You are on page 1of 7

$EVWUDFW-The partner selection in virtual enterprises

organization is one of the key issues corporate enterprises


experience nowadays. Based on the model of Ant Colony
Optimization Algorithm (ACA) in virtual enterprise partner
selection, in this paper, we fuse the genetic algorithm into ACA,
called fusion algorithm, in order to improve the effect of the
partner selection. The fusion algorithm has two steps: 1) it uses
the GA to optimize the model of partner selection and takes
advantages of rapid convergence of GA in initial search periods.
2) When GA search speed has become slow, the ACA takes over
the search process, in which it uses the candidates produced by the
GA as the seeds of pheromone used by ACA. By experimental
comparison with GA optimization and ACA optimization, it
shows that the fusion algorithm has performed better than the GA
and ACA optimization, respectively, in both speed and accuracy
under our selected numerical case. The fusion algorithm
presented in this study may be applicable to similar business
problems.

,QGH[ 7HUPV-ACO, Virtual Enterprise, Partner Selection, GA,
Hybrid Algorithm

I. INTRODUCTION
N the process Irom the industrial age transition into the
inIormation age, corporate entities are in a dilemma in the
process oI doing their core business and responding to
environmental changes. ThereIore, virtual enterprises have
been appeared and replaced entities enterprises in part to Iorm a
new organizational model compatible with the inIormation age.
Collaborative organizations are becoming more important in
business globalization, because the resources such human
experts, experiment laboratory etc. required by some kinds oI
projects can not be competent in one entity company. On the
other hands, individual companies seek eIIicient gains by
Iocusing on their core competences while outsourcing non-core
operations. This leads to the degree oI inter-Iirm transactions

Manuscript received December 15,2007. This work was supported in part by
the National Science Foundation oI China under Grant 70672020.
Z. Yao is with the School oI Economics and Management, BeiHang
University, Beijing 100083, China. (Tel: 8610-82317802; Iax: 8610-82328037;
e-mail: iszhyao buaa.edu.cn).
J. Liu is a graduate with the with the School oI Economics and Management,
BeiHang University, Beijing 100083, China. (e-mail: liujian
sem.buaa.edu.cn).
Y.-G. Wang is with the School oI Business, Nanjing University, Nanjing
210093, Jiangsu Province, China (e-mail: nkygwangsohu.com).
growing considerably. To complete one project, it is imperative
Ior core companies to select the collaborative partners, which in
turn call Ior the development and deployment oI decision
support models that assist companies in the selection and
management oI these relationships |9|. In the process oI
establishment oI virtual enterprises, how to select the partners
is one oI the key issues, which has drawn the attention oI some
scholars and conducted a Iew researches, see
|1||2||6||9||12||13||15|. However, the choice oI business
partners involved in a number oI Iactors, which makes partner
selection became a complex problem. Thus, there is an urgent
need to Iind an eIIective decision support model Ior
management reIerence.
In literature, various models and algorithms have been
proposed to optimize the problem oI partner selection |3|.
Establishing mathematical model and using classic
optimization method such as complex method to help virtual
enterprises select partners is the most commonly used, also
gained good results, such as an ally selection model based on
AHP and Iuzzy set in the virtual enterprise |1|. Cao et al.
present a multi-objective optimization model, taking the cost,
time and risk Iactors into account to minimize the risk oI
project Iailure and project completion time through Genetic
Algorithm |2|; Talluri et al. propose a quantitative Iramework
to design business union with two-stage analytical Iramework
|4|. However, due to the complexity oI the virtual enterprise
partner selection, no speciIic solution has been accepted
widely, the research is still in the initial stage. ThereIore, it is
necessary to continue to Iind more suitable partner selection
methods.
To eIIectively select partners Ior virtual enterprise, in this
paper, we present an integrated meta-heuristic algorithm, called
a Iused algorithm, i.e., Iusing the genetic algorithm (GA)
algorithm into the ant colony optimization algorithm (ACA) by
taking the advantages oI both algorithms. We have
experimentally Iound that GA algorithm has Iaster converging
characteristic in initial searching stage or earlier search process,
but with the search time increase, GA will suIIiciently reduce
their converging speed. On the other hand, the ACA has slowly
search speed in earlier searching stages because oI the lack oI
the initial pheromone. With the availability oI inIormation oI
pheromone, the ACA will speed up in searching. By Iusing two
algorithms, we not only can use the advantage oI GA that has
an initial speed-up convergence, but also can use the candidates
Fusing Genetic Algorithm and Ant Colony
Algorithm to Optimize Virtual Enterprise Partner
Selection Problem
Z. Yao, J. Liu, and Y.-G. Wang
I
3614
978-1-4244-1823-7/08/$25.00 c 2008 IEEE
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.

