Int. J.

Production Economics 74 (2001) 135}146

A genetic approach to the scheduling of preventive maintenance

tasks on a single product manufacturing production line
G. Cavory, R. Dupas*, G. Goncalves
Laboratoire en Organisation et Gestion de la Production, Faculte& des Sciences Applique& es, University of Artois, F-62408 Be& thune, France

Abstract

The present article deals with optimising the schedule of maintenance tasks of all the machines in a single product
manufacturing production line. This study was made in the context of one machine assigned to one operator. This
operator intervenes to change tools during a stoppage. Our goal is to increase the overall through-put of the line. We
"rstly formalised the problem and showed the di$culty of its analytical resolution. Then, we presented the software
environment that enables the resolution of this problem: it is made up of a simulator of the production line and an
optimiser using the genetic algorithms. Our approach to the scheduling of maintenance tasks was validated upon an
actual production line of car engines. We focused our study on the setting of parameters of a genetic algorithm. We
proceeded with a systematic approach inspired by the Taguchi method to "nd the best combination of levels for each
studied parameter and performed a statistical con"rmation of the results. Finally we validated the genetic approach as
against naive optimisation.  2001 Elsevier Science B.V. All rights reserved.

Keywords: Scheduling; Genetic algorithm; Discrete event systems simulation; Single product; Tools changes

1. Introduction the line is balanced, the machines have equal

The domain of this study is the manufacturing
production systems. More precisely, the line con-
cerned is a single product mass production line.
This line, as in Fig. 1, of a succession of machines
alternating with bu!ers of pieces.
Each machine unit in this line is characterised as
follows.
E The machines perform multiple operations, the
succession of which is a line of type `transfera. As
* Corresponding author. Tel.: #33-321028761; fax: #33-
321028758.
E-mail address: dupas@iut-gtr.univ-artois.fr (R. Dupas).
duration cycles.
E The preventive maintenance tasks on these ma-
chines are realised by the operator. Each task
occupies a certain amount of the operator's time.
The number of maintenance tasks varies accord-
ing to the machine.
There are two sorts of preventive maintenance
operations:
E a tool change, necessitating a machine stoppage;
E a control, which does not necessitate a machine
stoppage.
This work has been done in the following frame-
work.

