European Journal of Operational Research 185 (2008) 563–592

www.elsevier.com/locate/ejor

Production, Manufacturing and Logistics

A hybrid simulated annealing for solving an extended

model of dynamic cellular manufacturing system
N. Safaei *, M. Saidi-Mehrabad, M.S. Jabal-Ameli
Department of Industrial Engineering, Iran University of Science and Technology, P.C. 16844, Narmak, Tehran, Iran

Received 14 August 2006; accepted 22 December 2006

Available online 16 February 2007

Abstract

This paper develops a mixed-integer programming model to design the cellular manufacturing systems (CMSs) under
dynamic environment. In dynamic environment, the product mix and part demand change under a multi-period planning
horizon. Thus, the best designed cells for one period may not be efficient for subsequent periods and reconfiguration of
cells is required. Reconfiguration may involve adding, removing or relocating machines; it may also involve a change in
processing rout of part types from a period to another. The advantages of the proposed model are as follows: considering
the batch inter/intra-cell material handling by assuming the sequence of operations, considering alternative process plans
for part types, and considering machine replication. The main constraints are maximal cell size and machine time-capacity.
The objective is to minimize the sum of the machine constant and variable costs, inter- and intra-cell material handling,
and reconfiguration costs. An efficient hybrid meta-heuristic based on mean field annealing (MFA) and simulated anneal-
ing (SA) so-called MFA–SA is used to solve the proposed model. In this case, MFA technique is applied to generate a good
initial solution for SA. The obtained results show that the quality of the solutions obtained by MFA–SA is better than
classical SA, especially for large-sized problems.
 2007 Elsevier B.V. All rights reserved.

Keywords: Dynamic cellular manufacturing systems; Mixed-integer programming; Mean field annealing; Simulated annealing

1. Introduction

Cellular manufacturing is described as a manufacturing procedure which produces part families within a
single line or cell of machines serviced by operators and/or robots that function only within the line or cell.
Cell formation (CF) as the first and most important step in the designing of a cellular manufacturing system
consists of two fundamental tasks: part-family formation and machine-cell formation. Part-family formation
is to group parts with similar geometric characteristics or processing requirements to take advantage of their
similarities for the design or manufacturing purpose. Because of increasing variety of consumer goods and
decrease in product life cycles, manufacturing organizations often face fluctuations in product demand and

*
Corresponding author. Tel.: +98 9111751379; fax: +98 1712240855.
E-mail addresses: [email protected], [email protected] (N. Safaei).

0377-2217/$ - see front matter  2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2006.12.058
564 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

M1 M6

M3

M2
M2

M6
M5
M5
M4 Cell 1 Cell 2 Cell 3
Period 1

Inactive cell
M6 M4

M1
M3

Idle

M5
M5

M6
machine

M2
M3

Cell 1 Cell 2 Cell 3

Period 2

Fig. 1. A schema of machine relocation and changing the number of cells in a DCMS.

product mix leading to a dynamic or turbulent production environment (Rheault et al., 1995). In dynamic pro-
duction environments, a multi-period planning horizon is considered in which product mix and/or part
demand rate in each period may be different such as seasonal products demand. As frequently discussed in
literature, traditional CMS has many operational advantages over other manufacturing systems such as job
shop and flow shop, but there are also some disadvantages such as reducing shop flexibility and machine uti-
lization (Rheault et al., 1995). Traditional CMS ignores any changes in demand over time from product rede-
sign and other factors. It assumes that product mix and part demand is constant for the entire planning
horizon. Product mix refers to a set of part types to be produced at each period. In dynamic environment,
a planning horizon can be divided into smaller periods where each period has different product mix and
demand requirements. Consequently, the formed cells in a current period may not be optimal and efficient
for the next period. To overcome disadvantages of traditional CMS, the concept of dynamic cellular manu-
facturing system (DCMS) is introduced (Rheault et al., 1995). DCMS implicates to reconfiguration of man-
ufacturing cells includes part families and machine groups at each period. Reconfiguration involves swapping
existing machines between cells called machine relocation; adding new machines to cells include machine rep-
lication, and removing existing machines from cells. A schema of DCMS includes machine relocation for two
consecutive periods is shown in Fig. 1. In this figure, it is assumed that the maximum cell size is equal to four.
Because of the processing requirements, machine 1 must be relocated from cell 1 in period 1 to cell 2 in period
1, machine 6 from cell 3 in period 1 to cell 1 in period 2, and machine 5 from cell 3 in period 1 to cell 2 in
period 2. As a result, cell 3 will become empty in period 2 and called inactive cell. Thus, the number of cells
at two successive periods is different. On the other hand, machine 6 is idle in cell 2 in period 2. Therefore,
because of the maximal cell size limitation for cell 2, machine 6 must be relocated from cell 2 in period 1
to another cell (say cell 3), and afterward machine 1 is replaced by machine 6 in cell 2. Actually, cell 3 in period
2 can play role of the idle machines storage.

2. Literature review

The concept of DCMS was introduced by Rheault et al. (1995). Previous investigations have been done in
the content of modelling and methodology of the cell formation problem in dynamic environment by assum-
ing stable cells, i.e., without cell layout change. For instance, Harhalakis et al. (1990) applied a two-stage
design approach to obtain a cellular design with the minimum expected inter-cell movement cost over a system
design horizon which was broken down into elementary time periods. Vakharia and Kaku (1993) incorporated
long-term demand changes into their 0–1 mathematical programming cell design method by reallocating parts
to families to regain the benefits of cellular manufacturing systems. Similarly, new parts are allocated to exist-
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 565

ing cells. So, cells are not rearranged in their multi-period design. Kannan and Ghosh (1995) studied DCMS
from scheduling point of view and showed simplified scheduling in cell based production systems by DCMS.
Song and Hitomi (1996) developed a methodology to design flexible manufacturing cells. The method is to
integrate production planning and cellular layout in a long-run planning horizon. Their methodology can
determine the production quantity for each part and timing of adjusting the cellular layout with a dynamic
demand in a finite planning horizon period. Their objectives were to minimize the sum of inventory-holding
cost, group setup cost, material handling cost, and layout adjusting cost. Drolet et al. (1996) proposed a four-
stage methodology for designing DCMS from scheduling and operating points of view. Their objectives were
to minimize the machine constant and reconfiguration costs by assuming space capacity limitation. Marcoux
et al. (1997) show that DCMS is more efficient than classical CMS in terms of throughput time, machine uti-
lization and WIP criteria. Askin et al. (1997) proposed a four-stage algorithm that designs cells to handle var-
iation in the product mix. Initially, a mathematical programming-based method is used to assign operations to
machine types. Subsequent phases allocate operation-parts to specific machines, identify manufacturing cells
and improve the design. Experiments are also conducted to evaluate the effects of factors such as utilization
and maximum cell sizes on the effectiveness of the algorithm. Again, cells once designed are expected to remain
unchanged during the planning horizon.
Wilhelm et al. (1998) proposed a multi-period formation of the part family and machine cell formation
problem. Their objectives were to minimize the sum of inter-cell movement cost, investment in additional
machines, and cost of reconfiguration over the planning horizon. Chen (1998) modelled the DCMS with
the aim of minimizing reconfiguration and machine constants (operating and holding) costs by assuming
upper and lower bounds of cell size limitation and constant number of cells. He also proposed some of the
extensions for his proposed model. Wicks and Reasor (1999) modelled the DCMS aiming at minimizing
reconfiguration and machine constant costs by assuming lower bound of cell size and machine capacity lim-
itations. They considered machine replication and operation sequence within their model. Harlahakis et al.
(1994) also considered product demand changes during a multi-period planning horizon. Their design is based
on robustness, i.e. designing a cellular configuration that is effective over the ranges of expected demand over
multiple periods. Thus, once the cells are designed they are expected to remain unchanged during the multi-
period horizon. Balakrishnan and Cheng (2005) proposed a two-stage procedure based on the generalized
machine assignment problem and dynamic programming for cell formation problem under conditions of
changing product demand. The objectives were to minimize the material handling and machine relocation
(shifting) costs. They proposed in changing the cellular configuration periodically when the cost-benefit anal-
ysis favours such a move. In this way, the cellular layout will be better suited to the demand in each period and
thus be more effective and agile during the planning horizon. In addition, they proposed examining multiple
layouts when considering cell redesign in order to incorporate different qualitative and quantitative consider-
ations. Such an analysis can also highlight the need for easily movable machines to ensure that cellular man-
ufacturing is effective throughout the planning horizon. Tavakkoli-Moghaddam et al. (2005) developed the
model proposed by Chen with additional assumptions as: alternative process plans, sequence of operations,
machine capacity limitation and machine replication. In addition, the developed model assumes that the
inter-cell movements perform in batch with same cost. Their objectives were to minimize the sum of machine
constant/variable costs, inter-cell movement cost and reconfiguration cost. Saidi-Mehrabad and Safaei (in
press) developed the model proposed by Tavakkoli-Moghaddam et al. (2005) in which the number of formed
cells at each period may change (i.e., cell number flexibility) and then the developed model is solved by a neu-
ral network based heuristic. Jeon and Leep (2006) developed a methodology which can be used to form man-
ufacturing cells using both a new similarity coefficient based on the number of alternative routes during
machine failure and demand changes for multiple periods. Their methodology is divided into two phases.
Using a new similarity coefficient, which considers the number of available alternative routes when available
during machine failure, is suggested as Phase I. In Phase II, machines are assigned to part families by using a
mixed-integer programming approach for a given period to minimize the total costs related to the scheduling
and operational aspects including inventory holding, earliness/lateness penalty and machine investment costs.
Table 1 shows a brief review of works done in context of DCMS.
There is a growing interest in computational models inspired in physical and natural phenomena such as
quantum computing, DNA computing, evolutionary algorithms, and chaotic networks. This, together with
566 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Table 1
Brief review of DCMS literature
No. Author(s) Year Objectives (+InterMHC) Constraints Explanation and advantages
(+Cell size)
1 Harhalakis et al. 1990 A two stages approach, stable cells
2 Vakharia and Kaku 1993 Stable cells
3 Harlahakis et al. 1994 Stable cells + probabilistic demand
4 Rheault et al. 1995 RfC Introducing DCMS
5 Kannan and Ghosh 1995 Scheduling aspect
6 Song and Hitomi 1996 IC + SC + RfC Determining production quantity
7 Drolet et al. 1996 MCC1 + RfC SCC A four stages approach
8 Marcoux et al. 1997 TT + MU + WIP They show DCMS is more efficient
than CCMS
9 Askin et al. 1997 MVC + MCC1 Four stages approach + MAC
10 Wilhelm et al. 1998 RfC + MCC1
11 Chen 1998 RfC + MCC1
12 Wicks and Reasor 1999 RfC + MCC1 MCC2 + LPa + LMb OS + MR
13 Balakrishnan and 2005 RfC A two stages approach
Cheng
14 Tavakkoli- 2005 RfC + MCC1 + MVC MCC2 APP + OS + MR
Moghaddam et al.
15 Saidi-Mehrabad In RfC + MCC1 + MVC MCC2 APP + OS + MR + CNF
and Safaei press
16 Jeona and Leep 2006 IC + MVC + MCC1 + BC + MCC2 Two stages
earliness/tardiness penalty approach + MR + MF + APP + OS
(scheduling aspect)
17 Current research – RfC + MCC1 + MVC + IntraMHCl MCC2 APP + OS + MR + CNF
(only operational aspects)
APP = Alternative process plan OS = Operation sequence
BC = Budge constraint IC = Inventory cost
MAC = Machine available capability InterMHC = Inter-cell material handling cost
MCC1 = Machine constant cost IntraMHC = Inter-cell material handling cost
MCC2 = Machine capacity constraint RfC = Reconfiguration cost
MR = Machine replication/duplication SC = Setup cost
MF = Machine failure SCC = Space capacity constraint
MVC = Machine variable cost TT = Throughput time
MU = Machine utilization WIP = Work in process
CNF = Cell number flexibility
a
Minimum number of parts within each family.
b
Lower bound of cell size.

