Anda di halaman 1dari 6

2011 Workshops of International Conference on Advanced Information Networking and Applications

Adaptive Genetic Algorithm for QoS-aware Service Selection

Chengwen Zhang
Beijing Key Laboratory of Intelligent Telecommunications Software and Multimedia
Beijing University of Posts & Telecommunications
Beijing, P. R. China
zwjcbj2007@gmail.com

Abstract—An adaptive Genetic Algorithm is presented to select kinds of QoS properties calculation methods. To meet the
optimal web service composite plan from a lot of composite plans global constraints and to find the optimal combination are
on the basis of global Quality-of-Service (QoS) constraints. In this under the scope of combinatorial optimization, and QoS-based
Genetic Algorithm, a population diversity measurement and an web service selection is NP-hard problem [11], therefore,
adaptive crossover strategy are proposed to further improve the approximate algorithm is more suitable to solve optimization
efficiency and convergence of Genetic Algorithm. The probability combinatorial problems. Genetic Algorithm is a kind of
value of the crossover operation can be set according to the approximate algorithm. It is a good method to solve
combination of population diversity and individual fitness. The optimization combinatorial problems [12].
algorithm can get more excellent composite service plan because
Genetic Algorithm (GA) is a powerful tool to solve
it accords with the characteristic of web service selection very
well. Some simulation results on web service selection with global
combinatorial optimizing problems [12]. It is an iterative
QoS constraints have shown that the adaptive Genetic Algorithm procedure on the basis of population where each individual
can gain quickly better composition service plan that satisfies the describes a solution. Basic evolution operations, including
global QoS requirements. crossover, mutation and selection operations, make GA be apt
to very effectively perform global search [13].
Keywords- adaptive; genetic algorithm; qos-aware; web service The design of Genetic Algorithm operators and parameters
selection will have significant impact on itself [13]. In particular,
crossover operator is one of the most important evolution
I. INTRODUCTION operators. It will affect the convergence and efficiency of
Genetic Algorithm. At present, Genetic Algorithm’s efficiency
It is a very important thing how to create service is not enough and its speed of convergence is slow.
composition [1]. A composition service has specific functions At present, the proposed web service selection algorithms
that can be divided into some component functions. These based on Genetic Algorithm solved the service selection
component functions can be accomplished by some component problem to a certain extent. But in order to accelerate the
services respectively. With the rapid development of web convergence speed and improve the global search capability of
service technology, as well as intensified competition among Genetic Algorithm, some methods need to be used.
service providers, there are inevitably many service providers Based on the above analyses, this paper presents a Genetic
to provide services with same functionalities and different QoS. Algorithm. To accelerate the convergence speed of GA and
These services can combine tens of thousands composite improve the global search capability of GA, an adaptive
services with same functions and different QoS. That is, there Genetic Algorithm is introduced.
are many different combination plans. Therefore, in a service The remaining sections of this paper are as follows. Section
composition process, we need to choose service components two described researches of QoS-based web service selection
from massive services with same functions and different QoS computing. The proposed Genetic Algorithm was discussed in
based on user's QoS requirements. The service selection with detail in section three. Section four presented some simulation
global QoS constraints possesses a considerably big proportion works and discussed the simulation results. Section five came
in the problem of QoS-based web service selection. QoS-based to conclusions and noted that the next step in research content.
web service selection plays an important role [2-3] in the
combination of Web services. Researches in this area have
aroused widespread concern in academic circles [4-11, 14-21]. II. QUALITY-BASED WEB SERVICE SELECTION
QoS-based Web service selection problem is one of the hot COMPUTATION
research areas. A lot of international research organizations in Based on all global QoS constraints, to select the best plan
this field carried out relevant research work and have made from a large number of service composition plans is the area of
some research results. But there are still certain deficiencies. combinatorial optimization. To solve such problems, the
The existing QoS-based service selection algorithms include calculation methods based on QoS attributes are divided into
QoS-based semantic technology [4-6, 14] and calculation two categories. One category is exhaustive algorithm. In this
algorithms based on QoS properties [7-11, 15-21]. The QoS- kind of algorithm, all of candidate plans are calculated
based calculation methods can better meet the combination of according to certain rules in order to choose the best plan. The
services with the global QoS constraints. Exhaustive methods methods used in [7, 8, 9, 15] fall into this category. The other is
[7-9, 15] and approximate algorithms [10-11, 17-21] are two approximate algorithm. In this type of algorithm, an ideal