0925-5273/01/$ - see front matter  2001 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 5 - 5 2 7 3 ( 0 1 ) 0 0 1 2 0 - 7
136 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
Fig. 1. Manufacturing line.
Hypothesis 1. The assignment of operators to the
machine is restricted to the case of one machine
assigned to one operator: an operator is always
present on each machine of the line.
Hypothesis 2. The changing of tools is more critical
than the controls; since, through the stoppage of
the machine, it has a negative in#uence on the
number of pieces produced. In a normal situation,
the controls are done in `masked timea, i.e. without
stopping the machine. Consequently these controls
are not considered in this study.
The subject of this article is the scheduling of
preventive maintenance tasks, namely the changing
of tools on all the machines in the line. Anderson et
al. [1] have presented a review of scheduling in
production machines or in computers. They are
academic scheduling problems. Here in this article,
we aim to address the industrial problem de"ned in
Section 2.
2. Scheduling maintenance tasks on the
machines of the line
In this section the maintenance tasks will
be described. In the second part, the evolution of
these tasks in the course of time will be analysed;
then the problem of scheduling these tasks will be
described.
2.1. Dexnition of the maintenance tasks
We consider the maintenance tasks that are of
the preventive type. They are de"ned as follows for
the set of machines that form the line.
Let ¹ be the set of tasks:
*
¹ "t
i3[1, w], k3[1, n ],
* GI G
where t is the task k on machine i of the line,
GI
n the total number of tasks on the machinei, w the
G
total number of machines in line, and N" U n
G G
the cardinality (¹ ).
*
These tasks occur cyclically after a "xed time of
machining. So each task has multiple occurrences
in the course of production. They are de"ned by the
following three parameters.
The "rst parameter of a task is the period of
running after which the task is launched. It depends
on the wearing out of tools during the machining. It
can be naturally expressed in number of pieces
"nished and is de"ned by the parameter b%(t ) for
GI
the task t
GI
The second parameter is the shortened running
time before the launching of the "rst occurrence
of a task. The "rst occurrence is speci"c in that
it can be selected before production starts.
It is de"ned by the parameter f%(t ) for the
GI
task t . We have the following constraint on
GI
f%(t ): f%(t )(b%(t ). This parameter corres-
GI GI GI
ponds to a premature change of tools in the starting
phase of manufacturing.
The last parameter is the task duration expressed
in time units. It is de"ned by the parameter d (t )
GI
for task t . In order to express the three para-
GI
meters in the same unit, an approximation in the
number of pieces produced is worked out to express
the duration de"ned by parameter d% (t ). This
GI
approximation is based on machine cycle time
&&Tc'', which is the span of time between the output
of two consecutive pieces. The task duration is
precisely de"ned by the parameter d%(t )"
GI
d(t )/Tc  with the following notation: x, the
GI G
 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
integer closest to x and Tc , the duration cycle of
G
machine i.
2.2. Task analysis on one machine with multiple
maintenance tasks
This section describes the case of a machine in-
dexed i with n maintenance tasks. Here it is impor-
G a periodical function with a simple de"nition. Since,
tant to notice that the span of time between two
consecutive occurrences of the same task t varies
GI
since the machine is also stopped for other tasks to
be performed. The duration between the v and v#1
occurrences of the same task t is de"ned by the
GI
parameter a%(v, t ) with occurrence number v.
GI
The following two functions are used to repres-
ent the problem.
E &R+ , the running function of the machine: R(t)"0
if the machine is stopped at time t; R(t)"1 if the
machine is running at time t. Two causes can stop
the machine: either a maintenance task t is in
GI
progress, or the machine is idle, starved because of
the lack of pieces in the upstream bu!er or
blocked because the downstream bu!er is full.
E &P', the discrete running function which gives
a representation of the number of pieces produc-
ed by the machine.
Fig. 2. Parameters (b%, d%, f%) of a machine with multiple tasks t .
137
Fig. 2 shows three successive occurrences of
t task on a multiple maintenance tasks machine
GI
and the corresponding functions R and P. These
three occurrences are at di!erent time spans. How-
ever the number of pieces produced between two
consecutive occurrences, which is represented by P,
is constant with the value b%(t )"3.
GI
As a result, it appears that function P is not
in general cases, it is di$cult to give an analytical
formulation of the running function, it is also im-
possible to calculate the number of pieces produced
in a given space of time.
With the tasks precisely characterised and ana-
lysed we are now able to specify our scheduling
problem.
2.3. Scheduling the tasks of the set of machines
in the line
The scheduling of maintenance tasks consists of
the de"nition of the values corresponding to the
three parameters (d%, b%, f%) for each task t . So,
GI
the general form of our scheduling function S is
S%: ¹P-;-;-,
t C S%(t )"b%(t ), d%(t ), f%(t )
IG GI GI GI GI
GI
138 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
and M the matrix associated with the scheduling
1
function is de"ned as follows:
M "m(r, c)
r3[1, N], c31, 2, 3,
1 ing problem.
i3[1, w], k3[1, n ]
m(r, 1)"b%(t ),
G GI
m(r, 2)"d%(t ), m(r, 3)"f%(t ).
GI GI
