designs whose
design criterion requires the minimization of exceptional elements as well as balancing workloads in
each cell.
3. GA approach
GA is a meta-heuristic for solving combinatorial optimization problems (Goldberg, 1989). In this
section, we dene terminology and discuss design issues leading to the development of a GA for the
Cellular Manufacturing Systems problem.
3.1. Terminology and concept
In the GA employed for the work reported in this paper, a candidate solution is represented by a string
of numbers, which denote the position of a machine in a sequence. A candidate solution is known as a
string. A set of strings at time t is called a solution space S(t). The solution space at a given time is known
as a state. The potential of an individual string is determined by a score function. The score function
gives domain-specic information about the relative merit of each string. It maps strings into real
numbers, which are used to determine the strings that will be used to produce intermediate state for
the subsequent string. The GA procedure manipulates the solution space in such a way that better
solutions are obtained in subsequent strings. Since a solution space is a set of possible solutions, GA
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 131
applies the principle of parallelism to solve problems. For each solution space, the best solution is found,
and by applying some GA operators, new solution spaces are explored, leading to better results. With the
generation of each successive solution space, improvements in the quality of the individual solutions are
gained. In this way a GA can move to a successful outcome without being confounded in local optima
and without the need to examine every possible solution to the problem in a drastically small time. These
characteristics make GA a novel metaheuristic for solving NP-hard problems.
3.2. Representation
In the cellular manufacturing problem, the machine cell-membership number occupies a string bit.
For example, the rst string in Fig. 1 (2 3 1 1 2 3 1 1 3) represents 9-machine problem. The length of the
string represents the number of machines considered in the problem. The representation shows that
machines 3,4,7,8 are in cell 1; machines 1,5 are in cell 2; and machines 2,6,9 are in cell 3.
3.3. Initialization and evaluation
The initialization process is executed with a randomly generated solution space. An initial population
S(0) of the desired size (popsize) is generated randomly with population size popsize 10: GA-proce-
dure computes score function for each string of the solution space so that the string with the maximum
score function value is determined. The goal of optimization problems is to minimize some cost func-
tion. In GA approach, the cost function being optimized is usually mapped to a score function. We apply
the mapping process suggested by Goldberg (1989):
f
i
t
f
i
max
2g
i
t when g
i
t , f
i
max
0 if otherwise
_
7
where g(t) is the objective function of a string and f
max
is the largest objective function value in the
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 132
Fig. 1. Basic GA information.
current solution space. In the 9-machine cell formation of Fig. 1, the objective function value of the rst
string is 0.1942 and the value of f
max
is 0.5667 corresponding to that of the last string in the solution
space. Hence, the score function of the rst string is 0.3725 (0.5667 20.1942). Other score function
values are similarly computed.
3.4. Selection and recombination
There are six alternate selection schemes available in Goldberg (1989): deterministic sampling,
remainder stochastic sampling without replacement, remainder stochastic sampling with replacement,
stochastic sampling without replacement, stochastic sampling with replacement, and stochastic tourna-
ment. The remainder stochastic sampling without replacement has been found to be superior to other ve
strategies (Goldberg, 1989) and is the one used in the work reported here. In this strategy, the expected
count e
i
is calculated as usual
e
i
f t

f t=popsize
_ _
:
The fractional parts of e
i
values are treated as probabilities. One by one, weighted coin tosses are
performed using the fractional parts as success probabilities. The strings receive copies equal to the
whole parts of e
i
.