978-0-7695-4338-3/11 $26.00 © 2011 IEEE 273


DOI 10.1109/WAINA.2011.43
composition plan is infinitely close to the best one. At last, a an one-dimensional chromosome encoding method to describe
plan that meets all QoS requirements but is not the best one the combination of services, but the expression of gene loci
will be gained. The methods in [10, 11, 16-21] fall into this was different from [10]. [20-21] proposed a combination
category. service method based on Genetic Algorithm. Through Genetic
QoS properties calculation through the establishment of QoS Algorithm, it could be ensured that the results of services
matrix is a representative calculation method. [7] presented a choice met the restrictive conditions.
run-time services choice method in dynamic service At present, the proposed web service selection method based
composition, which was based on the definition and calculation on Genetic Algorithm solved the service selection problem to a
of QoS attribute matrix. But, based on the definition of QoS certain extent. In order to accelerate the convergence speed of
properties matrices, when the number of the QoS properties of Genetic Algorithm and improve the global search capability of
web service or the number of web services was very large, the Genetic Algorithm, a population diversity measurement and an
matrix will be quite huge and the number of computation adaptive crossover method are adopted in Genetic Algorithm to
would be greater. It could select a single better web service, but achieve fairly good results.
it could not meet the entire QoS requirements. That is, it was
only a local optimization algorithm, rather than a global III. ADAPTIVE GENETIC ALGORITHM
optimization algorithm. In this section, a Genetic Algorithm with a population
In [8, 9], a multi-constrained QoS service selection problem diversity measurement and adaptive crossover probability is
was modeled as an integer programming problem, and the presented.
integer programming expressions were presented for a variety
of QoS constraints. The most optimal solution of the multi- A. Population Diversity
constrained QoS service selection problem was gotten using In Genetic Algorithm, a population needs to maintain a
the integer programming method. Based on QoS properties certain degree of diversity. A spatial distribution variance of
matrix, a local optimization algorithm and a global individual fitness can be used to measure the population
optimization algorithm were proposed. The local optimization diversity.
algorithm could not take into account global QoS constraints, The mth generation population is composed by individuals
and was unable to reach a global optimal solution. The global X m1 , X m2 ,..., X mN whose fitness values are f m1 , f m2 ,..., f mN
algorithm had a large amount of calculation, especially in a
dynamic environment. However it could consider the global respectively. f m is the average fitness of all the above
QoS constraints, the global optimal solution could be reached.
individuals. Its formula is the following formula (1).
But when the size of composition services was large, the
compution of the global computational algorithm had increased
N
1
¦f
a lot. So these two algorithms are limited.
[15] expanded the methods in [9]. It analysed the conditions fm = i
m (1)
of triggering service re-selection in detail, gave the idea of the
N m =1

