Vous êtes sur la page 1sur 11

J Intell Manuf

DOI 10.1007/s10845-010-0434-z

A collaborative scheduling GA for products-packages service


within extended selling chains environment
Pedro Gomez-Gasquet · Raul Rodriguez-Rodriguez ·
Ruben Dario Franco · Angel Ortiz-Bas

Received: 7 July 2010 / Accepted: 8 July 2010


© Springer Science+Business Media, LLC 2010

Abstract The theory of network coordination provides the- parameters are tuned. Although the major contribution of this
oretical foundations to explain how companies can overcome work focuses on the algorithmic development of a proposal
organizational boundaries and constraints to jointly manage in the context of operations research that could help to solve
business processes across their selling chains. In particular, the problem also is discussed the environment in which this
this work focuses on Collaborative Scheduling, a collabora- occurs and that justifies our interest. In order to validate the
tion process whereby selling chain trading partners activate proposed solutions diverse configurations are presented and
either on-line or off-line inter-firm coordination mecha- the results obtained by means of the GA and some heuristics
nisms to jointly plan production activities in order to deliver rules are compared.
the final products to end customers each one of them, being
the delivery date as close to the date desired as possible. The Keywords Collaborative scheduling · Selling chains ·
problem of collaborative scheduling is formally defined by Multi-supplier scheduling · Genetic algorithm
means of a mathematical model. In the model, the defined
objective function has the goal to minimize the total weighted
tardiness of the package of products acquired by the clients Introduction
to be delivered in a specific date. The delivery date of each
Product-Package is conditioned by the latest date established This paper presents the problem of order management within
by each supplier for each product that forms part of the same extended collaborative selling chains, and an identification of
one. Besides, having different process times for each prod- the objectives, requirements and solutions of scheduling in
uct and different penalties for each Product-Package, each this environment. The main issue to be considered in this
supplier can offer a different mix of additional products with context is the production scheduling, whose final goal is to
different due date. Due to the complexity of the problem a satisfy the delivery dates agreed with final customers by using
Genetic Algorithm has been the approach taken for its reso- a collaborative strategy among the suppliers.
lution. The GA elements and procedures are defined and the The scheduling problem presented is not only interesting
from an operational point of view but also from the business
P. Gomez-Gasquet (B) · R. Rodriguez-Rodriguez · R. D. Franco · environment in which it occurs. For this reason, this work com-
A. Ortiz-Bas prises both views. In an Make-to-Order (MTO) environment
Centro de Investigación de Gestión e Ingeniería de la Producción
where companies can develop a line of business in collabo-
(CIGIP), Universidad Politécnica de Valencia, Cno. de Vera s/n,
Valencia, Spain ration with others (collaborative selling) without neglecting
e-mail: pgomez@cigip.upv.es their traditional business and commitments and where such a
R. Rodriguez-Rodriguez collaboration means that several suppliers produce elements
e-mail: raurodro@cigip.upv.es or parts, products for us, that belongs to a single sales unit
R. D. Franco named product-package. It should be emphasized that in this
e-mail: dfranco@cigip.upv.es process customers establish deadlines for product-packages,
A. Ortiz-Bas which, from a productive system point of view, takes us to
e-mail: aortiz@cigip.upv.es decompose such a product-pack into its indivdual products

