Similar to Ying hua, c. (2010): adopting co-evolution and constraint-satisfaction concept on genetic algorithms to solve supply chain network design problems
Similar to Ying hua, c. (2010): adopting co-evolution and constraint-satisfaction concept on genetic algorithms to solve supply chain network design problems (20)
Call Us ✡️97111⇛47426⇛Call In girls Vasant Vihar༒(Delhi)
Ying hua, c. (2010): adopting co-evolution and constraint-satisfaction concept on genetic algorithms to solve supply chain network design problems
1. Expt'n Systems WJth Applic,UIOns 37 {2010) 6919-6930
Contents lists available at ScienceOirect
Expert Systems with Applications
jou rna I hom ep age: www. elsevier .co m/locate /e swa
Adopting co-evolution and constraint-satisfaction concept on genetic algorithms
to solve supply chain network design problems
Chang Ying-Hua •
Deparrmenr of /nformarian Management. TomKang Vnrversrry, 151 Ymg-fhuan Rood. Tamsui. Taipei County 251 J7, Taiwan, ROC
ART ICLE
INFO
Keywords:
Supply ch<ain network design
Genetic Algorithms
Co-evolution concept
Consrraint~sat i sf<action
concept
Op11miution
ABSTRACT
With the r.apid glob.lliz.ation of m.arkets, integrating supply chain technology h.as become increasingly
complex. That is. most supply chains are no longer limited to a particular region. Because the numbers
of branch nodes of supply chains have increased, products and raw materials vary and resource constrainrs differ. Thus, integrating planning mechanisms should include the capacity to respond to change.
In the past. mathematical programming <md a general heuristics algorithm were used to so lve globalized
supply chain network design problems. When mathematical programming is used to solve a problem and
the number of decision variables is too high or constraint conditions are too complex, computation time
is long, resulting in low efficiency, and can easily become trapped in partia l optimum solution. When a
general heuristics algorithm is used and the number of variables and constraints is too high, the degree
of complexity incre.1ses. This usu.ally results in .an in.ability of people to think about resource constraints
of enterprises .and obtain an optimum so Iutton.
Therefore. this study uses genetic algorithms with optimum search features. This work combines the
co-evolutionary mode. which is in accordance with various criteria and evolves dynamically. and consrrainH<Hisfaction mode capacity to narrow the search space, which helps in finding rapidly a solution
that, solves supply chain integration network design problems. Additionally, via mathematical programming, .a si mple genetic .J:Igorithm. co-evolutionary genetic .algorithm. constr.aint-satisfaction generic algorithm and co-evolut ionary constraint generic algorithm are used to compare the experiments result and
processing time to con finn the perform.mce of the proposed method.
e 2010 Elsevier ltd. All rights resetved.
1. Introduction
With the rapid development of a global economy, info rmation
and manufacturing technology. the needs of consumers. production and sales have changed. Enterprises must now think globally.
Independent operation modes are no longer suitable for markets
undergoing upheaval s. The integration of ind ustries and the division of work has become the best so lution for enterprise suJVival.
To create an efficiem globa l supply chai n. resources. the supply
cha in and all fa ctories must be tightly integrated. Additionally,
firms must respond to customer requirements efficiently, offer
high-quality products, reduce operational cos ts and increase customer sa ti sfaction.
Beca use most supply chain nerwork designs have multiple layers. members. periods and products. and a comparative resource
constraint exists between different layers. For example, a factory
may have constra ints on its productivity. and a distribution center
may have limited storage ca pacity. These problems typically in• Tel.: +886 911 23IS3G: f<ilc:: +886 2 26209737
E-mail address: yhch;mgGJ'mdii.fku.('du.rw
crease as the numbe r of supply chain layers increase, the time period increases. and the number of prod ucts and purchase orders
increase. These cause the network search space and time required
to obtain a sol ution to increase markedly. Thus, the supply cha in
network design problem is an NP-complere problem (lbaraki &
Katoh. 1988). Many studi es adopted mathema tica l programming
or a heuristics algorithm to solve such problems. For example.
Robinson .1nd S<1tterficld ( 1998) establ ished a mixed- imeger program mi ng solution for a two-echelon, multip le products storage
problem. Hou and Ch~ng (2 003 ) use the evo lution principle to
so lve a production and distribution problem. When faced with a
simple supply chain system, mathematical programming ca n
easily find the optimum solution; however. mathematical program mi ng is not suitable for large mixed -in teger programming
problems. By adopt in g a single poi nt r.:tndom sedfch, the number
of decision variables increases, which extends the rime required
to find the solution and the method can easily become trapped
in a local optimum solution (Wu. 2002). When adopting a general
heuristics algorithm, such as simulated .:tnneali ng, the search for
the optimum so lution cannot be reached completely and efficiently
(Ait ipannak. Gen. & Lin. 2005; Wang, 2002).
0957-4174{S- see front m<atter ~ 2010 Elsevier Lid. All rights reserved.
doi: 10.10 6 /J.CSW<a.201 O.oJ.o30
2. 6920
Y.-H. Chang/Ex~rr Systems ""'irh App/icalions 37 (1010) 6919- 6930
This study uses the features of ge netic algorithms (parallel
searc hing and can evolve in an environment ) combined with the
co-evolutionary mode (cons idering multiple criteria. search speed
.:~nd avoiding converge before the right time) and the constraintsatisfaction mode (narrow the sea rch space accelerate the speed
at which the optimum solution is acquired ) ro establ ish sup ply
chain system for all manufacturing branches. The minimum cost
of a supply chain network is adopted as the most suitable plan
for developing an algorithm for a supp ly chain network distribution. The proposed method is a rapid way of find ing a solution
and obtaining optimum searc h capacity. To verify the accuracy
and efficacy of the proposed algo rithm. LINGO software was used
in mathematica l programming mode. The experimental result
was compared with the algorithmic result to determine the difference in cost and calculation time. This comparison c.:m verify the
efficiency of the proposed algo rithm such that a nexib le and efficient supply chain integration network mode can be obtained. This
result can function as a reference for man ufacturers in daily operations and for management decisions.
2. Literature review
This study focuses on supply chain integration and its network
design problem. Relrtted literature is discussed in the following sequence: the sup pl y chain integration model: genetic algorithms;
co-evo lutionary mode; and. co nstraint-satisfaction mode.
2. 1. Supply chain integration mode
Supply chain nerworks obtain ra w materials, supplies, manufactures and distributes product to customers. All these planning,
organizing and controlling from the supply point to the demand
point include cash flows, information flows and materials circulation (M aloni & Benton. 1997}. By combining strategies to ach ieve
integration of ideality and rea li ry of different enterprises to form
a reticulation supply chain network. Bec.:wse different enterprises
in a supply chain may have different or conflicting targets, and supply chains must adju st to changes. integration of a supply chain is
not easily achieved (Hayashi. Koguro, & Murakami, 2005; liu &
Hsieh. 2005; Su. 2000; Sue, 1999; Vaidyanathan & Hasan. 2001;
Viswanathan & Piplani, 2001 }. Thus. supply chain members must
be coordinated and cooperate to achieve a common target. Additionally. a dynamic integration mode must be established that
can respond effectively to changes in global markets.