service re-selection and gave the constraint expression for a


stateful web service selection. f m,max is the largest fitness. Its formula is the following
As a result, we can see the limitations of exhaustive formula (2).
calculations. It needs to figure out all possible solutions in
order to obtain the optimal solution. The web service selection
problem based on QoS belongs to NP-hard problem [11], so the f m,max = max{ f mi , i = 1ˈˈ
2 …ˈN} (2)
exhaustive combinatorial optimization method is poor
scalability and has large calculation. Heuristics method can be
used to obtain an approximate solution. f m,min is the minimum fitness. Its formula is the following
In [16], a multidimensional 0-1 knapsack problem model formula (3).
was used for multiple QoS constraints selection. A method
based on branch-and-branch was proposed for solving MMKP f m,min = min{ f mi , i = 1ˈˈ
2 …ˈN} (3)
(Multi-dimension Multi-choice Knap sack Problem) optimal
solution and heuristic-based method for solving second-best
solution. In the field of combinatorial optimization, there is Dm is the population fitness variance of the mth generation
also a random search algorithm based on probability. Using the
population. Its formula is the following formula (4).
random search algorithm, the optimal solution or sub-optimal
solutions can be gained. Genetic Algorithms fall into this
N
1
¦( f
category. Genetic Algorithm is suitable for solving such
problems [12], and it can effectively prevent exhaustive Dm = i
m − f m )2 (4)
algorithm limitations. The solution based on Genetic Algorithm
N i =1

is a novel global optimization one. [10, 11, 17-21] used Genetic


Algorithm for the optimization of Web service composition. Dmax is the maximum fitness variance from the first-
[10] used Genetic Algorithm to solve the QoS-based service generation population to the mth population. Its formula is the
selection. It used one-dimensional chromosome encoding following formula (5).
method to describe the combination of services. [11] also used

274
D max = max{D j , j = 1ˈˈ
2 …ˈP} (5) Among the formula (7), Pc−old is an old crossover

ȍ is the diversity value of the mth population. Its formula is probability value. fi is the fitness of individual xi in a
the following formula (6).
generation of population, fMax and fMin are the maximum
and minimum value of individuals fitness in the generation of
Ω = D m / Dmax (6)
population respectively, k
c is a constant and can be
The greater the value of ȍ is, the higher the level of determined according to the actual situation, ȍ is the
population diversity is. population diversity.
B. Adaptive Crossover Strategy C. Adaptive Genetic Algorithm(AdGA)
A cross-operation is that a pair of chromosomes exchange In this section, we present a novel genetic algorithm named
some of their genes each other in a certain probability ( Pc , AdGA in order to solve quality-driven selection, including the
called the crossover probability) so as to form new individuals. design of a population diversity and an adaptive crossover
The crossover operation is a characteristic of Genetic strategy. AdGA is available in figure 1.
Algorithm that is different from other evolutionary algorithms.
It plays a key role in Genetic Algorithm and is the main method
to generate new individuals. And it determines the global
search ability of Genetic Algorithm. The greater the value of
Pc is, the more easy it is to find new individuals and the
stronger the algorithm's global search ability is and the more
difficult it is to fall into local minimum. But, it will result in
slower convergence. If Pc is smaller value, good individuals
are not easy to be destroyed and convergence speed becomes
faster. But, it is easy to fall into local minimum.
Simple Genetic Algorithm uses fixed crossover probability,
so it can not achieve good results in the global search ability
and convergence speed at the same time. For an issue, it is
difficult to find a suitable fixed crossover probability to meet
the needs of the various stages of Genetic Algorithms.
The probability value of a crossover operation can be set
according to the combination of population diversity and
individual fitness. When the population diversity declines, the
crossover probability should be increased in order to increase
the capacity to detect new individuals. When the degree of the
population diversity increases, we should reduce the crossover
probability so as to increase the convergence speed of Genetic
Algorithm. When the fitness of an individual is lower, in order Figure 1. AdGA.
to accelerate the convergence speed and improve the global
search capability and the optimal individual search capacity, The adaptive crossover handling is an effective approach to
higher crossover probability is required to quickly generate prevent GA from being trapped into a local optimal solution.
new better individuals. When the fitness of an individual is
higher, in order to reduce the possibility of that the genes of
individual are damaged and create new better entities in the IV. TESTS AND ANALYSES OF ADAPTIVE GENETIC
vicinity of good individuals to improve local search ability, ALGORITHM
lower crossover probability is needed. The proposed service selecting algorithm in this paper
Based on the above analyses, this paper used adaptive improves simple Genetic Algorithm. Through the improvement,
crossover probability based on the fitness as well as the adaptive Genetic Algorithm has better search ability and faster
diversity of population. convergence speed. Here are the tests and test analyses through
which the capacity and efficiency of presented Genetic
fi − fMin 1 Algorithm will be validated.
Pc = Pc−old ×[1− × ] (7) In order to verify the effect of services choice done by the
fMax − fMin 1+ exp(−kcΩ) adaptive Genetic Algorithm, some comparison tests between
simple Genetic Algorithm and the adaptive Genetic Algorithm
were made.
In order to fairly test the two algorithms, they would run in
the same hardware and software operating environment,