In the context of this study, parameters b% and
d% are assumed to be constant. Actually b% is
related to mechanical know-how, wearing out of
tools and d% refers to methods of organising main-
tenance. Methods, derived from the concept of
just-in-time, such as the single minute exchange of
dies (SMED) aim to optimise this latter parameter.
Consequently, scheduling function S% can be
simpli"ed in the following way:
S%: ¹P-,
t C S%(t )"f%(t, k)
GI GI
and V the vector associated with scheduling func-
1 The speci"c scheduling problem studied in this
tion S%. The de"nition of this scheduling function
corresponds to the time setting of premature
change of tools in the starting phase of manufactur-
ing.
3. Position of the problem
The above-mentioned scheduling problem may
be characterised according to [2], once the terms
&&tasks'' and &&resource'' have been de"ned.
E The maintenance tasks to be scheduled are the
changes of tools, the whole of them being de-
scribed as set ¹. These tasks, called generic tasks,
are cyclically done. The occurrence of order v of
the task t is called a%(v, t ). Again, it is worth
GI GI
noticing that unlike with the mainstream ap-
proach, we are not looking at directly scheduling
production tasks but maintenance tasks.
E The resources required for a given instance of
a task to take place are, on the one hand, the
machine in the line whereupon the tools are to be
changed and, on the other hand, the operator
who has to change them. This operator is a dis-
junctive resource since he performs only one
operation at a time.
Moreover the tool changing operation is unin-
terrupted and cannot be broken down into partial
sequences. It is a cyclical, non-pre-emptive schedul-
Hanen et al. [3] gave an overview on the prob-
lems of cyclical scheduling. The problem of &&Basic
Cyclic Scheduling'' is to determine the starting time
for each instance of generic task, taking into ac-
count the precedence constraints between them.
The problem we are studying is therefore a particu-
lar case of `Basic Scheduling Problema in which
the instances of tasks have a periodical law of
occurrence in the referential timing of pieces. The
precedence constraints between tasks instances do
not exist. So, the unique parameter to be deter-
mined is the timing of the "rst instance, called
f%(t ).
GI
4. General approach to solving the problem
article was exposed above. Section 4.1 introduces
the software environment we have developed to
improve performances in the manufacturing system
and permits us to solve our scheduling problem.
Section 4.2 presents the optimal scheduling de"ni-
tion and the method for determining it.
4.1. Software environment for performance
improvements of manufacturing systems
The software architecture presented in Fig. 3 is
made up of four main modules that are described
subsequently.
The simulator is the central module of this envi-
ronment, and is presented below.
4.1.1. Simulator
This module aims to evaluate each scenario of
simulation requested by the optimiser and analyser
modules. As this simulator is repeatedly utilised by
other modules, it was decided to use a simulation
tool that is &&lighter'' and as e$cacious as possible.
That is why we chose to develop a new simulator,
rather than utilise a commercial one.
This simulator was implemented utilising the
C## classes library called C##SIM [4]. This
 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
Fig. 3. Software architecture for performance improvement.
library aims at simulation of discrete events sys-
tems, using the so-called process approach. The
object oriented approach of C##SIM permits
the construction of derived classes through
specialisation of classes dedicated to the simula-
tion. For instance, to simulate the production line
studied, the main application classes created were:
milling machine, bu!er, operator, change of tools.
The complete design model [5] of this simulator
was expressed with the uni"ed modelling language
(UML) [6].
4.1.2. Model editor
This module o!ers the user a simple modelling
language oriented toward discrete manufacturing
system. This editor proposes basic activities such
as, for instance, bu!ering and milling. In the "rst
step of modelling each activity is graphically se-
lected and connected to the other activities in order
to construct the whole manufacturing #ow of the
line. In the second step, the user must give a value
to the attributes of each activity. For instance,
the bu!er activity has an attribute corresponding to
the maximum capacity of the bu!er.
This module was written in Java language be-
cause of its ability to create graphical interfaces,
and because its relative lack of e$cacy does not
penalise the performance of this software architec-
ture.
4.1.3. Optimiser
The objective of this module is to improve the
performance of the production system. It is neces-
sary to choose one or more parameters to optimise,
and also de"ne one criterion of optimisation. The
139
Fig. 4. Genetic algorithm.
optimiser makes a request to the simulator when-
ever it has to assess a given scenario.
The problems tackled, thanks to this environ-
ment, are often NP-hard. The evaluation function
is generally non-linear and often not even explicit.
For these reasons we used evolutionary algorithms
and more precisely the genetic algorithms [7,8].
The family of genetic algorithms is one of the
machine learning models whose modus operandi
derives from imitation of certain mechanisms ob-
served in biological evolution. The principle of gen-
etic algorithms is mentioned in Fig. 4. Starting from
an initial population of N random solutions or
chromosomes, each iteration or generation consists
in choosing N (N )N) pairs of parent chromo-
 
