Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Int. J.

Production Economics 64 (2000) 91}100

Flow simulation and genetic algorithm as optimization tools


F. Fontanili!,*, A. Vincent!, R. Ponsonnet"
!GRPI, IUT Cergy, 95, rue Vale% re Collas, 95100 Argenteuil, France
"GRPI, IUT St Denis, Place du 8 mai 1945, 93000 St Denis, France

Abstract

This paper presents the use of #ow simulation and genetic algorithm to optimize the management parameters in
production. To start with, this study reviews the main parameters to be taken into account for the management of an
assembly line with bypass workstations. Then, we develop two cases of optimization to check the validity of the
combination between #ow simulation and a genetic algorithm for obtaining of optimal parameters. The "rst case
corresponds to a production of only one type of product. The second case is more complex: several di!erent types of
products are simultaneously produced. ( 2000 Elsevier Science B.V. All rights reserved.

Keywords: Flow simulation; Optimization; Assembly line; Genetic algorithm

1. Presentation of the experimental system excepted, each product is realised in only one
round, without necessarily going on all the work-
In order to illustrate #ow simulation and process stations.
optimization, we have chosen to apply our research The assembly routing of a product is constituted
on a system presently in operation in our laboratory. of `Non-permutate and non-redundant phasesa. For
This system has a widespread architecture for instance, a product's assembly path leads from by-
industrial applications called Free Modular Trans- pass-workstations 1 to 2, 2 to 5, and 5 to 6, in this
fer. Three types of functional modules are used order only.
(Fig. 1): loading/unloading line (]1), main line For the rest of this paper, the products will be
(]1), and bypass workstations (]6). regarded as pallets, the pallets on which the prod-
The links between the di!erent modules help to uct is assembled during manufacture.
obtain the structure shown in Fig. 2. The chosen architecture o!ers one major advant-
This line is able to produce simultaneously up to age over the classical architecture with worksta-
ten di!erent products, released in variable quantit- tions in series: the pallets #ow is independant (or
ies, and with di!erent routings and running times. asynchronous) with respect to the operations tak-
The products move on the line according to the ing place at the di!erent stations. Moreover, it
generalised Flow Shop: occupation of a workstation permits the simultaneous manufacture of products
with di!erent routings. Another advantage is found
in the system's reliability and ease of maintenance;
* Corresponding author. these features result from the independently con-
E-mail addresses: franck.fontanil@u-cergy.fr (F. Fontanili), trolled workstations which ensure continuous pro-
arnaud.vincent@u-cergy.fr (A. Vincent) duction even when a given workstation fails.

0925-5273/00/$ - see front matter ( 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 5 - 5 2 7 3 ( 9 9 ) 0 0 0 4 7 - X
92 F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100

2. Presentation of the management parameters

2.1. Access rule for pallets onto a bypass workstation

Pallet access to a bypass workstation is deter-


mined by a rule that takes into account parameters
that depend on the pallet itself (routing, defect,
completion of prior operations), as well as a double
condition algorithm involving the workstation
itself (availability, load of upstream stock). Fig. 3
illustrates the access algorithm.

Fig. 1. Functional modules.

Fig. 2. Assembly line structure.

The management of such line, particularly when


di!erent products are in process, is very di$cult.
The questions we must answer are

f What is the interval of time between the release


of two consecutive pallets?
f What is the batch size for each fabrication order?
f What priority rules must we use at the exit of
each workstation?
f What is a `bottlenecka workstation?
f What is the maximum number of pallets that we
can release without saturation of the line?
f What is the best shop scheduling?
f etc. Fig. 3. Access algorithm.
F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100 93