123
J Intell Manuf

that are managed individually. The main difficulty that arises production scheduling tasks that should be collaboratively
is how to establish the productive scheduling of all providers carried out regarding how to properly manage orders. The
so that each one can deliver the product on time and thus can novelty of this approach is not in its conception but in its
meet the delivery date of product-package agreed with the cus- execution and its linkage to the several collaborative supply
tomer. From a general point of view, suppliers should also take chains.
into account the scheduling of products that are part of a prod- This paper contains five sections beyond the Introduction.
uct-package (collaborative selling), being therefore of com- The following section introduces the trends in supply chains
mon interest, as they might have to be included within their regarding scheduling approaches; section “An example of
other products and also picked up from other business line. collaborative scheduling in extended selling chains” presents
Consequently, this problem may easily turn into a sched- an example of a collaborative scheduling problem within the
uling problem, where experienced people from the differ- extended selling chains; section “Problem definition” defines
ent enterprises should take care of carrying out the pertinent the proposed problem formally; section “A genetic algorithm
scheduling and re-scheduling of manufacturing orders. The for collaborative scheduling within extended selling chains”
scheduling problem can be defined as follows: There are n describes and studies the proposed GA approach by conduct-
product-packages each one comprising a maximum of s prod- ing an experimental analysis; and the last section concludes
ucts. Each product j(s) is available at time zero to be pro- with a short review of the results reached in the research.
cessed on one and only one specific supplier s, which can
process at most one product at a time. Each supplier pro- Literature review
cesses no more than one product by product-package. Prod-
uct—package j has a processing time p j,s and a due date The issue is interesting because globalization, together with
d j . The tardiness of product-package j is defined as T j = national, European and worldwide competitiveness, has
max(0; C j −d j )_where C j is the maximum completion time promoted the creation and consolidation of the so-called
of product-package j in any supplier, C j = max(C j,s ). The Extended Enterprises and Collaborative Networked Orga-
objective is to determine a job schedule δ such that the total nizations, which transcend the single enterprise domain and
n
weighted tardiness T δ = j=1 w j T j is minimized. It is build meta-enterprises (Dyer 2000). Basic concepts and con-
assumed that the processing time of the product-package j in siderations have been described in several previous works
the supplier s, p j,s, can be 0 in one or more suppliers. In the (Byrne 1993; Camarinha-Matos et al. 2005,b; Camarinha-
particular case in which the p j,s > 0 in only one supplier, Matos and Afsarmanesh 2005a; Davidow and Malone 1992).
this is a product and not a product-package. Most of the existing “extended enterprises” (EE) have ini-
This problem could be extended to a make-to-stock (MTS) tially been focused on reinforcing the links and flows among
environment, in which the products would be replaced by the companies that are involved in the same value chain
batches that would not be exempt from certain restrictions in (Macbeth et al. 1998). However, recently, this phenomenon
the configuration. has shifted towards the creation of EE which cross the bound-
The proposed problem is interesting because it is easy to aries of a single value chain, and link together different com-
think of several manufacturers that launch a product pack plementary value chains thus building extended collaborative
to the market and such a product pack becomes a success, selling chains (ECOSELL domain). Many industrial compa-
as customers like the extended value proposition obtained nies have desired, or actually achieved, to sell and jointly
by acquiring it. Then, the system that manages all the order distribute their aggregated products for a long time and now
management process should be run, in every enterprise, by face the challenge of the complete integration of their value
people able to carry out important scheduling decisions, giv- chains (Burton and Boeder 2003). The starting point of any
ing preference to those product orders that have to do with meta-value chain winning approach is, consequently, collab-
such a product pack. Thus, it is possible that an enterprise oration, where this concept inherently implies both agility
decides to reject some new orders to favour the fabrication and learn to learn capacity (Christopher 2003; Kramer and
of products inside the product pack; or they may even decide Tyler 1995; Meier 1995).
to postpone some other orders already scheduled in order to Within a value chain, the downstream supply chain, here-
re-allocate their manufacturing capacity to manufacture the after called selling chain, plays a key role in building and
new product pack. Anyway, it should be taken into account delivering competitive value to the customers, as it repre-
that these decisions will affect to the others providers in sents the direct nexus between the value chain and the cus-
the case of product-packages, being also relevant decisions tomer. The selling chain focuses on the management of the
regarding the remaining products orders life cycle, from the initial order placement to the deliv-
It is thus possible to conclude that what started as a busi- ery and provision of physical goods, covering the part of
ness-related problem (product and product-package) finds the value chain that goes from manufacturers to end cus-
its solution at the most operational layer; namely, by solving tomers (including logistics platforms/operators, distribution,

123
J Intell Manuf

Fig. 1 Business domains

intermediaries, resellers and retailers). Similarly, the meta- problem, which in the present case is simple to solve, as it is
selling chain (see Fig. 1), defined as the aggregation of selling considered that there is a single resource available for each
chains within a meta-value chain, becomes a key player to product. One of the techniques employed in these cases is
build and deliver competitive extended value propositions to the application of heuristics such as EWDD (Baker 1995;
customers (Putnik and Cunha 2007; Seuring et al. 2003). Cheng and Sin 1990). However, each problem would be
These different policies of scheduling, rejecting or post- NP-Hard and it would not take into account the interdepen-
poning orders together with others are of high interest within dence among the products forming a product-package. The
the production context, as the decisions made will affect the consideration of this restriction could significantly improve
future performance of the organisations and will delimit their the results in our problem.
competitive position. Additionally, there are more factors to Other way to approach the proposed problem is to classify
take into account when making both scheduling and global it as a classic problem of parallel machines, where machines
decisions at the supply chain and the meta-supply chain lev- are suppliers. Assuming that processing times are indepen-
els such as human factors, training factors, structural factors, dent of the machine/supplier of choice, since there is only
organisational factors, etc. The coordinated management of one possibility, one can consider the proposals for identi-
all these factors will result in an integrated approach of the cal, proportional or unrelated machines. In this problem, two
meta-supply chain and will therefore lead to provide the closely related aspects should be solved: the allocation and
desired extended value proposition to customers. sequencing. In this area, several contributions that address
Regarding the scheduling focus, which is precisely to be a problem with the objective to minimize weighted tardi-
addressed here, when addressing total weighted tardiness ness were analyzed (Azizoglu and Kirca 1998; Liaw et al.
with arbitrary processing times situations, there is no hope
 of 2003; Lee and Pinedo 1997; Park 2000; Ko et al. 2004),
finding polynomial algorithms because the problem 1  T j as well as those works that consider the machines eligibil-
is NP-Hard (Brucker 2007). Thus it follows that the under- ity constraint, usually considering makespan and identical
lying problem keeps, at least, the same level of complexity machine constraint (Su 2009; Hu et al. 2010); however, sci-
when not more. entific research addressing due date related problems are not
For classification purposes, it can be said that this problem abundant (Sheen and Liao 2007). It should not be of inter-
is difficult to be classified within in the traditional operations est how that solves the assignment problem, since this is
research. It has not be found any taxonomy which identifies not the problem, and it generally increases the complexity.
the problem clearly and specifically, so multiple approaches However, the problems of parallel machine, like those of
for finding alternatives to solve it had to be carried out. single machine, do not take into account interdependenc-
The first and simplest approach is to consider multiple sin- es among the products belonging to a product-package. For
gle machine scheduling problems, one for each supplier. This this reason, this approach was considered less interesting than
approach is based on a previous solution of the assignment single machine problem.