3.5. Genetic operators: crossover and mutation
Crossover is a mechanism for diversication. This diversication mechanism enables the GA to
examine unvisited regions and to generate solutions that differ in various signicant ways from those
seen before. The string to be crossed and the crossing points are selected randomly. The 2-point cross-
over technique has been adopted in this study. The crossover is done with a probability pcross. The
crossover operator generates two new strings as follows:
String 1:2
^
15
^
43 New String 1: 2 3 2 4 3
String 2:1
^
32
^
45 New String 2: 1 1 5 4 5
The crossover operation may cause the occurrence of a cell number to be lost as observed from the new
strings. For example, in the new string 1, cells 1 and 5 are lost and for the new string 2, cells 2 and 3 are
lost. To overcome this problem, mutation is done with a mutation probability (pmutate). Two random
integers r
1
, and r
2
are selected from strings 1 and 2, respectively such that 1 #r
1
,r
2
#n and r
1
r
2
. The
GA-procedure then swaps jobs with string bits designated as r
1
and r
2
. The process is repeated till all
cells are represented in the string. For example, if r
1
3 and r
2
1; then the previous new strings 1 and
2 become:
New string 1:2 3 1 4 3
New string 1:2 1 5 4 5
Repeating the mutation process with r
1
5 and r
2
3; we have new strings:
New string 1:2 3 1 4 5
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 133
New string 2:2 1 3 4 5
Applying the crossover and mutation operations, it has become possible to generate new strings
different from string 1 and string 2. The power of GA lies in its possession of these characteristics.
The mutation operation is essentially an intensication mechanism. During intensication, the search
focuses on examining neighbors of possible solutions. Successful crossovers and mutations lead to
generation of neighborhoods. In this example, new string 1 and new string 2 are neighborhoods of
string 1 and string 2, respectively.
3.6. Replacement strategy
In every crossover and mutation operation, new strings are created. The new strings may be better or
worse. This means that performing strings are to be replaced using a replacement strategy. Several
strategies have been suggested in the past by Liepins and Hilliard (1986). These include the following:
probabilistic replacement, crowding strategy, and elitist strategy. In this study, combination of these
strategies has been formulated.
3.7. Convergence policy
In GA application it is observed that as states proceed, the solution space will be converging to a
certain solution. A possible outcome is that the convergence may take place before the optimum solution
is obtained, this is known as premature convergence. A large number of subsequent strings for one type
of string may deprive the rest of the strings from producing new subsequent string resulting in a rapid
loss of diversity and premature convergence causing a problem called genetic drift. This results in the
loss of the best solution and should be avoided. To avoid the problem of premature convergence, a
number of proposals have been suggested such as, increasing the solution space size, or solution space.
Grefenstette (1986) suggested an entropic measure H
i
in a solution space of strings.
3.8. The entropic measure
For each machine i, the entropic measure is found thus
H
i

c
j1
n
ij
=plogn
ij
=p
logc
8
where, n
ij
is the number of strings in which job i is assigned a position identify j in the current solution
space; p is the solution space size; c is the number of parts.
The divergence, H, then is calculated as
H
i

m
i1
H
ij
=m 9
As the solution space proceeds the divergence parameter, H approaches zero. Therefore, this parameter
(preset at 0.491) was used in this study to monitor the diversity of each and every state.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 134
3.9. Diversication
When diversity falls below the predetermined value the whole solution space is diversied using the
mutation operator, until the diversity is above the present value.
The diversication results in creation of new strings, which may be better or worse and this may result
in the loss of better strings. This is a great disadvantage inherent in the convergence policies discussed
previously.
3.10. Preventing loss of best solution
It is important that the best strings in the old solution space should be maintained. The diversied
population is evaluated using the objective function and the best performing strings are compared with
those of the previous undiversied solution space. The best between the diversied and the undiversied
is advanced into the next state. This prevents the loss of the best solution. The technique used here was
tested and found to produce better results than the former methods by Gupta, Gupta, Kumar and
Sundaram (1986).
4. GA for cellular manufacturing
The GA procedure for a cellular manufacturing system-problem is developed incorporating the design
issues discussed in the previous section. These steps are now discussed in detail. Model 1 is for cell load
variation, model 2 is for inter-cellular moves, and model 3 is the multi-objective function.
4.1. Models 1 and 2 algorithms
Step 1. Initialization. Select the initial input parameters and generate randomly an initial diversied
population.