275
including CPU, memory, OS, development language and IDE, selection problem. The presented Genetic Algorithm in this
etc. paper accelerates the convergence rate through the use of
In the comparison tests, the two algorithms solved the adaptive crossover strategy. As described in Table 2,
service selection problems with the same size of services comparison of data can fully verify that the adaptive Genetic
combination. In every comparison test, the number of tasks in Algorithm has faster convergence than the simple Genetic
service composition was same. Without affecting the outcome Algorithm and can get the final results of service selection in a
of the comparative experiment, each service had four QoS shorter period of time.
attributes in the experiments. The values of specific QoS
attributes were randomly generated within a certain range. B. Tests and Analyses of Search Capability
Some global limits for a part of QoS properties were Search capability is that the algorithm can find the optimal
randomly generated. The overall limits were applied to all solution in a solution space. It can be measured by the quality
specific service compositions through the penalty function of the solution that the algorithm searches. In Genetic
method. Algorithm, the algorithm search capability can be measured
In addition, the simple Genetic Algorithm and the hybrid through the fitness value of the final selected individual.
algorithm used initialization parameters as following. The Adaptive Genetic Algorithm took an adaptive crossover and
population diversity to improve Genetic Algorithm. In order to
population size is 500. The mutation probability is 0.1. Pc−old verify these strategies, simple Genetic Algorithm and adaptive
is 0.7. Genetic Algorithm were run for 50 times. The average value of
Based on the above preparation of test data, simple Genetic the final fitness values at all running time was taken. A few set
Algorithm and the adaptive Genetic Algorithm were run of test data are listed in Table 3.
respectively. The test results were analyzed from both
convergence speed and search capabilities. TABLE III. AVERAGE VALUE OF MAXIMUM FITNESS

A. Tests and Analyses of Convergence Speed Task average fitness average fitness
number value of value of Adaptive GA
Algorithm convergence rate refers to the number of Simple GA
population generation to reach the final solution. In order to 10 0.138 0.148
validate whether the proposed Genetic Algorithm increased the 25 0.074 0.132
convergence speed, simple Genetic Algorithm and the adaptive 30 0.046 0.119
Genetic Algorithm were run for 50 times respectively. The
average running generation was taken. A few of result data are As shown in Table 3, when in the face of the selection
listed in Table 1. problem with the same size of combination services, the
adaptive Genetic Algorithm can get higher average final fitness
TABLE I. AVERAGE GENERATION WHEN A MAXIMUM FITNESS VALUE IS
OBTAINED value than the simple Genetic Algorithm. When the scale of the
composition problem is small, the advantage of the adaptive
Task Simple GA Adaptive GA Genetic Algorithm is not clear. But, when there are a larger
number
10 463 459
number of tasks in a combined service flow, adaptive Genetic
25 494 338 Algorithm can get much better solutions than the simple
30 447 313 Genetic Algorithm. In the test conditions of this article, when
the number of tasks is more than 10, the adaptive Genetic
Algorithm clearly has stronger search capabilities. This shows
As described in Table 1, comparison of data can fully verify that the adaptive Genetic Algorithm has better search
that the adaptive Genetic Algorithm has faster convergence capabilities, especially, that the search capabilities are more
than the simple Genetic Algorithm. The reason is that the prominent in the larger scale of service selection.
adaptive crossover strategy can get quicker convergence speed
than a fixed crossover method. V. CONCLUSIONS
As a kind of distributed computing model, academia and
TABLE II. COMPARISON OF AVERAGE RUNNING TIME
industry have been greatly concerned about web services in
Task Simple GA Adaptive GA recent years. With the web services technologies have become
number (ms) (ms) more sophisticated, more and more easily used web services
10 287 176
with the stability characteristics are shared on network. But a
25 879 568
30 3125 2665
single atomic web services can provide limited functionalities.
In order to more fully utilize the shared web services, it is
necessary to combine shared web services to form a new
As can be seen from Table 2, when the number of tasks is
combination of web services to provide more powerful service
added, the running time will increase. For the same number of
functions.
tasks, the adaptive GA has less running time than the simple
With the progressive development of web services
GA. This means that a dynamic and adaptive crossover
technology and application, it is inevitable for a function to
probability will help to shorten the time when a maximum
appear a large number of candidate services with the same
fitness value is gotten.
function properties and different non-functional attributes
The efficiency of simple Genetic Algorithm is still
(mainly referring to QoS attributes). It has become an urgent
unsatisfactory, although to a certain extent it solved the service