123
J Intell Manuf

Another interesting apporach is one that addresses produc- An example of collaborative scheduling in extended
tion batches. One of the major branches in the batch schedul- selling chains
ing research area addresses family scheduling models (Baker
1995; Potts and Kovalyov 2000). In family scheduling mod- The Sales Network is composed of several selling points
els, jobs are clustered into families according to their simi- which are used as storefronts of the ECOSELL environment.
larity, so as to reduce setup times. No setup is required for When a customer asks for a product-package this is selected
a job if it belongs to the same family of the job previously and the final delivery date is arranged, and then all involved
processed. Two variants are studied within the family sched- actors receive the information needed to fulfil the order.
uling models depending on when the jobs become available. Figure 2 represents an example, which is extended in Gomez-
In a family model with batch availability Kim et al. (2003) Gasquet et al. (2009), where the product-package is a bath-
provides a two-level batch scheduling heuristic (TH) that room composition (tiles + furniture) and where the actors are
is applied where jobs in the same batch can be distributed named logistic operators, tiles and furniture suppliers.
to several different machines. It overcomes the main diffi- The process starts when a final customer walks into a sell-
culty found when family models are analyzed, that is, usu- ing point and identifies a set of products that constitute a
ally job that belongs to batches is processed in the same product-pack for his bathroom. Such items must be deliv-
machine. A family scheduling model with batch availabil- ered and installed at the customer’s house in a specified due
ity for parallel machine as is very close but their approach date. The products could be available in stock or a replenish-
does not provide a fundamental restriction on the problem ment supply could be needed. If enough stock does not exist,
raised: machine eligibility. In this approach, each job (prod- the corresponding manufacturer should provide an alterna-
uct for us) is free to be allocated to any resource while the tive and suitable delivery date; i.e., a date that allows to the
problem proposed respect and keep an earlier assignment. seller to install the product-pack in the date desired by the
In any case, it is necessary to take into account the fact that client. The process shown involves several selling chains and
it is all the works’ completion the one that influences the is identified as the ECOSELL environment.
objective function, and not when the individual completion Once the order has been sent, the suppliers can have a
date. numerical situation like the one shown in the Table 1. It can
According to the analysis, it seems interesting to provide be seen that there are 2 product-package (PP-1 and PP-2) and
a new focus when designing the method of solving the prob- two products (P-1 and P-2) remaining to be scheduled on the
lem. Then, a Genetic Algorithm (GA) will be implemented. Tile Supplier and the same for the Furniture Supplier with
GAs are usually a suitable tool for scheduling problems and other two products (P-3 and P-4).
an specific GA will be designed to overcome the gaps found To simplify the data structure, all items have arisen as
in the approaches of the single machine problem and the a product-package, and there is only one due date and one
family scheduling model. time-slot penalty for delay per item. The difference between

Items/Package

Logistic Operator

POS

Orders
Tiles
Supplier
ORDER PLANNING ORDER SCHEDULING
POS

ECOSELL
ORDER MONITORING Customer
Environment

Ecosell Platform POS

Items
Sales
Network
Furniture Logistic Operator
Supplier

Fig. 2 Actors involved in the collaborative environment

123
J Intell Manuf

Table 1 Basic data for scheduling example


PP-1 PP-2 P-1 P-2 P-3 P-4

Due date (dd j ) 30 60 5 55 30 60


Weight (w j ) 25 20 5 10 5 10
Process time ( p j,1 ) 10 20 10 30 0 0
tiles supplier
Process time ( p j,2 ) 25 15 0 0 5 20
furniture supplier

Fig. 4 Schedule reached using modified dd j /w j sorting

involved because all of them may have heterogeneous


scheduling systems.
– Interdependencies between production processes which
are performed at different plants must be considered.
– Existing (local) scheduling systems for individual plants
that accomplish the local realisation in term of service of
global requirements should be integrated.
– The process of scheduling could be performed on-line as
Fig. 3 Schedule reached using dd j /w j sorting customers are requesting orders or could be executed off-
line after several requests have been consolidated. In any
case, it should be assumed that there is a time-horizon
a product and a product-package is determined by that in the in which orders for each supplier are uncertain because
first case, all processing times are 0 except for a single sup- they are subject to change. Obviously, it is not possiblt to
plier. The goal is to find that scheduling in which the total assume that this is the whole time-horizon.
weighted tardiness be as low as possible. A schedule could
be the one carried out without applying any kind of rule, i.e. At this point, it should be understood the complexity of the
sequencing in Tile supplier p-1, pp-1, pp-2, pp3 and in Fur- problem as a whole. From this section, concentrating efforts
niture supplier p-3, p-4, pp-1 pp-2, reaching a penalty of 775 in scheduling problem abstracting a little environment where
whereas an alternative scheduling (sort ascending the prod- this occurs.
ucts each supplier based on the result dd j /w j ) is shown in
Fig. 3, in which the penalty reaches a total of 350, signifi-
cantly better than the first approach. Problem definition
However, it seems natural that applying a logical rule will
output better results. The important thing issue is to realize The problem defined in the Introduction (paragraph 1), con-
that an improvement in the objective function can help to sisting in minimizing the sum of the total weighted tardiness
improve the degree of collaboration among suppliers avoid- in selling chains where orders are processed in parallel in a
ing that inappropriate action from a supplier affected others. single stage, and where such requests are made by product-
In this sense, we can see how a small change in the rule used packages, can be formally expressed as follows:
in the second case, consisting of exchanging the item PP-2
and P-3 (see Fig. 4) in the sequence of furniture supplier, Objective Function
reduces the penalty, and thus prevents the product—package
2 which had already been completed by the supplier Tile is 
n