(i) Test for upper and lower bounds of machines test for constraints;
(ii) Set the values for population size (popsize), number of generations (maxgen), probabilities of
crossover (pcross) and mutation (pmutation), and number of cells (no_cells);
(iii) Create randomly, an initial population, oldpop.
Step 2. Selection and recombination. Select strings using stochastic sampling without replacement.
(i) Evaluate strings by objective function, tness, and expected count calculation;
(ii) Create a temporary pool (temppop) using the integer parts of expected count and use the
fractional parts as success probabilities.
Step 3. Crossover/recombination. Apply the multi-crossover operator to tempop to create a selection
pool of size, popsizepcross poolsize:
(i) Candidates for crossover are selected randomly by remainder selection without replacement;
(ii) The multi-crossover operator is applied to two chosen strings, with a probability, pcross;
(iii) Apply mutation to subsequent strings normal mutation regardless of constraints;
(iv) Test for constraints; if violated, apply mutation until constraints are sufced;
(v) Evaluate by calculating objective function values;
(vi) Sort out the selection pool in increasing order of objective function values.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 135
Step 4. Replacement strategy. Compare each of the strings of sorted selection pool with corresponding
old population strings.
(i) Compare corresponding strings in selection pool and old population, and take the best in turn;
(ii) Take the one that fares better than the other in each turn;
(iii) For other offspring, a random selection is made with a probability, 0.6491.
Step 5. Diversicating. Diversify the population if diversity falls below a predetermined minimum.
(i) Calculate the diversity, H, of the population;
(ii) Compare with the given minimum, the acceptable diversity. If less, then apply mutation
operator to the whole population, otherwise proceed. Repeat this process until diversity is satised.
(iii) Calculate the objective function values of the diversied population;
(iv) Sort out the new pool of strings in increasing order of objective function values.
Step 6. New Generation. The value of the current generation-number determines the next step.
(i) If generation is less than current population then current population becomes old population and
the process is repeated;
(ii) Otherwise, stop. The solution is string with lowest objective function value in the current
population. Print out the clustered solution.
4.2. Model 3 algorithm
Step 1. Initialization. Select the initial input parameters and generate randomly an initial diversied
population.
(i) Test for upper and lower bounds of machines test for constraints;
(ii) Set the values for population size (popsize), number of generations (maxgen), probabilities of
crossover (pcross) and mutation (pmutation), and number of cells (no_cells);
(iii) Depending on type of input available, either (1) input processing and available times, and
demands or (2) input the workload values on each machine induced by each part;
(iv) Create randomly, two populations, oldpop, and oldpop2.
Step 2. Reproduction/selection
(i) Select strings using stochastic sampling without replacement to form two pools, tempop and
temppop 2 being recombined. Evaluate strings by calculating the objective function values, F1 for
oldpop and F2 for oldpop2;
(ii) Create two temporary pools (temppop and temppop2) being recombined using the integer parts
of expected count and use the fractional parts as success probabilities.
Step 3. Crossover/recombination. Apply the multi-crossover operator to temppop to create two
selection pools each of size popsizepcross poolsize:
(i) Two candidates for cross over are selected randomly for multi-crossover, (1) one from temppop,
and the other (2) from temppop2 by remainder selection without replacement;
(ii) The multi-crossover operator is applied to the two chosen strings, with a probability pcross;
(iii) Apply mutation to subsequent strings normal mutation regardless of constraints;
(iv) Test for constraints; if violated, apply mutation until constraints are sufced;
(v) Evaluate by calculating objective function values;
(vi) Sort out the selection pools separately, in increasing order of objective function values.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 136
Step 4. Replacement strategy. Compare each of the strings of sorted selection pools with correspond-
ing old population strings.
(i) Compare corresponding strings successively in selection pool spoolpop1, and old solution space
oldpop, and take the best in turn;
(ii) Take the one that fares better than the other in each turn;
(iii) For other offspring, a random selection is made with a probability, 0.6491.
Step 5. Diversicating. Diversify each of the populations if diversity falls below a predetermined
minimum.