As supply chain management is a complex decision problem,
many rese.uchers adopted mathematica l programming or a heuristics algorithm to solve this problem. For instance, Robinson and
Sa tterfield ( 1998) established a mixed-integer programming model that maximizes profit and is designed to find the solution for a
single period and single product in a supply chain network. Koray
and Marc (1999) proposed a multiple products mixed-integer programming mode for minimum cost that included manufacturing
cost. storage cost and transport cost. They used Bender's decomposition of integer programming to find the solution. Melachrinoudis
et al. (2000) used mult iple objective approaches to find a solution.
Ross {2000) employed a pertormance-based strategic resource
allocation to solve su pply nerwork design problems. Syam (2002 )
applied Lagrangian relaxation and simulated annealing to a supply
chain network with multiple levels and settings; minimize cost
was the target. Syarif. Yun, and Gen (2002} used genetic algorithms
and a spanni ng tree to find the solution for su pply chain network
with three levels and single product with the aim of minimizing
cost. Altipannak. Gen. Lin. and Karaoglan {2007) employed steady-state genetic algorithms to obtain the solution for the design
of a supply chain nerwork with multiple products and phases,
and compare their algorithm with linear programming, Lagrangian
relaxation and simulated annealing. This study aims to minimize
purchasing cost, transportation cost. fixed cost and manufacturing
cost. Based on the above literature. some studies rake maximize
profit as thei r objective, whe reas most supply chain network models take minimize cost as their measure of performance and consider many costs. including purchasing cost, trans portation cos t
and storage cost Altiparmak et al. (2007 ) proposed a supply chain
network model that covered almost all of these costs. Thus. this
study adopts Alti parmak's model as the nonn in the mathematical
mode.
2.2. Genetic algorithms
Genet ic algorith ms first developed by Holland ( 1975 ). use computer programs to simulate the evo lutionary process with the
ch romosome as the solution to the solved problem. Based on the
environmental adaptation of chromosomes, researchers identified
a fitness val ue such that a researcher could determine whether a
chromosome wo uld survive to the next generation. The evolutionary process continues umil the ta rget has been mer. via self-adaptation and an iteration threshold. the algorithm has the ability to
evolve to the optimum solution for a problem.
The parallel search feature is hidden in generic algorithms. This
feature is suita ble for handling mu ltip le space dimensions, and
non-linear complex NP-IMrd problems (Goldberg, 1989). Genetic
algorithms is faster than the simp le point search function of traditional algorithms and has a wide application in such areas as stock
market analysis (Szeto & Fang. 2000), vehicle routing problems
(Ting & Huang. 2005), production allocation problems (Hou and
Chang, 2002), tourism itinerary planning {Christophe & Hugues,
2007), communications network design {Chou, Premkumar. &
Chu. 200 1), weather prediction (Wong, Yip. & Li, 2008). locationall ocatio n problems Uaramillo. Bhadury. & Batta. 2002: Zhou &
Liu. 2003). personnel training (Juang, Li n. & Kao, 2007), and bus network optimization (Bielli. Caramia. & Carotenuto. 2002). However,
genetic algorithms have two shortcomings. If the fitness value is
set improperly, convergence can occur prematurely. The second is
that it is not suitable for highly non -linear problems (Hi lli s. 1992).
2.3. Co-evolutionary mode
To overcome these shortcomings. Hillis developed the co-evolutionary scheme in 1992. In this scheme a living being (chromosome ) and environment {m ultiple criteria ) interact and c<H."volve.
The genes are improved continuously for survival and the environment changes with the living being. For instance, when an eagle
hunts a rabbit, the rabbit must run fast to suiVive. and the eagle
must also ny fast to catch the rabbit. This means that both the
chromosome and multiple criteria constraint conditions must
evolve. Restated, as the chromosome must ma tch the fitness value,
criteria constraint conditions must be evaluated by a fitness function. The evalua ti on cri teria for next evaluation are selected and
based on th e degree of fitness of criteria. In th is way, the shortcomings of traditional genetic algorithms can be overcome and search
speed accelerated.
In the operation mode. all chromosomes must be evaluated to
determine their degree of march with the mu ltiple criteria constraint cond itions. The ease with which these constra int conditions
are met is directly related to the size of the fitness values of the criteria. Conversely. the fitness value can be large such that the criteria have the opportunity to be the evaluation function for the next
evaluation. Fig. 1 presents this co-evolutionary mode. After this
process, the evolutionary di recrion of a chromosome will be that
for which conditions are not easily met. Thus. rhe near optimum
3. 692 1
Y. -H Chang/ Expert 5ysrtms with Applications 37 (20 10) 6919-6930
3. Supply chain integratio n mode
Chromosome I
Chromosome .!
AJapti v.:
Chromo..;ome n
Evaluate
Criterion I
3.1. Supply chain archirecrure
Cntcrion 2
Cnte 1i o nm
~
Fig. 1. Co-evolution.uy modt-.
solution can be acquired rapidly and convergent immaturity can be
avoided.
2.4. Conscraint-sacis[action mode
Ge net ic algorithms are re liable. but do not have capac ity to process complex constraint conditions. Thus. generic algorit hms should
be integrated with another mechanism to solve the satisfactio n
problem of constraint conditions. Gen and Cheng (2000) proposed
four strategies to improve a genetic algorithm 's efficiency-reject
strategy, repair strategy. mod ify the genetic operator strategy and
penalty strategy.
The reject strategy is used when a result falls into a non-feasible
solution region: the chromosome can then be discarded directly.
The repair strategy is used when a gene disobeys constra int conditions such that it should be modified to feasible solution region.
The modify strategy means according to the different problem:
the chromosomes and fundam ental operat ions sho uld be modified
to avoid violating co nstraint condit ions during the calcu lation process. The penalty strategy adds a pena lty item to the origina l target
function. The degree of the penalty depends on degree to chromosome violates constraints. This concept, which is used to find the
so lution. is fast and efficie nt in findi ng the near optimum solu tion.
Simple genetic algorithms adopt multipoint parallel search
modes and find th e optimum solution gradually. Once the search
space becomes too large. the time spent finding the solution in creases, or the algorithm falls inro the local optimum solution left
graph in Fig. 2. However, the constraint-satisfact ion genetic algorithm n<~.rrow s down the soluti on spetce to feasible so lution region.
The chromosome solution that meets the constraint condition criteria is then created. Because it obta ins the near optimum solution
from feasible solutio n directly, its sea rch region is very sma ll (right
graph in Fi g. 2). Thus. the constrain t-satisfaction gene tic algorithm
is faster at finding the optimum solution than the sim ple genetic
algorithm (Hou and Chang. 2002 ).
This research adopts the model developed by Alti parmak e r a!.