produced by GA as the initial pheromone Ior ACA. This is our
main idea in this paper. By numerical experimentation, we Iind
the integrated algorithm (Iusing algorithm) has perIormed
better in the respect oI converging speed than those oI the GA
and the ACA algorithms, respectively.
The organization oI the reminder oI this paper is as Iollows:
Section II describes the virtual enterprise partner selection
issue. Section III reviews the Ant Colony Algorithm, the
Genetic Algorithm, the basic principles oI Iused algorithm, and
the necessity Ior applying Iused algorithm to the partner
selection problems. Section IV presents the basic model and
processes oI the Iused algorithm in the problem oI the partner
selection in detail. Finally, the perIormance oI new algorithm
has been demonstrated by a numerical experiment in Section V
and conclusions and Iuture research have been included in the
last section.

II. PROBLEM DESCRIPTION
Virtual enterprise, also known as the dynamic alliance,
widely recognized deIinition is as Iollows: A virtual enterprise
is a temporary alliance oI enterprises that come together to
share skills or core competencies and resources in order to
better respond to business opportunities, and whose
cooperation is supported by computer networks |16|

. To
achieve rapid response to meet the market demand objective,
virtual enterprise relies on electronic inIormation means to
contact and cooperate, and reorganizes human resources,
Iinancial resources and material resources distributed in the
diIIerent companies and geographic areas. Virtual enterprise is
widely existed in current internet age. For example, virtual
manuIacturing in supply chain reported in |17| describes a
logistics company who serves customers that include Ireight
Iorwarders, transportation buyers and suppliers, banks,
insurance companies and government agencies. Each oI these
partners is dynamic integrated according to the market
opportunity. As long as market opportunities exist, this virtual
enterprise will be to exist. II the market opportunity is
disappeared, the virtual enterprise will be disintegrated. More
examples Ior virtual enterprise can be Iound in |16|.
To create a virtual enterprise above deIined, the selecting
partners that satisIy the required core competence is key
strategic problem. The partner selection issue can be described
as Iollows: The main or sponsoring enterprise launches the
virtual enterprise project, analyzes the core competencies
required, and identiIies candidate enterprises Ior each
competency by announcing invite public bidding inIormation.
AIter collecting bidding inIormation Irom candidate
companies, the main enterprise assesses the candidate
enterprises through the relevant experts to determine the best
candidate enterprise portIolio. This process oI partner selection
can be presented in Fig.1.
Fig.1 shows that one virtual enterprise totally need J kinds oI
core competencies and each core competency including
prospective candidates as
) ( f
i
u , where f1, ., J and i1, ., I
f
.
According to standards set in advance, the sponsoring
enterprises select partners Irom the candidate enterprise in each
oI core competency category to compose a dynamic alliance.
Assuming one dynamic alliance needs J kinds oI core
competencies, Ior the f
th
core competency it has I
n
potential
candidate partners. Note that n may be diIIerent in various core
competencies, because each kind oI core competency may have
diIIerent candidate partners Ior selection. From the perspective
oI practical applications in the dynamic alliance, each kind oI
core competency needs at least 1, up to 2 partners. The number
oI partners combination Ior the f
th
competency is
2
) 1 (
2 1
+
= +
f
I
f
I
f
I
f
I
C C , and ) 1 (
1
2
1
+

=
f f
J
f
J
I I Ior entire
combination in one alliance. It is evident that there is a
combinatorial explosion problem with the increase oI J.
ThereIore, it is diIIicult to eIIectively select candidate partners
to satisIy the requirements oI all core competence.


Fig. 1. Partner Selection Process

In literature, virtual enterprise partner selection issue can be
expressed as multi-goals combinational optimization problems
|5||6||10||14|. In this paper, we Iollow the method used in |6|
Ior modeling the virtual enterprise partner selection problem, in
which the core competences required include the cost, time, and
risk. OI course, there are many Iactors need to investigate into
model, see |16|. However, the diIIerence between our research
and Xiong et al. |6| is that the later is based on genetic
algorithm to optimize the partner selection problem in virtual
enterprise, but our study is based on the Iusing algorithm that
integrates the genetic algorithm and ant colony algorithm. In
Iact, this paper is extended the work by Wu et al. |10| that uses
the ant colony optimization to achieve the partner selection.
The core competency assessment criteria are minimizing cost,
time and risk. SpeciIically, the member enterprises oI virtual
enterprise need to meet the Iollowing 3 Iunctions:
The Objective Function 1: Minimizing the cost oI the
dynamic alliance operation. Here the cost is composed oI two
parts: one is the inherent cost oI chosen individual and the other
is linked cost. Generally, the objective Iunction can be
expressed as Iormula (1):

+
+ =

= =
J
f
I
i
f
i
f
i
f f
i i
f
i
f
i
link in
H H C H C
C C C
1 1
"
"
'
'
) " ' (
" '
) (
2
1
) ( Min
) ( Min Min
(1)
2008 IEEE Congress on Evolutionary Computation (CEC 2008) 3615
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.