waiting longer than necessary. Min w j Tj


j=1
This scenario presents a set of characteristics that
increases considerably the complexity of its modelling such Index
as:
j, i = 1 to n indexes on product-package set.
– Companies must work cooperatively to properly coordi- s = 1 to S index on supplier set.
nate common scheduling activities from all the suppliers t = 1 to H index on time-slot with horizon in H.

123
J Intell Manuf

Data with an Intel Pentium 4, 3 GHz, 2 GB of memory. It was tested


with a dozen product-packages and the optimal outcome is
ps, j >= 0 processing time of product-package j in not feasible in a realistic environment because several hours
supplier s. were needed to reach a solution.
dj > 0 due date of product-package j.
wj > 0 weight of penalty of product-package j.
M denote a very large and positive number. A genetic algorithm for collaborative scheduling
within extended selling chains
Decision variables
This section proposes a new GA in order to cope with the
presented problem in a reasonable time, as described in the
xs, j,t 1 if operations starts in t-th instant on product-
previous section, and for a realistic size. The key features of
package j in supplier s, 0 otherwise.
the algorithm will be described, the process of adjusting its
c j,t 1 if last operation (in all suppliers) on product-
parameters will be discussed, and an experimental analysis
package j was finished in instant t −1, 0 otherwise.
that allows to compare the results obtained with the GA and
Tj Tardiness of product-package j. T j >= 0.
with some heuristic rules discussed in the section on literature
review will be presented.
Subject to:


H Definition
xs, j,t = 1 for all s, j (1)
t=1 Genetic algorithms (Holland 1975) have been known to offer
tt<t+ ps, j
   significant advantages over conventional methods because
xs,i,tt ≤ M ∗ 1 − xs, j,t they are able to use several search principles and methods
tt=t simultaneously. Nevertheless, GAs can suffer from slow con-
for all s, j, i, t/j = i and t ≤ H − Ps, j (2) vergence and premature convergence before providing an

H accurate solution because of their fundamental limitations;
c j,t = 1 for all j (3) namely, not exploiting local search information. To improve
t=1 the weaknesses of not exploiting the local search information

H 
H in GAs, various heuristic methodologies of GAs combined

ps, j + (t xs, j,t ) ≤ (t ∗ c j,t ) for all s, j (4) with conventional heuristics have been developed.
t=1 t=1 In order to find a good solution, it uses a rare population
S tt<t+
 ps, j structure and rules for the interpretation of solutions specif-
xs, j,tt− ps, j −1 ≤ S − c j,t for all j, t > 1 (5) ically designed for the problem. Two operators have been
s=1 tt≥ ps, j selected according with the genetic representation of solu-

H tions to cover the crossover and mutation. We will explain
Tj ≥ (t ∗ c j,t ) − d j for all j (6) these genetic operators and other basic elements of the pro-
t=1 posed GA in the following sections.

H
Tj ≥ d j − (t ∗ c j,t ) for all j (7) Genetic representation and population structure
t=1

Equation 1 states that there must be one and only one Figure 5 shows the genetic representation of a solution for 10
time-slot of initiation for each product regardless of the sup- product-packages (i.e., chromosome). The chromosome con-
plier from which to start the operation. Equation 2 prevents sists of a product-package sequence. In this case, the prod-
that two or more products are processed in the same supplier uct-package sequence is 2, 3, 9, 7, 8, 1, 10, 6, 4, 5. To obtain
simultaneously. Equation 3 states that there must be one and a schedule from this representation, it is necessary to apply
only one moment of completion of operations for each prod- a rule to convert it into a sequence by each supplier and also
uct-package. The usage of Eqs. 4 and 5 assures that the date consider the timing. In this case the allocation of products,
of completion of an operation should be exactly ps, j time- which are part of product-package, to suppliers are unique
slots after its starts (not pre-emptive). And the Eqs. 6 and 7 and therefore not a problem but a fact. This rule is an impor-
are used to calculate T j as the maximum value between 0 tant element and different alternatives have been proposed.
and C j − d j . Figure 6 shows the proposed structure of the population.
After modeling the problem, it starts the resolution of sim- This was designed by Zhong et al. (2004) and it clearly con-
ple cases with only 2 suppliers MPL using CPLEX v 4.11 tributes to local fight and a more intensive and expansive dis-