somes for the mating. The reproduction process
generates the o!spring through crossing-over and
mutations. The new generation of descended chro-
mosomes obtained in this way o!ers solutions
whose adaptation to the problem considered is still
better. The algorithm stops as soon as the perti-
nence of solutions ceased to improve. A good syn-
thesis of the use of these algorithms as concerns the
scheduling problems is to be found in [9].
The main two operators of Genetic Algorithm
are crossover and mutation.
E The "rst one recombines, with a probability
p the genetic material of two parents chromo-
A
somes to make two children. If no crossover
takes place, the o!spring is the exact copy of the
respective parents. For instance, in the classical
crossover called one-point crossover, the chil-
dren are made by cutting the parents at a ran-
domly selected point and exchanging parental
genetic material after the cut. In that way, the
"rst child of parents P1 and P2 is made up of P1
until the cutting point and P2 after.
140 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
E The mutation alters one or more genes of a
chromosome with a probability equal to the
mutation rates. For instance in the case of a
chromosome with binary genes, if a bit is selected
for mutation, it would be #ipped from zero to
one or conversely.
The above-mentioned pattern of genetic algo-
rithms must be customised to the problem to be
solved. Particularly, it is necessary to de"ne the
solutions of the problem under the form of chromo-
somes.
Our genetic algorithm implementation makes
use of Libga Freeware Library [10].
4.1.4. Analyser
This module assists the user in analysing the ef-
fects of the main constraints operating on the perfor-
mances of the production system. Those constraints
are, for instance: operators, maintenance operations,
breakdowns. This analysis is conducted in order to
focalise the performance improvement process on
the parameters associated with the aforementioned
constraints. This analysis step is very useful in the
context of complex production systems, where the
number of parameters potentially in#uent on the
performance of this system is very high. Once this
analysis has been conducted, it is possible to use the
optimisation module to perform studies of improve-
ment on the selected constraint.
The analyser module is based on the Taguchi
method [11] which proposes the experiments ne-
cessary to "nd the impact of the main constraints.
This method consists of a series of experiments,
planned in advance, in order to determine as quick-
ly and as exactly as possible the in#uence of a set of
factors on one resulting parameter. This approach
which has been successfully applied in the "eld of
quality control, is used here in the context of perfor-
mance improvement.
Starting from the factors and the output selected
by the user in charge of the analysis, this module
"rstly prompts the construction of the ad hoc or-
thogonal arrays, taken from those proposed by
Taguchi. It then secondly automatically performs
all the experiments de"ned in the orthogonal array.
Each trial or experiment contained in the ortho-
gonal arrays corresponds in our case to a scenario
evaluated by the simulator module. To complete
this analysis, the user must interpret the results
obtained.
For the same reasons as in the case of the editor
module, the module under development is written
in Java.
4.2. Optimal scheduling dexnition and search
In this section the notion of optimal scheduling is
introduced and the method to "nd a good or near-
optimal solution is presented.
The objective in scheduling the tasks of all the
machines in the line is to improve the throughput.
Throughput } called  } can be characterised in
more than one way, such as for example, the num-
ber of pieces produced by the machine in a certain
production time. Therefore, the task scheduling
consists in determining vector V which enables the
1
best  throughput. The optimal scheduling de"ni-
tion can be schematised as follows: given,
E  the line throughput,
E g the function that gives  in relation with V
1
"nd the scheduling vector V that gives the best
1
throughput: "g(V ), objective: max(g).
1
We said above that discrete running functions
are not easy to "nd for a single machine with
several tasks. This is even more so when it comes in
the general case of a line made up of several ma-
chines, to expressing g, which is dependent on the
running function of each machine. This is why the
environment presented in the previous section is
appropriate for the resolution of the aforemen-
tioned optimising problem. The search for good or
near-optimal scheduling is achieved through the
following two modules:
E The simulator assesses the "tness of a given
schedule de"ned by the vector V , starting from
1
the model of the line elaborated by the editor
module. It returns the worked-out  throughput
to the optimiser.
E The optimisation module searches for good
solutions to the scheduling problem.
As there are N variables, each one being in the
range of values between 1 and BM , the size of the
search space S of this problem is approximately
 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
equal to
S"(BM ),,
U LG
BM "1/N b%( j ),
GI
G I
where BM is the running duration mean of all the
tasks,and N the number total of tasks.
For a line with 10 machines, the possibilities are
of the order of a few thousand BM and a few hundred
for N. This exponential complexity is adapted to
the genetic approach used in the optimiser module
of our performance improvement environment.
5. Validation
The validation of this approach was performed
on an actual car engine production line. This line is
made up of six machines with as many as 61 tool
changes (N"61). The de"nition domain of the
parameter b% called &&initial running duration'' va-
ries from 300 to 10,000 depending on the tool
(mean BM "4200).
The principle of this validation consists in
measuring for each trial, the number of pieces pro-
duced at the end of the line during a given produc-