Where,

=
) (
) (
0
1
f
i
f
i
f
i
u decline otherwsie
u choose if
H ,
The
) ( f
i
C is the inherent cost Ior choosing
) ( f
i
u ,
)
' ' '
(
' ' '
f f
i i
C is
the linked cost between
)
'
(
'
f
i
u and ,
)
' '
(
' '
f
i
u where
| , 1 | and | , 1 | ,
' ' ' ' ' '
J f f I i i .
The Objective Function 2: In order to grasp the Iast-changing
market opportunities more quickly, the whole virtual enterprise
is required to reach the minimum response time to the market.
The overall response time oI virtual enterprise depends on the
enterprise structure and operational workIlows. In general, Ior
the vertical integration enterprises like supply chain (that is,
there is a serial process that each connection is mutually linked
between research and development, manuIacturing,
distribution and other aspects oI the operations), the overall
reaction time can be expressed as Iormula (2):

) ( Max Min Min
) (
| , 1 |
f
i
f
i
J
f
I i
H T T


= (2)

where,
) ( f
i
T is the reaction time Ior choosing
) ( f
i
u and
f
i
H is
the same as in (1). Because oI the existence oI parallel
operations, each parallel link could be taken as a new one,
where it is choosing the maximum reaction time oI the linkage
Ior each oI new overall reaction time.
The Objective Function 3: Minimize the overall risk oI
virtual enterprise operation. Generally, the overall operating
risk can be denoted by Iormula (3):

=

=
J
f
f
i
f
i
I i
H R R
1
) (
| , 1 |
) ( Max Min Min (3)
Where,
) ( f
i
R is the risk Ior choosing
) ( f
i
u , deIined as the
moment oI the loss expectations and the probability triggered
by each partner`s risk Iactor, that is, )) ( ), ( (
) ( ) ( ) ( f
i
f
i
f
i
r P r E R =
and
f
i
H is the same as deIined in (1).

III. THE GENETIC ALGORITHM AND ANT COLONY ALGORITHM
A. Background of GA and ACA
Genetic Algorithm is a bionic optimization algorithm. It is
Iirst proposed by Holland in 1975. It bases on Darwin`s
'survival oI the Iittest biological evolution theory and
Mendelian`s orderly genes in the chromosomes. It is imitating
biological evolutionary process to choose the optimum Ior
some complex problems. To use the GA Ior searching optimal
solutions, the problem will be transIormed to the binary codes
or digital series that Iorms a string called chromosome and is
placed into the problem environment. In accordance with the
'survival oI the Iittest principle, the strings will be reproduced
iI they are Iitted in the problem good enough. On the other
hand, the string will be split into two or more sub-strings called
gene strings that can be exchanged and mutated among gene
string to generates new generation (new string). New strings
that are Iitted in the environment will be kept as the new
sub-generation chromosomes. Thus process will be continued
generation-by-generation. Finally it converges to the most
Iitness series to the environment and obtains the optimal
solution |7|. In this process, algorithm doesn`t make use oI the
Ieedback inIormation enough. There are a large number oI
redundant iterations when the solution comes to a certain scope,
results in low eIIiciency

|13|.
The ant colony algorithm is kind oI simulative evolutionary
algorithm inspired by the collective behavior oI ant in the real
nature, Iinding the optimal solution through the exchange oI
inIormation between individuals and mutual cooperation |18|.
The essence oI the ACA as Iollow: 1) Selection mechanism, the
greater available the amount oI inIormation is, the greater the
probability that the item selected will be; 2) Updating
mechanism, the inIormation will grow aIter the ants passed
more, and gradually decreases as the time went by iI no ant
passes; 3) Coordination mechanism, the ants communicate and
work collaboratively by inIormation. However, a main
weakness oI ACO is the speed oI convergence will be slow iI
no available or lack oI inIormation (chromosome) in initial
searching stages.