2.2. Priority rule at bypass workstation exit cording to a FIFO queue. The "rst management
parameter to de"ne is therefore the inter-release
In order to prevent the pallets on the main line time (IRTi) which corresponds to the lapse of time
from colliding with the pallets leaving a bypass between the release of two consecutive pallets in the
workstation, it is essential to determine a right of queue. This parameter may either be constant or
priority. This problem, apparently simple, is prob- vary from pallet to pallet.
ably the most complex to solve, because of the At any bypass workstation, we can de"ne two
multitude of priority possibilities and their e!ects parameters:
on the global functioning of the line. An arbitrary
f the upstream stock capacity (UpSt): a bu!er sec-
rule has been chosen whereby correct operation is
tion located at the entrance of workstation,
monitored at each bypass workstation. Fig. 4 de-
f the downstream stock capacity (DoSt): a bu!er
scribes the corresponding algorithm.
section located at the exit of the workstation.
Other rules, from the simple to the more com-
plex, can be established to work out which pallet The maximal physical capacity of each of these
goes "rst on leaving a bypass workstation. Choos- bu!ers is six pallets. The UpSt value is a parameter
ing one of these rules is not easy, because it is taken into consideration in the access rule (Fig. 3)
di$cult to forecast the e!ects on the global yield. and the DoSt value is a parameter for the exit rule
The choice is all the more di$cult as di!erent rules (Fig. 4). Each bypass workstation may have di!er-
could be established for each bypass workstation ent UpSt and DoSt values.
according to other constraints.
One of the goals of our research is to determine
the optimal local rules associated with each work- 3. Limits of the study
station.
In this study, we shall limit ourselves to the rule The objective function depends on the market, in
described before (we are also researching and test- a make-to-order approach: the cycle time of a pallet
ing other rules). batch must be as short as possible.
It is also possible to "nd a compromise between
2.3. Flow management parameters the shortest lead time and the minimal number of
pallets being processed.
The pallets are sent into the line from the load- The quantity of parameters and the di!erent
ing/unloading area, where they are parked ac- values they can take made us realize that our pro-
duction system was particularly complex to man-
age.
Indeed, if we had tried to "nd the optimum
values for the parameters for a 25 pallet batch
production, a very high number of possible combi-
nations would have been found:
f concerning the inter-release time (IRTi), if we
consider that the value can be between 1 and 20
sec, and may vary from one pallet to another,
then there are 20(25~1) possible combinations!
or, more generally (this assumes integer times):
number of combinations
"[Max(IRTi)!min(IRTi)](/6."%3 0& 1!--%54~1);
f concerning the upstream stock capacity (UpSt)
Fig. 4. Priority rule. of each workstation and for a product needing to
94 F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100

stop on four workstations, there are 64 (i.e. 1296)


possible combinations. or more generally:
number of combinations"
Max(UpSt)/6."%3 0& 803,45!5*0/4.
In the rest of our study, we have decided to take
into account only one parameter:
the inter-release time (IRTi).
The other parameters are "xed on an arbitrary
value.

4. Simulation without optimization

To test di!erent values of the inter-release time


(IRTi), we have made a model of the line and used
#ow simulation with WITNESS software [1]. With
this tool, situations can be tested in a few seconds Fig. 5. Product routing.
rather than hours on the real system. The interest of
such a tool for research goes without saying [2}4].
So we used #ow simulation mostly, and the most f With 1 second as IRTi, the simulation gives a
interesting results were checked and proved right manufacturing lead time of 725 seconds.
on the real system. f With 8 seconds. as IRTi, the simulation gives a
manufacturing lead time of 519 seconds.
4.1. First case of production
f etc.
We want to launch a 25 product batch of the We can check each value of IRTi, but in this case, it
same type. The assembly routing is as follows: is easy to determine the best one: it's the value
W1[5]PW2[8]PW5[4]PW6[2] which is synchronized with the `bottlenecka work-
station, here workstation number 2. The inter-re-
with Wi[ j] the operation on Workstation i, opera- lease time (IRTi) must be identical to the cycle time
tion time"j seconds. of the `bottlenecka workstation, 8 seconds. In prac-
The theorical path of a product is the bold draw- tice, we add a constant of 4 seconds to this time,
ing in Fig. 5. which corresponds to the idle time between the
In practice, it is possible that some pallets loop arrival of two pallets at the workstation. The inter-
one or more times because they cannot access to release time (IRTi) must be
a workstation during the "rst loop.
The objective is to determine the best inter-re- 8#4"12 seconds.
lease time (IRTi), the one that leads to the minimal We run the simulation with an IRTi value of 12
manufacturing lead time (the makespan). seconds and we obtain a manufacturing lead time
Thanks to the simulation, it's easy to proceed of 475 seconds.
experimentally, from an initial arbitrary value of The equation of this time is
the inter-release time (IRTi).
First, we decide to run simulation with a con- MLT"FPLT#[(N!1)](BCT#4)]
stant value of IRTi between each pallet. "187#[(25!1)](8#4)]"475 seconds.
Results of diwerent runs: where MLT is the manufacturing lead time, FPLT
f With 3 seconds as IRTi, the simulation gives a the "rst pallet lead time, N the number of pallets,
manufacturing lead time of 633 seconds. and BCT the bottleneck cycle time.
F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100 95