tion time. This validation corresponds in our
application to about a week of production. This
line produces, in its pre-existing state, prior to an-
ticipation in the change of tools, 15,923 pieces. So,
we measured the production gain of our approach
against this particular value.
This validation was performed in four steps. The
"rst step is the adaptation of the general principle
of genetic algorithm to our application. The second
step consists in exploring the e!ect of parameters
on the results of the genetic algorithm (GA). The
third step of the validation consists in comparing
the results of the genetic algorithm with a naive
optimisation approach. In the last step, the whole
set of validation experiments is synthesised.
5.1. Adaptation to the problem
5.1.1. Coding the chromosome
In our application each individual represented in
Fig. 5, is described as an array. The dimension of
141
Fig. 5. Representing a chromosome.
the array re#ects the total number of maintenance
tasks `Na. Each gene of the chromosome is an
integer corresponding to a possible value for the
parameter f%(t ) of the task t
GI GI
5.1.2. Fitness function
As was described previously, in our application,
the evaluation function of solutions is implemented
by the simulator.
The next section describes the validation of
our general approach to solving the scheduling
problem.
5.2. Study of GA parameters
The choice of parameters in a genetic algorithm is
a major di$culty when it comes to their utilisation
[12]. Subjectively, we chose to leave certain para-
meters unchanged throughout all the experiments
and let other parameters vary. The unchanged gen-
etic operators are set in the following way:
Number of individuals by population: 100
One-point crossover probability: 0.8
Number of runs: 32
Stopping condition: Convergence of the population
It must be noted that the results given hereafter
are the mean of 32 runs for a given set of genetic
operators. That confers a certain robustness to the
results obtained through these genetic algorithms,
the comportment of which is stochastic.
5.2.1. Design of experiments
We study the e!ects of three parameters con-
sidered as important for the performance of the
GA: selection, initialisation and mutation. For this
purpose, we used the same approach of designed
experiments as the one adopted in the analyser
module of our performance improvement environ-
ment. The Taguchi method was used essentially to
142 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
construct the matrix experiments. More precisely,
this study considered the following parameters with
their associated levels:
Selection: Two selection methods were tested.
The "rst one is the tournament, which consists in
randomly drawing two individuals in the popula-
tion and choosing the individual best suited to
generating o!spring. The second is the classical
roulette wheel method in which each parent's
chance of being selected is directly proportional to
its "tness.
Initialisation: Two methods for creating an initial
population were tested. The "rst one initialises
a population with randomly chosen individuals.
The second initialisation method introduces into
a randomly chosen population, ten per cent of
individuals obtained by locally optimising each
machine independently of the others. This local
optimisation is rapid because the search space is
restricted to the tools of one machine.
Mutation: Four types of mutation were tested.
Numbers 1 and 2 consist of a random mutation
with respectively low and high mutation rates.
Numbers 3 and 4 replace the classical mutation
operator by a local search. In our case this local
search is a random search on one of the dimensions
of the scheduling vector.
Those di!erent levels are referred to by the num-
bers given in Table 1.
We used the matrix of experiments presented in
Table 2 to study the e!ects of the above-mentioned
parameters on the through-put. This orthogonal
array was derived from a standard Taguchi array.
It is a fractional array that limits the number of
experiments to only 50% of the possibilities. It is
important in our context where the evaluation of a
solution by simulation is costly in computation time.
Table 1
Reference numbers of the di!erent levels
Mutation
1 Low random (probability 0.001)
2 High random (probability 0.1)
3 Low local search (probability 0.001)
4 High local search (probability 0.1)
The results of these experiments are reported
in Table 2. The "nal column, called m, re-
presents the mean of the number of pieces produced
by the 32 runs conducted with the same set of
parameters.
The means of the experiments for each possible
level are shown in Table 3. This table enables
the determination of the best level for each
parameter. For instance, the mean of experiments
GA2, GA4, GA5 and GA7 in which the parameter
`Selectiona is set to level two is 17,001. This
value is better than the mean of the experiments
in which the parameter `Selectiona is set to level
one.
According to Table 3, the best combination of
levels for the three parameters is:
Mutation (2): high random;
Initialisation (2): local optimisation;
Selection (2): roulette wheel.
It was necessary to carry out this con"r-
mation experiment, which had not been done
in the previous matrix of experiments. This experi-
ment, called GAm, produced the results given in
Table 4.
This con"rms the conclusions of the matrix of
experiments as it produces a mean value of pieces
produced that is better than all the previous results.
Nevertheless, these conclusions were obtained from
a fractional design of experiments and ignored the
e!ects of possible interactions between parameters.
Consequently, it is not possible to claim that GAm
represents the best level combination of the three
parameters studied.
5.2.2. Statistical tests
It is necessary to con"rm the previous analysis
by the study of the statistical signi"cance of the
Initialisation Selection
1 Random 1 Tournament
2 Local optimisation 2 Roulette wheel
 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146 143