the practical successes attained by artificial neural networks, has raised some intriguing theoretical questions
on the links between discrete and analogy computation. Is there a relationship between the complexity of algo-
rithms in the continuous and discrete settings for solving the same problem? How do approximate solutions
affect the complexity? Computational complexity theory, as developed within computer science, has focused
on discrete computation, whereas analogy computation has mainly been studied using the tools of theoretical
physics. However, to try to answer the above questions, theories to study both types of computation under a
unifying perspective are needed, and these are only beginning to be developed. Meanwhile, several approaches
have been introduced to explore the links between discrete and continuous solutions to the same problem.
Mean field annealing (MFA) is one such approach.
MFA approach is proposed by Peterson and Anderson (Peterson and Anderson, 1987, 1988) and is devel-
oped by Peterson and Sodeberg (Peterson and Sodeberg, 1989) for solving the combinatorial optimization
problems. MFA combines the characteristics of neural networks, such as the Hopfield network, with the
annealing process of simulated annealing (SA). This approach is introduced as a new method to speed up
the Boltzmann machine learning algorithm. Originally, this method is derived from the well-known mean-field
theory approximation in quantum for spin systems. The technique is said to be superior to SA as it approx-
imates solutions deterministically rather than performing the slow stochastic hill-climbing search and, as a
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 567

result, leads to a faster convergence to its solution. However, it has been reported that the MFA algorithm is
much faster than the SA algorithm while they have the same quality of solutions (Bilbro et al., 1989; Van Den
Bout and Miller, 1990). MFA can be represented as an intrinsically parallel algorithm, which can be mapped
to a set of nonlinear differential equations called mean field equations. In MFA, discrete variables, called
spins, are used for encoding the combinatorial optimization problem. An energy function written in terms
of spins is used for representations of the objective function of problem. Then, using the expected values a
gradient descent type relaxation schema is used to find a configuration of the spins, which minimizes the asso-
ciated energy function. There have been only two studies cell forming by MFA in the literature: Lozano et al.
(1993) used MFA to cell forming with the aim of minimizing the intra-cell and inter-cell movement costs.
Their approach used information on the sequence of operations of different part types. Saidi-Mehrabad
and Safaei (in press) used a neural approach based on MFA and a self-organizing neural network introduced
by (Guerrero et al., 2002) to solve a developed model of DCMS (Table 1, research No. 15). In their approach a
heuristic interactive approach is used for updating spins stead the mean field equations. Thus, energy function
dose not have a principal role in convergence of spins. The obtained results show that the average Gap %
between optimal solution and their neural approach is nearly 16% that is not a promising result, but their
study is important from Hopfield network design point of view. Moreover, MFA has been successfully applied
to several practical combinatorial optimization problems such as graph partitioning (Van Den Bout and
Miller, 1990), Travelling Salesman Problem (Chong and Lee, 1992), satellite broadcasting schedules (Ansari
et al., 1995), object recognition (Kim et al., 2001) and cell placement in VLSI circuit design (Aykanat
et al., 1998). In contrast to other studies in the context of MFA, Bofill et al. (2003) described an experimental
comparison between SA and MFA as applied to the generation of balanced incomplete block designs
(BIBDs). They showed SA performed slightly better than MFA, also, difficult problems for SA were also dif-
ficult for MFA and easy problems for SA were also easy for MFA. As a result, there is no a certain opinion
about the performance of MFA and the quality of its solution in literature.
The work done by Chen (research No. 11 in Table 1) is one of the best works from operational aspect point
of view. He modelled the DCMS with the aim of minimizing the reconfiguration costs and machine operating/
holding costs by assuming upper and lower bounds of cell size and constant number of cells. Also, he sug-
gested some of extensions for his own proposed model. Thus, as mentioned earlier, we extended the Chen’s
problem by additional assumptions in our pervious works (research Nos. 14 and 15). Current paper develops
the mentioned previous works by considering simultaneous inter- and intra-cell material handling with differ-
ent batch sizes per part type and different unit costs. For instance, we can suppose that the part types are
shifted within cells by a robot or manpower and between cells by a conveyor, crane, truck, or AGV. Obvi-
ously, in this case, the capacity and unit cost of intra-cell movement are less than that ones in inter-cell move-
ment. We construct above advantages without adding any extra decision variable with respect to the previous
researches where the sequence of operations is directly considered in the inter/intra-cell material handling.
Intra-cell material handling is one of the most important operational aspects of CMS has not been considered
in previous works in the context of DCMS. Fig. 2 shows the direct effects of the sequence of operations on the
material handling. For example, operations 1 and 3 of part 11 must be performed on machine 7 in cell 1, also,
its operation 2 must be performed on machine 5 in cell 2. Thus, to observe the sequence of operations, the
processing route for part 11 requires two inter-call movements from cell 1 to cell 2 and vice versa. By a similar
reasoning, the processing route for part 3 in cell 2 requires two intra-call movements. From efficiency point of
view, considering simultaneously both inter- and intra-cell material handling causes a balance between sizes of
the formed cells. This is because very large cells involve a high intra-cell material handling cost and a low inter-
cell material handling cost. On the other hand, very small cells involve a high inter-cell material handling cost
and a low intra-cell material handling cost.
A hybrid SA so-called MFA–SA is introduced to solve the proposed model in which the initial solution is
generated by MFA algorithm and then is improved by SA. SA is a powerful optimization technique to solve
the combinatorial optimization problems, but it is computationally intensive, especially for more complicated
problems (Ansari et al., 1995). In other words, when the size of problem increases, the gap between the best
solutions found by SA and the optimum solution increases too. In this case, one of the key issues to improve
the performance of SA is providing a good initial solution, closest possible to the optimal solution. In studies
related to the meta-heuristic applications, either the initial solution is generated randomly or is generated by a
568 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Cell1 Cell 2 Cell 3

P1 P4 P10 P11 P12 P2 P5 P6 P7 P3 P8 P9
M2 1,2 1
M3 2 2
Cell 1
M7 3 1 1,3 2
M8 1 2
M4 3 1,3 1
Cell 2 M5 2 1 1 1 1 2 3
M9 2 2
M1 2
Cell 3 M6 2 3 3
M10 3 1
Fig. 2. The effect of sequence operation on inter/intra-cell material handling.

heuristic or meta-heuristic approach. There are many heuristic approaches such as multi-stage designing meth-
ods in literature can be used for generating the initial solution in CMS (Harhalakis et al., 1990; Drolet et al.,
1996; Askin et al., 1997). But, these approaches highly depend on the structure and assumptions of the prob-
lem and cannot guarantee a high quality solution, especially for a complicated problem considered in this
paper. As an alternative, MFA which provides a good trade-off between performance and computational com-
plexity is used to solve the combinatorial optimization problems (Bofill et al., 2003). It also can be used for
generating a good initial solution in a negligible time for other metaheuristics such as SA or Tabu search. This
paper indicates how MFA can improve the performance of SA with respect to the some of measures to solve
the proposed model especially to large-sized problems. Current research implements the MFA based on the
research done by Saidi-Mehrabad and Safaei (in press). However, there are principal differences between
the implementation of MFA in this paper and in the mentioned research.
The rest of this paper is organized as follows: the problem formulation is described in Section 3. Implemen-
tation of MFA is described in Section 4. Implementation of a classical SA is described in Section 5. The com-
putational results are reported in Section 6 and the conclusion is given in Section 7.

3. Problem formulation

In this section a new mixed-integer programming model of DCMS is formulated under following
assumptions.

3.1. Assumptions

1. Each part type has a number of operations that must be processed respectively as numbered.
2. The processing time for all operations of a part type on different machine types are known and
deterministic.
3. The demand for each part type in each period is known and deterministic.
4. The capabilities and time-capacity of each machine type are known and constant over the planning
horizon.
5. Constant cost of each machine type is known. The constant cost is independent of the workload allo-
cated to machine and implies rent or overall service costs. This cost is considered for each machine in
each cell and period irrespective of whether the machine is active or idle.
6. Variable cost of each machine type is known. Variable cost implies operating cost that is dependent on
the workload allocated to the machine.
7. Relocation cost of each machine type from one cell to another between periods is known. All machine
types can be moved to any cell. Relocation cost is the sum of uninstalling, shifting and installing costs
where installing and uninstalling costs are assumed to be same. Note that, if a new machine is added to
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 569

system, we have only installing cost. On the other hand, if a machine is removed from the system, we
have only uninstalling cost. Thus, we assume that the unit cost of adding a new machine or removing
a current machine is half of relocation cost.
8. Parts are moved between and within cells as batch. As mentioned earlier, inter- and intra-cell batches
have different costs and sizes. For reducing complexity, unit inter- and intra-cell movement costs are con-
stant for all moves regardless of the distance travelled. In other words, we assume that the distance
between cells is same, the distance between machines at each cell is same and also all machine types have
same dimensions. Thus, we consider a hypothetical layout such as the layout shown in Fig. 3. In this
figure, it is assumed that parts are moved between cell by AVG and within cells by robot. The quad
arrow shows the move path of AGV for travelling between cells. Also, the moving path of the robot
manipulator within cell is shown by a dashed directed curve. Fig. 3 implies a layout in which the physical
distance between cell locations and also machine locations at each cell is same.
9. The maximum number of cells can be formed in each period is specified in advance.
10. Maximal cell size is known in advance. A lower bound is not considered for the cell size, because we
assume that the smaller cells are preferable.
11. All machine types are assumed to be multipurpose. Thus, each machine type can perform one or more
operations without incurring a modification cost. Likewise, each operation-part can be performed on
different machine types with different processing times.
12. Backorders are not allowed. All demands must be satisfied in the given period.

3.2. Notation

Indexes
c index for manufacturing cells (c ¼ 1; . . . ; C)
m index for machine types (m ¼ 1; . . . ; M)
p index for part types (p ¼ 1; . . . ; P )
h index for time periods (h ¼ 1; . . . ; H )
j index for operations belong to part p (j ¼ 1; . . . ; Op )

3.3. Input parameters

P number of part types

Op number of operations for part p
M number of machine types
C maximum number of cells that can be formed
Dph demand for part p in period h

Fig. 3. Hypothetical layout with same distance between paired machine and cell locations.
570 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Binter
p batch size for inter-cell movements of part p
Bintra
p batch size for intra-cell movements of part p
cinter inter-cell movement cost per batch
cintra intra-cell movement cost per batch. For justification of CMS, it is assumed that
ðcintra =Bintra
p Þ < ðcinter =Binter
p Þ 8p
am constant cost of machine type m in each period
bm variable cost of machine type m for each unit time
dm relocation cost of machine type m
Tm time-capacity of machine type m in each period
UB maximal cell size
tjpm processing time required to perform operation j of part type p on machine type m
ajpm ¼ 1 if operation j of part p can be done on machine type m; 0 otherwise

3.4. Decision variables

Nmch number of machines type m allocated to cell c in period h

þ
K mch number of machines type m added in cell c in period h
Kmch number of machines type m removed from cell c in period h
X jpmch ¼ 1 if operation j of part type p is done on machine type m in cell c in period h.