(2007 ) co establish multiple products supply chai n network model
with multiple levels. Fig. 3 shows this supply chain system. The
supply chain model has four leve ls-raw material suppliers. manufacturers. distribution centers and retailers. The costs are purchasing cost. transportation cost. manufacturing cost and storage cosr.
The manufacturer and distribution center are fixed and non-dynamic invocation points. The same retailer can be controlled by
mu ltiple distribution centers. Thus, if productivity is defic ient,
other di st ribution centers can offer support. Fig. 4 shows the research mode. The mode takes the minimum coral cos t as the solu tion goal and assumes that all levels are under resource
constraints. and the production demand of a retailer can be satisfied. The fundamental assumptions in this study are as follows.
( 1) The demand and productivity of retailers and raw material
suppliers are known.
(2) The supply balance must meet demand such that no our-ofstock situation wi ll occur.
(3) Quantity an d productiv ity constraints of the raw material
supplier. manufacturer and distribution center are known.
Definition of the mathematic model in this study is as follows .
~~,~~, ·:·
'
....
~
~)i
_:il.ll"
~.
.
m.:. nu facturers
'~
retailers
~upp l i.::rs
-~
rn.:.tcrial
distributi on center:.
- pmducts
Fig. l . Supply chonn network system for an mdusrry.
soluuon space
' "'ib"~'"'"""''"' ~
Fig. 2. Se.Hch mode of sLmple genetic
~lgorithm
And
comtr~int-sAtlSfAction
generic
~lgorirhm
4. 6922
Y.-H. Chang/Expert
Sysrtms wich Applicacions 37 (2010) 6919-6930
( 1) Model Symbols
quantity of raw material sup pliers (s E S)
production capacity of manufacturers (J E F)
D
number of distribution centers (d E D)
M
number of retailers (r E R)
qu.mtity of raw material (m E M)
number of products (p t P)
(2) Model Parameters
suppliers. unit purchase cost of raw material m
unit rransporrarion cost of manufacturer f that
transport products to distribution cen ter d
unit rrans ponation cost of distribution cemer d that
transports products ro retailer r
unit manufacturing cost of manufacturer f for
produ ct p
w,
BOMpm
D"'
Ld
unit distribution cost of distribution center d
quantity of raw material m that product p needs
demand of retailer r for product p
upper quan tity bound of raw material m that can be
provided by raw material suppliers
productivity upper lim it of manufaclUrer Jfor
product p
distribution upper limit for distribution center d
(3) Decision Variables
~pd
quantity of product p that is transported from
distribution center d to retailer r
~PI
quantity of product p that is transported from
manufacturer f to distribution center d
£2tms quantity of raw material m that is purchased by
manufacturer f from suppliers
(4 ) Objective Function
Min Z =
L L L B,m QJ~- L L LMw
l2.pJ
f
- L L LT/,Qdpf+ L L L W,Q,.,
f
-LLLT,,Q,.,
I
m
d
J
d
'
P
r
p
p
p
d
d
(5) Constraints
LQ,., =D"'
(1)
LC2.,r = LQ,.,
(2)
d
I
'
L LQ!m> = L LQ,PIBOM,.
111
5
d
(3)
p
(4)
L:C2.,r <= L
w
(5)
d
LQfm> <= l,m
(6)
I
Q,., >= 0, Q,,, >= 0,
Qfm,
.>=
(7)
0
rurchase ..
supphcl"i
In the above mathematic model, item 1 in the objective function is
the purchase cost of the m.1nufacturer that purchases raw material
from the raw material supplie r. Item 2 is the transportation cost of
raw material that is transported by the raw material supplier to the
manufacturer. Item 3 is the manufacturing cost for products made
by the manufacturer. Item 4 is transportation cost of products
tran sported from the manufacturer to the distribution center. Item
5 is storage cost of products stored at the distribution center. Item 6
is transportation cost of products transported from the distribution
center to the retailer. Constrai nt I ensures that the products transported from the distribution center to the retailer meet retailer demands. Constraint 2 is the limit for products transported from the
manufacturer to the distribution center. such that the quantity of
products is the same as that transported from the distribution center to the retailer and storage in distribution center is controlled.
Constraint 3 is the limit for products transported to the retaile r:
the number of products should be the same as that manufactured.
Constraint 4 is explains the distribution constraints on the distribution center. Constra int 5 is the capacity constrai nt for the manufacturer. Constraint 6 is the quantity constraint of raw material the
material supplier purchases. Constraint 7 is the numerical value
constraint of the decision variable.
3.2. Co-Evolutionary and constrainr-sacisfacrion genetic algorithms
( 1) Chromosome coding and decoding
Because the supply chain network has many decision variables and these va ri ables are constrained as a positive integer. the chromosome is encoding in integer. This supp ly
ch.1in network has four levels {r.1w material supplier. manufacture r. distribution center .1nd retaile r). Thus. this study
uses three sections of chromosomes to demonstrate the relationship among the four levels. The research go<~. l is to find
the purchase decision solution for various supply chai n
members. such that the coding uses purchase decisions of
all members as the decode value of chromosomes. Suppose
there are 3 ret.1ilers. 2 distribution centers, 3 manufacturers,
2 raw material suppliers. 2 products and 2 raw materials.
Table 1 presents the parameters for this case. Fig. 5 shows
the corresponding supply chain network mode. where R, is
a retailer. D, is a distribution center, F, is the m.a.nufac[Urers,
S, is a sup plier.P, is a product, and M, is the raw m.a.terials.
The relations among the four levels in the supply chain network are mapped in a tree structure. Fig. 6 shows the chromosome coding form .
Phase I shows the demands of various retailers. and the
quantity delivered to all distribution centers (Fig. 6). The first
two nodes (0 1 . D2 ) (last row in Fig. 6) are th e demand of
retailer R1 for product P1 (total. 250 yards (69 + 181 )): distribution center D1 provides 69 yards and distribution center D2
provides 181 yards. Phase 2 shows that after all distribution
centers receive purchase orders from retailers. they gather
th is information and make purchase orders to manufac[Urers
( phase 2 in Fig. 6). The first three nodes (F1 . F2 and F3 ) of
Phase 2's last row mean distribution center purchases mate-
~ _lraMpo~auon., 1;::-:~_tr..msponauun .,
l.1!1.!.D
manuf!K'tun:rs
lmanuf~cturecost)
co~t
c;o-:;
co~t
di stri huuun c:c:nh:rs
(di.>tributioncost l
Fig. 4. Mode gr01ph of I he supply ch01in nerwork 111 I his srudy.
retailc:r
5. 6923
Y.-H. Chang / Expert Sys tl."ms with Applications 37 (2010) 6919 - 6930
To~b le
ers already meet the needs of the distriburion center. no further purchase is made from manufacturerf 3 • Ph.:1.se 3 in Fi g. 6.
the first two nodes (5 1 • S~ ) purchases raw mate rial M 1
( 14 kg) from supplier 51 and purchases 5 kg from supplierS~.