276
problem that how to fast and flexibly select a high-availability, [8] L. Zeng, B. Benatallah, M. Dumas, J. Kalagnanam, and Q. Z.
high reliability, high performance and the best services to meet Sheng, “Quality driven web services composition,” Proc. 12th
user’s needs from massive candidate services. Namely, it is Int’l Conf. World Wide Web (WWW), ACM Press, Budapest,
QoS-based service selection problem. Hungary, 2003, pp.411-421.
This paper presents a services selection algorithm based on [9] Liang-Zhao Zeng, B. Benatallah, A. H. H. Ngu, M. Dumas, J.
an adaptive GA. Based on the analyses of the diversity of GA, Kalagnanam, and H. Chang, “QoS-aware middleware for web
the simple Genetic Algorithm is improved itself. The services composition,” IEEE Transactions on Software
improvements include that the design of a population diversity Engineering, 30(5), 2004, pp. 311-327.
and an adaptive crossover strategy. In the result, the search [10] Liang-Jie Zhang, Bing Li, Tian Chao, and H. Chang, “On
ability and convergence speed can be improved at the same demand web services-based business process composition,”
time. IEEE International Conference on System, Man, and
Through the realization of the above-mentioned algorithm, Cybernetics (SMC'03), Washington, USA, 2003, pp.4057-4064.
testing and analyses of test results, some strong validations of [11] G. Canfora, M. Di Penta, R. Esposito, and M. L. Villani, “A
the proposed algorithm in capacity and efficiency effects were lightweight approach for QoS–aware service composition,” In
done. Proc. 2nd International Conference on Service Oriented
In the above experiments, the number of individuals in Computing (ICSOC'04), New York, USA, 2004, pp.36-47.
populations is the same in the face of different combination
[12] M. Srinivas, and L. M. Patnaik, “Genetic algorithm: a survey,”
sizes. If the populations with different sizes can be adopted for
IEEE Computer, 27(6), 1994, pp.17-26.
different composition scales, the efficiency of algorithm will be
greatly improved. Therefore, the next study will examine the [13] R. Ignacio, G. Jesús, P. Héctor, J. J. Merelo, P. A. Castillo, and
dynamic adaptive mechanism of population size. An approach G. Romero, “Statistical analysis of the main parameters involved
of Genetic Algorithm with ant colony algorithm is used to in the design of a genetic algorithm,” IEEE Transactions on
select concrete services in [23]. MMAS (Max-Min Ant System) Systems, Man, and Cybernetics—Part C: Applications and
has abilities of parallel processing and global searching. Reviews, 32(1), 2002, pp.31-37.
MMAS can put feedback information to Genetic Algorithm. So, [14] M. Sensoy, and P. Yolum, “Ontology-based service
another future work is about how Genetic Algorithm works representation and selection,” IEEE Transactions on Knowledge
together with MMAS in dynamic web service selection and Data Engineering, 19(8), 2007, pp.1102-1115.
environment. [15] A. Danilo, and P. Barbara, “Adaptive service composition in
flexible processes,” IEEE Transactions on Software Engineering,
ACKNOWLEDGMENT 33(6), 2007, pp.369-384.
The work presented in this paper was supported by the [16] T. Yu, Y. Zhang, and K.-J. Lin, “Efficient algorithms for web
National Natural Science Foundation project of China under services selection with end-to-end QoS constraints,” ACM
Grant No. 60872042. Transactions on the Web, 1(1), 2007, pp.1-26.

