Expert Systems with Applications 37 (2010) 4822–4829

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

A differential evolution algorithm for the manufacturing cell formation

problem using group based operators
Azadeh Noktehdan, Behrooz Karimi *, Ali Husseinzadeh Kashan
Department of Industrial Engineering, Amirkabir University of Technology, P.O. Box 15875-4413, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Keywords: Cellular manufacturing (CM) is an important application of group technology (GT), a manufacturing phi-
Cell formation problem losophy in which parts are grouped into part families, and machines are allocated into machine cells to
Grouping genetic algorithm take advantage of the similarities among parts in manufacturing. The target is to minimize inter-cellular
Differential evolution algorithm movements. Inspired by the rational behind the so called grouping genetic algorithm (GGA), this paper
proposes a grouping version of differential evolution (GDE) algorithm and its hybridized version with
a local search algorithm (HGDE) to solve benchmarked instances of cell formation problem posing as a
grouping problem. To evaluate the effectiveness of our approach, we borrow a set of 40 problem
instances from literature and compare the performance of GGA and GDE. We also compare the perfor-
mance of both algorithms when they are tailored with a local search algorithm. Our computations reveal
that the proposed algorithm performs well on all test problems, exceeding or matching the best solution
quality of the results presented in previous literature.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction lies in an approximate manner. Many of them are developed on the

A cellular manufacturing system (CMS) is a production approach
aimed at increasing production efficiency and system flexibility by
utilizing the process similarities of the parts. It involves grouping
similar parts into part families and the corresponding machines into
machine cells. This results in the organization of production systems
into relatively self-contained and self-regulated groups of machines
such that each group of machines undertakes an efficient production
of a family of parts. The significant benefits of cellular manufacturing
are a reduced setup time, reduced work-in-process inventory,
reduced throughput time, reduced material handling costs, im-
proved product quality, simplified scheduling, etc. (Wemmerlov &
Hyer, 1987).
The cell formation (CF) problem is the first step in designing of
cellular manufacturing systems (CMS). The main objective of CF is
to construct machine cells and part families, and then dispatch part
families to machine cells to optimize the chosen performance mea-
sures such as inter-cell and intra-cell transportation cost, grouping
efficiency and exceptional elements (Lei & Wu, 2005).
The machine–part cell formation (MPCF) problem is an NP-hard
problem (Ballakur & Steudel, 1998), thus extensive research has
been devoted to cell formation (CF) problems, with many methods
having been proposed for identifying machine cells and part fami-
* Corresponding author. Tel.: +98 21 66413034; fax: +98 21 66413025.
E-mail addresses: A.noktehdan@aut.ac.ir (A. Noktehdan), B.Karimi@aut.ac.ir
(B. Karimi), A.Kashani@aut.ac.ir (A. Husseinzadeh Kashan).
basis of heuristic clustering techniques to obtain approximate
solutions, but some of them may be far from optimum.
Up to now, the only algorithm that has been heavily modified to
consider the structure of grouping problems, such as cell formation
problem, is based on genetic algorithm which is called grouping
genetic algorithm (GGA). In this paper we apply the notion of
grouping representation and operators in the body of differential
evolution algorithm to obtain the grouping version of differential
evolution (GDE).
The reminder of the paper is organized as follows: In Section 2
we describe the problem definition. Section 3 provides a review of
previous solution approaches developed for cell formation prob-
lem. An overview on the differential evolution algorithm is given
in Section 4. The proposed grouping differential evolution algo-
rithm is introduced in Section 5. Section 6 deals with the computa-
tional results, and Section 7 concludes the paper.
2. Cell formation problem
Given 0–1 machine–part incidence matrix, cell formation task
involves rearrangement of rows and columns of the matrix to cre-
ate part families and machine cells. In this research we attempt to
determine a rearrangement so that the inter-cellular movement
can be minimized and the utilization of the machines within a cell
can be maximized. Two matrices shown in Fig. 1 are used to illus-
trate the concept. Fig. 1a is an initial matrix where no blocks can be
observed directly. After rearrangement of rows and columns, two