The full length of the chromosome is the sum of the lengths
of these three phases. Take tree schematic graph of a chromosome (Fig. 6) as an example, the total length of the chromo-
1
Pnameteruble.
Mode parameter
Symbol
Retailer
Distribution center
Manuf.lcturer
Raw material supplier
Product
R,(i• l . 2. 3)
D,(i•1.2)
f, (i• l. 2. 3)
S;{i•1,2)
P, (i•l. 2)
Raw material
M,(i•.2 )
Number
some is (3 • 2 • 2) + (2 • 2 "3 ) + (3 "2 "2)-36
(2) Evaluation
rial P1 (58 yards in total) from manufacturerf 1 : manufacturer
F~ the purchases 132 yards. Because these two manufactur-
ma~enal (M,)
product~(
P'~h<>>< ~l /?1'~1COSt
P)
product~(P,)
"'"' """"" •,....~ "'~'P"""';""~
"'
CO~
manufacmrc:~F1 1
~uppliers(S)
~
COSI
d1stribut10n (;C:ntc:n.f0 ,)
(manufacmrccosr)
(dJstrihutioncosl)
Fig. S. The relation graph for the three levels in the supply chain network.
Phase I :
Distnbutwn
Centc-t
Quamit}
Quantify
(ll<1}
(58)
(1!11)
(13::!)
(163)
fO)
(17)
(45)
(l~::!J
rl::!1)
( 155) 175)
(95f
1 76) (39-1) (5 )
(155)
{0)
(3)
(3)
(724)
(j)
(5)
(:~5())
(li J
( I)
(50)
(31}
(62)
{IJ
{::!::!5)
:::~::::'"'"~
s""'''"'~~
Qu~ntity
(14)
(5)
(lJ
(JO)
( I S)
genes.
The objective of this study is ro minimize cost. Thus, the
objective val ue and fitness value have an inverse relationship. That is. as cost increases, the fitness value decreases.
!ote: the grey part is a chromosome
Fig. 6. Tree schemouc groph of o chromosome.
6. Y.-H. Chang/ Expm Systems wirh Applicarions 37 (2010) 6919 - 6930
6924
To express the relationship between the fitness value and
objective value. this study defines the formula for fitness
value as the maximum cost of this generation's chromosome
plus a certain integer value, and deducts the cost of the chromosome such that the solution quality can be determined by
the final fitness value. The near optimum solution can then
be obtained. To make the final evolution result match the
supply upper bound of various levels and meet the demands
of manufacture. a penalty function is added to force the
result to evolve according to a constraint criterion. The rules
of the penalty formula are as follows.
a. When the value is lower than one level's demand, the cost
should plus the quantity of insufficient and multiply the
value of one level's highest unit cost. If demand exceeds a
level's demand. the penalty is waived because it has <~!ready
been included in the cost calculation.
b. When the value exceeds the upper bound, the cost should
plus the surplus qu<~ntities and multiply the unit cost of
all levels (e.g., distribution cos t, manufacturing cost and
purchasing cost).
(3) Selection. crossover and mutation
Selection me<ms th.lt only some chromosomes smvive to the
next generation from the parent generation based on a fitness value. Chromosomes with high fitness va lues Me most
like ly to su1vive. This study uses the roulette wheel method
for selection. which takes the fitness value of a chromosome
as the probability value; thi s method can be used to choose
the next gene ration fairly, and is the most widely used
reproduction method (Gen & Cheng, 2000).
Crossover switches some genes in two chromosome groups to
create a new filial generation of chromosomes. The goa l is to
enlarge the se.:uch space and increase the speed at which
the opti mu m solution is acquired. The crossover method in
this study adopts the common two-point crossover method.
Two tangent points are made first and the gene values of
two chromosomes are switched between two tangent points
(Fig. 7).
The purpose of mutation is to increase the diversity of chromosomes and prevent premature convergence and arriving
at a partial optimum solu tion during the search process.
The study adopts the single point mutation method (Fig. 8).
The system chooses a chromosome to mutate based on mutation probab ility. The chromosome selected can mutate again
based on the gene picked randomly from chromosome genes
by the system and add to, or deduct the muta tion value
which randomly generated by according to the protoge ne
value. For instance. the protogene value is 122. The system
picks the change value randomly (0 <change value :; :;: protogene v<~lue ). If ch.mge value is the deduction value and the
change value is 62. the gene value after mutation is 60
( 122 - 62).
(4) Co-evolutionary mode
The difference between a co-evolutionary and simple
genetic algorithm is that the co-evolutionary genetic algorithm takes the fitness value as the evaluation criterion
and. based on the interactive evaluation of chromosomes
and criteria, which make the chromosome of each generation evolve toward S.ltisfy the difficult contented criterion.
Thus, the evolutionary speed can be increased and the system can obtain the optimum solution. Table 2 li sts the study
criteria. The co-evolutionary steps are as follows.
a. Take the number of chromosomes that match the criterion
as the fitness value of the criterion. A criterion easily met
has a low fitness. The fitness value of a criterion equals
the value that the number of population size minus the
number inside that matches criterion.
b. Ta ke the fitness value as the probability val ue in the rou lette
wheel method and select the criterion. If a criterion is
selected more than once. it should be neglected. For
instance, if there .are 6 crite ria and the selection res ult is
1, 3, 2, 3, 1, 4, the selection result should be the first. second,
third and fourth criteria.
c. Based on above selec tion criteria, all chromosomes are
ev.alu.ated.
(Before Crossov er
Chromosome A
I ss
Chromosome B
I 20 I 36 I 48
11 32
Io
122 1121
Io I
3
3
60 154 1 17 1750 ~24
1724
I
3
Is
1259
I
I 22 I 19 I 35 I 75 I
(After Crossover)
Iss
Chromosome B' I 20
Chromosome A'
1132
I0
l6o ls4 117 1750
136 148 1122 1121
I3
1724
I3
1259 I
I0
Fig. 7. Schcm<ttic gr<tph of the two-point crossover operiltion.
[Before Mutation I
Chromosome
I 58
(After Mutation)
Chromosome'
I 58
lm I
lz59
I
~ mutation point
lm I 0 l6o 1121 I 0 I 3 I 3 lm I
l2s9
I
!132
I0
1122 11 21
0
Fig. 8. Schemiltic graph of the simple point mutiltion method.
7. Y.-H Chong / Expert SysrC"ms with Applications 37 (20 10) 691 9-6930
T.able 2
Sum mary of study co nslraints.
Number
Consrr.aim s
The quantsty ordered by a retasler must m.nch the qu.antity
purch.ased by • distribution center
The quantity purchased by the distribution center must march the
quantity manufactured
The raw material needed by the manuf.tcturer must match the
quantity it purchases from the supplirr
The stor~e constrainls of disuibutioo center must be s.atisfied
The productiOn C.Jpaciry of manufacture r must be S.ltssfied
The supply constro~ints of ro~w mo~rerio~l supplier mu st be satisfied
d. Repeat operations (a)-( c). Once the fitness values of a crite·
rion are close and exceed the threshold value. then use all
criteria to evaluate the chromosome until the near optimum
solution is found.