(i) Calculate the diversities, H, of the populations;
(ii) Compare with the given minimum, the acceptable diversity. If less then apply mutation operator
to the whole population, otherwise proceed. Repeat this process until diversity is satised.
(iii) Calculate the objective function values of the diversied populations, F1 for oldpop and F2 for
oldpop2;
(iv) Sort out the new pools of strings in increasing order of objective function values.
Step 6. New generation. The value of the current generation number determines the next step.
(i) If generation is less than current population then current population becomes old population and
the process is repeated;
(ii) Otherwise, stop. The optimal solution is string would be the common string to both current
populations. Print out the clustered solution if common string is found. Otherwise show both
solutions.
5. Computational analysis
The GA was written in PASCAL 7.0 and runs on a PC with a Pentium, 133 MHz processor. A GA
execution illustration is rst given, then two sets of problem types which were used are presented:
randomly generated block diagonalized incidence matrices, and published data sets.
5.1. Algorithm execution illustration
This section provides an illustration of all intermediate solutions arrived at in the solving of a machine
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 137
Fig. 2. Machine-component incidence matrix for Chan and Milner (1982).
grouping problem using the GA-scheme. The set of input data used in this illustration is from Chan and
Milner (1982). The input incidence matrix is shown in Fig. 2. Number of machines is 10; number of parts
is 15; minimum number of machines in a cell is 2; maximum number of machines in a cell is 6. The
option for simultaneous minimization of exceptional elements and cell load variation is considered
(multi-objective model). The objective function values correspond to the minimum number of excep-
tional elements and minimum cell load variation value. The results are plotted in Fig. 3.
The number of exceptional element drops from 9 at the initial solution-state to zero at the fteenth
solution-state and settles at that value. The cell load variation drops from a value of 18 at the initial
solution-state to 2.5 at the fth solution state and settles at that value. This means that in the actual sense,
the optimum solution was found at the fteenth solution-state even though the user indicated that the GA
algorithm search through 100 solution states. The machine part grouping for the solution corresponding
to the global is given in matrix form in Fig. 4. The clustering efciency is 92%.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 138
Fig. 3. The GA objective functions plots for Chan and Milner (1982).
Fig. 4. Final solution machine-component matrix (Chan & Milner, 1982).
5.2. Extensive computational results
The rst problem type set was generated using a support software to rst generate at random an
incidence matrix which is block diagonalized and then adding a predetermined number of exceptional
elements outside the blocks. Then the matrix was perturbed so that the diagonalised structure was lost.
This ensures that the global optimum solution is known a priori. In all the problems of this type, GA
accurately recovered the optimum solutions. Medium to large scale problems were solved for this
problem type. Table 1 shows the solution times and efciencies. The results show that GA is capable
of solving large scale problems in reasonable time while having very high quality result.
The second set of problems belongs to published results. Table 2 shows the clustering efciencies of
the GA scheme and those of HPH, IP, Tabu Search, Hybrid enetic Algorithm using similarity coefcient
and Traveling Sales Person combination procedures. For problem 1, the HPH procedure could not
cluster cells for machine cell size equal to 2, while GA performed best. Even for the same problem
and machine cell size equal to 3, GA performed best. All procedures tallied for problem 2, while GA
performed best together with HGA (SCTSP) compared to the other procedures for problem 3. For
problem 4, GA proved better than all others. However, when the cell number was increased to 4, GA did
not prove as effective as the 01 programming procedure. It should be noted that the problem sets were
small scale and both the HPH and 01 programming procedures can not be employed to solving large
scale size problems due to the prohibitive computation time involved. This set of problems is reserved
for metaheuristics.