3.5. Mathematical model

By using the above notations, the proposed model is now written as follows:
H X
X M X
C H X
X C X
P X
Op X
M
min Z¼ N mch am þ bm Dph tjpm xjpmch
h¼1 m¼1 c¼1 h¼1 c¼1 p¼1 j¼1 m¼1
& ’  
1X H X
P XOp 1 X
C
Dph inter X M XM 

þ c  x ðjþ1Þpmch  x jpmch 
2 h¼1 p¼1 j¼1
c¼1 Bp
inter  m¼1 m¼1

& ’
p 1 X
1X H X P O X C
Dph intra X M
þ c jxðjþ1Þpmch  xjpmch j
2 h¼1 p¼1 j¼1 c¼1 Bintra p m¼1
 !
X M XM 
 
 xðjþ1Þpch  xjpch 
 m¼1 m¼1


1X H X C X M
þ dm ðK þ 
mch þ K mch Þ ð1Þ
2 h¼1 c¼1 m¼1
s:t:
XC X
M
ajpm xjpmch ¼ 1 8j; p; h; ð2Þ
c¼1 m¼1
P X
X Op
Dph tjpm xjpmch 6 T m N mch 8m; c; h; ð3Þ
p¼1 j¼1

X
M
N mch 6 UB 8c; h; ð4Þ
m¼1

N mcðh1Þ þ K þ 
mch  K mch ¼ N mch 8m; c; h; ð5Þ
xjpmch 2 f0; 1g; N mch ; K þ 
mch ; K mch P0 and integer: ð6Þ
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 571

The objective function given in Eq. (1) is to minimize the total sum of the machine constant/variable costs,
inter/intra-cell movements and reconfiguration costs over the planning horizon. The first term represents
the constant cost of all machines required in all cells over the planning horizon. This cost is obtained by
the product of the number of machine type m allocated to cell c in period h and their associated costs. This
term dose not allow for the extra machine replication and force model to maximum machine utilization. The
second term is the variable cost of all machines required in all cells over the planning horizon. It is the sum of
the product of the time-workload allocated to each machine type in each cell and their associated cost. This
term causes a balance between the workload assigned to machines at each cell. It is noted that, each machine
type with low (or high) constant cost has not necessarily a low (or high) variable cost and vice versa. For
example, an inkjet printer has a low price (constant cost) but high operating (variable) cost because of the need
to quickly exchange the cartridge. On the other hand, a laser printer has a high price but low operating cost
because of longer lasing cartridges. Also, laser printers have more quality, less breakdowns and more life cycle
of cartridge. Thus, there is a trade off between the first and second terms of objective function with remaining
terms of objective function for choosing the machines with high capability and low relatively cost. In gener-
ally, the proposed model selects a machine type based on four criteria as: constant cost, variable cost, process-
ing capabilities and capacity. The third term computes the total inter-cell material handling cost. It is the sum
of product of the number of inter-cell transfers (i.e., Dph =Binter p ) resulting from both consecutive operation of
each part type and the cost of transferring an inter-cell batch of each part type (cinter). Likewise, the fourth
term of objective function computes the total intra-cell material handling cost. It is the sum of number of
the product of intra-cell transfers (i.e., Dph =Bintra
p ) resulting from both consecutive operation of each part type
and the cost of transferring an intra-cell batch of each part type (cintra). As mentioned earlier, the sequence of
operations directly affects on the inter/intra-cell movement costs. Because if two consecutive operations of a
part type, which must be processed, are allocated to different cells, then the system will incur a unit inter-cell
movement cost (cinter). On the other hand, if two consecutive operations of a part type are allocated to same
cell but different machines, then the system will incur a unit intra-cell movement cost (cintra). From modelling
point of view, if two consecutive
PM operations j and j þ 1 of part p causes a inter or intra-cell movement in period
h then we have ujpch ¼ m¼1 jxðjþ1Þpmch  xjpmch j ¼ 1. On the other hand, if two consecutive operations j and
jPþ 1 of part p causes only a inter-cell movement in period h then we have wjpch ¼
 M xðjþ1Þpmch  PM xjpmch  ¼ 1. Accordingly, if ujpch  w
m¼1 m¼1 jpch ¼ 1 then we encounter with an intra-cell move-
ment. Fig. 4 shows schematically how computing inter/intra-cell movements by the third and fourth terms of
objective function. Coefficient 1/2 in the third and fourth terms is embedded because each inter/intra-cell
movement taken into account two times in calculations. Because moving from a cell (or machine) is corre-
sponding to moving toward another cell (or machine). Last term of objective function computes the reconfig-
uration cost in dynamic environment. This term is the sum of the number of product of relocated/added and/
or removed machines and their associated cost. Coefficient 1/2 in last term is embedded, because, each relo-
cation taken into account two times in calculations. In other word, relocating machine m from cell c to cell d in
period h causes K  þ
mch ¼ 1 and K mdh ¼ 1. Thus, in the last term of objective function, we have

Fig. 4. How computing inter- and intra-cell material handling by assuming operation sequence.
572 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

   þ ðK þ  þ 
mch þ K mch Þ þ    þ ðK mdh þ K mdh Þ þ    ¼    þ ð0 þ 1Þ þ    þ ð1 þ 0Þ þ    ¼ 2, therefore, each relo-
cation taken into account two times in calculations.
Eq. (2) guarantees that each part operation is assigned to one machine and one cell at each period. Eq. (3)
ensures that machine capacities are not exceeded and satisfy the demand. Also, this equation determines the
required number of each machine type in each cell including machine duplication. Inequality (4) ensures the
maximal cell size is not violated. Eq. (5) is called balance constraint that ensures that the number of machines
in the current period is equal to the number of machines in the previous period, plus the number of machines
being moved in, and minus the number of machines being moved out. In other words, it ensures the conser-
vation of machines over the horizon and plays the role of the memory for available machine types in each
period. The last term of objective function and coupling constraint (6) establish a connection between periods.
If we remove the last term of objective function and coupling constraint 5, the model will be fully decomposed
into sub-problems corresponding to different periods (Chen, 1998). This decomposition approach can be used
to obtain a lower bound for objective function value.

3.6. Linearization of the proposed model

Eq. (1) is a nonlinear integer equation because of the absolute terms in the third and fourth terms of the
objective function. To transform the third term of the objective function into a linear term, non-negative vari-
ables z1jpch and z2jpch are introduced and the third term of the objective function is rewritten as follow:
& ’
p 1 X
1X H X P O X C
Dph inter 1
c ðzjpch þ z2jpch Þ;
2 h¼1 p¼1 j¼1 c¼1 Binter p

where below constraint must be added to the original model.

X
M X
M
z1jpch  z2jpch ¼ xðjþ1Þpmch  xjpmch :
m¼1 m¼1

Likewise, to transform the fourth term of the objective function into a linear term, non-negative variables
y 1jpmch and y 2jpmch are introduced and the fourth term of the objective function is rewritten as follow:
Introduce a new set of constraints:
& ’ ( )
XH X X p 1 X
P O C
Dph intra X M

intra
c ðy 1jpmch þ y 2jpmch Þ  ðz1jpmch þ z2jpmch Þ ;
h¼1 p¼1 j¼1 c¼1 Bp m¼1

where below constraint must be added to the original model

y 1jpmch  y 2jpmch ¼ xðjþ1Þpmch  xjpmch 8j; p; m; c; h:

3.7. A numerical example

The aim of this section is to verify the applicability of the proposed model by a hypothetical example with
randomly generated data. This example is solved by branch and bound (B&B) method under Lingo 8.0 soft-
ware on an Intel Celeron Mobile 1.3 GHz Personal Computer with 512 Mb RAM. This example is gener-

Table 2
Test problem generation
Parameter Value Parameter Value Parameter Value
H 2 Dph U(100, 1000) dm $am/2
M ðP =2Þ þ 2 tjpm U(0, 1) hours Binter U(10, 50) units
P p
Op 2 m ajpm 2 8j; p Bintra
p Binter
p =5
C 3 pffiffiffi am $U(1000, 2000) cInter $50
UB P P þ1 bm U(1, 10) $ cIntra $5
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592
Table 3
Typical test problem
Machine info. P1 P2 P3 P4 P5 P6 P7 P8
Tm am ($) bm ($) dm ($) 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3
(hour)
500 1200 6 600 M1 0.76 0.57 0.54
500 1400 3 700 M2 0.68 0.61 0.67 0.23 0.24 0.19 0.55 0.58 0.71 0.72 0.47 0.44 0.28 0.17
500 1500 7 750 M3 0.63 0.88 0.48 0.57 0.82 0.97 0.15
500 1400 4 700 M4 0.63 0.13 0.58 0.49 0.47 0.84 0.86
500 1200 2 600 M5 0.55 0.79 0.89 0.96 0.26 0.17 0.65 0.12 0.76 0.86 0.2
500 1600 8 800 M6 0.36 0.78 0.45 0.59 0.81 0.48
Dph Period 200 0 0 650 350 600 550 600
1
Period 500 450 0 500 0 500 200 450
2
Period 600 0 600 0 750 350 300 350
3
Binter
p 25 25 20 35 20 40 40 30
Bintra
p 5 5 4 7 4 8 8 6

573
574 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Table 4
Best obtained solution in details
Zbest Machine constant Machine variable Inter-cell movements Intra-cell movements Reconfiguration CPU
cost cost cost cost cost time
$91,794 $34,800 $37,706 $7366 $9296 $2625 3 hours

C1 C2 C3
C1 C2 C1 C2
P4 P6 P7 P1 P8 P5
P2 P4 P7 P1 P6 P8 P1 P3 P5 P6 P7 P8
C1 3 M2 1,2 1 1
2 2 M2 2,3 1,2 1,3 2 M2 1,2 2 3
1 M4 2 C1
C1 1 M4 2 3 2 M4 3 1,3 2
1 M1 2
1 M5 1 3 1 M1 2 3
1 M2 3 2 2
C2 1 M1 2 1 M2 1 1 2
1 M3 3 3 C2
1 M2 1 2 1 M4 2
1 M5 3 1 1 C2
1 M3 2 3 1 M5 1,3 3 1
C3 1 M5 3 1,3
1 M5 1 3 1
Period 1 Period 2 Period 3

Fig. 5. Best obtained cell configurations for typical test problem presented in Table 3.

ated according
P to the information provided in Table 2. In this table, term ‘‘U’’ implicates to the uniform dis-
tribution. m ajpm is equal
P to the number of alternative machines for processing operation j of part p. It is obvi-
ous that by increasing m ajpm , the solution space increases progressively because the number of alternative
process plans for each operation-part increases. The data set related to the considered example is shown in
Table 3. The considered example consists of eight part types, six machine types and three periods in which
each part type is assumed to have three operation must be processed respectively. Each operation being able
to performed on two alternative machines. Thus, each part type has 2  3 ¼ 6 process plans and there are 68
combinations to select a process plan for each part type in each period. Therefore, there are about (68)3 com-
binations, because H ¼ 3, for allocating operation-parts to machines in throughout the planning horizon.
First four columns of Table 3 include the machine information, i.e., capacity, constant cost, variable cost
and relocation cost. Quantity of demand for each part type and also the product mix for each period is pre-
sented in Table 3. Also, last two rows of Table 3 show inter- and intra-cell batch size for each part type. After
linearization, the proposed model consists of 3623 variables and 1167 constraints under the considered exam-
ple. Runtime is limited to three hours. Thus, the best solution found after three hours (Zbest) is reported in
Table 4 in detail. The cell configurations for three periods corresponding to the best obtained solution are
shown in Fig. 5. This figure shows some of the characteristics and advantages of the proposed model. For
example, one unit machine 2 is removed from cell 1 in period 2. Machine 5 is relocated from cell 3 to cell
1 in period 2. Machine 4 is duplicated in cell 1 in period 3. Also, the sequence of operations can be identified
by the numbers shown in the column associated with each part type. For example, the first and third opera-
tions of part 5 in period 1 must be performed on machine 5 in cell 3, but its second operation must be per-
formed on machine 4 in cell 1. Thus, the process plan allocated to part 5 needs two inter-cell movements
and one intra-cell movement according to the reasoning presented in subsection 3.6. Period 1 includes three
formed cells while periods 2 and 3 include two formed cell. As shown in Fig. 5a, cell 3 is formed because of the
maximal cell size restriction on cell 1. In other words, machine 5 in cell 3 should be allocated to cell 1 but cell 1
is fully saturated
pffiffiffi by four machines. It is noted that according to Table 2, the maximal cell size is obtained as
UB ¼ d P e þ 1 ¼ 4.