If the population size is 100, the number of criteria is 6.
Table 2 shows the criteria. After the system creates the initial popu lation and evaluates this population using the coevolutionary mode (Fi g. 9 ), the first step is to calculate
how many criteria all chromosomes match. When chromosomes matching criterirt 1- 6 are 20, 30, 70, 45, 40. and 0,
respectively, the fitness value of each criterion is
80( - 100 - 20), 70. 30. 55. 60, and 100. After retrieving the
fitness value of each criterion. the roulette wheel method
is applied to pick the evaluation criteria for this generation
as criteria I. 6. 1. 2, 6, and 1. Thus. if all chromosomes are
evaluated by criteria 1, 2, and 6. and is this circle in each
generation is retained until the fitness values of all criteria
are >90, then all criteria are used to evaluate the chromosomes unti l the system converge.
(5) Constraint-satisfaction mode
When the initial chromosome is created, one must determin e
whether it matches all constraints. If one of initial chromosomes violates the constraint, the chromosome should be
created again. The creation method is as follows. First. distribute the demand of last levels to next leve l. For example,
distribution center D1 demands 240 units thar shou ld be distributed ro th ree manufacturers. The first two manufacturers
are distributed randomly by the sysrem. If manufacturer f 1
has 100 units, manufacturer F2 has 60 units and the remain-
a.
b.
c.
cnfcriun
cri t<"nl>ll
nit..:rion
1-fit n c~ ·
!W
2 - titnot~><·
7!)
3-fi t n.:' ~:
30
55
60
~·nt.:ntiO
.J.-fi tnc" :
nitcrion !- rl tnes<
chmmn,;omr
c hrnmosom<.·
~·ntcnon o-i1 t nc'~
Ch ll•OUhOOI<'
100
]!)()
critcnon I. 2. o
chrum"sonlC
Chll llllOS(lffit:
1-
f1tnc~~. 10~000
2 · filti<.'S' 204!:!0
cn rcn.•n I
cnten.:• n 2
criteriu n6
(<. ah:u lale f•t n c":.; ,
hycnr.-na t.:::.M
n f <: h rt'>m.>• 0m<:~
Fig.. 9.
Co-evoluuon ~ ry e v o~ luarion
d.
(6)
mode.
6925
ing 80 (240 - 100 - 60 ) units are distributed to manufacturer f 3 . After initial population was created, the
chromosome that already meets the requirement must be
verified by production capacity and the upper bound of storage. If any chromosome does not match requirements. the
chromosome must be created again until all chromosomes
match the constr.1ints. This method ensures that the initial
chromosome must fall into the optimum solution region.
However. even if the initial population matches constraints,
after crossover and mutation, the chromosome can still be
outside the feasible region. Thus. the ch romosome that does
not match constraints should be adj usted such chat the
search space falls inside the feasible solution space. This
method narrows down the search space to find the near
optimum solution quickly. The adjustments to invalid chromosome are as follows.
When the distribution result cannot meet the demands of
the level s. pick an item randomly and add the insufficient
quantity directly to the distribution quantity. For example,
rete~il er R, demand s 250 unitsP, , and distribution center
D, di stributes 140 units. distribution center D2 distributes
100 units. However, the retailer is short 10 units
(250- 140- 100). Thu s, the system selects a discribution
center randomly and adds 10 units to the distribution quantity. If distribution center D2 is selected. the distribution
quantity of D2 must be increased to 110 units ( 100 + 10).
When the distribution amoum exceeds the demand of a
level, then select an item random ly and use the distribution
quantity to deduct the quantity exceeding that needed . If
the distribution quantity is still excessive, the next distribution quantity of the next item shou ld be deducted until the
distribution quantity matches demand. For instance. retailer
R, orders 250 units P, . distribu tion center D1 distributes 260
units and distribution center 0 2 distributes 100 units. However, the retailer receives an additional 11 0 units
(260 + I 00 - 250 ). Thus. the system selects a distribution
center randomly. If it se!ectsD 2 • the distribution quantity
of D2 is reduced by 100 units and the others should be
deducted 10 from next item D, 's distribution quantity. Thus,
the demand equals the distribution quantity. The distribution quantity of 0 1 is 250 units and the distribution quantity
of D2 is 0 un its.
When the total distribution resu lt exceeds the upper bound,
then select an item random ly and use the distribution quantity to deduct the excessive quantity. If the rota I distribution
quantity is sriH excessive. the disrriburion quamity of t he
next item should be deducted unt il the total distribution
quantiry equals the upper bound. For example, distribution
center D, of retailer R1 orders 150 unitsP 1 , and the distribution center 0 1 of retailer R, order 100 unitsP 2 • However. the
upper bound of distribution center 0 1 is 100 units. Thus, the
system then selects an item randomly. !fit se lects retailer D,
of retailer R1 order P2 , the distribution quantity should be
deducted 1 DO units, insufficient part should be deducted
50 from next item 0 1 's distribution quantity of retailer R,
order P1 • Thus. the distribution quantity is 100(150 - 50 ),
and the total distribution quantity (0 + 100) equals the
upper quamity.
Repeat operations (a}-(c) until total distribution quantity
equals demand and is less than or equal to the upper bound
quantity.
Co-evolutionary constraint mode
This mode combines the fearures of the co-evo lutionary
mode and constraint-satisfaction mode. The constraint-satisfaction mode was introduced in item 5. Because the adjustment mechanism of the constraint-satisfaction mode can
8. Y.-H. Cha ug j txpm Sys ttms wirll Applications 37 (1010) 6919-6930
6926
Tabl~
1
Summary constramts for co-evolutiOnary constraint genetiC algorit hm.
Number
Constraints
The
The
The
The
The
quantity of products transported from a distribution center to .1 ret.liler multiplied by th~ unit tr.lnsporution cost should be mimmized
distribution qu;mtity of a disrnbution center's products multiplied by the unit distribution cost should be minimized
quantity of products rransport~d from manufacturers ro a distribution center multiplit'd by the unit mmsportation cost should bf' minimiud
quantity of m,mufactured prod ucts multiplied by the unit manufacturing cost should be minimized
raw materi.tl quantity of a suppli~r's r.lw llldteridl multiplied by the unit purchase cost should be minimized
the search space for a feasib le solution that reduces unnecessary
search cost. Then, the co-evolutionary mode and const rai nt are
combined as the tensile force. As the fitness value of a constraint
increases, the power of evolution force increases. and the constraint has more needed to dominant its evolvement. This concept
is opposite to chromosome. The system ca n use the interactive
evaluation between constraint and chromosome to find the near
optimum solution. Fig. 10 shows the concept co-evolutionary constraint mode.