We further tested the results of our algorithm with those of the TSP-based heuristic (Cheng et al.,
1998) based on genitor software (Whitley, 1989) and zodiac (Chandrasekharan & Rajagopalan, 1987)
for 25 data sets from the literature. The detailed results of the further experiments are presented in
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 139
Table 1
Solution clustering efciency and CPU times for 1st problem type
Problem set Machines Parts Cells Maximum cell size Clustering efciency (%) CPU time (s)
1 5 15 2 4 68.75 0.000
2 3 68.75 0.000
3 3 78.125 1.700
3 2 78.725 0.380
2 10 20 3 6 69.825 13.740
3 5 69.825 8.410
4 4 87.719 14.170
5 4 78.947 19.270
3 15 30 5 7 62.281 36.750
5 6 62.55 38.230
6 5 83.333 26.310
6 4 83.333 47.120
4 20 45 6 8 57.087 82.610
6 7 58.849 104.460
7 6 64.795 117.600
7 5 67.249 122.370
Table 3. Two measures frequently used in the literature are used in the present comparative study. One is
grouping efciency (Chandrasekharan & Rajagopalan, 1986), which is dened as follows:
h qh
1
11 2qh
2
;
where
h
1
e
d
_

k
i 21
P
i
Q
i
;
h
2
1 2e
o
_
_
mn 2

k
i1
P
i
Q
i
_
where m is the number of machines; n the number of parts; P
i
the number of machines in the rth cell;
Q
i
the number of parts in the rth cell; e the number of 1's in the incidence matrix; e
d
the number of
1's within the diagonal blocks; e
o
the number of 1's outside the diagonal blocks; e
v
the number
of voids; k the number of cells; h the grouping efciency; G the grouping efcacy; q the weighting
factor (0 #q #1).
For q 0:5; we have dene grouping efciency as:
h 2 2
e
v
a
1
e
o
mn 2a
_ _ _ _
=2:
Kumar and Chandrasekharan (1990) proposed grouping efcacy (G) as:
G
a 2e
o
a 1e
v
Table 3 shows that our algorithm favorably competes with that of Cheng et al. (1998) and that both
algorithms produce better solutions than zodiac, which is the commonly used reliable algorithm in
comparative studies. In some cases our algorithm produces the best results. For example, our algorithm
produces the best solution in problems 4, 8, and problem 11, for small to medium sized-problems. For
the same category of problems, the TSP-based algorithm is best for problems 1, 12, and 13. For the large
sized problems, the TSP-based algorithm perform best in a number of problems, followed by the present
GA-based procedure, while zodiac performs least.
The drawback in the presentation of the published results in Cheng et al. (1998) is that they did not
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 140
Table 2
Clustering efciency values for test experiments for GA and other algorithms
No Problem Machines Parts Cells Matrix
density
GA Hybrid GA
(SCTSP)
Tabu
search
HPH
method
01 integer
programming
1 (De witte, 1980) 12 19 2 0.33 45 42 42
3 45 37 37 0.37
2 (Chandrasekharan &
Rajagopalan, 1986)
8 20 3 0.38 85 85 85 0.85
3 (Burbidge, 1975) 16 43 5 0.18 44 44 40 0.42
4 (Kumar & Vannelli, 1987) 30 41 4 0.105 30 31 27 0.16
specify the number of cells considered, as this affects the solutions obtained. The fact that cell-size
affects the solution of cell formation problem is evident in Tables 1 and 2. In our algorithm, the cell size
is automatically generated, but there is the exibility for the user to interactively specify different cell
size(s) during experimentation. The solutions from our algorithm in Table 3 are based on automatically
generated cell size and there is the possibility that we can obtain better solutions by specifying different
cell sizes and cell size capacities. Another point that needs to be addressed is that the authors of the TSP-