4. Mean field annealing implementation

MFA is the deterministic counterpart of SA. Actually, deterministic mean field equations are replaced by
probabilistic hill-climbing algorithm of SA but the annealing process is so used by MFA. The concept of
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 575

energy function in Hopfield neural networks is used to create a link between the mean field equations and the
considered problem. MFA uses the mean field equations and the annealing process of SA to find the global
minimum of the energy function associated with the Hopfield network. In other words, in the MFA algorithm,
the aim is to find the spin values minimizing the energy function of the system. Energy function is same objec-
tive function in which state variables (spins) are replaced with common binary decision variables. The state
variables are the continuous form of the binary decision variables in considered optimization problem. When
MFA algorithm is finished, binary variables are become convergent to zero or one. The required stages for
implementation of MFA are as follows.

4.1. Designing Hopfield network

The first stage for implementation of MFA is designing the Hopfield neural network associated with the
considered problem. The weights of this network are same state variables or spins. These weights are updated
by mean field equations in each iteration until the network achieves to a stable state. The stable state for the
network means that the associated energy function achieves to a local or global optima. With regard to the
proposed model, the architecture of the Hopfield network in each period h is shown in Fig. 6. This architecture
consists of three layers in each period as follows:
Input layer: This layer presents the input data including operation-part’s
P processing time (i.e., tjpm) and
demand of parts (i.e., Dph). The input layer has totally n ¼ H  p Op neurons.
Hidden layer: This layer determines the allocation of operation-parts to machines in cells (i.e., xjpmch). The
hidden layer has totally H  M  C neurons. Each neuron ðm; c; hÞ is associated to each machine type in each
cell in each period. The net input of neuron ðm; c; hÞ is equal to the total workload assigned to machine m in
cell c. Eq. (7) shows how computing the net input for neuron ðm; c; hÞ. The inputs of hidden layer, i.e., spin
wjpmch, is the continuous form of the decision variables xjpmch in the proposed model. wjpmch can be interpreted
as the degree of membership of operation j of part p to machine m in cell c in period h. At first, wjpmch are
randomly generated at interval (0, 1). Then, MFA algorithm guarantees wjpmch are become convergent to
about 0 or 1 while the energy function decreasing. The output of neuron ðm; c; hÞ specifies the required number
of machine type m in cell c in period h (i.e., decision variable Nmch) with respect to the allocated workload and
available time-capacity. Eq. (8) shows how computing the output of neuron ðm; c; hÞ. Note that the third con-
straint of the proposed model, i.e., Inequality (3), is automatically satisfied by Eq. (8). After computing Nmch,
values K þ 
mch and K mch are updated by Eq. (9).
Output layer: This layer specifies the degree of infeasibility of the current solution with respect to the max-
imal cell size restriction. This layer has totally H  C neurons.
PThe net input of each neuron c in period h is the
total number of machines assigned to cell c in period h, i.e., m N mch . The output P of each neuron c in period h
is the number of extra machines assigned to cell c in period h, i.e., V hc ¼ maxf0; m N mch  UBg as shown in
Eq. (10)

Network output: degree of violation of the maximal cell size constraint

h
V c
1 2 C Output layer
Nmch
1 2 … M 1 2 … M 1 2 … M Hidden layer

wjpmch

1 O1 1 O2 1 Op Input layer
tjpm
Part 1 Part 2 Part P
Network input: processing time of part-operation

Fig. 6. Architecture of Hopfield network in each period for the considered problem.
576 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

P X
X Op
I h½mc ¼ tjpm  wjpmch  Dph ; ð7Þ
p¼1 j¼1
& h ’
I ½mc
N mch ¼ ; ð8Þ
Tm
Kþ mch ¼ maxfN mch  N mcðh1Þ ; 0g; Kmch ¼ maxfN mcðh1Þ  N mch ; 0g; ð9Þ
8 P
< 0; N mch 6 UB;
m
V hc ¼ P P ð10Þ
: N mch  UB; N mch > UB:
m m

4.2. Energy function

As mentioned earlier, there is an energy function corresponding to each Hopfield network. Energy function
is interpretation of Lagrange’s function in continuous space. Lagrange’s function indicates the objective func-
tion plus the some of penalty terms corresponding to the model’s constraints. With regard to the objective
function of the proposed model, the energy function associated to the designed network in Section 4.1 is writ-
ten as Eq. (11). This equation is the sum of a modified version of the original objective function with respect to
the designed network plus the penalty terms corresponding to the violation of constraints 2,4 and also the pen-
alty term associated with the saturation criterion (see Subsection 4.8). Parameters A1, A2, and A3 are penalty
coefficients. Also, coefficients n, C  H and g are embedded to normalize the penalty values and smoothing the
energy function. n is equal to the number of neurons in input layer. C  H is thePnumber of neurons in output
H PC PM PP POp
layer and g is the number of nonzero elements in state W ¼ ½wjpmch  where g ¼ h¼1 c¼1 m¼1 p¼1 j¼1 ajpm .
W implicates to a possible state for the considered system or a solution for the considered problem. As men-
tioned earlier, wjpmch is continuous form of xjpmch where wjpmch 2 [0, 1]. A solution is feasible if the three last
terms of the energy function reduce to zero. Also, the network achieves to the stable state if the last term
of objective function reduce to zero.
X
H X
M X
C H X
X C X
P X
Op X
M
EðW Þ ¼ N mch am þ bm Dph tjpm wjpmch
h¼1 m¼1 c¼1 h¼1 c¼1 p¼1 j¼1 m¼1
& ’ O 1 !2
1X H X P
Dph Xp XC
inter
XM X M
þ  c wðjþ1Þpmch  wjpmch
2 h¼1 p¼1 Binter
p j¼1 c¼1 m¼1 m¼1
& ’ 0 !2 1
p 1 X
1X H X P OX C
Dph intra @X M XM XM
wjpmch A
2
þ c ðwðjþ1Þpmch  wjpmch Þ  wðjþ1Þpmch 
2 h¼1 p¼1 j¼1 c¼1 Bintra
p m¼1 m¼1 m¼1

1X H X M
þ ðmaxfN mch  N mcðh1Þ ; 0g þ maxfN mcðh1Þ  N mch ; 0gÞdm
2 h¼1 m¼1
2 !2 3
1 XH X P X Op XC X M
þ A1 4 ajpm wjpmch  1 5
n h¼1 p¼1 j¼1 c¼1 m¼1
" ( )#
1 X H X C XM
þ A2 max N mch  UB; 0
H  C h¼1 c¼1 m¼1
" #
1X H X P X Op XC X M
þ A3 ajpm wjpmch ð1  wjpmch Þ : ð11Þ
g h¼1 p¼1 j¼1 c¼1 m¼1

There are some differences between Eq. (11) and the original objective function in Eq. (1). First, the absolute
terms are presented as the quadratic terms because of easy computing the derivative of energy function. Sec-
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 577

ond, the last term of the objective function is revised according to the requirements of the designed network in
Eq. (9). In other words, the decision variables K þ 
mch and K mch are computed in terms of the decision variables
Nmch and N mcðh1Þ . Finally, three penalty terms are inserted to the objective function (i.e., last three terms of
Eq. (11)). The first penalty term is aimed at imposing the constraint (2) that guarantees that each machine is
assigned to one and only one cell in each period. The second penalty function is aimed at imposing the con-
straint (4) that guarantees the maximal cell size limitation. The third penalty term is aimed at imposing the
integrality constraints on variables wjpmch or satisfying the saturation criterion (see Section 4.8).
One of the existent obstacles for implementation of MFA is the usage of the penalty functions in the
energy function. The penalty functions cause the shape of energy function become uneven and therefore
MFA cannot satisfy all constraints fully. Bofill et al. (2003) showed that even when the search strategy
may have some influence, the main factor determining the difficulty of a problem is the shape of its energy
landscape, and deciding which is actually the best strategy depends on the problem considered. They sug-
gested that significant improvements in performance should be oriented to smoothing or simplifying the
energy surface, rather than trying to improve the optimization technique itself. Thus, we need to rank the
priority of the penalty functions (Ansari et al., 1995; Lozano et al., 1993). As an experiment, in Eq. (11),
the first and third penalty functions are superior over the second penalty function and thus the first and third
penalty functions should be weighed heavier than the second penalty function. As an alternative, Lozano
et al. (1993) used a greedy heuristic to guarantee satisfying the maximal cell size limitation. However, when
we used a similar greedy heuristic to satisfy the second penalty function, the machine costs increased extre-
mely. Thus, our strategy is empirically A1 ¼ A3 > A2 where A1 is determined in terms of the size of problem
but A2 stays fixed.

4.3. Mean field equations

In MFA algorithm, the weight vector or state of Hopfield network, W, is updated by the mean field equa-
tions at each iteration. In mean field theory, instead of concerning the decision variables directly, the means of
decision variables in a continuous form are considered as wjpmch ¼ hxjpmch i ¼ 1  Prðxjpmch ¼ 1Þþ
0  Prðxjpmch ¼ 0Þ in which Prðxjpmch ¼ 1Þ follows Boltzman distribution. According to the mean field theory,
the probability that the hypothetical system lies in state W is equal to Pr(W) where Pr(W) is the density func-
tion of the Boltzman distribution that is defined by
X EðW Þ
~Þ ¼ 1 e T ; Z ¼
EðW Þ
PrðW e T ;
Z W

where Z is called ‘‘partition function’’ and T is the current temperature of system. Similar to the SA, the tem-
perature is a controller parameter in MFA. To estimate the value of wjpmch, we need to define the concept of
the ‘‘local field’’ or ‘‘thermal average’’. The local field of wjpmch, say ujpmch, is defined as the partial differential
of energy function with respect to wjpmch in current temperature
1 oEðW Þ
ujpmch ¼ ;
T owjpmch

ujpmch can also be interpreted as the decrease in the overall solution quality by assigning operation j of part p to
machine m in cell c in period h. By using the above definitions and the some of advanced computations, the
mean field equations are obtained as Eq. (12)
8 u
e jpmch
< wjpmch ¼ hxjpmch iT ¼ PC P M u ;
e jpm0 c0 h
c0 ¼1 m0 ¼1
ð12Þ
:u 1 oEðW Þ
jpmch ¼ T owjpmch :

A detailed description and the derivation of mean field equations can be found in (Peterson and Anderson,
1988; Peterson and Sodeberg, 1989). Eq. (12) guarantee when convergence is complete, each operation-part
is assigned to only one machine in one cell in each period. Also, the mean field equations guarantee the value
of energy function is reducing in each iteration while the temperature decreasing.
578 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