123
J Intell Manuf

Fig. 5 Chromosome corresponding with the genetic representation


proposed

Fig. 7 Application of rule 2

est to the right of the due date where space is avail-


able. It is important to point out that none product
Fig. 6 Structure of population with a specific (worse) position in the sequence can
cause that a product with a better position, and pre-
tribution of the genetic knowledge accumulated. As it shown, viously located, creates or increases the tardiness
it has got a square matrix structure where each cell, which of the last one.
is occupied by an individual, is characterized by its chromo- A rf (rule factor) is used to select either rule 1 or rule 2.
some and the value of the evaluation function, which in this
case is the total weighted tardiness. In this matrix only indi- Generation of population
viduals relate to their 4 neighbours (left, right, up and down).
In our experiment the matrix has been established with a size The initial population is generated creating a percentage
of 10 × 10. equal to gf (generation factor) of individuals that are cre-
ated using a heuristics; other individual chromosomes are
Rules: from sequence to schedule created randomly.
The heuristic rules that can be used are known as earliest
To convert a sequence of product-packages in a Schedule for weighted due date (EWDD) modified from EDD (Jackson
each supplier the following 2 rules were applied: 1995) and EDD properly. EWDD consists on ascending the
order of the product-package based on the result of the oper-
Rule 1. Each supplier process each product as soon as pos- ation d j ∗ U j . U j is a value between 1 and Vmin > 0.5, so
sible following a common sequence for all sup- that Vmin is assigned to the product-package with the highest
pliers, the product-package sequence. The product w j and 1 is assigned to the product-package with the lowest
order is those has been assigned to the product- w j . The remaining being distributed between the two val-
package that they belong to. ues in proportion to the value of w j . EWDD heuristic can
Rule 2. Each supplier’s products are processed as close as apply Rule 1 or 2 for the interpretation of the sequence. A hf
possible to the due date (see Fig. 7) and the prod- (heuristic factor) is used to select either EWDD or EDD.
ucts are placed by following the sequence of the
product-package and using the position assigned Selection, crossover and mutation
to the product-package that they belong to. When
a product attempts to settle at a position which is The algorithm does not maintain the classic concept of iter-
wholly or partially occupied by another product, it ation in the proposed GA an iteration it is to go over every
then first try to move the product that was previ- single individual in the array. For each of the individuals a
ously located in this position to the left (compact- genetic operation has been performed as follows:
ing); if this movement does not create the necessary
space to be allocated without generating tardiness, Step 1. The individual X is randomly selected, which has a
the new product will be placed in the position clos- percentage cf (crossover factor) of being selected

123
J Intell Manuf

for crossover operation. If selected go to step 2 Other approaches


(crossover) else go to step 5 (mutation).
Step 2. We analyze the evaluation value of the 4 neighbours In order to compare the results obtained with the proposed
of the individual X, selecting the individual that has GA, some heuristics, which have proved efficient in sim-
lower value. Go to step 3. ilar cases, have been implemented. In particular, we have
Step 3. It makes the crossing between the individual X and adapted to this problem two versions of the heuristic Process-
Y using the method Similar Block 2-point Order ing Weighted Shortest Time (SWPT) to which we have called
Crossover (SB2OX) (Ruiz et al. 2006), which are SWPT-PP (SWPT-Product-Package) and SWPT-P (SWPT-
obtained by 2 subjects children to replace the indi- Product) due to SWPT has been applied in two different
vidual candidate X. Go to step 4 ways. The already mentioned EWDD heuristics, and a rule
Step 4. The individual X with an assessment value Vx is randomly (RANDOM) used as reference and iterated as GA.
replaced by the child with a lower assessment value EWDD is the most commonly used in single machine when
Vx y if and only if Vx y < af * Vx . af (approach fac- the objective is to minimize the total weighted tardiness.
tor) parameter is set at 1.05, and it allows the X EWDD and RANDOM can employ either Rule 1 or 2 for
individual to renew the chromosome if the child is the interpretation of the sequence.
better or even as much as 5% worse. Select a new Although the SWPT rule is the best and the most com-
X individual and go to step 1. monly used dispatching rule when the objective is to mini-
Step 5. The individual X is randomly selected, which has mize the total weighted flow-time, some carried out test have
a percentage mf (mutation factor) of being selected achieved better results than with EWDD. This rule applies
for mutation operation. If selected go to step 6 ascending sorting of items based on the result of the operation
(mutation) else select a new individual and go to p j /w j . As in the case shown there is no single process time
step 1. for each product-package two strategies have been used. The
Step 6. It runs along the X chromosome and each allele heuristic SWPT-PP is to generate a single sequence of prod-
of the individual is swapped with the next with a uct-package, on the processing time of each product-pack-
probability mf. Once all the individual alleles have age that corresponds with the highest among their products.
been considered, the mutation process is completed The heuristic SWPT-P is to generate a different sequence for
obtaining a mutated X’ chromosome. Go to step each supplier, considering directly in this case the processing
7. time of the product that corresponds to each product-pack-
Step 7. The individual X with an assessment value Vx is age. When using SWPT-PP heuristic then rule 1 or 2 provided
replaced by the mutated one with an assessment for the interpretation of the sequence could be used. And if it
value Vx  if and only if Vx  < af * Vx . Using af employs heuristics SWPT-P then rule 1 is used for the inter-
parameter as in the crossover process. Select a new pretation of the sequence. It makes no sense to use rule 2
X individual and go to step 1. since it is the traditional application of the rule SWPT.
It was also considered to use the TH rule, mentioned in
the literature review, but after studying their adaptation to this
Regeneration and finish condition problem it was observed that this was reduced to a EWDD
and it was therefore discarded.
The algorithm continues running until a solution with a value
of 0, which is the best one, is found, or until it reaches an Experimental study
established iteration, in this case 5.000. However, in order
to avoid to conider a local optimum, a regeneration sys- This study is based on the production data obtained from a
tem of the population has been established. This consists in semi-real case used in the ECOSELL project environment
verifying whether the value that corresponds to the best (European Commission GRD1-2001-40692, Extended Col-
solution has remained unimproved uf times (unimproved laborative Selling Chain). The processing time for each prod-
factor) and, if so, the population is modified as follows: uct depends on the supplier assigned, and it is generated from
If an individual has a value Vx that is larger that the U[10,20] (slot time unit). We found that the variations of the
value of the best solution multipled by the approach factor, processing times do not affect our experimental results.
af * Vb , is replaced by a new one, but is replaced with Due dates of jobs are integer values generated from
a probability of 70%. Obviously the best solution is never U[MK(1−τ −ρ/2), MK(1−τ +ρ/2)], as suggested by (Potts
replaced. and Van Wassenhove 1982) and used in (Kim et al. 2003).
The new individuals are obtained by applying the same MK, τ , and ρ control makespan, priority factor, and due date
method, as it was done in the initial generation of the popu- range factor, respectively. Because MK cannot be calculated
lation. accurately, it is estimated to be the maximum value of sum