based algorithm under the section of comparative study say: It is not appropriate to use our objective
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 141
Table 3
Performance of the algorithm compared to zodiac and TSPGA
Number Source Size Grouping efciency (%) Grouping efcacy (%)
zodiac TSPGA GA zodiac TSPGA GA
1 Waghodekar and Sahu
(1984)
5 7 72.20 77.10 73.85 56.52 68.00 62.50
2 Seifoddini and Wolfe (1986) 5 18 86.76 86.76 86.76 77.36 77.36 77.36
3 Kusiak and Cho (1992) 5 7 87.50 87.50 87.50 76.92 76.92 76.92
4 Kusiak and Chow (1987) 7 11 65.01 73.70 74.24 39.14 46.88 50.00
5 Boctor (1991) 7 11 86.08 86.08 86.08 70.37 70.37 70.37
6 Chandrasekharan and
Rajagopalan (1986)
8 20 95.83 95.83 95.83 85.24 85.24 85.24
7 Chandrasekharan and
Rajagopalan (1986)
8 20 71.88 71.88 64.60 58.33 58.33 55.91
8 Mosier and Taube (1985) 10 10 85.29 85.29 85.76 70.59 70.59 72.79
9 Stanfel (1985) 14 24 83.90 84.10 88.93 65.55 67.44 63.48
10 Chan and Milner (1982) 15 10 96.00 96.00 96.00 92.00 92.00 92.00
11 King (1980) 16 43 80.20 79.42 83.73 53.76 53.89 86.25
12 Mosier and Taube (1985) 20 20 53.05 75.61 62.81 21.63 37.12 34.16
13 Carrie (1973) 20 35 87.81 88.00 84.83 75.14 75.28 66.30
14 (Boe and Cheng (1991) 20 35 77.56 79.60 73.35 51.13 55.14 44.44
15 Kumar et al. (1986) 23 40 66.97 81.41 67.55 38.66 46.62 39.02
16 Chandrasekharan and
Rajagopalan (1989)
24 40 100.00 100.00 100.00 100.00 100.00 100.00
17 Chandrasekharan and
Rajagopalan (1989)
24 40 95.20 95.20 95.20 85.11 85.11 85.11
18 Chandrasekharan and
Rajagopalan (1989)
24 40 90.84 90.84 90.84 73.03 73.03 73.03
19 Chandrasekharan and
Rajagopalan (1989)
24 40 77.31 83.75 71.76 20.42 49.37 37.62
20 Chandrasekharan and
Rajagopalan (1989)
24 40 72.43 85.29 70.42 18.23 44.67 34.76
21 Chandrasekharan and
Rajagopalan (1989)
24 40 69.33 81.10 72.60 17.61 42.50 34.06
22 Kumar and Vannelli (1987) 30 41 68.14 82.43 75.93 33.46 53.80 40.96
23 Stanfel (1985) 30 50 75.35 85.95 79.35 46.06 56.61 48.28
24 Stanfel (1985) 30 50 62.92 82.18 73.41 21.11 45.93 37.55
25 Chandrasekharan and
Rajagopalan (1987)
40 100 95.07 95.10 95.00 83.92 84.03 83.90
function to evaluate the other solutions. This means that their objective function is solely dependent on
the distance measures that belong to the family of Minkowski metric. In practice, a manufacturing
systems engineer may want to consider only total cell load variation as objective function, or only
intercellular moves as objectives function, or a combination of both. Other objective functions are
possible as already discussed (Onwubolu & Songore, 2000). This is what we have built into our
algorithm: exibility for the manufacturing systems designer.
6. Conclusions
In this paper, the cell formation problem was modeled with three options of objective functions:
minimization of intercellular movements; minimization of cell load variation, combination of the above
options. Computational results of the GA objective function values were found to score better than those
obtained from published algorithms. The hybrid GA result published in the literature did not show any
added advantage when compared with our model. A previous publication of the TSP-based algorithm for
the cell formation problem yielded better results than the well-known algorithm in the literature. The
results of the present study show that our approach competes with the TSP-based algorithm. Moreover
our algorithm has added advantage of exibility, which has the potential of yielding improved results.
We have also compared our GA-based procedure with those of tabu search and our algorithm performed
better.