4.4. Derivation of the energy function

As shown in the second term of Eq. (12), we need to compute the gradient of the energy function. This value
is used in the MFA algorithm to compute the direction of maximum energy decrease, and the expected values
of the spins are updated accordingly. Thus, we use the direct differential equation (11) as follows:
oEðW~Þ oN mch
¼ am þ bm Dph tjpm
owjpmch owjpmch
& ’ " ! !#
Dph inter XM XM X M XM
þ inter c  wðjþ1Þpmch  wjpmch þ wjpmch  wðj1Þpmch
Bp m¼1 m¼1 m¼1 m¼1
& ’ ( !
Dph intra X M X M
þ intra c ðwðjþ1Þpmch  wjpmch Þ þ wðjþ1Þpmch  wjpmch þ ðwjpmch  wðj1Þpmch Þ
Bp m¼1 m¼1
!)  
XM X M
dm o maxfN mch  N mcðh1Þ ; 0g o maxfN mcðh1Þ  N mch ; 0g
 wjpmch  wðj1Þpmch þ þ
m¼1 m¼1
2 owjpmch owjpmch
!  P
2A1 ajpm X C X M
A3 o max m N mch  UB; 0 A2
þ ajpm wjpmch  1 þ þ ð1  2wjpmch Þ;
n c¼1 m¼1
H C owjpmch g
ð13Þ
where according to Eqs. (7) and (8) and the principal definition of derivative,
 h 
oN mch I mc ðwjpmch þ eÞ
 N mch ðwjpmch þ eÞ  N mch ðwjpmch Þ ¼  N mch
owjpmch Tm
 h 
I mc ðwjpmch Þ etjpm Dph
¼ þ  N mch ¼ Dmch P 0:
Tm Tm
For computing fifth term of Eq. (13), we have oN mch =owjpmch P 0, thus,
o maxfN mch  N mcðh1Þ ; 0g=owjpmch ¼ maxfN mch þ Dmch  N mcðh1Þ ; 0g  maxfN mch  N mcðh1Þ ; 0g;
o maxfN mcðh1Þ  N mch ; 0g=owjpmch ¼ maxfN mcðh1Þ  N mch  Dmch ; 0g  maxfN mcðh1Þ  N mch ; 0g:
Also, seventh term of Eq. (13) is computed as follows:
( ), ( ) ( )
X X X
o max N mch  UB; 0 owjpmch ¼ max N mch þ Dmch  UB; 0  max N mch  UB; 0 :
m m m

4.5. Initial temperature

Determining the initial temperature is a critical problem in MFA. The initial temperature is found by find-
ing a critical point where the energy function decreases significantly. Finding this parameter is important to
avoid unnecessary computation. On the other hand, with small value of initial temperature, MFA converges
quickly to poor solutions. However, there is no general technique to estimate the initial temperature. We con-
sider the initial temperature (T0) experimentally as a multiple of g. Therefore, the initial temperature is con-
sidered as T 0 ¼ bg; b > 0 that will be dependent to the size of problem. As will be shown in Section 5, when the
size of problem grows, parameter b approaches to 1. This mean that the value of g is a good initial temperature
for large-sized problems.

4.6. Temperature rule updating

Similar to the SA, MFA needs to a suitable temperature schedule to guarantee that the algorithm is become
convergent to a good solution. We consider the classical rule updating of SA as T k ¼ aT k1 where a is the rate
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 579

of cooling or decrement factor. In general, the convergence speed and solution quality depend to a. A strategy
with two phases is used to control the cooling schedule process. In the slow cooling phase, temperatures are
decreased using a ¼ 0:95 until T < T 0 =1:5. Then, in the fast cooling phase, a is set to 0.85. The cooling process
continues until the spins are converged to about 0 or 1 (Aykanat et al., 1998).

4.7. Equilibrium condition (convergence criterion)

Equilibrium condition implies that the value of energy function cannot be improved further in current tem-
perature. In other words, the satisfaction of equilibrium condition means the network achieves a stable state in
current temperature and it needs temperate update. Thus, at each temperature, mean field equations are iter-
ated until the following criterion is met:

1X H X C X M X P X Op
ðtÞ ðt1Þ
ajpm jwjpmch  wjpmch j < e; ð14Þ
g h¼1 c¼1 m¼1 p¼1 j¼1
ðtÞ
where wjpmch indicates the value of wjpmch in temperature t. Also, parameter e is an arbitrary small number, e.g.,
e ¼ 0:001. Inequality (14) is also known as convergence criterion.

Set α , ε , T0, A1, A2 and A3

Let n = 0 and k = 0

Generate initial solution (state) W 0 and compute E(W 0).

(Set randomly wjpmch ≈ 0.5)

Update decision variables by using mean filed equations and generate neighbor
solution Wnew and compute E(Wnew)

Equation (14) is satisfied? No

Let n = n+1
and Wn = Wnew

Yes

Let k = k+1 and Tk =αTk-1

Equation (15) is satisfied? No

Yes

Stop

Fig. 7. Flowchart of the MFA algorithm.

580 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

4.8. Stoppage condition (saturation criterion)

The annealing process is stopped when all state variables are within range ½0:0; e or within the range
½1  e; 1 without exceptions. Thus, annealing process is iterated until Inequality (15) is met. Obviously,
wjpmch ð1  wjpmch Þ  0 means that wjpmch  1 or wjpmch  0. As shown in Eq. (11), the left-hand side of Eq.
(15) is added to the energy function as a penalty function

1X H X C X M X P X Op
ajpm wjpmch ð1  wjpmch Þ < e: ð15Þ
g h¼1 c¼1 m¼1 p¼1 j¼1

Inequality (15) is also known as saturation criterion. After satisfying the saturation criterion, a more improve-
ment in the energy function is not possible, thereby, algorithm must be stopped.

4.9. MFA algorithm

Similar to the SA, MFA algorithm consists of two loops: (1) inner loop: this loop is iterated until the equi-
librium condition is satisfied, (2) Outer loop: outer loop handles the annealing process and it is iterated until
the stoppage condition is satisfied. In each temperature, the state variables are updated by Eq. (12) until
Inequality (15) is met. The flowchart of the applied MFA algorithm is shown in Fig. 7. Parameters n and k
are counters of inner and outer loops, respectively. As shown in Fig. 7, in contrary SA algorithm, in MFA
algorithm’s inner loop, a deterministic process is replaced by the stochastic hill-climbing search.

5. Simulated annealing implementation

In this section, a SA algorithm is developed to improve the solutions obtained by MFA. The principal
advantage of SA is to escape from local optima by accepting non-improver solutions by a certain probability
in each temperature. We apply the SA algorithm with a new matrix schema and novel operators. The steps for
development of SA for the considered DCMS problems are as follows.

5.1. Solution representation schema

To determine the solution representation schema, we use two matrices ½X RP and ½Y RP in each period as
shown in Fig. 8, where R ¼ maxp fOp g. The matrix [X] denotes the allocation of operation-part to machine and
the matrix [Y] denotes the allocation of operation-part to cell. xhjp is the machine that operation j of part p must
be performed on, where xhjp 2 Wjp and Wjp ¼ fmjajpm ¼ 1g. Also, y hjp is the cell that operation j of part p is allo-
cated to, where 1 6 y hjp 6 C. Because of Op 6 R8p, some of the entries in the solution representation are inher-
ently zero.

5.2. Neighborhood search strategy

To generate neighborhood solutions, we use four heuristic operators with respect to the proposed structure
shown in Fig. 8 as follows:

Fig. 8. Solution representation schema in each period h.

 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 581

Single-mutation on [X] matrix: For implementing this operator, an operation of a part in a period (e.g.,
operation j of part p in period h) is randomly selected and then randomly allocated to another machine as
0 0
xhjp ! xhjp where xhjp 2 Wjp .
Single-mutation on [Y] matrix: For implementing this operator, an operation of a part in a period (e.g.,
operation r of part p in period h) is randomly selected and then randomly allocated to another cell as
0 0 0
y hjp ! y hjp where y hjp 6¼ y hjp and 1 6 y hjp 6 C.
Multi-mutation on [Y] matrix: For implementing this operator, a part in a period (e.g., part p in period h) is
randomly selected and then all its operation are randomly allocated to the same cell as y hjp ! c8j where
1 6 c 6 C.
Multi-period implementation: In this case, each one of the above operators is implemented on all periods.
Thus, we have totally six operators for exploring the solution space.
These six operators are experimentally used with the same rate. In other words, when the tuning of oper-
ators are changed, it cannot be seen as a significant improvement in the performance of SA. After the imple-
mentation of each operator, the value of the decision variable Nmch must be updated with respect to the entries
of matrices [X] and [Y]. This is done by Eq. (8), and then the decision variables K þ 
mch and K mch are updated by
Eq. (9).

5.3. Fitness function

Due to complexity, the improvement of the infeasible solutions become computationally more intensive if
the size of problem increases. In other word, with respect to the machine capacity and maximum cell size con-
straints, we need to have a complex subroutine for repairing the infeasible solutions. Moreover, because of the
significant changes in input solution, the output of this reparation subroutine is often a new solution out of the
input solution’s neighborhood. Thus, it is straightforward to fitness function is considered as the sum of ori-
ginal objective function plus the penalty function
PH PCrelated P to the violation of constraint (4) as Fitness
M
ðX Þ ¼ ZðX Þ þ AQ in which Q ¼ ½1=ðH  CÞ h¼1 c¼1 maxf m¼1 N mch  UB; 0g is the degree violation of
the maximal cell size constraint and A is the penalty coefficient. Note that, as mentioned in Section 4.5, con-
straint (2) is automatically satisfied by Eq. (8). As an experimental result, A will be considered equal to A2 in
Eq. (11).

5.4. Initial temperature

According to the fundamental concepts of SA, non-improver solutions are accepted in the primary itera-
tions with high probability. Thus, we set the initial temperature in such a way that the non-improver solutions
are accepted with a probability of about 95% in the primary iterations. The related pseudo-code is shown in
Fig. 9.

5.5. Initial solution

The initial solution is generated according to a stochastic heuristic in which matrices [X] and [Y] are con-
structed in each period at random but is depended on the machine-cell allocations of immediate preceding

Sub InitTemp()
Do
Generate two solutions X1 and X2 at random
LOOP UNTIL (F(X1) ≠ F(X2))

− F ( X1 ) − F ( X 2 )
Set T0 =
ln(0.95)
End Sub

Fig. 9. Pseudo code of the initial temperature generation subroutine.

582 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

FOR p = 1 To P
FOR h = 1 To H
IF Dph > 0 THEN
FOR j = 1 To Op

'***************************** Constructing matrix [X]

IF h > 1 And Dp(h-1) > 0 THEN
Generate y→U(0,1)
IF y > e THEN
DO
m = [RND * (M - 1)] + 1
LOOP UNTIL (m ≠xh-1jp)
h
x jp = m
ELSE
xhjp = xh-1jp
END IF
ELSE
m = [RND * (M - 1)] + 1
xhjp =m
END IF

'***************************** Constructing matrix [Y]

IF h > 1 And Dp(h-1) > 0 THEN
Generate y→ U(0,1)
IF y > e THEN
DO
c =[RND * (C- 1)] + 1
LOOP UNTIL (c ≠yh-1jp)
h
y jp = c
ELSE
yhjp = yh-1jp
END IF
ELSE
c = [RND * (C - 1)] + 1
yhjp = c
END IF
NEXT j
END IF
NEXT h
NEXT p

Fig. 10. Pseudo code of the initial solution generation subroutine.

period by a certain probability e, as shown in the pseudo-code given in Fig. 10. In other words, with proba-
bility e, the value of xhjp (or y hjp ) is generated randomly independent of the value of xh1 h1
jp ðy jp Þ. Also, with prob-
h h h1 h1
ability 1-e, the value of xjp (or y jp ) is set to xjp (or y jp ). However, matrices [X] and [Y] are completely
generated at random in the first period. Function RND generates a random number in the interval ½0; 1.

5.6. Stopping rules

A single stopping criterion is used as the maximum number of consecutive temperature trials (N). The SA
algorithm is stopped if the outer loop of algorithm, i.e., cooling process, is iterated N times.