0957-4174/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2009.12.033
 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829
(a) Parts
P1 P2 P3 P4
M1 1 0 0 1
M2 0 1 1 0
Machines M3 1 0 0 1
M4 0 1 1 0
M5 1 0 0 1
Fig. 1. Rearrangement of rows and columns of matrix to create cells: (a) initial matrix and (b) matrix after rearrangement.
blocks can be obtained along the diagonal of the solution matrix in
Fig. 1b.
There have been developed several measures of goodness of
machine–part groups in cellular manufacturing in the literature
(Sarker, 2001). Two measures frequently used are the grouping
efficiency (Chandrasekharan & Rajagopalan, 1986a) and the group-
ing efficacy (Kumar & Chandrasekharan, 1990). Grouping efficiency
g is defined as follows:
g ¼ qg1 þ ð1  qÞg2
where g1 is the ratio of the number of 1s in the diagonal blocks to
the total number of 0s and 1s in the diagonal blocks, g2 is the ratio
of the number of 0s in the off-diagonal blocks to the total number of
elements in the off-diagonal blocks, and q is the weight factor
(Chandrasekharan & Rajagopalan, 1989). Those 1s outside the diag-
onal blocks, are called ‘exceptional elements’; while those 0s inside
the diagonal blocks are called ‘voids’.
Although grouping efficiency has been used widely, it was ar-
gued for its low discriminating capability in some cases affected
by the size of the problem. To overcome this weakness, Kumar
and Chandrasekharan (1990) proposed another measure, the
grouping efficacy s, that can be defined as:
e  eo
s¼
e þ ev
where e is the total number of 1s in the matrix. eo is the total num-
ber of exceptional elements, and ev is the total number of voids.
Grouping efficacy ranges from 0 to 1, with 1 being the perfect
grouping. As grouping efficacy has been widely accepted in recent
studies on CF problem, we will use it to measure the quality of a gi-
ven solution.
3. Literature review
A large number of studies related to GT/CM have been per-
formed both in academia and industry. Reisman, Kumar, Motwani,
and Cheng (1997) gave a statistical review of 235 articles dealing
with GT and CM over the years 1965–1995.
Comprehensive summaries and taxonomies of studies devoted
to part–machine grouping problems were presented in Wemmer-
lov and Hyer (1989) and Kusiak and Cho (1987). Methods for part
family/machine cell formation can be classified as design-oriented
or production-oriented. Design-oriented approaches group parts
into families based on similar design features. An overview of de-
sign-oriented approaches based on classification and coding was
presented in Askin and Vakharia (1991). Production-oriented tech-
niques are for aggregating parts requiring similar processing. These
approaches can be further classified into cluster analysis, graph
partitioning, mathematical programming, artificial intelligence
(AI), and heuristics based approaches (Joines, Culbrethe, & King,
1996). Mathematical programming is widely used for modeling
CMS problems. The objective of the mathematical programming
model is often to maximize the total sum of similarities of parts
4823
(b) Parts
P5 P2 P3 P5 P1 P4
0 M2 1 1 1 0 0
1 M4 1 1 1 0 0
0 Machines M1 0 0 0 1 1
1 M3 0 0 0 1 1
0 M5 0 0 0 1 1
in each cell, or to minimize inter-cell material handling cost. Pur-
check (1975) applied linear programming techniques to a group
technology problem. Kusiak and Cho (1987) proposed the general-
ized p-median model considering the presence of alternative rout-
ings. Shtub (1989) used the same approach and reformulated the
problem as a generalized assignment problem. Wei and Gaither
(1990) developed a 0–1 programming cell formation model to min-
imize bottleneck cost, maximize average cell utilization, and min-
imize intra-cell and inter-cell load imbalances. Rajamani, Singh,
ð1Þ and Aneja (1992) proposed three integer programming models to
consider budget and machine capacity constraints as well as alter-
native process plans. Askin and Chiu (1990) proposed a cost-based
mathematical formulation and a heuristic graph partitioning pro-
cedure for cell formation. Shafer and Rogers (1991) applied a goal
programming method to solving CMS problems for different sys-
tem reconfiguration situations: setting up a new system and pur-
chasing all new equipments, reorganizing the system using only
existing equipments, reorganizing the system using existing and
some new equipments. Shafer and Kern (1992) presented a math-
ematical programming model to address the issues related to
exceptional elements.
Due to their excellent performance in solving combinatorial
optimization problems, meta-heuristics algorithms such as genetic
algorithms, simulated annealing, neural networks and tabu search
ð2Þ make up a class of search methods that has been adopted to effi-
ciently solve the CF problem and its variants with good results ob-
tained. Sun, Lin, and Batta (1995) presented a short-term tabu
search-based algorithm for solving the CF problem with the objec-
tive of minimizing the inter-cellular parts flows, while Lei and Wu
(2005) maximizes the parts flow within cells using long-term tabu
search-based algorithm. Aljaber, Baek, and Chen (1997) proposed a
tabu search approach to deal with this problem by modeling it as a
shortest spanning path problem with respect to both parts and ma-
chines. The resulting spanning paths for parts and machines are
then decomposed into subgraphs representing machine groups
and part families, respectively. Cheng, Gupa, and Lee (1998) formu-
lated the CF problem as a traveling salesman problem (TSP) and
proposed a solution methodology based on genetic algorithm,
while Dimopoulos and Mort (2001), presented a hierarchical clus-
tering approach based on genetic programming. Onwubolu and
Mutingi (2001) developed a genetic algorithm which accounts for
inter-cellular movements and the cell-load variation. Goncalves
and Resende (2004) presented a hybrid algorithm combining a lo-
cal search and a genetic algorithm with very promising results re-
ported. Wu, Chang, and Chung (2007) developed a simulated
annealing algorithm that uses the similarity coefficient for design-
ing cells. Boulif and Atif (2006) proposed a branch-and-bound-en-
hanced genetic algorithm for the manufacturing cell formation
problem which is combination of a genetic algorithm with the
branch-and-bound approach. Brown and James (2007) presented
a hybrid algorithm combining a local search and a grouping genetic
algorithm. Recently many researchers apply grouping genetic algo-
rithm for solving cell formation problems. Vin, Delit, and Delcham-
ber (2005) and Brown and James (2007) are some of the
researchers that apply the GGA to cell formation problems.
4824 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829
4. An introduction to DE
The differential evolution (DE) algorithm introduced by Storn
and Price (1995) is a novel parallel direct search method and is
one of the latest evolutionary optimization techniques for optimiz-
ing continuous nonlinear functions.
Currently, there are several variants of DE. The particular vari-
ant used also in this research is DE/rand/1/bin scheme. This
scheme will be discussed shortly here. Generally, the function to
be optimized, F, is of the form FðXÞ : RD ! R. The optimization tar-
get is to minimize the value of this objective function FðXÞ by opti-
mizing the values of its parameters X = {x1, x2, . . . , xD}, X 2 RD, where
X denotes a vector composed of D objective function parameters.
Usually, the parameters of the objective function are also subject
to lower and upper boundary constraints.
4.1. Initialization
As with all evolutionary optimization algorithms, DE works
with a population of solutions. Population P at generation T con-
tains NP solution vectors called individuals and each vector repre-
sents a potential solution for the problem.
In order to establish a starting point for optimum seeking, the
population must be initialized. Often there is no more knowledge
available about the location of a global optimum than the bound-
aries of the problem variables. In this case, a natural way to initial-
ize the population P(0) (initial population) is to seed it with random
values within the given boundary constraints.
4.2. Mutation
The population recombination scheme of DE is different from
the other evolutionary algorithms. From the first generation on-
ward, the population of the subsequent generation PðTþ1Þ is ob-
tained on the basis of the current population P ðTÞ . A mutated
parameter vector V Tþ1 ¼ ðv Tþ1
1i ; v 2i ; . . . ; v Di Þ of candidate vectors
Tþ1 Tþ1
i
for the subsequent generation is generated as follows:
v jiTþ1 ¼ xTj;r 3
þ FðxTj;r1  xTj;r2 Þ 8i ¼ 1; . . . ; NP 8j ¼ 1; . . . ; D
where r1 ; r 2 ; r 3 2 f1; . . . ; NPg and F > 0.
Three randomly chosen indexes, r1 ; r 2 , and r 3 refer to three ran-
domly chosen vectors of population. They are mutually different
from each other and also different from the running index i.
4.3. Crossover
In order to increase the diversity of the perturbed parameter
vectors, crossover is introduced. To this end, the trial vector:
U Tþ1
i
Tþ1
¼ ðu1i ; uTþ1 Tþ1
2i ; . . . ; uDi Þ
is formed, where
( Tþ1
Tþ1
v ji if randðjÞ 6 CR or j ¼ randintðiÞ
uji ¼ 8i ¼ 1;...;NP 8j ¼ 1;...;D
xTþ1
ji otherwise
rand(j) is the jth evaluation of a uniform random number generator
with outcome 2 ½0; 1. CR is the crossover constant 2 ½0; 1 which has
to be determined by the user. randint(i) is a randomly chosen index
in {1, . . . , D} which ensures that U Tþ1
i gets at least one parameter
from V Tþ1
i (Storn & Price, 1995).
F and CR are DE control parameters. Both values as well as the
third parameter, NP (population size), remain constant during the
search process. F is a real-valued factor that controls the amplifica-
tion of differential variations. CR is a real-valued crossover factor.
Generally, both F and CR affect the convergence rate and robust-
ness of the search process. Their optimal values are dependent
both on objective function characteristics and on the population
size, NP. Usually, suitable values for F, CR and NP can be found by
experimentation after a few tests using different values. Practical
advice on how to select control parameters NP, F and CR can be
found in Storn and Price (1995).
4.4. Selection
The selection scheme of DE also differs from other evolutionary
algorithms. On the basis of the current population P(T) and the tem-
porary population, the population of the next generation P(T+1) is
created as follows:
(
U Tþ1 if FðU Tþ1 Þ 6 FðX Ti Þ
X iTþ1 ¼ i i
ð6Þ
X Ti otherwise
Thus, each individual of the temporary or trial population is com-
pared with its counterpart in the current population. The one with
the lower value of cost-function F(X) will propagate the population
of the next generation. As a result, all individuals of the next gener-
ation are as good as or better than their counterparts in the current
generation (Storn & Price, 1995).
5. The proposed grouping based differential evolution
algorithm (GDE)
This section describes how DE meta-heuristic is modified to
solve the cell formation problem which is discrete in nature. The
original DE algorithm can only optimize problems in which the ele-
ments of the solution are continuous real numbers. Therefore, sev-
eral approaches have been used to deal with discrete optimization
by DE. Most of them round off the variable to the nearest available
value before evaluating each trial vector. In the following subsec-
tions we introduce a new DE algorithm adapted based on the struc-
ture of the MPCF problem. The new algorithm called GDE, owns a
ð3Þ mutation equations analogous to the classical DE’s mutation which
enables us to maintain all major characteristics of DE.
5.1. The encoding
The algorithm developed in this study (GDE) uses the encoding
strategy used in GGA (Falkenauer, 1992), namely the grouping
encoding. The grouping encoding scheme uses a variable length
chromosome (solution) that includes the items to be grouped
along with an additional section denoting the actual groups pres-
ent in the solution. Consider for example the individual ABCB that
encodes the solution where the first object is in group A, second in
ð4Þ
B, third in C, and fourth in B. The grouping encoding related to this
individual could be ABCBjBAC. Note that the order in which the
groups are listed does not matter (the order BAC) (Falkenauer,
1992). This representation is crucial to the design of the GDE, as
the modified operators for crossover and mutation are designed
to manipulate the group portion of the individuals.
ð5Þ
The encoding scheme used for the machine–part cell formation
(MPCF) problem is a natural adaptation of this strategy. The chro-
mosome representation consists of three sections: one represent-
ing the parts, one representing the machines, and the additional
group section that may be variable length (Brown & James,
2007). The individuals used for the MPCF problem can be repre-
sented as shown in (7) where pi denotes what group part i is as-
signed, for parts 1, . . . , P; mj denotes what group machine j is
assigned, for machines 1, . . . , M; and gk denotes the group numbers
for groups 1  K
 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829
p1 p2 p3 p4    pp jm1 m2 m3 m4    mM jg 1 g 2 g 3 g 4    g K
Considering the example given in Fig. 1b, the related solution can be
encoded as follows:
2 1 1 2 1j2 1 2 1 2j1 2:
The solution consist of two cells with the cell 1 containing parts
{2, 3, 5} and machines {2, 4}, and cell 2 containing parts {1, 4} and
machines {1, 3, 5}. Note that the part and machine portions of the
individuals are fixed in length based on the size of the problem.
For this example, there are five parts and five machines. The group
portion of the individuals can vary in length depending on the num-
ber of cells into which the machines and parts are grouped.
5.2. Initial solutions
Since the CF problem considers the grouping of parts and ma-
chines, an intuitive solution approach is to decompose the entire
problem into two subproblems dealing with parts assignment
and machine assignment, respectively. When parts assignment is
firstly determined, followed by a proper assignment of machines,
generation of an initial solution is hence completed (Wu et al.,
2007).
5.2.1. Parts assignment
Minkowski’s metric (Heragu, 1997) is used to evaluate dissimi-
larity between parts, as a measure to assign parts to cells. The dis-
similarity measure dij is defined as:
" #
XM
dij ¼ jaik  ajk j
k¼1
where