4.2. Second case of production We apply immediatly the principle of synchroni-


zation between the bottleneck workstation and the
We want to launch a manufacturing campaign, inter-release time. In this case, the di!erence is that
which means a set of successive product batches, the bottleneck workstation is not necessarily the
with di!erent sizes and di!erent routings. Each same on each batch, and the cycle time of a given
batch corresponds to a manufacturing order (MO), workstation can vary from one batch to another.
with a routing and a quantity of products. In our Fitted to this case, the inter-release time settles to
case, the release of the campaign is scheduled as 8#4"12 seconds for the MO1 batch of "ve pallets;
follows: 5#4"9 seconds for the MO2 batch of three pallets;
15#4"19 seconds for the MO3 batch of 10 pallets;
1st"MO1("ve pallets) 2nd"MO2(three pallets) 5#4"9 seconds for the MO4 batch of seven pallets.
The theorical manufacturing lead time of this
3rd"MO3(ten pallets) 4th"MO4(seven pallets). campaign (533 seconds) can be determined from the
Gantt chart of Fig. 7.
The matrix in Fig. 6 presents the di!erent routings We can hope that the simulation of this case will
associated with each Manufacturing Order. give a lead time close to the theorical value.
The objective of this campaign release is identical
to the "rst case: to determine the optimal inter- Simulation result: 601 seconds
release time between each pallet to obtain the min- This result is over the theorical lead time and so
imal manufacturing lead time. we decide to check all the values of inter-release

Fig. 6. Routings matrix.

Fig. 7. Gantt chart of manufacturing lead time.


96 F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100

time between 1 and 35 seconds. To limit the num- optimum of an objective function comprising
ber of combinations, the inter-release time is "xed numerous variables, without testing all possible
to a constant value between each pallet of a cam- combinations.
paign. The results of the simulations of these com- This optimization method must be able to
binations are presented in Fig. 8. handle even more complex situations, such as sev-
The analysis of this graphic shows us a discon- eral products with di!erent simultaneous routings
tinuous variation of the lead time, with several Our investigation of this problem has led us to
peaks and hollows. Contrary to the "rst case, it is use a genetic algorithm [5,6]. This type of algorithm
impossible to associate a mathematical function to must be tested with the real system, or even better,
determine the lead time and so the minimal value is a model of the real system. Contrary to other op-
given from the graphic. Moreover, we have two timization methods, such as linear programming or
values of inter-release time which give the minimal quasi-Newton method, it is not necessary to have
lead time (627 seconds): 6 and 14 seconds. If we a mathematical formulation of the model [7,8].
have to choose one of these two values, we must Our #ow simulation model can then be used as the
take a selection criterion. support of the genetic algorithm.
But it is not necessary to develop anything in this An important point to be noticed is that the #ow
case because the minimal value of the manufactur- simulation model is a stochastic model, capable of
ing lead time (627 seconds) is over the value that we taking unpredictable events into account (break-
have found in the "rst simulation, with di!erent downs, setups, changes in cycle times, etc.), and the
values of IRTi between each manufacturing order genetic algorithm can adapt to model variations by
(601 seconds.). itself. So, it seemed relevant to us to test the com-
bined use of #ow simulation and genetic algorithm
[9,10].
5. Simulation with optimization In order to check this algorithm, we decided to
start by working out the optimum inter-release
We are interested in "nding a quick and e!ective time (IRTi), assuming that the other variables were
optimization method, capable of determining the constant.

Fig. 8. Manufacturing lead time"f (IRTi).