123
J Intell Manuf

Table 2 Relative average total weighted tardiness


Suppliers Product- Due-date loose (L) Due-date tight (T)
package
EWDD GA SWPT-P SWPT-PP EWDD GA SWPT-P SWPT-PP

2 10 2.02 0.92 3.61 3.04 2.09 0.89 4.47 3.58


20 1.70 0.86 2.43 2.00 1.92 0.81 3.17 2.09
30 1.34 0.55 3.54 2.81 1.50 0.48 4.17 3.03
4 10 1.6 0.87 3.54 2.65 1.82 0.80 5.17 3.83
20 1.51 0.61 4.42 3.30 1.72 0.61 5.54 3.69
30 1.82 0.52 3.84 2.71 1.62 0.47 5.21 3.27
8 10 3.07 0.77 8.58 5.64 2.19 0.85 5.04 3.14
20 2.06 0.65 4.56 2.70 1.50 0.66 5.13 2.58
30 1.80 0.47 6.43 3.75 1.77 0.50 6.16 3.51

processing time of each supplier for a product-package. ρ is the solutions. At the end, in the final analysis we only used
fixed to be 0.6, and τ is 0.6 (tight: T) and 0.5 (loose: L). data for the average of 10 replicates, and also in the case of
Underlining the urgency, other factor to analyze is the EWDD, RANDOM and SWPT-PP only select data from the
number of items that each supplier receives. In this experi- experiment that uses Rule 1 as it proved to be the best.
ment it should be interpreted as if it indicates a certain load, Table 2 summarizes the relative performance of all the five
for example, 10 product-packages, that means that the 50% methods. The relative performance is computed by dividing
of load that each supplier must process, 5 items in the exam- the average TWT of each method by the average TWT of
ple, are part of equal number of product-package and the rest, RANDOM. Thus, the ratio indicates the deviation from the
and other 5 items are standalone products that do not belong solution produced by RANDOM
to any product-package. The results show that the GA clearly provides the best
Four heuristics have been examined, namely the EWDD, solution in all the cases, whereas SWPT-P provides the lower
SWPT-PP. SWPT-P, and RANDOM, besides the designed bound. It was observed that none of the heuristics used in the
GA. As mentioned later on, some data pre-treatment pro- problems that traditionally have been considered closer to
cesses were performed with non-deterministic methods in the proposed in the paper, can be considered valid because
order to unify the amount and type of data, and likewise they are always worse than the RANDOM rule. This shows,
the calibration was also performed. Five combinations of the firstly, that the proposed GA is an acceptably good solution to
number of suppliers (SP = 2, 4, 8), the number of Product- the problem analyzed, and, secondly, that there are not ade-
Packages (PP = 10, 20, 30) and due dates (τ = T(0.6), L(0.5)) quate tools to deal with this problem as necessary yet. The
are considered, which implies 5 × 3 × 3 × 2 = 90 sce- statistical significance of performance between methods is
narios. Thus, considering five algorithms, we ran a total of analyzed by using ANOVA over a TWT transformation by
5 × 90 = 450 treatments. Algorithms were written in .NET, means of a logarithm (LOG).
and an Intel Pentium 4, 3 GHz, 2 GB of memory was used to Regarding the CPU times consumed by the different meth-
perform these computations. ods, there are no relevant findings. It has been observed that
In order to simplify the final analysis refers to data from the GA is generally the slowest of all, as assumed on the basis
the 540 mentioned treatments, some actions were carried out. of complexity. The heuristic rules are more than 100% faster
First, before executing the experiment described in the pre- on average.
ceding paragraphs, a study of the behavior of GA in a similar The ANOVA table decomposes the variability of LOG
environment that the one described with several levels in fac- (TWT) into contributions due to various factors. Since 8
tors as cf (50%, 20%), mf (10%, 5%), gf (30.10), uf (200, P-values are less than 0,05, these factors have a statistically
50), rf (1.2) and hf (EWDD, EDD) was performed. The GA significant effect on LOG (TWT) at the 95,0% confidence
calibration process concluded with the establishment of the level. However, the most important effect for this experi-
following values cf = 20%, mf = 5%, gf = 30, uf = 50, rf = 2 ment is related to the methods used. In this sense, the main
and hf = EWDD. Second, during the execution of the exper- effect of the Method factor confirms the results discussed
iment, there were 10 replicates for methods RANDOM and on Table 2. Finally, as shown in Fig. 8, in relation to the
GA, as they are nondeterministic procedures. In addition, the interactions between factor Method and other factors, noted
heuristics EWDD, RANDOM and SWPT-PP were analyzed in the ANOVA as the DD factor, related to the urgency of
for the case of using Rule 1 or 2 for the interpretation of the product-package, it is not significant and therefore there