Table 2 Table 4
Orthogonal array of the GA parameters Con"rmation experiment

Trial Mutation Initialisation Selection m Mutation Initialisation Selection m

GA1 1 1 1 16,706 Gam 2 2 2 17,088

GA2 2 1 2 17,003
GA3 3 1 1 16,719
GA4 4 1 2 16,993
GA5 1 2 2 17,007
GA6 2 2 1 17,076
based on the statistics reported in Table 6, which
GA7 3 2 2 17,003 corresponds to a standardised normal deviate. For
GA8 4 2 1 17,079 a unilateral statistical test, the table shows the
following.
E Experiments GA6, GA8 and GAm are signi"-
Table 3 cantly better than the other experiments at the
Means of experiments for each level 0.5% con"dence level. The critical value of this
Levels Mutation Initialisation Selection
statistic being z equals 2.58 at this degree of
con"dence.
1 16,856 16,855 16,895 E The di!erences between the three experiments
2 17,040 17,041 17,001 considered, GA6, GA8 and GAm are not signi"-
3 16,861 * * cant at this level of con"dence and we cannot
4 17,036 * *
reject the null hypothesis.
A "rst partial conclusion con"rms that the re-
observed deviations between the means of the dif- sults of the GA are better with a non-random
ferent experiments. The standard deviations was initialisation. Concerning the levels of mutation
calculated and reported in Table 5. tested, it is clear that the best results are obtained
In order to limit the number of couples of com- with a high level of mutation, either with the pure
parisons, this deviation analysis was focused on the random mutation, or with the use of a local search.
three experiments GA6, GA8, GAm producing the However, at this step of validation analysis, it is not
best values for the maximum number of pieces. possible to decide which selection method is the
A signi"cance test of the null hypothesis was used, best.