1 if part i requires processing on machine k
aik ¼
0 otherwise
and M is the total number of machines. After calculating the dissim-
ilarity value for each pair i and j of parts, we are now able to gener-
ate the initial parts assignment by using the following greedy rule:
the lower dissimilarity measure a pair of parts has, with the higher
priority, they should be placed in the same cell. This process is re-
peated until all parts being assigned to cells.
5.2.2. Machine assignment
We used the same strategy as above for assigning machines to
groups, and the dissimilarity measure is defined as:
" #
XP
dij ¼ jaki  akj j
k¼1
where
 got in the created individual (C).
1 if machine i has an operation on part k
aki ¼
0 otherwise
In this measure, P is the total number of parts. After calculating the
dissimilarity value for each pair of machines, we generate the initial
machine assignment.
The potential of any initial solution of (MPCF) depends on the
number of initial cells (L). Starting with large number of initial cells
would be undesirable. Thus we use a truncated geometric distribu-
tion to generate randomly the number of initial groups (Hussein-
zadeh Kashan, Karimi, & Jolai, 2006). Using a geometric
distribution to simulate the number of initial groups ensures that
the probability of starting with large number of groups would be
small, against the high probability for starting with less number
of groups. The following relation gives the random number of ini-
tial groups
4825
& ’
ð7Þ lnð1  ð1  ð1  pÞl ÞRÞ
L¼ þ 1; L 2 f1; 2; 3; . . . ; lg ð10Þ
lnð1  pÞ
where L is a random number of initial groups in an initial solution,
distributed by a truncated geometric distribution, R is a uniformly
distributed variable belonging to [0, 1], p is the probability of suc-
cess, and l, is the maximum number of cells. To generate an initial
individual, first the number of cells (L) is produced by (10) and parts
and machines are assigned accordingly.
5.3. The GDE mutation
The mutated parameter vector in GDE algorithm is generated
via equation (11) as follows:
 