123
J Intell Manuf

Table 3 Analysis of variance


for LOG(TWT)—type III sums Source Sum of squares Df Mean square F-ratio P-value
of squares
MAIN EFFECTS
A:DD 0.0695979 1 0.0695979 0.46 0.4963
B:SP 69.9132 2 34.9566 232.90 0.0000
C:PP 294.386 2 147.193 980.67 0.0000
D:Method 237.233 4 59.3081 395.14 0.0000
INTERACTIONS
AB 3.06485 2 1.53243 10.21 0.0000
AC 2.28554 2 1.14277 7.61 0.0006
AD 0.268314 4 0.0670785 0.45 0.7747
BC 5.46099 4 1.36525 9.10 0.0000
BD 3.3767 8 0.422088 2.81 0.0048
CD 3.68663 8 0.460828 3.07 0.0023
All F-ratios are based on the RESIDUAL 61.8392 412 0.150095
residual mean square error TOTAL (CORRECTED) 681.584 449

Interaction Plot Interaction Plot


12,7 Method 14,1
Method
EWDD
12,2 EWDD
GA 13,1 GA
RANDOM
RANDOM
LOG(TWT)

SWPT-P
LOG(TWT)

11,7
SWPT-PP 12,1 SWPT-P
SWPT-PP
11,2
11,1
10,7

10,2 10,1

9,7 9,1
2 4 8 10 20 30
SP PP

Fig. 8 Interaction between factor Method and SP factor Fig. 9 Interaction between factor Method and PP factor

is not influence on the selection the method used. However,


although both interactions Method-Method-SP and Method- transaction per supplier. The dependence between the opera-
PP are significant, it is only interesting the behavior of the tions of elements (product) that belong to a single deliverable
GA in the latter interaction. It also notes that the interaction (product-package) is the essence and the main difficulty of
PP-Method GA stabilizes the TWT, and it does not grow the problem, and it is rarely mentioned/treated within the
proportionally as the number of product-package increases academic literature.
to 30. This change in the trend of the line represents that EWDD, SWPT-P and SWPT-PP were outperformed by
the GA suggests that, in situations of any complexity, it can the GA, providing the lower bound for the given problem.
increase further its efficiency with the rest. This might be due to the fact that they focus on the prod-
uct-package sequencing only and not on the product level
sequencing. The GA produced a fairly good solution in a
Conclusions few seconds. It worked especially well when the problem
size was more than 20 product-packages per supplier. Thus,
This paper presents the results of a study seeking to minimize the GA can be used as a handy tool for the practical appli-
TWT of a product-package scheduling within Extended Sell- cations of the collaborative scheduling in Extended Selling
ing Chains. The problem of operations with sets of products Chains such as automotive or bathroom or kitchen composi-
that form a single family sales or product-package compli- tions retailers.
cates the scheduling problem even in a situation where it is Further research would include the comparison of the pro-
not considered dependent set-up or processing time, only one posed heuristics with some other heuristics discussed in the