B. The Idea of the Fusing GA and ACA
To compensate Ior the shortcomings oI GA and ACA, Xiong
et al. |8| Iound the two algorithms generally have the
speed-time curve shown in Fig.2. GA with a higher rate oI
convergence to optimal solution in early search stages (t
0
~ t
a

time), but aIter t
a
time there is a signiIicant reduction in solving
eIIiciency. While, the ACA searches slowly in the initial period
(t
0
~ t
a
time) due to the lack oI inIormation, but the speed oI
convergence is rapidly improved with the pheromone
accumulated to a certain intensity (aIter t
a
time). ThereIore,
integrating two algorithms into one is a useIul attempt Ior
improving the perIormance oI meta-heuristics. The basic idea
oI the dynamic integration oI the GA and ACA is: Adopting
GA to generate initial available solutions beIore the best point
(a), and aIter that, ACA takes over the search process in which
ACA will use the available solutions produced by GA as its
initial pheromone.

Fig. 2. The DiIIerence oI Speed-time Curve oI the GA and ACA|8|

Wu et al. |15| proposed the integrated solving strategy Ior
3616 2008 IEEE Congress on Evolutionary Computation (CEC 2008)
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.

the general optimization problem, and used the travelling
salesman problem (TSP) as a numerical case to demonstrate
their new strategy. However, they set Iixed number oI iterations
Ior genetic algorithm, which causes genetic algorithm too early
(t
d
moment) ending or too late (t
e
moment) ending and can not
guarantee the eIIective integration. ThereIore, Wu et al. |15|
has noted that it is important problem to choice the integration
time between GA and ACA. To remedy this problem, Xiong et
al. |8| proposed a dynamic integration strategy in which the
time taken over search strategy by ACA Irom GA can be
dynamic and thereIore can obtain at an appropriate jointing
time. The method is: 1) Set the minimum Gene
min
(t
b
moment)
and maximum number Gene
max
(t
c
moment) oI iteration Ior the
genetic algorithm. 2) Set the minimum oIIspring evolutionary
rate Gene
min-impro-ratio
aIter taking a statistic. 3) In the range oI
iterations number pre-setted, iI the evolution rate is lower than
the Gene
min-impro-ratio
Ior Gene
die
generations, the eIIiciency oI
the genetic algorithm is low enough to be paused to enter the
ant colony algorithm |8|.

IV. PARTNER SELECTION BASED ON THE FUSED ALGORITHM
IN THE VIRTUAL ENTERPRISES

A. The Model of GA in Partner Selection
We Iirst Iollow Feng et al. |6| proposed using GA to solve
the partnership selection problem. Based on Feng et al. |6|
method, we modiIy the Feng et al. method by importing a
preliminary Iilter in order to enhance the precision oI the group
selection.

B. The Model of ACA in Partner Selection


In practice, each dynamic alliance requires at least one and at
most two candidates to Iorm dynamic partnerships. Wu et al.
|10| have reported using ACA to optimize the partner selection
problem in virtual enterprise. This paper improves the
algorithm proposed by Wu et al. |10| Ior partner selection issue
and modiIied model shows as in Fig.3.

Fig. 3. Improved Model oI Partner Selection

In Fig. 3, the connections between enterprises are taken as
ant paths, and the ant paths only exist in two adjacent
enterprises. II one enterprise has been chosen twice Ior one type
oI core competence, it means only one candidate participating
in the alliance; otherwise, it can be used as two candidates
participating in one alliance.
As noted above, the partner selection is a multi-objective
decision-making problem, while the ACA can only optimize
one objective Iunction. ThereIore, it is necessary to transIorm
multi-objective problem into a single objective
decision-making issue. In this paper, we use the Iitness Iunction
to construct the single-objective Iunction. We Iirst deIine a
positive ideal point ) , , (
+ + +
R T C and a negative
point ) , , (

R T C in three Iunctions space. Then the objective
Iunction (Iitness Iunction) can be constructed as Iollows:
+

+
=
d d
d
t f ) (
(4)

where
) / ) ) ( ( , / ) ) ( ( , / ) ) ( (( ) (
+ + +
= R R t R T T t T C C t C t d
and
, ) / ) ) ( ( , / ) ) ( ( , / ) ) ( (( ) (
+ + + + + + +
= R R t R T T t T C C t C t d

where, means norm, d

is the distance to the negative ideal


point, d

is the distance to the positive ideal point.
In order to get the greatest possible number oI solutions in
each round search, the number oI ants participating in search
should be as much as possible. Ants always start a search Irom
type 1. Considering the practice application, the type oI
enterprise that includes the largest number oI ants should be set
as type 1 and each enterprise will be placed one ant, so ant
number is I
1
.
At moment t, the amount oI inIormation reserves on the path
between enterprises
)
'
(
'
f
i
u
and
)
"
(
"
f
i
u
can be expressed by
) (
)
" '
(
" '
t
f f
i i