6. Computational results

In order to verify the performance of the MFA–SA approach, we have solved 33 random instances (23
small/medium-sized + 10 large-sized) corresponding to seventeen groups of problems. We have considered
a time criterion to determine the number of solved instances. For small/medium-sized problems, we have
started from a small-sized problem and increased gently the size of problem by a specific rule until the exact
approach could not reach the feasible space within a predetermined run time. In a similar fashion, for large-
sized problem, we have started from the largest medium-sized problems and increased the size of problem by a
specific rule until the heuristic approach could not reach the feasible space within a predetermined run time.
These problems are generated randomly according to Table 1 based on consideration of similar data in the
literature. Also, problems are solved under conditions discussed in Section 3.7. For simplicity, we assume that
the capacity of the machines are independent of their type, but depend on the length of the planning horizon.
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 583

For this purpose, we assume that the planning horizon is a three months period or one season. Also, each
period includes 75 workdays and each workday includes 8 hours. Therefore, each period is equals to
75 · 8 = 600 hours. Consequently, by taking into account the controllable and uncontrollable reasons for
interrupting production activities, we consider a 500-hour effective capacity for all machine types. Each prob-
lem is allowed 5400 seconds (1.5 hours). However, because of the computational complexity, the proposed
model cannot be optimally solved within 5400 seconds or even more time for medium and large-sized
instances. Thus, to solve the small and medium-sized problems, we consider a possible interval for optimum
value of objective function (F*) that constructed by the Fbound and Fbest values that are introduced by Lingo
software where F bound 6 F  6 F best . According to the Lingo software’s documents, The Fbest indicates the best
feasible objective function value (OFV) found so far. Fbound indicates the bound on the objective function val-
ues. This bound is a limit on how far the solver will be able to improve the objective. As mentioned in Sub-
section 3.4, Fbound can be also obtained by using the decomposition method. At some point, these two values
may become very close. Given that the best objective value can never exceed the bound, the fact that these two
values are close indicates that Lingo’s current best solution is either the optimal solution, or very closes to it.
At such a point, the user may choose to interrupt the solver and go with the current best solution in the inter-
est of saving on additional computation time. As mentioned earlier, we interrupt the solver within 5400 sec-
onds. This section is divided to three subsections in which MFA, SA and MFA–SA are verified individually.
Finally, the performance of SA and MFA–SA will be compared together.

6.1. MFA results

In this section, the performance of MFA will be verified. As pointed in literature, the quality of solution
provided by MFA depends on the initial value of state variables, initial value of temperature and of the pen-
alty coefficients A1, A2 and A3 (Lozano et al., 1993). In order to find a good value of parameter b to determine
the initial temperature, each problem is solved by different values of b in interval ½1; 100 by step 5, until the
MFA algorithm cannot satisfy the saturation criterion or cannot reach to a stable state and become so-called
non-convergence. As an important result, as shown in Table 6, by increasing the size of problems, the value of
b approaches to value 5. In other words, the initial temperature T 0 ¼ 5g is a suitable initial temperature for
large-sized problems. In order to find a good value for a, two medium-sized problems (i.e., problems No.
19 and 23 in Table 6) are selected and solved with a values 0.85, 0.9, 0.95 and 0.99 and the moderate cooling
rate a ¼ 0:95 is selected empirically. In a similar manner, coefficient A2 is fixed as 104 for all problems but coef-
ficients A1 and A3 are dependent to the size of problem and selected as 105 for small-sized problems, 106 for
medium-sized problems and 107 for large-sized problems. As mentioned in Subsection 4.2, since each opera-
tion-part at each period should be assigned to one and only one machine in a cell and also all spins should be
convergent to 0 or 1, the first and third penalty functions should be weighed heavier than the second penalty
function. Thus, we consider A1 ¼ A3 > A2 . The MFA parameter setting is summarized in Table 4.
Table 6 indicates the comparison between the results obtained by B&B and MFA corresponding to 23 small
and medium-sized problems. Each problem is run 20 times and the mean of OFV (Zmean), best OFV (Zbest),
and also mean of CPU time (TMFA) is reported in Table 6. The relative gap between the best OFV found by
B&B (Fbest) and Zmean found by MFA is shown in column ‘‘Gmean’’. The Gmean is calculated as:
Gmean ¼ ½ðZ mean  F best Þ=F best   100. Also, the relative gap between Fbest and Zbest is shown in column ‘‘Gbest’’.
In a similar manner, The Gbest is calculated as: Gbest ¼ ½ðZ best  F best Þ=F best   100. Column ‘‘Inf’’ indicates the
mean degree of infeasibility for obtained solutions in 20 runs. Parameter ‘‘Inf’’ for each obtained solution is
defined as the mean grade violation of the maximal cell size constraint associated with each cell in each period
that is calculated by
( )
1 X H X C X
Inf ¼ max N mch  UB; 0 :
H  C h¼1 c¼1 m

Above equation is the same second penalty function in Eq. (11). It is notable that Inf is only related to the
violating constraint (4) and other two penalty functions associated with the violating constraint (2) and the
saturation criterion (i.e., Eq. (15)) are fully satisfied in the all solved problems.
584 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Fig. 11. A schema of spins convergence.

As an example, Fig. 11 shows how the typical spins wjpmch for j ¼ 1, p ¼ 1, h ¼ 1 and different values of m
and c, related to problem No. 20 in Table 6, converge to 0 or 1 after 13 iterations. As shown in Fig. 11, only
one spin converges to 1 while other spins converge to 0. This means that the operation j ¼ 1 of part p ¼ 1 in
period h ¼ 1 is allocated to one and only one machine in a cell and therefore constraint (2) is satisfied. Fig. 12
shows the behaviour of the energy function versus the different values of b, i.e., b ¼ 1; 2; 3; . . . ; 10. Energy
function precociously converges to a poor solution for b ¼ 1. Also, for b ¼ 10, MFA could not satisfy Eq.
(14) and therefore could not achieve to a stable state. As discussed earlier, the best result is obtained under
b ¼ 5. Likewise, Fig. 13 shows the behaviour of the energy function versus the different value of a (i.e.,
a ¼ 0:85; 0:90; 0:95 and 0.99). As mentioned earlier, the best result is obtained with a ¼ 0:95.
As depicted in the last row of Table 6, the average of Gmean and Gbest is obtained as 21.78 and 15.39%,
respectively. Also, MFA could not achieve to the feasible solution space because violating constraint (4).
At the first glance, it seems that the obtained results are not a favorable and promising one. But, by attention
to the amount of reducing in the energy function value within the reported CPU time, it seems that the
obtained results are acceptable and reasonable. For instance, as shown in Fig. 12, the energy function value
is reduced from about $1,200,000 to about $200,000 (i.e., a decrease about 500%) within only 1.44 seconds
while constraint (1) and also saturation criterion are fully satisfied.

1200000
b=9
1000000
800000 b=2
OFV

600000
b=1
400000
200000
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43
Iteration

Fig. 12. Behaviour of the energy function under the different values of parameter b.

1200000
1000000
α =0.95
800000 α =0.85
OFV

600000
α =0.99
400000 α =0.9
200000
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43
Iteration

Fig. 13. Behaviour of the energy function under different values of parameter a.
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 585

6.2. SA results

In this section, the performance of SA developed in Section 5 will be verified. The SA parameter setting is
shown in Table 5. Coefficient A like A2 is fixed to 104 for all problems. Parameter K is the number of accepted
solution in each temperature to achieve equilibrium. Also, parameter N is the maximum number of consecu-
tive temperature trails until stopping the SA algorithm. Both parameters K and N are set to 100 empirically.
By attention to whether the mean and best OFV found by SA lie in interval [Fbound, Fbest] or not, six measures
for judgment on the effectiveness of SA is defined as

1. Gmean ¼ Gap between Fbest and Zmean. We assume that if Z mean < F best then the associated gap will be a neg-
ative number.
2. Gbest ¼ Gap between Fbest and Zbest. We assume that if Z best < F best then the associated gap will be a neg-
ative number.
3. AvgðjGmean  Gbest jÞ ¼ average of the absolute difference between Gmean and Gbest.
4. Nmean = the number of problems their corresponding Zmean lie in interval [Fbound, Fbest].
5. Nbest = the number of problems their corresponding Zbest lie in interval [Fbound, Fbest].
6. CPU time.

Table 5
MFA and SA parameter settings
Approach parameter MFA SA
e b 2 Nþ A1 A2 A3 K N a A
5 6 7 4 5 6 7
Value 0.001 [1, 100] 10 , 10 , 10 10 10 , 10 , 10 100 100 0.95 104

Table 6
Comparison between B&B and MFA runs for small and medium-sized problems
No. Problem info. B&B MFA Gap
best bound mean best
P M H C UB F F TB&B Z Z Inf TMFA b Gmean (%) Gbest (%)
1 4·4·2 3 3 20,156 20,156 1 29,719 26,122 0.02 0.10 100 47.4 29.5
2 6·5·2 3 4 46,074 46,074 62 56,464 52,830 0.12 0.21 70 22.5 14.6
3 8·6·2 3 4 44,591 44,591 100 58,584 52,514 0.18 0.27 50 31.3 17.7
4 10 · 7 · 2 3 5 75,373 75,373 721 92,490 87,317 1.11 0.41 40 22.7 15.8
5 12 · 8 · 2 3 5 57,397 57,397 4050 75,679 71,243 0.65 0.50 20 31.8 24.1
6 14 · 9 · 2 3 5 97,084 86,877 >5400 116,940 110,327 1.46 0.63 20 20.4 13.6
7 16 · 10 · 2 3 5 91,098 81,593 >5400 113,741 108,876 1.84 0.65 10 24.8 19.5
8 18 · 11 · 2 3 6 95,081 84,839 >5400 119,481 111,304 1.67 0.80 10 25.6 17
9 20 · 12 · 2 3 7 144,242 126,078 >5400 160,193 154,461 2.04 1.03 10 11 7
10 22 · 13 · 2 3 7 174,030 153,498 >5400 194,656 187,887 3.07 1.05 10 11.8 7.9
11 24 · 14 · 2 3 7 128,170 117,355 >5400 158,196 148,829 2.58 1.42 10 23.4 16.1
12 26 · 15 · 2 3 10 146,769 126,327 >5400 170,582 162,639 1.80 2.14 10 16.2 10.8
13 26 · 15 · 2 3 9 146,570 126,476 >5400 170,119 163,648 2.64 2.09 10 16 11.6
14 26 · 15 · 2 3 8 153,761 126,903 >5400 169,691 159,857 3.39 2.10 5 10.3 3.9
15 26 · 15 · 2 4 6 148,420 126,195 >5400 179,793 174,532 3.20 1.94 1 21.1 17.5
16 28 · 16 · 2 3 10 182,771 158,648 >5400 209,738 205,662 1.82 2.07 5 14.7 12.5
17 28 · 16 · 2 3 9 183,621 159,600 >5400 210,207 203,044 2.63 2.05 5 14.4 10.5
18 28 · 16 · 2 3 8 180,392 158,498 >5400 211,405 201,662 3.60 2.05 5 17.1 11.7
19 28 · 16 · 2 4 6 183,298 157,803 >5400 220,206 216,544 3.36 2.52 1 20.1 18.1
20 30 · 17 · 2 3 10 165,638 151,749 >5400 208,035 200,201 1.89 2.35 5 25.5 20.8
21 30 · 17 · 2 3 9 172,324 151,708 >5400 209,828 196,950 2.70 2.36 5 21.7 14.2
22 30 · 17 · 2 3 8 176,327 151,624 >5400 208,537 200,975 3.54 2.25 5 18.2 13.9
23 30 · 17 · 2 4 6 170,818 150,832 >5400 227,422 214,932 3.49 2.12 10 33.1 25.8
Average 21.78 15.39
586 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Table 7
Comparison between B&B and SA runs for small and medium-sized problems
No. Problem info. B&B SA Gap
P M H C UB Fbest Fbound TB&B Zmean Zbest Inf TSA Gmean(%) Gbest (%) jGmean  Gbest j (%)
1 4·4·2 3 3 20,156 20,156 1 20,156 20,156 0 5.1 0.00 0.00 0.00
2 6·5·2 3 4 46,074 46,074 62 47,858 46,120 0 19 3.87 0.09 3.78
3 8·6·2 3 4 44,591 44,591 100 45,367 44,591 0 42 1.74 0.00 1.74
4 10 · 7 · 2 3 5 75,373 75,373 721 78,211 77,318 0 57 3.76 2.58 1.18
5 12 · 8 · 2 3 5 57,397 57,397 4050 60,748 58,558 0 102 5.83 2.02 3.81
6 14 · 9 · 2 3 5 97,084 86,877 >5400 99,750 96,733 0 75 2.74 0.36 3.1
7 16 · 10 · 2 3 5 91,098 81,593 >5400 94,307 90,238 0 128 3.52 0.95 4.47
8 18 · 11 · 2 3 6 95,081 84,839 >5400 98,259 93,835 0 144 3.34 1.32 4.66
9 20 · 12 · 2 3 7 144,242 126,078 >5400 145,947 141,185 0.1 103 1.18 2.16 3.34
10 22 · 13 · 2 3 7 174,030 153,498 >5400 175,788 171,284 1.4 354 1.01 1.60 2.61
11 24 · 14 · 2 3 7 128,170 117,355 >5400 133,993 129,007 0 226 4.54 0.65 3.89
12 26 · 15 · 2 3 10 146,769 126,327 >5400 144,729 141,693 0 236 1.40 3.58 2.18
13 26 · 15 · 2 3 9 146,570 126,476 >5400 145,599 142,539 0 128 0.66 2.82 2.16
14 26 · 15 · 2 3 8 153,761 126,903 >5400 153,013 140,588 1.5 159 0.48 9.36 8.88
15 26 · 15 · 2 4 6 148,420 126,195 >5400 153,670 144,108 1.8 146 3.53 2.99 6.52
16 28 · 16 · 2 3 10 182,771 158,648 >5400 181,912 176,772 0 292 0.47 3.39 2.92
17 28 · 16 · 2 3 9 183,621 159,600 >5400 185,338 178,451 0 221 0.93 2.89 3.82
18 28 · 16 · 2 3 8 180,392 158,498 >5400 186,183 178,925 1.8 242 3.21 0.81 4.02
19 28 · 16 · 2 4 6 183,298 157,803 >5400 192,981 189,720 2.4 218 5.28 3.50 1.78
20 30 · 17 · 2 3 10 165,638 151,749 >5400 175,205 171,339 0 511 5.77 3.44 2.33
21 30 · 17 · 2 3 9 172,324 151,708 >5400 177,148 171,284 0 377 2.79 0.60 3.39
22 30 · 17 · 2 3 8 176,327 151,624 >5400 176,689 174,087 0.2 830 0.20 1.28 1.48
23 30 · 17 · 2 4 6 170,818 150,832 >5400 185,818 177,195 1.3 261 8.78 3.73 5.05
Average 0.45 212 2.56 0.78 3.35