V Tþ1
i ¼ X Tr3  X Tr1  X Tr2 ð11Þ
The definitions of the operators, used in the body of (11) are gi-
ven next.
The subtract operator ðÞ. Differences between two individuals
represented in form of (7) can be presented by an array of elements
in which each element shows that whether the content of the cor-
responding element in the first individual is different from the sec-
ond one or not. If yes, that element gets its value from the first
individual. Fig. 2 illustrates the manner in which  operator per-
forms. More precisely, the number of elements that have not the
same value in both A and B are equal to the Hamming distance be-
tween A and B. Worth to mention that when X Tr1 is exactly equal to
ð8Þ X Tr2 , we omit the terms ðX Tr1  X Tr2 Þ, since it is a null array.
To reassigns orphaned parts or machines (0’s in the array re-
sulted by  operator), a repair heuristic, given by Brown and
Sumichrast (2001), is utilized.
It is important to note that since we are developing an algo-
rithm which is capable to be applied on grouping problems, and
that, all operators employed in an algorithm modified to suit the
structure of grouping problems (i.e. a grouping algorithm) should
work with groups (cells, in our case) rather than items, thus it is
required that ðÞ operator be essentially a grouping operator
(examples of grouping operators are crossover and mutation oper-
ators in GGA). If we consider a group (cell) as a set of different parts
and different machines, then the result of ðÞ operator would be
equal to the result of applying the subtraction operator on two sets,
sequentially. Given the fact that the set subtraction operator can be
regarded as an operator which is applied on two sets (groups), thus
ðÞ would be a grouping opearator. For example in Fig. 2 group 1
from individual A contains parts {3, 4, 5} and machine {1} while
ð9Þ group 1 from individual B contains parts {3, 6} and machine {1}.
Applying the set subtraction operator, we have group 1 containing
parts {4, 5} and no machine and this is the same result as we have
The add operator ðÞ. This operator is a crossover operator that
typically is used in genetic algorithms. Crossover is viewed as the
operator that has the mission of interchanging structural informa-
tion developed during the search. The crossover operator utilized
in the current algorithm is consistent with the implementation de-
scribed by Falkenauer (1998). The crossover operator works like
X rT1 ) = 2 3 1 1 1 2 3 2 | 1 2 3 3 | 1 2 3.
X rT2 ) = 3 2 1 2 4 1 3 4 | 1 2 3 4 | 1 2 3 4.
B=23011202|0003|123.
Fig. 2. Illustration of the manner in which  operator performs.
4826 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829

two-point crossover (Seifiddini, 1989), but the cross-points sepa-
rate the group portion of the individual into segments, not the part
or machine portions. Specifically, the crossover takes the following
steps (Falkenauer, 1998):
By renumbering the groups, the resulting new child appears as:
C 1 ¼ 3 2 3 4 2 3 1j2 2 3 3 2 3 4 1 3 4 3j1 2 3 4 ð18Þ
This repair operator is applied to all orphaned machines or parts.