,
the degree oI inIormation decay can be expressed by 1.
The ant k takes its path as shown in Fig.3. The Ieasible set oI
the next enterprises the ant will pass can be expressed by
allowed
k
.
The probability Ior the ant k moving Irom
)
'
(
'
f
i
u
to
)
"
(
"
f
i
u
at t
moment is
) (
)
" '
(
" '
t p
f f k
i i
, and

=


, 0
) , ( ) (
) , ( ) (
) (
;
) (
) ( ) ' ' ' (
'
) ' ' (
' '
) ' ' ' (
' ' '
) ' ' ' (
' ' '
otherwise
u tabu f t
u tabu f t
t P k
allowed
r
r
u
r
r k
f f
r i
f
i k
f f
i i
f f k
i i

(5)
where, .
)
" '
(
" ' k
f f
i i
allowed

k
tabu is the current solution Ior ant k,
) , (
)
"
(
"
f
i
k
u tabu f
is the
objective Iunction value Ior the dynamic alliance oI
)
"
(
"
f
i
u
and
enterprises in set
k
tabu . and stand Ior the diIIerent role oI
) (
)
" '
(
" '
t
f f
i i

and
) , (
)
"
(
"
f
i
k
u tabu f
in the ant path chosen, respectively.
AIter n moments, ant k has Iinished one round search by
traveling all types oI enterprises. At the same time, the amount
oI inIormation on the path should be updated. In order to speed
up the convergence rate, this paper takes the amount oI
inIormation updated strategy with Max-Min Ant System |11|,
that is, only to update amount oI inIormation Ior the optimal
solution path in each round. ThereIore, the amount oI
inIormation is
2008 IEEE Congress on Evolutionary Computation (CEC 2008) 3617
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.


) 1 , 0 ( ), ( ) ( ) (
)
" '
(
" '
)
" '
(
" '
+ = +
best
f f
i i
f f
i i
tabu f t n t (6)

Where, ) (
best
tabu f is the objective Iunction value oI the Iittest
solution
best
tabu in current search round.

C. Integration of the GA and ACA
To make use oI merits both oI GA and ACA, we integrate
two algorithms into one algorithm as in |8|. At the early stage
the GA takes the searching process and aIter a time the ACA
will be Iollowed. As we have mentioned above, the interIace
between the two algorithms is an important step that lead to the
Iollowing two issues:
(1) Choose the proper integrated time, that is, what is the
ending condition oI the GA. We use the probability oI
searching new results by GA is less than top 10 optimum that
has been selected beIore new search as the ending conditions.
See step (6) in D. Algorithm Process oI this section.
(2) Set the initial pheromone Ior ant colony algorithm: AIter
the GA ending, N solutions are obtained (N Ior the population
size), calculating the Iitness oI each solution and recorded as
N
f f ~
1
respectively. According to requests by ACA Ior the
partner selection model, initializing the amount oI inIormation
(pheromone) on the path corresponding as
n
f f
i i
f n = ) (
) ' ' ' (
' ' '
, n
means the n
th
path, where | , 1 | N n .

D. Algorithm Process
// Genetic Algorithm with Pre-Iilter
(1) Initialize the control parameters Ior the genetic algorithm,
including population size N, crossover probability P
c
and
mutation probability P
m
;
(2) Set the end condition oI GA, that is,
set
min
Gene ,
max
Gene ,
ratio impro
Gene
min
and
die
Gene ;
(3) Generate the initial population P(0) in line with the
restrictive conditions at random, and set the number oI
generations as g0;
(4) Calculate the value oI C, T and R Ior the individual in
P(0), with comparison to ) , , (