Table 5
Standard deviation of the 32 runs of each experiment

Trial GA1 GA2 GA3 GA4 GA5 GA6 GA7 GA8 GAm

 9.82 46.49 22.33 53.09 41.23 27.07 29.02 30.76 30.66

Table 6
Statistics for comparing the maximums of GA experiments

GA1 GA2 GA3 GA4 GA5 GA6 GA7 GA8 GAm

GA6 70.54 7.51 55.79 7.65 7.76 * 10.16 0.32 1.53

GA8 63.31 7.48 51.85 7.65 7.69 0.32 9.84 * 1.15
Gam 63.21 8.32 52.19 8.41 8.59 1.53 10.88 1.15 *
144 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
Fig. 6. Performance against the number of solutions explored.
5.2.3. Performance versus computation cost
Our goal is to "nd a good solution with a reason-
able computation time cost.
As a matter of fact, the time necessary to evaluate
a solution by simulation is not negligible and gen-
erates an important convergence time of the genetic
algorithm. For instance, the mean convergence
time of the GA is about one hour and a half on
a Pentium II 350 under Linux. So, for each experi-
ment, the mean number of explored solutions was
measured. This number can be considered an indi-
cator of computation time. Fig. 6 brings out the
following points:
E GA6 and GA8 are better than GAm because the
number of solutions explored is divided by two.
E The di!erence between the numbers of solutions
explored in GA6 and GA8 is small.
Table 7 reports the statistics corresponding to
a standardised normal deviate concerning the com-
parisons of the number of explored solutions.
This table con"rms that the computation cost of
GA6 is not signi"cantly lower than that of GA8
with a level of con"dence of 99.5%. This table also
con"rms the lower quality of GAm comparatively
to GA6 and GA8.
The best compromise between performance and
computation cost is obtained with experiments
GA6 and GA8. The conclusion of the previous
paragraph can now be re"ned with the choice of the
tournament selection method common to the ex-
periments GA6 and GA8, which are less costly in
computation time. However, the interest of the lo-
cal search mutation was not proved.
Table 7
Statistics for comparing the computation time
GA8 GAm
GA 6 1.71 12.24
GA 8 * 9.93
5.3. Comparing GA with a naive approach
In order to validate the genetic approach, we
compared it with a random search. Fig. 7 presents
the convergence curve of experiments GA6, GA8
and GAm compared with the random search
This random search was performed a number of
times equal to the number of solutions explored by
the genetic algorithm. These three experiments are
better than the random search. But it is worth
noticing that two experiments of the genetic algo-
rithm, namely GA1 and GA3, produced results
inferior to the random search
Table 8 contains the mean and the standard
deviation of the 32 runs in the Random search for
the maximum number of pieces produced. Table 9,
deduced from Table 8, presents the statistics related
to the comparison of the maximum between ran-
dom search and the experiments GA6, GA8 and
GAm. This table con"rms the better results ob-
tained by the GA with a con"dence level of 99.5%.
5.4. Validation: Synthesis
When it comes to assessing our approach to the
scheduling of the changes of tools by a genetic
 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146 145

Fig. 7. Comparison of the random search and GA.