(1) Select two cross-points from the group portion of one of the 5.5. Local search
individuals.
(2) Inject the cross-section from this contributing individual The local search employed in the current algorithm is given by
into the other one. Goncalves and Resende (2004). It uses the partial efficacy given in
(3) Modify the assignment portion(s) of the individual to reflect Eq. (19). The partial efficacy, lc , is the efficacy value obtained by
the group assignments from the contributing individual. assigning either a part to a machine group or a machine to a part
(4) Apply a local problem-dependent heuristic in order to adapt family/group. In Eq. (19), e is the total number of 1s in the MP ma-
the resulting groups. trix, eo;c is the number of exceptions for that part/machine in rela-
tion to the machine group/part family being considered and ev ;c is
This crossover method does not simply switch two portions of the number of voids for that machine/part, also in relation to the
the individuals like traditional two-point crossover, but instead in- machine group/part family being considered (Brown & James,
serts groups from one individual into the other one. As step (4) of 2007)
crossover indicates, modifications of the resulting groups via some e  eo;c
repair heuristic may be necessary to create feasible individuals. De- lc ¼ ð19Þ
e þ ev ;c
tails of the problem-dependent heuristic employed by GDE are gi-
ven in Section 5.4. Taking the machine grouping of the incoming solution, first the par-
An illustrative example of the crossover operator, similar to that tial efficacy is calculated for each part using Eq. (19) and then each
was presented by Brown and James (2007), is as follows. Consider part is reassigned to a group with the largest partial efficacy value. If
the following two parental solutions for a seven parts and 11 ma- the modified solution is better than the original solution, the origi-
chines problem nal solution is replaced and the process is restarted with this time
taking the part families rather than the machine groupings. After
P 1 ¼ 3 1 1 1 2 3 2j1 2 3 3 2 3 1 2 3 1 1j1 2 3 calculating the partial efficiency for each machine, it is reassigned
ð12Þ
P 2 ¼ 2 4 2 3 1 2 4j1 1 2 2 1 2 3 4 2 3 2j1 2 3 4 to a part family with the largest partial efficacy value. The algorithm
iterates by returning to the further reassignment of the parts in the
The crossover begins by selecting two cross-points from the group
third iteration and further reassignment of the machines in the
portion of the second parent P2; for example, CP1 ¼ 0 and CP 2 ¼ 3
forth iteration and so on, until the quality of the new solution does
P2 ¼ 2 4 2 3 1 2 4j1 1 2 2 1 2 3 4 2 3 2j½1 2 3 4 ð13Þ not exceed the quality of the last solution. To illustrate, consider the
offspring, C1, created in the above. Fig. 4 gives the resulting partial
In this example, groups 1, 2 and 3 from second parent are inserted
efficacy values for the local search in which the parts are being con-
into first parent, which produces the offspring C1 after the corre-
sidered for reassignment. The partial efficacy matrix given in Fig. 4
sponding machine/part modifications are made to reflect the ma-
results in the following repaired individual:
chine/part assignments of the inserted groups from the
contributing parent. The modified values are underlined C 1 ¼ 4 2 3 4 4 3 1j2 2 3 3 2 3 4 1 3 4 3j1 2 3 4 ð20Þ
C 1 ¼ 2 1 2 3 1 2 2j1 1 2 2 1 2 3 2 2 3 2j1 2 3 1 2 3 ð14Þ
As can be seen, the original group 3 from P1 no longer contains
either parts or machines, so it can be eliminated, resulting in: P(i)/M(j) 1 2 3 4 5 6 7 8 9 10 11
C 1 ¼ 2 1 2 3 1 2 2j1 1 2 2 1 2 3 2 2 3 2j1 2 1 2 3 ð15Þ 1 1 0 1 0 0 0 1 0 0 0 1
2 1 1 0 0 0 1 0 0 0 0 0
The original group 1 from P1 contains one part (part 2) but no ma- 3 0 1 0 0 0 1 0 0 1 0 0
chines, and is thus infeasible. We employ a repair heuristic to place 4 0 0 0 1 1 0 0 0 0 1 0
part 2 into a group that contains at least one machine. 5 0 0 1 0 0 0 1 0 0 0 0
6 0 0 1 1 0 0 0 0 0 0 1
5.4. Repair heuristic 7 0 0 0 0 1 0 0 1 0 1 0

To fix infeasibilities arise when applying GDE operators, and Fig. 3. Machine–part incidence matrix.
reassign orphaned parts or machines, a repair heuristic given by
Brown and Sumichrast (2001), is adopted. Recalling the previous
example:
Group 2 Group 3 Group 4
C 1 ¼ 2 1 2 3 1 2 2j1 1 2 2 1 2 3 2 2 3 2j1 2 1 2 3 ð16Þ
Parts Required 8 1,2,5 3,4,6,9,11 7,10
Group 1 contains part 2, but no machines. Therefore, part 2 must be
Machines
reassigned to a new group. Considering the MP incidence matrix gi-
ven in Fig. 3, we locate all the machines part 2 needs, ma- 1 1,3,7,11 77.27 78.26 79.17 81.82
chines = {1, 2, 6}. The machine section of C1 is then examined to 2 1,2,6 81.82 90.91 76.00 78.26
locate in which groups machines 1, 2, and 6 are assigned. In this 3 2,6,9 81.82 82.61 83.33 78.26
example, machine 6 is in group 2, while machines 1 and 2 are in 4 4,5,10 81.82 82.61 76.00 86.36
group 1. As group 1 contains more of the required machines for part 5 3,7 86.36 79.17 80.00 90.91
2, part 2 is reassigned to group 1 and group 1 is eliminated, result- 6 3,4,11 81.82 75.00 91.30 78.26
ing in the following solution: 7 5,8,10 90.48 82.61 69.23 86.36
C 1 ¼ 2 1 2 3 1 2 2j1 1 2 2 1 2 3 2 2 3 2j2 1 2 3 ð17Þ
Fig. 4. Partial efficacy matrix.
Table 1
Computational results on 40 benchmarked problems.

#Problem source Size GGA GDE HGGA HGDE

Min. Avg. Max. Avg. Min. Avg. Max. Avg. Min. Avg. Max. Avg. Min. Avg. Max. Avg.
sol. sol. sol. time sol. sol. sol. time sol. sol. sol. time sol. sol. sol. time
1 King and Nakornchai (1982) 57 82.35 82.35 82.35 – 82.35 82.35 82.35 3.12 82.35 82.35 82.35 – 82.35 82.35 82.35 3.55
2 Waghodekar and Sahu (1984) 57 60.87 66.22 69.57 – 69.57 69.57 69.57 3.35 69.57 69.57 69.57 – 69.57 69.57 69.57 3.55
3 Seifiddini (1989) 5  18 57.53 73.39 79.59 – 79.59 79.59 79.59 3.86 79.59 79.59 79.59 – 79.59 79.59 79.59 3.96
4 Kusiak and Cho (1992) 68 76.92 76.92 76.92 – 76.92 76.92 76.92 3.79 76.92 76.92 76.92 – 76.92 76.92 76.92 4.04
5 Kusiak and Cho (1987) 7  11 56.00 58.27 60.87 – 60.87 60.87 60.87 4.52 60.87 60.87 60.87 – 60.87 60.87 60.87 5.13
6 Boctor (1991) 7  11 70.37 70.60 70.83 – 70.83 70.83 70.83 4.39 70.83 70.83 70.83 – 70.83 70.83 70.83 4.83
7 Seifiddini and Wolf (1986) 8  12 63.83 67.02 69.44 – 69.44 69.44 69.44 4.64 69.44 69.44 69.44 – 69.44 69.44 69.44 4.99
8 Chandrasekharan and Rajagopalan 8  20 66.67 79.50 85.25 – 85.25 85.25 85.25 5.08 85.25 85.25 85.25 – 85.25 85.25 85.25 5.30
(1986a)
9 Chandrasekharan and Rajagopalan 8  20 52.08 54.68 55.32 – 58.72 58.72 58.72a 4.44 58.72 58.72 58.72 – 58.72 58.72 58.72 5.11
(1986b)