F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100 97

The main di$culty in using a genetic algorithm These two individuals are the selected PARENTS
was the encoding of the problem. We resorted to for the reproduction and the creation of the next
viewing the pallets as individuals carrying one or generation.
several genes corresponding to the parameters to
test. In this case, and within the limits we had set, 5.1.4. Reproduction
each individual carries only one gene whose code is Hypothesis: The number of individuals is the
the inter-release time (IRTi) between each pallet. same within each generation (eight individuals/gen-
For example, for the release of a 7 pallet batch, we eration in our example).
thought in terms of one individual whose genes
include 6 elements corresponding to 6 inter-release (a) The selected parents in the previous generation
times. are the "rst two individuals of the active gen-
eration: it is the elitist strategy of our algo-
5.1. Simple example of a genetic algorithm rithm.
(b) The two parents' genes are randomly crossed
5.1.1. Random creation of an eight individual to obtain six CHILDREN in the active genera-
generation (i.e. eight genes) tion.
Example: for the release of a seven pallets batch: Example:
Gene 1: 05 - 12 - 02 - 17 - 11 - 09P genetic encoding Parent 1"Gene 2: 16 - 10 - 08 - 03 - 13 - 02.
of IRTi for the "rst batch, which means that there Parent 2"Gene 7: 12 - 04 - 07 - 19 - 17 - 13.
are 5 seconds between the release of the "rst pallet Child 1"Gene C1: 16 - 10 - 07 - 19 - 17 - 13:
and the second pallet, 12 seconds between the sec- crossing of the "rst two elements of parent 1's gene
ond and the third, etc. and the last four elements of parent 2's gene.
Gene 2: 16 - 10 - 08 - 03 - 13 - 02P second batch. Child 2"Gene C2: 03 - 13 - 02 - 12 - 04 - 07.
............................... .........................
................................ .........................
Gene 8: 07 - 18 - 03 - 04 - 09 - 19. Child 6"Gene C6: 12 - 05 - 07 - 20 - 13 - 03:
Mutant gene which makes possible the generation
of an individual capable of adapting to a modi"ed
5.1.2. Evaluation of the xtness (or objective) function environment. In our algorithm, the mutation con-
of each individual sists in the incrementation or the decrementation of
every other element of the gene.
Example: The objective is to minimize the lead time
of a product batch. 5.1.5. Evaluation of the xtness function of the new
Gene 1: lead time"193 seconds. generation
Gene 2: lead time"168 seconds. Loop back to steps 1}4 for k generations, k being
Gene 3: lead time"260 seconds. determined empirically (in our experiments
....................... 20(k(50 generations).
Gene 8: lead time"219 seconds.
The result of the "tness function (i.e. lead time) is 5.1.6. Hardware and software used in the experi-
obtained by exchanging the data with the #ow ments
simulation model.
f The genetic algorithm was encoded in PASCAL
5.1.3. Selection of the two individuals with the best with BORLAND DELPHI 2 using WINDOWS
xtness function 95.
f The simulation model was done with the #ow
Example: simulation software WITNESS 8.0c.
Gene 2: lead time"168 seconds and Gene 7; lead f The data exchange (IRTi and Delay) between the
time"189 seconds. two applications above was achieved through
98 F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100

MICROSOFT OLE Automation links. WIT- minimal lead time (475 seconds), is obtained with
NESS (and the simulation model) was used a gene completly di!erent from the optimal: 12-12-
as a OLE server and DELPHI 2 as a OLE 12-12- .......
customer. Second case: Release of a campaign with diwerents
f The implementation of those applications was products and routings
done on a PENTIUM 166 MHz computer.
1st"MO1(5 pallets) 2nd"MO42(3 pallets)
5.1.7. Results 3rd"MO3(10 pallets) 4th"MO4(7 pallets).
First case: Release of a 25 products batch with the f Theorical minimal lead time: 533 seconds
same routing (Fig. 7);
f Manufacturing lead time without any optimiza- f Manufacturing lead time with synchronous IRTi
tion: 606 seconds with a constant IRTi of 3 sec- (12 seconds for MO1, 9 seconds for MO2, 19
onds. seconds for MO3, 9 seconds for MO4): 601 seconds
f Manufacturing lead time with synchronous IRTi (a) First optimization with genetic algorithm: con-
of 12 seconds: 475 seconds. stant value of IRTi on each manufacturing order, but
f Fig. 9 shows research of the minimal lead time variable between each order.
with a genetic algorithm during 50 generations. The genetic algorithm gives a result better than
The minimal lead time (479 seconds) is obtained for 601 seconds (Figs. 10 and 11): 558 seconds.
the gene: This minimal lead time (558 seconds) is obtained
for the gene:
12-11-9-6-1-12-4-6-4-4-18-3-9-11-5-1-6-15-3-8-3-
10-2-15 10-10-10-10-10-10-10-10-22-22-22-22-22-22-22-22-
22-22-9-9-9-9-9-9
which corresponds to the inter-release time be-
tween each pallet, variable from one pallet to an- which corresponds to the inter-release time be-
other. We can notice that this result, close to the tween each pallet.