Table 8 by a naive optimisation such as a random

Mean and standard deviation of the 32 runs for random search search.
E The parameter setting of the genetic algorithm
m 
determines its performance. A statistical study
Random search 16,925 26.88 proves the interest of utilising a non-random
initialisation. On the other hand, this statistical
analysis con"rms the superiority of high
Table 9
mutation rates. In the end, the tournament
Statistics for comparing the maximums of random search selection method proved to be better than
against those of GA the roulette wheel in respect to the speed of
convergence.
GA6 GA8 GAm

Random search 21.43 20.27 21.18

6. Conclusion

The method proposed o!ers a theoretical inter-

approach, we can now say:
E The GA enables production gains that are about
7.5% better than the approach without anticip-
ated changes of tools.
E This approach is validated because it signi"-
cantly improves the production gains o!ered
est, veri"ed by simulation. The gain obtained is
considerable, given the price of the pieces produced.
Indeed, the industrialists of the sector concerned
admit that a solution might be implemented if the
resulting gain exceeded 1%.
A feasibility study is currently under way with
our industrial partner to assess the practicality of
146 G. Cavory et al. / Int. J. Production Economics 74 (2001) 135}146
this approach on an actual production line. The
problems are the material constraints derived from
the anticipated changes of tools. The consequences
of breakdowns on improvement is also under
evaluation.
The pursuit of this study will be oriented towards
the following points:
Firstly, assessing the cost of each solution in
order to choose the most pro"table one. The pro-
duction gains must be compared with the extra cost
entailed by the premature change of tools. A multi-
criteria analysis may be useful to connect the
several optimisation objectives: throughput maxi-
misation, cost minimisation.
Secondly, there is the question of assigning the
operators to the machines. One must study the
general case of multiple operators assigned to a set
of machines. Both their location in relation to the
machines and their level of training must be taken
into account. The use of constraints system pro-
gramming could be helpful.
At last, it must be noticed that the approach
presented in the case of a single product manufac-
turing production line can easily be extended to
a multiple products line.
Acknowledgements
Many thanks to my father, A. Dupas, for helpful
comments and translation.
References
[1] E.J. Anderson, C.A. Glass, C.N. Potts, Discrete mathemat-
ics and optimisation, in: E. Aarts, J.K. Lenstra (Eds.), Local
Search in Combinatorial Optimisation, Wiley, New York,
1996, pp. 361}414 (Chapter 11).
[2] Gotha, Les problèmes d'ordonnancement, RAIRO-
Recherche OpeH rationnelle 27 (1) (1993) 77}150.
[3] C. Hanen, A. Munier, Cyclic scheduling on parallel proces-
sors: An overview, in: P. Chretienne, E.G. Co!man, J.K.
Lenstra, Z. Liu (Eds.), Scheduling Theory and its Applica-
tions, Wiley, New York, 1994 (Chapter 7).
[4] C##SIM User Guide, Department of Computing
Science, Newcastle upon Tyne, Arjuna Project, 1994.
[5] R. Dupas, G. Cavory, G. Goncalves, U.M.L. design model
of a simulator for discrete event production system,
LabOGP, Technical Report no. 99-2, FaculteH des Sciences
AppliqueH es, UniversiteH d'Artois, France, 1999.
[6] G. Booch, J. Rumbaugh, I. Jacobson, The Uni"ed Modelling
Language User Guide, Addison-Wesley, Reading, MA, 1998.
[7] D.E. Goldberg, Genetic Algorithms in Search, Optimisa-
tion and Machine Learning, Addison-Wesley, Reading,
MA, 1989.
[8] Z. Michalewicz, Genetic Algorithms#Data Structures"
Evolution Programs, Springer, Berlin, 1992.
[9] EvoStim, The EvoNet Working Group on Evolutionary
Scheduling and Timetabling: The state of the art in evolu-
tionary approaches to timetabling and scheduling, URL
http://www.dcs.napier.ac.uk/evonet/Coordinator/wg }
f.html, 1998.
[10] A. Corcoran, Freeware LIBGA, URL http://
enter.mcs.utulsa.edu/&corcoran/libga.html, 1993.
[11] J. Alexis, Pratique industrielle de la meH thode Taguchi, les
plans d'expeH rience (AFNOR), 1995.
[12] M. Sebag, M. Schoenauer, Contro( le dun algorithme geH n-
eH tique, Revue Intelligence Arti"cielle 10 (2}3) (1996)
389}428.