A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829

10 Mosier and Taube (1985a) 10  10 72.41 74.48 75.00 – 75.00 75.00 75.00 4.91 75.00 75.00 75.00 – 75.00 75.00 75.00 5.07
11 Chan and Milner (1982) 10  15 92.00 92.00 92.00 – 92.00 92.00 92.00 4.82 92.00 92.00 92.00 – 92.00 92.00 92.00 5.00
12 Askin and Subramanian (1987) 14  24 67.09 69.93 72.06 – 72.06 72.06 72.06 6.44 72.06 72.06 72.06 – 72.06 72.06 72.06 6.68
13 Stanfel (1985) 14  24 68.12 70.05 71.83 – 70.13 71.17 71.83 6.36 71.83 71.83 71.83 – 71.62 71.79 71.83 6.69
14 McCormick et al. (1972) 16  24 42.72 48.35 51.58 – 53.26 53.47 53.85a 6.89 52.75 52.75 52.75 – 53.33 53.37 53.41a 7.31
15 Srinivasan et al. (1990) 16  30 61.11 66.42 68.61 – 67.65 68.66 68.99a 7.16 68.99 68.99 68.99 – 68.99 68.99 68.99 7.53
16 King (1980) 16  43 47.48 51.72 55.48 – 57.06 57.40 57.53a 8.25 57.53 57.53 57.53 – 57.43 57.51 57.53 8.81
17 Carrie (1973) 18  24 50.00 53.47 57.43 – 57.06 57.42 57.73a 7.42 57.43 57.70 57.73 – 57.73 57.73 57.73 7.90
18 Mosier and Taube (1985b) 20  20 36.36 38.65 40.74 – 41.79 42.67 43.26a 6.89 42.74 42.93 43.18 – 42.42 42.95 43.45a 7.53
19 Kumar et al. (1986) 20  23 41.77 45.83 49.65 – 48.32 49.44 50.81a 7.31 50.81 50.81 50.81 – 50.81 50.81 50.81 7.94
20 Carrie (1973) 20  35 72.15 75.49 77.02 – 77.58 77.81 77..91a 7.90 77.91 77.91 77.91 – 77.91 77.91 77.91 8.43
21 Boe and Cheng (1991) 20  35 53.93 55.72 57.14 – 57.59 57.82 57.98a 8.14 57.98 57.98 57.98 – 57.98 57.98 57.98 8.72
22 Chandrasekharan and Rajagopalan 24  40 96.18 99.62 100.00 – 100.00 100.00 100.00 9.25 100.00 100.00 100.00 – 100.00 100.00 100.00 9.64
(1989)
23 Chandrasekharan and Rajagopalan 24  40 82.01 84.08 85.11 – 85.11 85.11 85.11 9.25 85.11 85.11 85.11 – 85.11 85.11 85.11 9.65
(1989)
24 Chandrasekharan and Rajagopalan 24  40 67.97 71.27 73.51 – 73.51 73.51 73.51 9.26 73.51 73.51 73.51 – 73.51 73.51 73.51 9.76
(1989)
25 Chandrasekharan and Rajagopalan 24  40 45.27 50.70 52.41 – 52.56 53.22 53.29a 9.51 53.15 53.27 53.29 – 53.29 53.29 53.29 10.28
(1989)
a
26 Chandrasekharan and Rajagopalan 24  40 43.42 45.14 46.67 – 47.71 47.88 48.95 9.70 48.55 48.66 48.95 – 48.61 48.75 48.95 10.73
(1989)
a
27 Chandrasekharan and Rajagopalan 24  40 41.43 43.32 45.27 – 41.88 44.55 47.26 9.96 46.90 47.16 47.26 – 46.85 47.16 47.26 10.46
(1989)
28 Chandrasekharan and Rajagopalan 27  27 43.79 48.65 52.53 – 54.13 54.48 54.82a 8.25 53.41 53.84 54.02 – 54.82 54.82 54.82a 8.97
(1989)
29 Kumar and Vannelli (1983) 30  41 56.52 59.57 61.39 – 59.49 61.51 62.59a 10.61 62.24 62.74 63.31 – 62.42 62.85 63.31 11.87
30 Stanfel (1985, Fig. 5) 30  50 54.49 55.96 57.95 – 54.49 56.70 58.89a 12.76 59.77 59.77 59.77 – 59.77 59.77 59.77 13.75
31 McCormick et al. (1972) 37  53 43.23 49.27 52.47 – 58.03 59.38 60.40a 12.90 60.48 60.57 60.64 – 60.48 60.59 60.64 15.39
32 Chandrasekharan and Rajagopalan 16  30 34.84 38.06 45.77 4.70 51.18 51.64 52.07a 6.70 52.05 52.05 52.05 7.35 52.07 52.17 52.29a 7.68
(1989)
33 Chandrasekharan and Rajagopalan 16  30 51.09 56.30 61.16 4.33 61.84 62.19 63.04a 6.36 62.99 62.99 62.99 7.37 62.29 63.03 63.04a 7.60
(1989)
34 Chandrasekharan and Rajagopalan 16  30 58.33 61.95 64.81 4.45 68.38 68.38 68.38a 5.57 68.38 68.38 68.38 7.14 68.38 68.38 68.38 7.41
(1989)
a a
35 Chandrasekharan and Rajagopalan 16  30 46.13 46.60 48.18 4.57 48.19 49.16 50.00 6.79 49.29 49.60 49.65 7.48 49.62 49.75 50.00 7.71
(1989)
36 Chandrasekharan and Rajagopalan 16  30 51.41 62.78 70.54 4.55 72.37 72.37 72.37a 5.47 72.37 72.37 72.37 7.08 72.37 72.37 72.37 7.30
(1989)

(continued on next page)

4827
4828 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829

The local search is applied to 10% of generated individuals.

Based on preliminary computations, we also choose the following

7.34

8.43
7.41

9.80
setting for parameters: the population size is considered to be

time
Avg.
100; the number of generations is considered to be 50; and CR is
equal to 1.