R T C . II any oI the value is
less than the negative ideal value, ignore this individual, and
then Iill the absence with the individual oI the highest Iitness
value;
(5) Calculate individual`s Iitness value in P(0);
(6) Repeat the operations above until meeting the ending
conditions oI genetic algorithm as Iollows;
(a) According to individual Iitness value and roulette choice
strategy set P
i
as the choice probability oI each individual in
P(g);
(b) For (k0; k N; k k 2)
i. Choose two Iathers according to P
i
in P(g);
ii. r random|0, 1|;
iii. II (
m
P r )
Implement mutation operation on the two Iathers chosen,
produced two generations, amend in accordance with the
amendment mechanism and insert them into new
group ) 1 ( + g P ;
Else iI (
c m
P P r + )
Implement crossover operation, amend in accordance with
the amendment mechanism and insert them into new
group ) 1 ( + g P ;
Else
Insert the two unchanged Iathers into new group ) 1 ( + g P ;
}
(c) Calculate the individual Iitness value in P(g), choose top
10 individuals with high Iitness value to Iorm a new group
P(g1), and gg1;
// The InterIace oI GA and ACA
(7) According to the Ioregoing strategy about setting up the
initial pheromone Ior ant colony algorithm, set up the value oI
each edge in the model based on ant colony algorithm
as ) 0 (
) ' ' ' (
' ' '
f f
i i
;
// Improved Ant Colony Algorithm
(8) nc1 ( nc is the round number Iro search); f 0 ( f is the
optimal value oI the objective Iunction); place an ant colony on
all candidate partners, overall

=
=
J
f
f
I m
1
ant colonies, m is the
number oI ant colonies;
(9) Set up the initial solution set and Ieasible set Ior the m ant
colonies. Assumed ant colony k`s starting point is
) ' 2 (
1
u , and
then its initial solution set is } {
) ' 2 (
1
u tabu
k
= , its initial Ieasible
set is } {
) 3 (
3
) 3 (
2
) 3 (
1
, , ,
I k
u u u allowed = ;
(10) Ant colony k moves to the next enterprise according to
probability ) (
) ' ' ' (
' ' '
t P
f f k
i i
, put the enterprise selected into set tabu
k
,
and update the Ieasible set allowed
k
;
(11) AIter n movements, all the ant colonies complete a
round search by traveling overall 2J types. Then
calculate ) (
k k
tabu f f = , | , 1 | m k , record the maximum
k
f
as
max
f
, and the corresponding solution as tabu
k
;
(12) Update the amount oI inIormation Ior all paths
as
) (
)
" '
(
" '
n t
f f
i i
+
, and set ncnc1;
(13) II f f >
max
and
max
nc nc < (
max
nc
is the maximum
number oI iteration), then
max
f f = , return all ant colonies back
to their starting points, go to (2); otherwise, iI f f =
max
and
min
nc nc Ior 3 times consecutively, remove the repeat
enterprise in tabu
k
, and print the whole tabu
best
, determine the
whole process.

V. NUMERICAL EXPERIMENTS
In order to test the Ieasibility oI the integrated algorithm to
choose partners based on the genetic algorithm and the ant
colony algorithm. We use Matlab |19| program to conduct
simulation based on the Iollowing example.

A. The Parameters Description
3618 2008 IEEE Congress on Evolutionary Computation (CEC 2008)
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.

This paper chooses the case in Wu and Su |12| as numerical
experiment data: A middle-scale manuIacturing enterprise got a
large mold injection order. It needed a large scale milling
machine and grinder processing mold rack to conduct the
related mold processing, a Iew high precision CNC machine
tools and cores, and a large scale injection molding machine to
test mold. This enterprise did not have all oI these resources.
ThereIore, virtual enterprise will be organized in order to
complete this task. Core competencies required in this task can
be decomposed as Iollowing:
Mission 1: product 3D modeling and tools design, including
cavity, the core and electrode design, requires experienced
mold designer and related CAD soItware package.
Mission 2: mold rack design, mold manuIacturing process
design, related mold processing, large milling machine and
grinder are required.
Mission 3: cavity and core process design, NC programming
and insert the rough machining, CNC machine tools are
required.
Mission 4: electrode process design, NC programming and
machining, precise NC machine tools are required.
Mission 5: insert the EDM.
Mission 6: other parts design and manuIacturing, die
assembly, experienced mold workers are required.
Mission 7: test mold, large injection molding machine is
required.
Mission 8: mold repair according to the historical test
records, experienced mold workers are required.
The company decided to complete mission 1 by themselves
and other missions by partners. The company themselves also
can be considered as a partner. AIter gross evaluation oI tender
books and selection, mission 2 has Iive candidate partners,
mission 3 has Iive, mission 4 has Iive, mission 5 has Iive,
mission 6 has Iour, mission 7 has Iive and mission 8 has 4.
This injection mold manuIacturing dynamic alliance
comprises eight kinds oI core competencies (missions). The
number oI candidate partners Ior each core competencies is
respectively 1, 5, 5, 5, 5, 4, 5, 4. In practice, each core
competencies requests choosing at least one, up to two
companies to Iorm dynamic alliance. ThereIore, in theory, the
overall possible portIolios oI partner selection are
7
1
2 1
10 6 . 7 ) 1 (
2
1
+ = +
=
J
f
f f J f
I
f
I
I I C C
, which is unIeasible Ior
calculation. Now the integrated algorithm will be used to solve
this issue, with multi-objective Iunctions min C, min T and min
R aIorementioned.