4. Experiments and verification
Fig. 10. Conceptual graph of the co-evolurion.Jry constraint genetic
<~lgorithm .
modify chromosomes tha t do not match const raims. if adopt
constrain evaluation method on co-evolutionary geneti c
algorithm. the degree of chromosome that matches constrc:lints canno t be distinguished. Thus. const raint-s<Hisfaction mode can be adjusted as follows.
a. When the demands of all levels and resource constraints are
mer. the minimum cost is the basic rule (Table 3).
b. Take the chromosome with highest cost as the fitness value
of th e constraint. For instance. if the purchase cost of one
chromosome is highest, the fitness value of constraint 5 is
purchase cost plus 1. Fitness of the constraint equals the
sum of all chromosomes with the highest cost. it shows that
higher the cost is. more need of evo lvement to the
constraint.
c. Take the fitness value of constra ints as the probability va lue
of roulette wheel method.
d. According to the selected constraint s, evaluate all chromo-somes and calc ulate their fitness values. The fitness of a
chromosome can be calcu lated as described in Section
3.2(2). Only calcula te the fitness values of related costs that
are re levant to this constraint select ion.
e. Re peat operations (a)-{d) until the near optimum solution is
found.
A beneficial feature of the co-evolutionary constraint genetic algorithm is that it uses constraint-satisfaction mode to narrow down
This srudy uses the above mathematical model to simulate a
supply chain network de sign problem. To make the experiment
closely resemble practical supply chain operat ion, the Taiwanese
textile export industry is adopted as the experimental object. Th is
study adopts proportionaliry information acquired from interviews
with manufacturers to set parameter values for the model. This is a
real supply chain mechanism ro a large extent. Because the supply
chain network scales in the textile industry differ, this study divided the test case into 10 large and small-scales (supply chai n
network nodes below 1000 is small-scale. the rest are large-scale )
ro ensure the model can be applied ro various scales. Table 4 shows
the test case. With mathematical programming. the simple genetic
algorithm, constraint-satisfaction genetic algorithm. co-evolutionary genetic algorithm. and co-evolutionary constraint genetic algorithm a re used to find the solution that verifies solution search
efficiency and accu racy. Table 5 presenrs parameters of the text ile
industry.
In experimen ts. population size is 100, crossover rate is 1, and
mutation rate is 0.01. The terminati on condi tion when fi tness values of chromosomes in a population are the same. When the number of generations exceeds 100,000, searchi ng will stop
automatically.
This experiment is written using the Delphi development tool,
and is executed on a Pent ium rv 2.8G with lG RAM to continuously
test 10 times as a benchmark for comparison. The mathematical
programming method, however, uses LINGO 9.0 software for comparison. and it is commonly used for non -linear integer programming with the same computer Table 6 presents comparison
results. Fig. 11 compares the time costs in obtaining a solution.
Table 4
Tes1 cases .lnd parameter table.
C.Jse
R~tailer
Distribution center
M<~nufilcturers
Supplier
Product
MCJterial
Smaii-SC.llt'S
Combination number of
n~rwork
.ucs
12
96
324
768
t..uge-sules
10
10
12
14
16
18
20
10
10
12
14
16
18
20
10
12
14
16
18
20
10
12
14
16
18
20
!SOO
10
2592
4116
6144
8748
12.000
'
9. I
Y.- H. Ch ang/ Expert Sys rems wuh Applicatio ns 37 (20 10) 6919-6930
6927
Table 5
P01rameter romdom range of the textile industry.
Code
Parameter
Range
Unit
RetAiler dem<~nd
Textile r.aw m.ueri o~l
Unit purc.hase cost of textile raw materi.al by a rexrile supplier
Unit manuf<~cturing cost of .a product
100- 10,000
0.09-0.3
60-120
5-12
Yard
kg
Unit distribution cost of a disrribution center
0.9-1.6
dollil r/kg
doll.lr/
V.1rd
0.6-7.5
dollar/
Y.Jrd
dollar/
Unit tra nsport.alion cost of .1 product uansported from a distribution center to a re u iler
1-8
v.ud
dollar/
Troxtile supplier Tdw ffidteri.ll up~ r bound of supply
Manufacturer upper bound of productivity
Distribution upper bound of .a distribution center
1.5-2 times of m.anufdctur~r's production cap.lCity
2- 3 times of distribution cente(s capacity
1.5-2.5 rimes of retailer's total demand quantity
Unit tr.Jnsport.ltion cost of a product
tro~uspone<l
from manuf.acturers to distribution cenrers
Y~rd
10
T~ble
Y.m!
6
Perform~nce
Case
of UNGO •nd
rh~
genetic
10
Case
~lgonthms.
Simple CA
UNGO
Time
Cost
937.546
3,710.342
9.868,523
22.526.384
43.206.351
63,557.214
135.063.824
175.395. 128
356.951,353
395.298,246
Avg. cost
268
1632
23,141
37,682
50.924
93,951
173.952
313,825
937,546
3,7 10.342
9,868.523
22.526.384
43,206.351
67.072.073
142.577.049
187.904.091
378.041 .78 1
426.351,982
Co-evoluriotury GA
Avg. time
937.546
3.710.342
9,868.523
22.526,384
43.206,351
66.569.586
141.297,516
185.910.201
377.086.922
4 21.965,2 15
Avg. gene.
Avg. cost
68
937,546
3.710.342
9,868.523
22,526.384
43.206.351
66,826.246
142.038.588
187,333,036
377.879.078
425.841.874
Ill
! 58
75 1
654
1351
2133
3649
4997
6824
Constraint-satisfaction GA
Avg. cost
10
kg
Yard
15.049
24.973
32.089
67.842
82.684
127.682
179.72 1
184,932
Avg. time
Avg. gene
129
602
594
11 98
2033
3249
4970
6592
43
78
12,492
20,439
24.03 1
52.968
76,843
119.192
168.518
176.821
Co-evolutionary Constraint GA
Avg. time
Avg. gene.
Avg. cost
Avg. time
Avg. gene.
4
153
501
583
1105
1900
2594
4733
6413
28
70
10,726
14.392
20,195
44.395
73.824
105.835
156.418
166.492
937.546
3,710,342
9,868.523
22.526,384
43.206,351
66.153,035
141.11 5.368
185,476.283
366.865.767
409,754.413
84
342
540
983
1503
2347
4008
6348
20
69
9038
10,832
19.824
37,521
70,952
96,382
148.254
159,824
Nare : the unit of the target value is dollars and time is in seconds.
is the abbrevi<ltion of "average·.
"gene." is the ~bbrev1ation of "generaliOil ~.
"avg.~
This study analyzes the time and solution search cost fo r large
and small-scale cases. Fig. 11 and Table 6 shows the solution
search rime fo r the small -scale case. Cases 1 and 2 are smallest.