73.24

72.37
77.31

77.30
Max.
sol.
6. Computational results

73.24

72.37
77.31

77.00
Avg.

In this section, 40 test problems taken from literature are used

sol.

to evaluate the computational characteristics of the proposed heu-

73.24

76.51

72.37
77.31

ristic with and without local search. Results are compared with
HGDE
Min.
sol.

GGA algorithm of Brown and James (2007). The proposed algo-

rithm was coded in Matlab7.4 and implemented on a laptop com-
puter with 2.5 GHZ CPU speed and 2 GB of main memory.
time

7.12

8.13

9.54
Avg.

7.10

On the first 31 problems we just state the results obtained by

GGA as reported in Brown and James (2007). For the further nine
problems we report the results obtained by GGA based on our
73.24

72.37
77.31

77.30
Max.

implementation of this algorithm. Computational results on all

sol.

40 problems are given in Table 1. The worst solution, average solu-

tion, and best solution found in 10 runs are presented for both
73.24

72.37
77.31

76.60
Avg.

algorithms. Average computational time in seconds is also re-

sol.

ported. It can be verified that GDE outperforms GGA in terms of

73.24

76.51

72.37
77.31

solution quality, on 23 problems. For the remaining 17 problems,

HGGA
Min.
sol.

GDE matches the solution quality obtained by GGA. Furthermore,

the hybridized version of the GDE, called HGDE, is shown to out-
perform the HGGA in terms of solution quality on problems 14,
5.53

6.65

7.72
5.61
time
Avg.

18, 28, 32, 33 and 35. For the remaining problems, HGDE matches
the solution quality obtained by HGGA.
73.24a

72.37
77.31

77.30
Max.
sol.

7. Conclusions
77.31

73.24

76.64

72.37

In this paper, we present a new solution approach for the ma-

Avg.
sol.

chine–part cell formation problem based on the grouping repre-

sentation used in the grouping genetic algorithm. We proposed
77.31

73.24

75.74

72.37

an equation analogous to that of the mutation equations in the

Min.
GDE

sol.

classical differential evolution algorithm (DE) yielding a grouping

version of DE (GDE).
Results obtained from extended computational efforts, justifies
time
4.52

4.46

4.86

6.12
Avg.

the competitive performance of GDE compared to a recently pro-

posed algorithm, i.e., the grouping genetic algorithm (GGA). We
77.31

71.43

76.19

72.37

found that GDE algorithm outperforms GGA in terms of solution

Max.
sol.

quality in 23 out of 40 benchmarked problem instances. For the

rest of problems both algorithms provide the same results. We also
compared the hybridized version of GDE and GGA (HGDE and
69.44

63.55

66.64

65.58
Avg.
sol.

HGGA, respectively), where the same achievement was observed.

However, there were only six problems on which HGDE outper-
formed HGGA.
67.23

56.25

58.89

61.38
Min.
GGA

sol.

For future research the extension of our approach could be

investigated for other type of grouping problems, e.g. bin packing
16  30

16  30

24  40
20  35

problem, graph coloring problem, etc.

Size

Denotes a case where GDE outperforms GGA.

References
and Rajagopalan

and Rajagopalan

and Rajagopalan

and Rajagopalan

Aljaber, N., Baek, W., & Chen, C. (1997). A tabu search approach to the cell formation
problem. Computers and Industrial Engineering, 32, 169–185.
Askin, R. G., & Chiu, K. S. (1990). A graph partitioning procedure for machine
assignment and cell formation in group technology. International Journal of
Production Research, 28, 1555–1572.
Askin, R. G., & Subramanian, S. P. (1987). A cost-based heuristic for group
technology configuration. International Journal of Production Research, 25,
Chandrasekharan

Chandrasekharan
Chandrasekharan

Chandrasekharan
#Problem source

101–113.
Askin, R. G., & Vakharia, A. J. (1991). Group technology – Cell formation and
Table 1 (continued)

operation. In D. I. Cleland & B. Bidanda (Eds.), The automated factory handbook:

(1986a)

Technology and management (pp. 317–336). New York: TAB Books.

(1989)
(1989)

(1989)

Ballakur, A., & Steudel, H. J. (1998). A within–in cell utilization based heuristic for
designing cellular manufacturing system. International Journal of Production
Research, 25, 639–655.
37

38

39

40

Boctor, F. F. (1991). A linear formulation of the machine–part cell formation

a

problem. International Journal of Production Research, 29, 343–356.