B. Simulation Results and Analvsis
To solve the example in 4.1, using genetic algorithm, ant
colony algorithm and integrated algorithm respectively, we get
results as shown in Fig. 4.


Fig. 4 Comparison oI three algorithms objective Iunction value

From the Fig. 4 we can see that, in the GA, aIter 80 iterations,
the largest Iitness partner portIolio is stable in 0.8219. In the
ACA, the objective Iunction value Ior partner portIolio is stable
in 0.7504 aIter about 40 iterations. While in the integrated
algorithm, the objective Iunction value reaches the maximum
0.8219 aIter 42 iterations.
Detailed analysis and comparison Ior these three algorithms
Ior the optimal solutions are as Iollows.
(1) The Iitness value oI the GA, described by diamond line:
During the early 20 iterations, the Iitness value (optimal
solutions) climbs up Irom 0.38 to 0.8098. It is obvious that the
GA has a great convergence rate with a high eIIiciency in early
stage oI searching. However, on the other hand, aIter 60
iterations, the optimal solution Iitness rises only Irom 0.8098 to
0.8219, with a gentle increase in the diamond line. It can be
concluded that the GA has redundant iterations in later stages
and its searching eIIiciency is relatively low.
(2) The Iitness value oI triangle line, standing Ior the optimal
values oI the ACA: Each oI the solution process is relatively
shorter than the GA and the integrated algorithm. The number
oI iterations is between 20 and 50, which means the ACA can
quickly converge to a local optimum solution. The results given
here is somewhat ideal solutions. It is a Ilaw that the ant
algorithm itselI will Iall into the local optimization due to
parameters improperly set. However, aIter the analysis oI the
parameters used in algorithm, it was Iound that the change in
parameters did not bring about a signiIicant improvement Ior
solving quality.
(3) The square line oI integrated algorithm: The Iormer 34
iterations implement the GA, and the later 8 iterations
implement the ant colony algorithm. The optimal solution
Iitness rises Irom 0.46 to 0.8126 in early 20 iterations with
rapid rising trend. During the 21
st
to 32
nd
iterations, the optimal
solution Iitness rises Irom 0.8126 to 0.8152, with gentle rising
trend. According to the aIorementioned ending condition Ior
the GA, it can be drawn that GA convergence rate has become
too slow and the solving eIIiciency is relatively low. ThereIore,
aIter the 34
th
iteration the ACA starts to search optimal
solutions.
During the Iinal 8 iterations, the optimal values rise Irom
0.8152 to 0.8219. It can be Iigured out that algorithm eIIiciency
is high in this period, but the convergence rate can not be
judged.
By comparing Iitness values oI GA and integrated algorithm,
2008 IEEE Congress on Evolutionary Computation (CEC 2008) 3619
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.

conclusion can be drawn that in the same circumstance, the
integrated algorithm is superior to the genetic algorithm on
solving time and the eIIiciency. In the integrated algorithm, the
ACA converges to the optimal solution by only 8 iterations. As
previous experiments have Iound, the algorithms based on the
ant colony to solve partner selection is more easily converging
into local optimum. However, with the ant colony algorithm in
the integrated algorithm, the integrated algorithm can
eventually converges to a better solution, which is somewhat
unexpected. This may be because in the integrated algorithm,
the genetic algorithm implemented in initial search stages has
produced a relatively good solution that provides the initial
pheromone Ior the ant colony algorithm, which greatly reduces
the blindness search in the initial search state oI the ant colony
algorithm. The good candidates oI initial pheromone make the
ants being able to compare the right search directions, thereby
to be quickly seeking a better solution.

VI. CONCLUSION AND FUTURE RESEARCH
Virtual enterprise partner selection problem is a very
complicated issue due to selection involving a number oI
Iactors. Studies show that using diIIerent methods Ior diIIerent
examples may result in diIIerent conclusions. Although
experiments in this paper have proved that our integrated
algorithm Ior partner selection method is Ieasible and in some
respects is superior to genetic algorithm and ant colony
algorithm, many unresolved issues need to Iurther
investigation. For example, more example data need to be used
to test the Iusing algorithm, and considering more Iactors in
core competence into the search space is another direction.
ThereIore, Iurther research work mainly Iocuses on using
diIIerent virtual enterprise partner selection examples to veriIy
integrated algorithm. On the other hands, how to identiIy the
best integration time Ior starting ant colony algorithm and
ending genetic algorithm is more interested in.