In measures 1–4, Fbest is a base to evaluate the performance. Thus, a negative value implicates to a better
performance. Table 7 indicates the comparison between the results obtained from B&B and SA algorithm cor-
responding to the 20 three small and medium-sized problems considered in Table 6. The two last columns of
Table 7 show the values of Gmean and Gbest resulted from SA, respectively. In general, the smaller values of
Gmean, Gbest and AvgðjGmean  Gbest ZjÞ are more favorable and implicate to the solutions with higher quality.
Obviously, the measure AvgðjGmean  Gbest jÞ is directly dependent to the standard deviation of OFV, d, in 20
times run. As shown in the last row of Table 7, the average values of Gmean, Gbest and AvgðjGmean  Gbest jÞ
resulted from SA are obtained as 2.56%, 0.78% and 3.35%, respectively. In other words, Zmean is worse aver-
age 2.56% than Fbest while Zbest is better average 0.78% than Fbest and also there is average 3.35% difference
between Zmean and Zbest. Moreover, we have N mean ¼ 5 and N best ¼ 16, thus, in 21.73% of problems, Zmean lie
in interval [Fbound, Fbest] and in 69.56% of problems, Zbest lie in interval [Fbound, Fbest]. The average of CPU
time is obtained as 212 seconds that is a promising result in comparison to the CPU time reported for

Fig. 14. Comparison between B&B and SA results (Table 7): (a) (Zmean found by SA) vs. (Fbound and Fbest) and (b) (Zbest found by SA) vs.
(Fbound and Fbest).
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 587

B&B. The average degree of infeasibility is obtained as 0.45 that is negligible. Fig. 14a and b shows the behav-
iour of Zmean and Zbest values obtained by SA versus Fbest and Fbound according to the information provided in
Table 7. In these figures, the problems are arranged by terms of Fbest values. The obtained results show that
the SA is a relatively suitable approach to solve the small and medium-sized problems. As will be shown in the
next subsection, MFA–SA can improve all mentioned measures in comparison to SA and provide more con-
fident conditions to solve the large-sized problems.

6.3. MFA–SA results

In this section, at first, the performance of MFA–SA will be verified and then a comparison between the
performance of SA and MFA–SA to solve the large-sized problems will be presented. The measures defined
Table 8
Comparison between B&B and MFA–SA runs for small and medium-sized problems
No. Problem info. B&B MFA–SA Gap
P M H C UB Fbest Fbound TB&B Zmean Zbest Inf TMFA–SA Gmean Gbest jGmean  Gbest j
(Seconds) (%) (%) (%)
1 4·4·2 3 3 20,156 20,156 1 20,362 20,156 0 3.8 1.02 0.00 1.02
2 6·5·2 3 4 46,074 46,074 62 47,475 46,120 0 25 3.04 0.09 2.95
3 8·6·2 3 4 44,591 44,591 100 44,920 44,591 0 34 0.73 0.00 0.73
4 10 · 7 · 2 3 5 75,373 75,373 721 78,068 77,181 0 51 3.57 2.39 1.18
5 12 · 8 · 2 3 5 57,397 57,397 4050 60,135 58,507 0 74 4.77 1.93 2.84
6 14 · 9 · 2 3 5 97,084 86,877 >5400 97,088 94,508 0 74 0.00 2.72 2.72
7 16 · 10 · 2 3 5 91,098 81,593 >5400 92,971 90,469 0 90 2.0 0.69 2.69
8 18 · 11 · 2 3 6 95,081 84,839 >5400 96,719 93,883 0 183 1.72 1.27 2.99
9 20 · 12 · 2 3 7 144,242 126,078 >5400 142,509 138,305 0.7 128 1.21 4.29 3.08
10 22 · 13 · 2 3 7 174,030 153,498 >5400 171,662 166,602 1.5 175 1.37 4.45 3.08
11 24 · 14 · 2 3 7 128,170 117,355 >5400 133,804 131,250 0 185 4.39 2.40 1.99
12 26 · 15 · 2 3 10 146,769 126,327 >5400 143,846 140,809 0 201 2.03 4.23 2.2
13 26 · 15 · 2 3 9 146,570 126,476 >5400 144,373 139,710 0.1 127 1.52 4.91 3.39
14 26 · 15 · 2 3 8 153,761 126,903 >5400 147,177 139,237 1 173 4.47 10.43 5.96
15 26 · 15 · 2 4 6 148,420 126,195 >5400 151,700 144,834 1.6 152 2.20 2.47 4.67
16 28 · 16 · 2 3 10 182,771 158,648 >5400 180,776 178,143 0 183 1.10 2.59 1.49
17 28 · 16 · 2 3 9 183,621 159,600 >5400 182,454 177,599 0 179 0.63 3.39 2.76
18 28 · 16 · 2 3 8 180,392 158,498 >5400 183,163 180,182 1.4 241 1.53 0.11 1.64
19 28 · 16 · 2 4 6 183,298 157,803 >5400 188,365 184,724 2.7 210 2.76 0.77 1.99
20 30 · 17 · 2 3 10 165,638 151,749 >5400 171,403 164,038 0 372 3.48 0.97 4.45
21 30 · 17 · 2 3 9 172,324 151,708 >5400 172,183 167,013 0 403 0.08 3.17 3.09
22 30 · 17 · 2 3 8 176,327 151,624 >5400 175,183 170,604 0.5 451 0.65 3.35 2.7
23 30 · 17 · 2 4 6 170,818 150,832 >5400 178,815 174,500 2 282 4.68 2.15 2.53
Average 0.5 173 0.99 1.70 2.70

Fig. 15. Comparison between B&B and MFA–SA results (Table 8): (a) (Zmean found by MFA–SA) vs. (Fbound and Fbest) and (b) (Zbest
found by MFA–SA) vs. (Fbound and Fbest).
 588
N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592
Table 9
Comparison between MFA–SA and SA runs for large-sized problems
No. Problem info. MFA SA MFA–SA
P M H C UB Zmean Zbest Inf d dT MFA Zmean Zbest d TSA Zmean Zbest d TMFA–SA
1 30 · 17 · 2 5 4 233,389 226,663 3.80 4111 3.41 190,904 181,624 5319 358 181,495 178,758 2356 283
2 40 · 22 · 2 3 15 339,561 326,040 3.03 5331 4.10 299,570 293,080 4075 777 296,402 292,299 2386 705
3 40 · 22 · 2 4 11 353,873 344,626 3.15 6788 4.38 305,451 299,212 4257 884 300,432 289,479 5392 520
4 40 · 22 · 2 5 9 364,572 356,915 2.64 4038 5.48 313,733 306,034 5305 694 306,709 299,792 4771 697
5 50 · 27 · 2 3 20 391,198 380,570 2.35 3744 4.08 336,579 331,569 3154 1060 334,785 329,414 3231 972
6 50 · 27 · 2 4 15 404,740 391,582 2.95 6023 7.10 343,132 338,524 3801 1094 339,549 334,079 3552 1202
7 50 · 27 · 2 5 12 415,959 407,584 2.98 5232 9.01 352,200 339,853 9679 1281 344,295.3 337,785 3199 1702
8 60 · 32 · 2 3 20 424,282 409,893 3.98 6948 6.29 353,643 345,952 5580 1355 352,340 338,722 6897 1017
9 60 · 32 · 2 4 15 433,636 417,008 4.30 7538 11.6 368,045 357,254 4321 1651 357,624 352,175 5050 1328
10 60 · 32 · 2 5 12 455,494 440,394 4.12 7417 14.17 369,549 358,906 7902 1950 363,502 357,470 4076 1744
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 589

in the previous subsection will be used to verify this comparison. Table 8 indicates the comparison between the
results obtained from B&B and the proposed MFA–SA algorithm corresponding to the 23 small/medium-
sized problems considered in Table 6. The average of Gmean, Gbest and AvgðjGmean  Gbest jÞ are obtained as
0.99%, 1.70% and 2.70% respectively. The comparison between SA and MFA–SA results indicates that mea-
sures Gmean, Gbest and AvgðjGmean  Gbest jÞ are improved by amount of 1.57%, 0.92% and 0.45% respectively.
Moreover, we have N mean ¼ 10 and N best ¼ 17. Thus, in 43.47% of problems, Zmean lie in interval [Fbound, Fbest].
Also, in 73.91% of problems, Zbest lie in interval [Fbound, Fbest]. Thus, the probability that Zmean lie in interval
[Fbound, Fbest] increases to 0.43 from 0.21 and also the probability that Zbest lie in interval [Fbound, Fbest]
increases to 0.73 from 0.69 in comparison to the performance of SA. Therefore, Nmean and Nbest are improved
by amount of 21.74% and 3.95%. By a similar reasoning, CPU time is also improved by 17.88% in comparison
to the SA run. Fig. 15a and b shows the behaviour of Zmean and Zbest values obtained by MFA–SA versus Fbest
and Fbound found by B&B, according to the information provided in Table 8. As a general result, MFA–SA
dominates the SA with respect to the all defined measures. Also, MFA–SA affects more on Zmean and CPU
time and less on Zbest and the standard deviation of OFV. The superiority of MFA–SA over SA can be ranked
with respect to the defined measures as ‘‘ CPU time  Z mean  Z best  d’’. Thus, MFA tangibly improve the
speed convergence of SA while leads to relatively better solutions.
To evaluate the performance of the MFA–SA to solve the large-sized problems, ten random instances cor-
responding to four groups of problems are solved by MFA, SA and MFA–SA and the obtained results are
presented in Table 9. The summery of paired comparisons MFA vs. SA and SA vs. MFA–SA is shown in
Table 10. Also, the superiority of MFA–SA over SA, with respect to three measures Gmean, Gbest, CPU time
and also the size of problem is summarized in Table 11. Noted that, measures Nmean and Nbest and
AvgðjGmean  Gbest jÞ are not available for large-sized problems because we do not have any information about
interval [Fbound, Fbest]. The problems presented in Table 10 are said to be large-sized because it is not possible
to obtain a lower bound or feasible solution within the considered maximum runtime, i.e., 5400 seconds. It is
noted that by increasing C or deceasing UB, the computational difficulty of the proposed model increases pro-
gressively. As shown in the last row of Table 10, the percent improvement of measures Gmean, Gbest by MFA–
SA in comparison to SA is obtained as 1.90% and 1.38% respectively. Also, CPU time is improved by amount
of 15.09%. Thus, as can be derived from Table 11, concerning all defined measures, MFA–SA shows better