A major another advantage inherent in the GA procedure described in this paper is the choice of
objective functions to apply depending on the design criteria. The minimization of exceptional
elements model proves to be ideal for an optimization problem that is focused on minimization of
intercell movement of parts during manufacture. Intercellular movements lead to transportation costs in
a manufacturing rm, and should be reduced as much as possible. Therefore, minimizing exceptional
elements leads to minimizing transportation cost between cell and consequently leads to reduced manu-
facturing cost. The minimization of cell load variation is ideal where balancing of workloads within
cells is a design criteria or where there is a restriction on the available working time per period. Inclusion
of several objective function options in the GA developed makes the approach exible for designing
cellular manufacturing systems. In our program-implementation, the designer if offered the liberty to
specify the maximum cell size or number of machines in a cell and the number of cells. The GA method
also allows the user to specify the number of cells. With such a facility the manufacturing systems design
engineer is able to arrive at an optimum machine allocation to cells that respects the desired design
criteria.
Acknowledgements
Dr G. Srinivasan of Indian Institute of Technology (IIT), Madras provided most of the data sets for
experimentation.
References
Askin, R. G., & Subramaniam, S. P. (1987). A cost-based heuristic for group technology conguration. International Journal of
Production Research, 25 (1), 101113.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 142
Boctor, F. F. (1991). A linear formulation of the machine part cell formation problem. International Journal of Production
Research, 28, 185198.
Boe, W. J., & Cheng, C. H. (1991). A close neighbour algorithm for designing a cellular manufacturing system. International
Journal of Production Research, 29, 20972116.
Burbidge, J. L. (1975). The introduction of group technology, New York: Wiley.
Carrie, A. S. (1973). Numerical taxonomy applied to group technology and plant layout. International Journal of Production
Research, 11, 399416.
Chan, H. M., & Milner, D. A. (1982). Direct clustering algorithm for group formation in cellular manufacture. Journal of
Manufacturing Systems, 1 (1), 6476.
Chandrasekharan, M. P., & Rajagopalan, R. (1986). An ideal seed non-hierarchical clustering algorithm for cellular manu-
facturing. International Journal of Production Research, 24 (2), 451464.
Chandrasekharan, M. P., & Rajagopalan, R. (1987). Zodiac-an algorithm for concurrent formation of part-families. Interna-
tional Journal of Production Research, 25, 835850.
Chandrasekharan, M. P., & Rajagopalan, R. (1989). Groupability: an analysis of the properties binary data matrices for group
technology. International Journal of Production Research, 27, 10351052.
Cheng, C. H., Gupta, Y. P., Lee, W. H., & Wong, K. F. (1998). A TSP-based heuristic for forming machine groups and part
families. International Journal of Production Research, 36 (5), 13251337.
De witte, J. (1980). The use of similarity coefcients in production ow analysis. International Journal of Production Research,
18 (4), 503514.
Goldberg, D. E. (1989). Genetic algorithms: in search, optimization and machine learning, Reading, MA: Addison Wesley.
Grefenstette, J. J. (1987). Incorporating problem specic knowledge into genetic algorithms. In L. Davis, Genetic algorithms
and simulated annealing, London: Pitman.
Gupta, Y., Gupta, M., Kumar, A., & Sundaram, C. (1986). A genetic algorithm-based approach to cell composition and layout
design problems. International Journal of Production Research, 34, 447482.
Harhalakis, G., Nagi, R., & Proth, J. M. (1990). An efcient heuristic in manufacturing cell formation for group technology
applications. International Journal of Production Research, 28, 185198.
Hindi, K. S., & Hamam, Y. M. (1994). Solving the part families problem in discrete-parts manufacture by simulated annealing.
Production Planning and Control, 5 (2), 160164.
Hsu, C. M., & Su, C. T. (1998). Multi-objective machine-component grouping in cellular manufacturing: a genetic algorithm.
Production Planning and Control, 9 (2), 155166.
Hyer, N. L., & Wemmerlov, U. (1984). Group technology and productivity. Harvard Business Review, 62 (4), 140149.
Kaparthi, S., & Suresh, N. C. (1992). Machine-component cell formation in group technology: a neural network approach.
International Journal of Production Research, 30 (6), 13531368.
King, J. R. (1980). Machine-component grouping in production ow analysis: an approach using rank order clustering algo-
rithm. International Journal of Production Research, 18 (2), 213232.