ACKNOWLEDGMENT
The authors would like to thank three anonymous reIerees
Ior their constructive suggestions and comments.
REFERENCES
|1| D. Wang, X.-C. Yang, and G.-R. Wang, 'Implementation oI Partner
Selection in Virtual Enterprise Based on Fuzzy-AHP (in Chinese).
Journal of northeastern universitv, 2002, Vol. 21, No. 6, pp 606~609.
|2| H.-Y. Cao, and D.-W. Wang, 'A Genetic Algorithm Ior A
Multi-Objective Optimization Model Ior Partner Selection in Virtual
Enterprise (in Chinese), Information and Control, 2001, Vol. 30, No. 4,
pp 348~351.
|3| X.-B. Liu, X.-W. Huang and T.-Y. Gao, 'The Method and Realizing
System oI Selecting Enterprise Partner Ior Virtual Enterprises (in
Chinese), Industrial Engineering Journal, 2001, Vol. 4, No. 2, pp10~13.
|4| S. Talluri and R. C. A. Bake, 'Quantitative Iramework Ior designing
eIIicient business process alliance, in International ConIerence on
Engineering Management and Control, 1996, Vancouver: Lonely Planet
Publications Ltd., pp656-660.
|5| E. Qi, Z.-F. Mao and Y.-F. Huo, 'Study on Summarization oI Virtual
Enterprise Development (in Chinese), Manufacture Technologv and
Tool, 2004, Vol. 9,NO. 1, pp37-41
|6| W.-D. Feng, J. Chen, and C.-J. Zhao, 'Partner selection process and
optimization model Ior virtual corporations based on genetic algorithms
(in Chinese), Journal of Tsinghua Universitv, 2000, 40(10): 120-124
|7| J. H. Holland, Adaptation in Natural and Artificial Svstems, Detroit: Ann
Arbor University oI Michigan Press, 1975
|8| Z.-H. Xiong, S.-K. Sikun, and J.-H. Chen, 'Hardware/SoItware
Partitioning Based on Dynamic Combination oI Genetic Algorithm and
Ant Algorithm (in Chinese), Journal of Software, 2005, Vol. 16, No. 4,
pp503-512
|9| T. Jarimo and A. Salo, 'Optimal Partner Selection in Virtual
Organizations with Capacity Risk and Network Interdependencies,
Working-paper, 2007, Available:http://www.sal.hut.Ii/Publications/
pdI-Iiles/mjar07.pdI
|10| Z..-J. Wu and Z. Yao, 'An Improved Model and Algorithm Based on Ant
Colony Algorithm Ior Partner Selection (in Chinese), Management
Science of Chinese, 2007, Vol. 20(supplement), pp52-55
|11| T. Stutzle, and H. H. Hoos, 'Max-Min Ant System, Future Generation
Computer Svstem, 2000, Vol. 16, No. 1, pp889~914.
|12| N.-Q. Wu, and P. Su, 'Selection oI partners in virtual enterprise
paradigm, Robotics and Computer-Integrated Manufacturing, 2005 Vol.
21, pp119-131
|13| J.-L. Ding, Z.-Q. Chen and Z.-Z. Yuan , 'On the Combination oI Genetic
Algorithm and Ant Algorithm(in Chinese), Journal of Computer
Research and Development, 2003, Vol. 40, No.9, pp1351-1356
|14| L. Zhao and W.-W. Ma, 'Review oI Western Virtual Enterprise (in
Chinese),` Chinese and Foreign Entrepreneurs, 2002, Vol. 10, No. 4, pp
62-66
|15| B. Wu and Z.-Z. Shi, 'An Ant Colony Algorithm Based Partition
Algorithm Ior TSP (in Chinese), Chinese Journal of Computers, 2001,
Vol. 24, No. 12, pp1328-1333.
|16| L. M. Camarinha-Matos and H. AIsarmanesh. Infrastructures for Jirtual
Enterprises - Networking Industrial Enterprises, Kluwer Academic
Publishers, 1999.
|17| A. Gunasekaran and E. W. T. Ngai, 'Virtual supply-chain management,
Production Planning & Control, Vol. 15, No. 6, September 2004, 584
595.
|18| M. Dorigo & T. Sttzle, 2004. Ant Colonv Optimi:ation, MIT Press,
USA.
|19| Mathworks Inc. MATLAB 6, 2006.

3620 2008 IEEE Congress on Evolutionary Computation (CEC 2008)
Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on May 14,2010 at 06:23:55 UTC from IEEE Xplore. Restrictions apply.