 A. Noktehdan et al. / Expert Systems with Applications 37 (2010) 4822–4829
Boe, W. J., & Cheng, CH. (1991). A close neighbor algorithm for designing cellular
manufacturing systems. International Journal of Production Research, 29(10),
2097–2116.
Boulif, M., & Atif, K. (2006). A new branch-and-bound-enhanced genetic algorithm
for the manufacturing cell formation problem. Computers and Operation
Research, 33, 2219–2245.
Brown, E., & James, T. (2007). A hybrid grouping genetic algorithm for the cell
formation problem. Computers and Operation Research, 34, 2059–2079.
Brown, E., & Sumichrast, R. (2001). CF-GGA: A grouping genetic algorithm for the
cell formation problem. International Journal of Production Research, 36,
3651–3669.
Carrie, A. S. (1973). Numerical taxonomy applied to group technology and plant
layout. International Journal of Production Research, 399–416.
Chandrasekharan, M. P., & Rajagopalan, R. (1986a). MODROC: An extension of rank
order clustering for group technology. International Journal of Production
Research, 24, 1221–1233.
Chandrasekharan, M. P., & Rajagopalan, R. (1986b). An ideal seed non-hierarchical
clustering algorithm for cellular manufacturing. International Journal of
Production Research, 24, 451–463.
Chandrasekharan, M. P., & Rajagopalan, R. (1989). GROUPABILITY: An analysis of the
properties of binary data matrices for group technology. International Journal of
Production Research, 27, 1035–1052.
Chan, H. M., & Milner, D. A. (1982). Direct clustering algorithm for group formation
in cellular manufacturing. Journal of Manufacturing Systems, 1, 65–74.
Cheng, C., Gupa, Y., & Lee, W. A. (1998). TSP-based heuristic for forming machine
groups and part families. International Journal of Production Research, 36,
1325–1337.
Dimopoulos, C., & Mort, N. (2001). A hierarchical clustering methodology based on
genetic programming for the solution of simple cell-formation problems.
International Journal of Production Research, 39, 1–19.
Falkenauer, E. (1992). The grouping genetic algorithms-widening the scope of the
GAs. Belgian Journal of Operation Research, Statistics and Computer Science, 33–79.
102.
Falkenauer, E. (1998). Genetic algorithm for grouping problems. New York: Wiley.
Goncalves, J., & Resende, M. (2004). An evolutionary algorithm for manufacturing
cell formation. Computers and Industrial Engineering, 47, 247–273.
Heragu, S. S. (1997). Facility design. Boston, MA: PWS Publishing Company.
Husseinzadeh Kashan, A., Karimi, B., & Jolai, F. (2006). Effective hybrid genetic
algorithm for minimizing makespan on a single-batch-processing machine with
non-identical job sizes. International Journal of Production Research, 44,
2337–2360.
Joines, J., Culbrethe, C., & King, R. (1996). Manufacturing cell design: An integer
programming model employing genetic algorithms. IEE Transactions, 28, 69–85.
King, J. R. (1980). Machine-component grouping in production flow analysis: An
approach using a rank order clustering algorithm. International Journal of
Production Research, 213–232.
King, J. R., & Nakornchai, V. (1982). Machine-component group formation in group
technology: Review and extension. International Journal of Production Research,
20(2), 117–131.
Kumar, C., & Chandrasekharan, M. P. (1990). Grouping efficacy: A quantitative
criterion for goodness of block diagonal forms of binary matrix in group
technology. International Journal of Production Research, 28, 233–243.
Kumar, K. R., Kusiak, A., & Vannelli, A. (1986). Grouping of parts and components in
flexible manufacturing systems. European Journal of Operational Research, 24,
387–397.
Kumar, K. R., & Vannelli, A. (1983). Strategic subcontracting for efficient
disaggregated manufacturing. International Journal of Production Research, 25,
1715–1728.
Kusiak, A., & Cho, W. S. (1987). Efficient solving of the group technology problem.
Journal of Manufacturing system, 117–124.
4829
Kusiak, A., & Cho, M. (1992). Similarity coefficient algorithm for solving the group
technology problem. International Journal of Production Research, 30(11),
2633–2646.
Lei, D., & Wu, Z. (2005). Tabu search approach based on a similarity coefficient for
cell formation in generalized group technology. International Journal of
Production Research, 19, 4035–4047.
McCormick, W. T., Schweitzer, P. J., & White, T. W. (1972). Problem decomposition
and data reorganization by a clustering technique. Operation Research, 20,
993–1009.
Mosier, C. T., & Taube, L. (1985a). The facets of group technology and their impact on
implementation. OMEGA, 13, 381–391.
Mosier, C. T., & Taube, L. (1985b). Weighted similarity measure heuristics for the
group technology machine clustering problem. OMEGA, 13, 577–583.
Onwubolu, G. C., & Mutingi, M. (2001). A genetic algorithm approach to cellular
manufacturing systems. Computers and Industrial Engineering, 39, 125–144.
Purcheck, G. F. K. (1975). A linear programming method for the combinatorial
grouping of an incomplete power set. Journal of Cybernetics, 51–76.
Rajamani, D., Singh, N., & Aneja, Y. P. (1992). Selection of parts and machines for
cellularization: A mathematical programming approach. European Journal of
Operational Research, 62, 47–54.
Reisman, A., Kumar, A., Motwani, J., & Cheng, C. H. (1997). Cellular manufacturing: A
statistical review of the literature (1965–1995). Operations Research, 45,
508–520.
Sarker, B. (2001). Measures of grouping efficacy in cellular manufacturing systems.
European Journal of Operation Research, 130, 588–611.
Seifiddini, H. (1989). A note on the similarity coefficient method and the problem of
improper machine assignment in group technology applications. International
Journal of Production Research, 27, 1161–1165.
Seifiddini, H., & Wolf, P. M. (1986). Application of the similarity coefficient method
in group technology. IIIE Transaction.
Shafer, S. M., & Kern, G. (1992). A Mathematical programming approach for dealing
with exceptional elements in cellular manufacturing. International Journal of
Production Research, 30, 1029–1036.
Shafer, S. M., & Rogers, D. F. (1991). A goal programming approach to cell formation
problems. Journal of Operations Management, 10, 28–43.
Shtub, A. (1989). Modeling group technology cell formation as a generalized
assignment problem. International Journal of Production Research, 27, 775–782.
Srinivasan, G., Narendran, T., & Mahadevan, B. (1990). An assignment model for the
part-families problem in group technology. International Journal of Production
Research, 28, 145–152.
Stanfel, L. E. (1985). Machine clustering for economic production. Engineering Costs
and Production Economies, 9, 73–81.
Storn, R., & Price, K. (1995). Differential evolution – A simple and efficient adaptive
scheme for global optimization over continues spaces, Technical Report TR (pp.
95–102).
Sun, D., Lin, L., & Batta, R. (1995). Cell formation using tabu search. Computers and
Industrial Engineering, 28, 485–494.
Vin, E., Delit, P., & Delchamber, A. (2005). A multiple-objective grouping genetic
algorithm for the cell formation problem with alternative routings. Journal of
Intelligent Manufacturing, 16, 189–205.
Waghodekar, P. H., & Sahu, S. (1984). Machine-component cell formation in group
technology MACE. International Journal of Production Research, 22, 937–948.
Wei, J. C., & Gaither, N. (1990). A capacity constrained multi-objective cell formation
method. Journal of Manufacturing Systems, 9, 222–232.
Wemmerlov, U., & Hyer, N. (1987). Research issues in cellular manufacturing.
International Journal of Production Research, 25, 413–431.
Wemmerlov, U., & Hyer, N. L. (1989). Cellular manufacturing in the US industry: A
survey of users. International Journal of Production Research, 27(9), 1511–1530.
Wu, T., Chang, C., & Chung, S. (2007). A simulated annealing algorithm for
manufacturing cell formation problems. Expert System with Applications.