Table 10
Summary of paired comparison
No. Problem info. MFA vs. SA SA vs. MFA–SA
P M H C UB Gmean (%) Gbest (%) Gmean (%) Gbest (%) CPU time
1 30 · 17 · 2 5 4 22.25 24.79 5.18 1.60 26.50
2 40 · 22 · 2 3 15 13.34 11.24 1.06 0.26 10.21
3 40 · 22 · 2 4 11 15.85 15.17 1.67 3.36 70.00
4 40 · 22 · 2 5 9 16.20 16.62 2.29 2.08 0.43
5 50 · 27 · 2 3 20 16.22 14.77 0.53 0.65 9.05
6 50 · 27 · 2 4 15 17.95 15.67 1.05 1.33 8.98
7 50 · 27 · 2 5 12 18.10 19.92 2.29 0.61 24.73
8 60 · 32 · 2 3 20 19.97 18.48 0.36 2.13 33.23
9 60 · 32 · 2 4 15 17.82 16.72 2.91 1.44 24.32
10 60 · 32 · 2 5 12 23.25 22.70 1.66 0.40 11.81
Average of improvement 18.09 17.61 1.90 1.38 15.09

Table 11
Superiority of MFA–SA over SA bases on the defined measures and problem size
Problem size Measure
Gmean (%) Gbest (%) AvgðjGmean  Gbest jÞ (%) Nmean (%) Nbest (%) CPU time (%)
Small and medium-sized 1.57 0.92 0.45 21.74 3.95 17.88
Large-sized 1.90 1.38 – – – 15.09
590 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Fig. 16. Comparison between SA and MFA–SA results (Table 9): (a) (Zmean found by SA) vs. (Zmean found by MFA–SA) and (b) (Zbest
found by SA) vs. (Zbest found by MFA–SA).

Fig. 17. Comparison between CPU times related to the SA and MFA–SA: (a) small and medium-sized problems (Tables 6 and 7) and
large-sized problems (Table 9).

performance over SA. Therefore, MFA–SA is superior to the SA for solving large-sized problems with respect
to the three mentioned measures. However, concerning CPU time, the MFA–SA for small and medium-sized
problems shows slightly better performance over the MFA–SA for large-sized problems. Generally, as can be
derived from Table 11, when the size of problem increases the superiority of MFA–SA to SA is more tangible.
Fig. 16a shows the behaviour of Zmean values obtained by MFA–SA versus Zmean found by SA according to
the information provided in Table 9. Likewise, Fig. 16b shows the behaviour of Zbest values obtained by
MFA–SA versus Zbest found by SA. Fig. 17a shows the behaviour of Zmean values obtained by SA versus
Zmean values found by MFA–SA according to the information provided in Table 9. Likewise, Fig. 17b shows
the behaviour of Zbest values obtained by SA versus Zbest values found by MFA–SA according to the infor-
mation provided in Table 9. Also, Fig. 17a and b shows the schematic comparison between CPU times
resulted from SA and MFA–SA for small/medium and large-size problem respectively.

7. Conclusion

In this paper a new optimization model of dynamic cellular manufacturing system (DCMS) is introduced
along with a hybrid solution approach based on mean field annealing (MFA) and simulated annealing (SA)
so-called MFA–SA. The advantages of the proposed model are as follows: simultaneous considering inter-
and intra-cell material handling by assuming the sequence of operations, alternative process plans for part
types, and machine replication. Main constraints are the maximal cell size and machine time-capacity. The
objective is to minimize the sum of the machine constant and variable costs, inter/intra-cell material handling
cost, and reconfiguration cost. The formulated mathematical is still open for considering other issues such as
carrying inventory between periods, partial or total subcontracting, considering distance between machine and
cell locations and like that suggested for future research.
 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592 591

The main difference between the classical SA and the proposed MFA–SA is generating a high quality initial
solution by MFA for SA. The obtained results show that MFA can provide a good initial solution in a neg-
ligible time where the average Gap between the quality of the initial solution found by MFA and the best solu-
tion found by branch and bound (B&B) method is nearly 22%. The performance of MFA–SA is evaluated and
compared to the performance of a classical SA by 23 small/medium-sized problem and 10 large-sized prob-
lems, with respect to the some of defined measures. It is observed that quality of the results obtained by
MFA–SA is better than SA with respect to the all defined measures. The superiority of MFA–SA over SA
can be ranked with respect to the considered measures as follow: CPU time  mean of OFV  best
OFV  standard deviation of OFV. Thereby, the average CPU time and mean of OFV are significantly
improved when MFA is used to generate the initial solution for SA. Thus, MFA tangibly improve the speed
convergence of SA while leads to the relatively better solutions. Especially, as the size of problem increases, the
superiority of MFA–SA to SA is more tangible. Thus, MFA can be replaced by the traditional initial solution
generation methods in more difficult problems. However, the availability of feasible space is an open question
in implementation of MFA and designing the energy function. Also, MFA can be used to generate the initial
population in the evolutionary algorithms that suggested for future research.

References

Ansari, N., Hou, E.S.H., Yu, Y., 1995. A new method to optimize the satellite broadcasting schedules using the mean field annealing of a
Hopfield neural network. IEEE Transactions on Neural Networks 6, 470–483.
Askin, R.G., Selim, H.M., Vakharia, A.J., 1997. A methodology for designing flexible cellular manufacturing systems. IIE Transactions
29, 599–610.
Aykanat, C., Bultan, T., Haritaoglu, I., 1998. A fast neural-network algorithm for VLSI cell placement. Neural Networks 11, 1671–1684.
Balakrishnan, J., Cheng, C.H., 2005. Dynamic cellular manufacturing under multi-period planning horizons. Journal of Manufacturing
Technology Management 16 (5), 516–530.
Bilbro, G.L., Snyder, W.E., Mann, R., Van den Bout, D., Miller, T., White, M., 1989. Optimisation by mean field annealing. Advanced
Neural Information Process Systems 1, 91–98.
Bofill, P., Guimera, R., Torras, C., 2003. Comparison of simulated annealing and mean field annealing as applied to the generation of
block designs. Neural Networks 16, 1421–1428.
Chen, M., 1998. A mathematical programming model for systems reconfiguration in a dynamic cell formation condition. Annals of
Operations Research 77 (1), 109–128.
Chong, S.Y., Lee, W.D., 1992. Parallel mean field annealing neural network for solving travelling salesman problem. International Joint
Conference on Neural Networks 4, 532–536.
Drolet, J., Abdulnour, G., Rheault, M., 1996. The Cellular Manufacturing Evolution. Computers & Industrial Engineering 31 (1), 139–
142.
Guerrero, F., Lozano, S., Smith, KA., Canca, D., Kwok, T., 2002. Manufacturing cell formation using a new self-organizing neural
network. Computers & Industrial Engineering 42 (2), 377–382.
Harhalakis, G., Nagi, R., Proth, J., 1990. An effective heuristic in manufacturing cell formation for group technology applications.
International Journal of Production Research 28, 185–198.
Harlahakis, G., Ioannou, G., Minis, I., Nagi, R., 1994. Manufacturing cell formation under random product demand. International
Journal of Production Research 32 (1), 47–64.
Jeon, G., Leep, H.R., 2006. Forming part families by using genetic algorithm and designing machine cells under demand changes.
Computers & Operations Research 33, 263–283.
Kannan, V., Ghosh, S., 1995. Using dynamic cellular manufacturing to simplify scheduling in cell based production systems. Omega 23
(4), 443–452.
Kim, J.H., Yoon, S.H., Lee, C.W., Sohn, K.H., 2001. A robust solution for object recognition by mean field annealing techniques. Pattern
Recognition 34, 885–902.
Lozano, S., Onieva, L., Larraneta, J., Teba, J., 1993. A neural network approach to part-machine grouping in GT manufacturing. In:
Takamori, T., Tsuchiya, K. (Eds.), Robotics, Mechatronics and Manufacturing Systems. North-Holland, Amsterdam, pp. 619–624.
Marcoux, Y., Drolet, J., Abdulnour, G., 1997. Studying the performance of a dynamic cellular manufacturing system. Computers &
Industrial Engineering 33 (1), 239–242.
Peterson, C., Anderson, J.R., 1987. A mean field theory learning algorithm for neural networks. Complex Systems 1 (5), 995–1019.
Peterson, C., Anderson, J.R., 1988. Neural networks and NP-complete optimization problems: A performance study on the graph
bisection problem. Complex Systems 2, 3–22.
Peterson, C., Sodeberg, B., 1989. A new method for mapping optimization problems onto neural networks. International Journal of
Neural Systems 1 (1), 3–22.
Rheault, M., Drolet, J., Abdulnour, G., 1995. Physically reconfigurable virtual cells: A dynamic model for a highly dynamic environment.
Computers & Industrial Engineering 29 (1–4), 221–225.
592 N. Safaei et al. / European Journal of Operational Research 185 (2008) 563–592

Saidi-Mehrabad, M., Safaei, N., in press, A new model of dynamic cell formation by a neural approach. International Journal of
Advanced Manufacturing Technology, doi:10.1007/s00170-006-0518-2.
Song, S., Hitomi, K., 1996. Integrating the production planning and cellular layout for flexible cellular manufacturing. International
Journal of Production Planning and Control 7 (6), 585–593.
Tavakkoli-Moghaddam, R., Aryanezhad, M.B., Safaei, N., Azaron, A., 2005. Solving a dynamic cell formation problem using
metaheuristics. Applied Mathematics and Computation 170 (2), 761–780.
Vakharia, A.J., Kaku, B.K., 1993. Redesigning a cellular manufacturing system to handle long-term demand changes: A methodology and
investigation. Decision Sciences 24 (5), 909–929.
Van Den Bout, D.E., Miller, T., 1990. Graph partitioning using annealed neural networks. IEEE Transactions Neural Networks 1, 192–
203.
Wicks, E.M., Reasor, R.J., 1999. Designing cellular manufacturing systems with dynamic part populations. IIE Transactions 31, 11–20.
Wilhelm, W., Chou, C., Chang, D., 1998. Integrating design and planning considerations in cell formation. Annals of Operations Research
77 (1), 97–107.