123
J Intell Manuf

existing literature. Another interesting effort can be made to Kim, D. W., Na., D. G., & Chen, F. (2003). Unrelated parallel machine
extend the GA. scheduling wit setup times and total weighted tardiness objec-
tive. Robotics and Computer Integrated Manufacturing, 19, 173–
181.
Ko, H. H., Baek, J. K., Kang, Y. H., & Kim, S. S. (2004). A scheduling
References scheme for restricted parallel machines with cycling process. Jour-
nal of Korean Institute Industrial Engineering, 30, 107–119.
Azizoglu, M., & Kirca, O. (1998). Tardiness minimization on Kramer, R. M., & Tyler, T. R. (1995). Trust in organizations: Frontiers
parallel machines. International Journal of Production Eco- of theory and research. Berkeley, CA: Sage Publications.
nomic, 55, 163–168. Lee, Y. H., & Pinedo, M. (1997). Scheduling jobs on parallel machines
Baker, K. (1995). Elements of sequencing and scheduling. In N. with sequence-dependent setup times. European Journal Opera-
H. Hanover (Ed.), Amos tuck school of business administration tional Research, 100, 464–474.
dartmouth college. Liaw, C. F., Lin, Y. K., Cheng, C. Y., & Chen, M. (2003). Sched-
Brucker, P. (2007). Scheduling algorithms (5th ed., pp. 33–94). Berlin: uling unrelated parallel machines to minimize total weighted
Springer-Verlag tardiness. Computer Operations Research, 30, 1777–1789.
Burton, T. T., & Boeder, S. M. (2003). The lean extended enterprise: Macbeth, D. K., Boddy, D., & Wagner, B. (1998). Partnering strategy
Moving beyond the four walls to value stream excellence. Florida, implementation in the supply chain. In Bittici, U. S., & Carrie,
USA: J Ross Publishing Inc. A. S. (Eds.), IFIP Conference Proceedings 129, 291–304, Pro-
Byrne, J. A. (1993). The Virtual Corporation. BusinessWeek, 8th ceeding of the International Conference of the Manufacturing
February 1993, New York, USA. Value-Chain on Strategic Management of the Manufacturing
Camarinha-Matos, H., & Afsarmanesh, L. (2005). Collaborative net- Value Chain. Kluwer.
works: A new scientific discipline. Journal of Intelligent Manu- Meier, J. (1995). The importance of relationship management in
facturing, 16(4–5), 439–452. establishing successful interorganizational systems. Journal of
Camarinha-Matos, L., Afsarmanesh, H., & Ollus, M. (Eds.). (2005). Strategic Information Systems, 4(2), 135–148.
Virtual organizations: Systems and practices. UK: Springer. Park, M. W. (2000). A genetic algorithm for the parallel-machine total
Camarinha-Matos, L., Afsarmanesh, H., & Ortiz, A. (Eds.). (2005). Col- weighted tardiness problem. Journal of Korean Institute Industrial
laborative networks and their breeding environments IFIP Engineering, 26, 183–192.
186. Berlin: Springer. Potts, C. N., & Kovalyov, M. Y. (2000). Scheduling with batching: A
Cheng, T. C. E., & Sin, C. C. S. (1990). A state-of-the-art review review. European Journal of Operational Research, 120, 228–249.
of parallel-machine scheduling research. European Journal of Potts, C. N., & Van Wassenhove, L. (1982). Decomposition Algo-
Operational Research, 77, 271–292. rithm for the single machine total tardiness problem. Operational
Christopher, M. (2003). Logistics and supply chain management: Strat- Research Letters, 1, 177–181.
egies for reducing cost and improving service (2nd ed.). New York, Putnik, G. D., & Cunha, M. M. (2007). Knowledge and technology
USA: Financial Times-Prentice Hall. management in virtual organizations: Issues, Trends, Opportuni-
Davidow, W., & Malone, M. (1992). The virtual corporation: Structur- ties and Solutions. IGI Publishing.
ing and revitalizing the corporation for 21th Century. New York, Ruiz, R., Maroto, C., & Alcaraz, J. (2006). Two new robust genetic algo-
USA: Harper Collins. rithms for the flowshop scheduling problem. Omega, 34(5), 461–
Dyer, J. H. (2000). Collaborative advantage: Winning through extended 476.
enterprise supplier networks. New York: Oxford University Press. Seuring, S., Müller, M., Goldbach, M., & Schneidewind,
Gomez-Gasquet, P., Franco, R. D., Rodriguez, R. & Ortiz, A. (2009). U. (2003). Strategy and organization in supply chains. New York,
A scheduler for extended supply chains based on a combinatorial USA: Physica-Verlag.
auction, Journal of Operations and Logistics, 2(1), V1–V12. Sheen, G. J., & Liao, L. W (2007). Scheduling machine-depen-
Holland, J. H. (1975). Adaptation in natural and artificial systems. dent jobs to minimize lateness on machines with identical
Report of the Systems Analysis Research Group SYS–1/92, speed under availability constraints. Computer & Operations
University of Dortmund, Department of Computer Science. Uni- Research, 34(8), 2266–2278.
versity Michigan Press, Ann Arbor. Su, L. H. (2009). Scheduling on identical parallel machines to mini-
Hu, X. F., Bao, J. S., & Jin, Y. (2010). Minimising make- mize total completion time with deadline and machine eligibil-
span on parallel machines with precedence constraints and ity constraints. International Journal of Advanced Manufacturing
machine eligibility restrictions. International Journal of Produc- Technology, 40(5–6), 572–581.
tion Research, 48(6), 1639–1651. Zhong, W. C., Liu, J., Xue, M. Z., & Jiao, L. C. (2004). A multi-
Jackson, J. R. (1995). Scheduling a production line to minimize agent genetic algorithm for global numerical optimization. Ieee
maximum tardiness. Research Report 43, Management Science Transactions on Systems Man and Cybernetics Part B-Cybernet-
Research Project, University of California, Los Angeles. ics, 34(2), 1128–1141.

123

Vous aimerez peut-être aussi