With both LINGO and the genetic algori thms . solution time is
<6 s. However. as the number of leve ls increases in cases 3 and 4,
time cost for LINGO increases from several seconds to roughly
30 min ( 1632 s); the simple generic algorithm only takes 12 min
(751 s): and the co-evolutionary const raint genetic algorithm takes
6 min (342 s). Thus. genetic algori thm is fas ter than LINGO in for
the small-scale cases. For the small -sca le cost (Fig. 12 ). the genetic
algorithm and LINGO solution search qualities are almost the same.
For the large-scale cases, as the case parameter increases. complexity increases. The time cost for LIN GO increases dramatically
(Fig. 13). For case 5, LINGO needs over 6 h; for case 7. LINGO needs
over 14 h. Although ti me cost for the genetic algorith ms also increases (e.g.• the simp le generic algo rithm needs 35 min and coevo luti onary constraint genetic algori thm only needs 25 min ).
but rema ins consi derably less than that by LINGO. For case 8. time
cost fo r LINGO inc reases by days. For case 10, LI NGO needs over
3 days to search the solution. Thus. LINGO is very ineffi cie nt when
supply chain management needs every second. However. all genetic algorithms reta ined high performance levels. Time cost was
within 1.8 h for all genetic algorithms. Thus, th e average tim e for
so lution search for all genetic algorithm s was better than that with
LINGO, especially the co-evolutionary constra int genetic algorithms. which performed best in complex problems.
Fig. 14 li sts the cost for the large-sca le case. Although case complexity increases and the solution searc h quality of the ge ne ric
algorithms and LINGO are almost the same. the difference is 8%.
In the most complicated case, case 10. the co-evolutionary constraint genetic algorithm only has a 3.6% difference from LINGO.
Therefore, the co-evolutionary constraint generic algorithm has
some efficiency for large-scale solution cost
The comparison of so lution search efficiency and quali ty for
la rge and small cases indicates that although costs by the generic
algori thms are sli ght ly more than those for LINGO, time cost
for the genetic algorithms is much less than for LINGO. Norably,
the co-evolutionary constraint ge neti c algorithm us es the
10. Y. ~H. Chang { Ex~rt
6928
Systems wir/1 Applicarions 37 (2010) 6919- 6930
:~I
1,4))
f
1,300
!
''i
!
1,200
1.100
I,OOJ
j:
//
'
..,
F100
1
/i
EIL----------------~·~'_:.~J
2::~:=·-~-~-·~-·_;_':_·/_._:_:_
Fig. 11. Se.-.rching solution time cost for sm.-.IJ .sc.-.Je uses.
l,OOO
1,000
1,600
11,400
§ 1,200
~
I,OOl
.s ""
600
~ QA
- - -· ~OA
--~oomryccmlrMt CA
Fig. 12. Search soluuon cost for small-scale cases.
co-evo lutionary mode to accelerate evo lution. and the constraintsati sfaction mode can narrow down the search space to obtain
the opti mu m solution. Co-evolutionary co nstraint generic algorithm is more efficient in searching the near optimum solution
than the simp le genetic algorithm, co-evolutionary genetic algorithm and constraint-satisfaction generic algorithm. The cost for
co-evolu tionary constraint genetic algorithm is also less than that
of othe r generic algorithms. Thus, the co-evolutionary constraint
genetic algorithm performs best for the supp ly chain network
design problem.
5. Conclusions and future research
Market competition has increased and thi s challenge has caused
ente rprises to change their operational modes and enhance the
efficiency of supply chain management to enhance their competitiveness. One must consider multiple products, levels and resource
constraints to establish the best integration mode that matches
supply chain manage ment needs. The co-evolutionary constraint
genetic algorithm can solve the sup ply chain nerwork design problem. Ma th ematical programming. the simple genetic algorithm, coevo lutionary genetic algorithm, constraint-satisfaction generic
algorithm were compared in terms of solution search quality and
efficiency. Th e co-evolutionary cons traint genetic algorithm narrows down the search space with the constraint-satisfaction mode.
Additionally, the co-evolutionary mode can adjust evaluation constraints dynamically to match a complex reality, has excellent
solution search quality and speed when searching for a solution
to supp ly chain management production and sale mode problems.
and performs better than other meth ods. Analysis of real textile
11. 6929
Y.-H. Chang / Expert Sysrems wirh Applications 37 (2010) 6919-6930
,-----
mmm ----
i!SQOOO
mm~ -:
.-·,
!
1
~
.;! 100.000
!0.(00
'·""
"""
'if4,oo:J
"' 3,00)
~
""2,(ro
1,001
l""""""""'s=;~~~:::::::::~~~:;;;;;:::;::;;;,..--·
Fig. 13. Time cost for solution searching in the .J.rge-sc.J.le case.
~ OA
--co-e~~ 'I'C<)(IS(r mt C-A
Fig. 14. Cost for solution searching in the large-sco.le co.se.
industry cases indicates that this mode reduces supply chain management cost and increases the con nectio n among enterprises.
Thus, the experiment result can be a decision-making reference
for Taiwan's manufacturing industry when developing global
plans. and remind decision-makers that planning should consider
production planning and not neglect resource integration of all
entities in their supply chain. The integrated power of supply chain
management is the bes t channel for enterprises to acquire profit
and competence.
This study takes the textile industry as its expe rimemal case.
Because of the scale of supply chain networks, member structures
of various industries, products. raw materials and resource constraints differ. rn future research. the supply chain netwo rk struc-
tures of different industries can be compared to verify solution
search efficiency of various constraint-satisfaction genetic algorithms. This study referred ro the model structure in Altiparmak
et a. (2007 ). Because the demands of supply chain modes vary,
modes can be adjusted based on supply chain demands such that
the mode is close ro actual supply chain operation. For instance.
the model can increase the choices in different periods, storage
control. batch discounts. and emergency purchase processing. In
a supply chain network, different enterprises have different targers. This study only adopted total minimum cost as the measure
of supply chain operations. Multiple constraints will be added as
the decision-making standard for evaluation in the future. Moreover, genetic algorithms are nexible and have many parameters
12. 6930
Y.-H. Chang/£tpm Sysums with Applications 37 (1010) 6919- 6930
to adjustment (e.g., encoding method, evaluation means. and termination condit ion). All these factors influence algorithm performance. Simulations can be done with different parameters or
integr.:ne additional field theories ro improve ex perimental efficiency and combine othe r a lgorit hms for function integration and
result comparison.
References
Bielli. M~ C.uamia, M.. & C.uorenuto. P. (2002 ). Genetic algorithms tn bus nerwort
optimization. Transportation Research. Parr C. 10. 19- 34.
Chou, H., Premkumar. G., & Chu, C. H. {2001 }. Geneti( algorithms for
commumcations network design - An emp1ncal study of the factors that
mnuence perfonnance. IEEE Transactions on Et•olurionory Computation. 5(3).
236-249.
Chnsrophe. P. & Hugues. B. (2007). A gestalt genetic algorithm: less detailS for
~ner se.uch. In Genetic and evolutiOnary computation conferenu (pp. 13281334 }.