REFERENCES [17] C. M. Gao, and M. L. Cai, “Chen Huowang. QoS-aware service


composition based on tree-coded Genetic Algorithm,” 31st
Annual International Computer Software and Applications
[1] F. Curbera, R. Khalaf, N. Mukhi, S. Tai, and S. Weerawarana, Conference (COMPSAC 2007), 2007, vol.1, pp.361-367.
“The next step in web services,” Commnincation of the ACM,
46(10), 2003, pp. 29-34. [18] C. W. Zhang, S. Su, and J. L. Chen, “DiGA: population
diversity handling Genetic Algorithm for QoS-Aware web
[2] D. A. Menascé, “QoS issues in web services,” IEEE Internet services selection,” Computer Communications, Elsevier, 30(5),
Computing. 6(6), 2002, 72-75. 2007, pp.1082-1090.
[3] D. A. Menascé, “Composing web services: a QoS view,” IEEE [19] Y. H. Yan, and Y. Liang, “Using genetic algorithms to navigate
Internet Computing, 8(6), 2004, pp.88-90. partial enumerable problem space for web services
[4] M. Tian, A. Gramm, H. Ritter, and J. Schiller, “Efficient composition,” 3rd International Conference on Natural
selection and monitoring of QoS-aware web services with the Computation (ICNC 2007), China, 2007, pp.475-479.
WS-QoS framework,” IEEE/WIC/ACM International [20] G. Canfora, M. Dipenta, R. Esposito, and M. L. Villani, “A light
Conference on Web Intelligence (WI'04), Beijing, China, 2004, weight approach for QoS-aware service composition,”
pp.152-158. Proceedings of the 2th International Conference on Service
[5] A. Soydan Bilgin, and M. P. Singh, “A DAML-based repository Oriented Computing, New York, 2004, pp.232-239.
for QoS-aware semantic web service selection,” Proceedings of [21] G. Canfora, M. Dipenta, and R. Esposito, “An approach for
the IEEE International Conference on Web Services (ICWS QoS-aware service composition based on genetic algorithms,”
2004), California, USA, 2004, pp.368-375. Proceedings of the 2005 Conference on Genetic and
[6] Chen Zhou, Liang-Tien Chia, and Bu-Sung Lee, “DAML-QoS Evolutionary Computation. Washington, 2005, pp.1069-1075.
ontology for web services,” IEEE International Conference on [22] J. Yen, C. Liao, B. Lee, D. Randolph, “A hybrid approach to
Web Services (ICWS 2004), California, USA, 2004, pp.472-479. modeling metabolic systems using a genetic algorithms and
[7] Y. Liu, A. H. Ngu, and L. Zeng, “QoS computation and simplex method,” IEEE Trans. on Systems, Man, and
policing in dynamic web service selection,” In Proceedings of Cybernetics-part B: Cybernetics, 28(2), 1998, 173-191.
the 13th International Conference on World Wide Web (WWW), [23] Cang-hong Jin, Ming-hui Wu, Tao Jiang, and Jing Ying,
ACM Press, New York, USA, 2004, pp.66-73. “Combine automatic and manual process on web service

277
selection and composition to support QoS,” 12th International Design (CSCWD 2008), Xi’an, China, 2008, pp.459-464.
Conference on Computer Supported Cooperative Work in

278

Anda mungkin juga menyukai