Fig. 9. 1st case: lead time with a genetic algorithm.


F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100 99

Fig. 10. 2nd case: lead time with genetic algorithm and constant IRTi.

Fig. 11. 2nd case: lead time with genetic algorithm and variable IRTi.
100 F. Fontanili et al. / Int. J. Production Economics 64 (2000) 91}100

(b) Second optimization with genetic algorithm: between #ow simulation and genetic algorithms
Variable IRTi between each pallet. is the only reliable way to work out the adequate
The minimal lead time obtained with the genetic IRTi.
algorithm is 514 seconds. This result is smaller than f As the real system requires important response
the theorical lead time (533 seconds, Fig. 7). This delays for each experiment, the use of #ow simu-
result is particularly interesting because the genetic lation is absolutely justi"ed. In this case, suppos-
algorithm has found a better solution than our ing that the mean delay to process 25 pallets is
objective. The gene which corresponds to this lead 560 seconds on the real system, the genetic algo-
time is very `chaotica: rithm would have needed 62 hour to "nd the
optimum solution without the simulation!
8-5-4-16-5-7-5-9-9-16-8-3-1-6-5-7-17-5-8-12-3-8-9-4.
The results obtained so far have led us to envis-
This result shows us that the genetic algorithm
age the further study of:
seems to be very e!ective for problems with lot of
combinations and possible solutions. f optimization of the upstream and downstream
stock size on each workstation;
f optimization of priority rule at the exit of each
6. Analysis of the results and conclusions workstation;
f optimization of the pallets dispatching/schedul-
ing, in the case of several di!erent routings;
f The results obtained with the genetic algorithm
f etc.
associated with the #ow simulation are very
promising because they are better than all the
results that we have found. References
f The calculation/simulation time is proportional
to the number of generations, but does not de- [1] WITNESS ver. 8.0, Manuel de programmation
pend on the number of pallets. With the com- } SERETE.
puter we used, we obtained a calculation/ [2] A.M. Law, W.D. Kelton, Simulation Modeling and Analy-
sis, McGraw-Hill, New York, 1991.
simulation time of 150 seconds/generation, or [3] P. Chau, P.C. Bell, Decision support for the design of
125 min to obtain the optimum lead time for a new production plant using visual interactive simulation,
25 pallets. The increasing performances of com- Journal of the Operational Research Society 45 (11) (1994).
puters lead us to believe that in the near future, [4] R. Caprihan, Simulation studies of alternative control
we will get even shorter times. But those times strategies for FMS, Ph.D. Dissertation, 1995.
[5] D.E. Goldberg, Genetic Algorithms, Addison-Wesley,
are quite acceptable, all the more so in an indus- Reading, MA, 1991.
trial environment where the optimization [6] J. Holland, Adaptation in Natural and Arti"cial Systems,
process can take place during the production 2nd Edition, MIT Press, Cambridge, MA, 1992.
of another batch. [7] E.L. Lawler, Combinatorial Optimization, Holt, Rinehart
f On the other hand, the optimum inter-release and Winston, New York, 1976.
[8] M. Syslo, N. Deo, J.S. Kowalick, Discrete Optimization
time (IRTi) given by the genetic algorithm may Algorithms with Pascal Programs, Prentice-Hall, Engle-
appear disappointing because of their random wood Cli!s, NJ, 1983.
variations. Indeed, not many production man- [9] M.A. Wellman, D.D. Gemmill, A genetic algorithm
agers would agree to release a batch of products approach to optimization of asynchronous automatic as-
with such chaotic parameters, especially pro- sembly systems, International Journal of Flexible Manu-
facturing Systems 7 (1995) 27}46.
vided by a computer! But owing to the use of [10] L. Tautou, H. Pierreval, Using evolutionary algorithms
#ow simulation, we can prove that they make and simulation for the optimization of manufacturing sys-
sense. To us, it seems that this combination tems, IEEE, 1995.

You might also like