King, J. R., & Nakornchai, V. (1982). Machine-component group formation in group technology: review and extension.
International Journal of Production Research, 20 (2), 117133.
Kumar, C. S., & Chandrasekharan, M. P. (1990). Grouping efcacy: a quantitative criterion for goodness of block diagonal
forms of binary matrices in group technology. International Journal of Production Research, 28, 233243.
Kumar, K. R., & Vannelli, A. (1987). Strategic subcontracting for efcient disaggregate manufacturing systems. International
Journal of Production Research, 25, 17151728.
Kumar, K. R., Kusiak, A., & Vannelli, A. (1986). Grouping parts and components in exible manufacturing systems. European
Journal of Operations Research, 24, 387397.
Kusiak, A., & Cho, M. (1992). Similarity coefcient algorithm for solving the group technology problem. International Journal
of Production Research, 30, 26332646.
Kusiak, A., & Chow, W. S. (1987). Efcient solving of the group technology problem. Journal of Manufacturing Systems, 6,
117124.
Liepins, G. E., Hilliard, M. R. (1986). Genetic algorithms as a paradigm for machine learning. Paper presented at the ORSA/
TIMS Joint National Meeting, Miami, FL.
McCormick, W. T., Schweitzer, P. J., & White, T. W. (1972). Problem decomposition and data reorganisation by a clustering
technique. Operations Research, 29 (5), 9931009.
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 143
Moon, Y. B. (1990). Forming part families for cellular manufacturing: a neural network approach. International Journal of
Advance Manufacturing, 11 (3), 149159.
Mosier, L., & Taube, C. (1985). Weighted similarity measure heuristics for the group technology clustering problem. Omega,
13, 577579.
Onwubolu, G. C., &Songore, V. (2000). A tabu search-based approach to cellular manufacturing systems. Production Planning
and Control, 11 (2), 153164.
Rajagopalan, R., & Betra, J. L. (1975). Design of cellular production systems a graph theoretic approach. International
Journal of Production Research, 13, 567579.
Sarker, R. B., & Xu, Y. (1998). Operations sequences-based cell formation methods: a critical survey. Production Planning and
Control, 9 (8), 771783.
Seifoddini, H., & Wolfe, P. M. (1986). Application of the similarity coefcient method in group technology. IIE Transactions,
18, 271277.
Shafer, S. M., & Rogers, D. F. (1991). A goal programming approach to the cell formation problem for cellular manufacturing.
Journal of Operations Management, 10 (1), 2843.
Stanfel, L. E. (1985). Machine clustering for economic production. Engineering Costs and Production Economics, 9, 7381.
Vannelli, A., & Kumar, K. R. (1986). A method for nding minimal bottleneck cells for grouping part-machine families.
International Journal of Production Research, 24 (2), 387400.
Venugopal, V., & Narendran, T. T. (1992). A genetic algorithm approach to the machining grouping problem with multiple
objectives. Computers and Industrial Engineering, 22 (4), 469480.
Waghodekar, P. H., & Sahu, S. (1984). Machine-component cell formation in group technology mace. International Journal of
Production Research, 22, 937948.
Wei, J. C., & Gaither, N. (1990). An optimal model for cell formation decisions. Decision Sciences, 21 (2), 416433.
Wemmerlov, U., & Hyer, N. L. (1986). Procedures for the part family/machine group identication problem in cellular
manufacture. Journal of Operations Management, 6 (2), 125148.
Wemmerlov, U., & Hyer, N. L. (1989). Cellular manufacturing in the US industry: a survey of users. International Journal of
Production Research, 27 (9), 15111530.
Whitley, D. (1989). The genitor algorithm and selection pressure: why rank-based allocation of reproductive trials is best
Proceedings of the Third International Conference of Genetic Algorithms, Don Mateo, California, Los Altos, CA: Morgan
Kaufmann (pp. 116121).
G.C. Onwubolu, M. Mutingi / Computers & Industrial Engineering 39 (2001) 125144 144

You might also like