Altip;~rmdk. F.. Gen. M~ & lm. L {2005). A genO"tic .Jigomhm for supply chdln
network d0'51gn. In lnlemaflonal conference on computers and indtu!nal
engmeer111g (pp. 11 1-116).
Alnpannak. F.. Gen. M., Un, L. & Knilogliln. I. (2007). A steady-sure genenc
illgonthm for multi-product supply chain network design. Computas and
lndustnal Engrneering. d01: I0 10 16/J.Cit'.l007 .05.012
Gen. M .. & Cheng. R. (2000). Gt>11etic algorithms and enginerring optimtzatton. New
York: Addison-Wesley.
Goldberg. D. E. (1989). Genetic algorithms in search. optimizadan and machine
learning. Boston. MA: Addison-Wesley Longman Publishmg Co.. Inc.
H>~ y>~.~h1. K Koguro. N.. & Mur.:~kami. R. (2005). lndustndl .:~ nd pr~ctlcdl experience
..
tro'lck paper session 2: Diversified SCM st.a ndard for the jdpilnese retail industry.
In Specwl mterest tracks and posters of the 14th internat ional conference on Warld
Wide Wrb (pp. 856-863).
H1llis. W. D. ( 992). Coevolvmg p.1r~sites improve simulated evolution as an
optimiution procedure. In C. G. Ungton, C. T.1y!or.j. D. Foarmer, & S. Rasmussen
(Eds.). Arnficiallift II (pp. 313-324). Redwood City, CA: Addison-Wesley.
Hollilnd. J. H. ( 1975 ). Adaptation m natural and arrifina/ systtms. Annilrbor.
Michigan: The University of Michigan Press.
Hou, Y. C. & Chang Y. H. {2002}. Research on co-evolution.ny electroniC commerc~·s
negotiatmg mech.:~msm. In Anthology of industry electronic opera/ions
managtment academic and prawcal conference (pp. 1-9).
Hou. Y. C.. & Chilng. Y. H. (2002b}. The new efficient hit'r.Jrchy combm.ltion
encoding method of evolution srrategtes for prochKtion atlociltJOn problems.
Computers and lndustnol Engintering, 43(3). 577-589.
Hou. Y. C., & Chang. Y. H. (2003). Dynilmic programming vari.Jnt in t'Volution
srr.uegH~S for production <~ilounon problems. MIS Rev1ew, 12. 1-16.
lb.Jraki, T., & Katoh, N. { 1988). Resource Allocation Problems-Aigontltm1c Approaches.
lhe MIT Press.
j.lr<~millo, j. H.. Bhildury. j., 8r B.ltu. R. (2002). On the use of genetic .lllgonthms to
solv!1' location probl~ms. Cumpurers a11d Operanons R.-srarch. 29.76 1- 779.
juang. Y. S.. Lm, S. S., & KilO, H. P. (2007). An ad.llptive scheduhng system with
genetic algorithms for arr.anging employee tr<lining progr<~ms. /nttmational
journal. JJ. 642-651.
Koray, D., & Marc, G. (1999}. A prim.1l decompoSitiOn method for the mtegrated
d~sign of multi-period production-distribution sysr~ms./IE Transactions. 31( 11 ).
1027-1036.
liu. P. Y Hsieh. Y. C. {2005). 828 e-commerce and enrerpris~ integrJtion: A study
..
bdsed on thO" v.Jiue system for the inrerilction of the multi-tiered supply ch<~ln
under the trend of e-business. In Proc~:edings of rhe 7th internanonal conference
on electronic commerce (pp. 385-392).
Maloni. 1.. & Benton, W. C. (1997). Supply chain p.Ht'nerships; Opportunities for
operations rese.Jrch. European journal af Operational Rest"arch, 10(3). 419429.
Mei.J(hrinoudis. E.. & Min. H. (2000). The dynamiC reOCiltion and phJse-out of il
hybnd. two-echelon plant{warehousing facihty: A mul tiple objective appro.Jch.
European journal of Operational Research. 123( I ). 1-15.
Robmson, E. P.. & S.1tterfield, R. K. ( 1998). 'Designing distribution sySt11'ms to support
vendor stril teg1es in supply cham m.Jn.Jgement'. Decurou Scii:UCe, 29. 685705.
Ross. A D. (2000). Perfonn.1nce-b.:~sed str.1teg1 resource <~11oc.:~t10n in supply
c
networks. lmemotionul journal of Production Economics. 63(2 ). 255- 266.
Su. X. Y. (2000). Materials circulacion and operations mouogemrnt ( 1st ed.). T01ipe1:
Hw.l T.1i Publishing.
Sue. A H. (1999}. Network design in supply chain m.lnilgement lnternatiOnaljoumal
of Agrlr Managem ent Sysrems, 1(2). 99-106.
Syam. S. S. (2002). A model .:~nd methodologies for the location problem With
logistic .'II components. Computers and Operanons Research. 29. 1171193.
Sy•rif. A.. Yun, Y., & Gen. M . (2002 ). Study on mul ti-sl.lge logistics ch.lin network: A
spanning tree-b.Jsed genetic illgorithm .:~pp rwc h . Compurers and lndustnal
Engineering, 43.299-314.
Szeto. K. Y.. & Fong. L Y. {2000). How adaptive agents in stock markrr perform in the
presence of rundom news: A genrric algorithm approach. Lecwre Notes in Compu ter
Scitnce. Berlin/Heidelberg: Springer (pp. 505-5 10).
Ting, C. j .. & Huang. C. H. (2005 ). An improved genetic algorithm for vehicle routing
problem with time windows. lnt emacirmal joun1a/ of Industrial Engineering.
12(3), 216-226.
V.a1dyanathan. j., & H~ san. P. (2001 ). Pl.lnnmg .Jnd coordm.Jtlon of production .and
distribution fac il ities for multiple commodities. European journal of Operational
Res~orch, I 33(2 ), 394-403.
Viswanat han, S., & Pipl ilni, R. (2001}. Coordin.ltmg supply chilm through common
replenishment epochs. EuroiXan journal of Optrot1onal Research, 129. 277286.
W.1ng. S. Q. (2002). Purchase ordrrs distribution mode of mulnple facrory seers·
production systtm - LCD monitor industry os the e)l.ample. Master Dissertiltion.
Institute of Industrial Engineering .:~ nd Business Oper.ltlon. Tungha1 UniverSity.
Wong. K. Y_Yip. C. L. & U. P. W. {2008). Automatic tropic.lll cyclonl1' eye fix using
genet1c algorithm. Expert 5yst!1'm5 With ApplicatiOns. J4{ 1 ). 643- 656.
Wu. Z. X. (2002}. Research ofg~nttic algon'thm's oppltcatioll in SCN productiOn and salt
optimum modt. Master Dissertation. Institute of Enterpnse Man.llgemenr. Shih
Chien University.
Zhou, j., & liu, B. (2003 ). New srochasric models for c;~paci t .ated lociltlon-dllocation
problem. Computtrs and Industrial Enginttring. 111 - 125.