Automatically Estimating Software Effort and Cost using Computing Intelligence Technique

Sundarapandian et al. (Eds): CoNeCo,WiMo, NLP, CRYPSIS, ICAIT, ICDIP, ITCSE, CS & IT 07,
pp. 381–392, 2012. © CS & IT-CSCP 2012 DOI : 10.5121/csit.2012.2437
Automatically Estimating Software Effort
and Cost using Computing Intelligence
Technique
Jin-Cherng Lin , Han-Yuan Tzeng, and Yueh-Ting Lin
Dept. of Computer Science & Engineering
Tatung University
Taipei 10452, Taiwan
jclin@ttu.edu.tw
Abstract
In the IT industry, precisely estimate the effort of each software project the development cost
and schedule are count for much to the software company. So precisely estimation of man
power seems to be getting more important. In the past time, the IT companies estimate the work
effort of man power by human experts, using statistics method. However, the outcomes are
always unsatisfying the management level. Recently it becomes an interesting topic if computing
intelligence techniques can do better in this field. This research uses some computing
intelligence techniques, such as Pearson product-moment correlation coefficient method and
one-way ANOVA method to select key factors, and K-Means clustering algorithm to do project
clustering, to estimate the software project effort. The experimental result show that using
computing intelligence techniques to estimate the software project effort can get more precise
and more effective estimation than using traditional human experts did.
Keywords
Software Effort Estimation, Project Clustering, Computing Intelligence, Particle Swarm
Optimization, K-Means Clustering.
1. INTRODUCTION
The software industry are fast to grow up, the cost of software will inevitably become one of the
topics of concern. Software cost estimation affects the success of software projects. However it
has many problems. The software cost estimation is with high degree of uncertainty.
Overestimation or underestimation might occur to happened. These problems led to the
development team in budget not enough and manpower shortage, and also caused development
schedule delay and the problem of poor software quality. In practice, the accuracy of software
effort estimation is one of the keys to project success. For example, according to Standish Group's
EXTREME CHAOS Report 2001, of 30,000 applications development projects, 23% of project
failures, 49% of the project was being challenged, only 28% of the project is successful [2]. The
failure of software projects for IT companies will lead to the financial losses and loss reputation

382 Computer Science & Information Technology ( CS & IT )
of company. Therefore, how to accurately assess the effort of the project development is an
important key to the successful projects.
In the past time, the IT companies estimate the work effort of software projects by human experts,
and use some statistics method and fixed parameters to make the result more precisely. Earlier
study for software effort estimation, COCOMO mostly bases on the three modes as the cluster,
namely Organic Mode, Semidetached Mode, and Embedded Mode [1]. However, the outcomes
are always unsatisfying the management level.
In recent years, the computing intelligence techniques have been grown. It becomes an interesting
topic if computing intelligence techniques can do better in this field. Currently, some researches
combined with neural networks [11], genetic algorithms, differential evolution algorithm [15],
and intelligent gray theory [14] apply to software effort estimation and calculation of parameters
optimization. However, their approaches still based on the original COCOMO model‘s law
expert classified the cluster of three modes,
Our research also focuses on applying computing intelligence to software effort estimation.
However, the clustering results will be different from the COCOMO model’s law expert
classified the cluster of three modes, but through the K-Means clustering of the group. After
project clustered, the Particle Swarm Optimization (PSO) algorithms [8] is applied for parameter
optimization. Comparing with genetic algorithms and differential evolution algorithm [15],
Particle Swarm Optimization has no complex mating and mutation. It is natural selection, simple,
and fast convergence. The experimental result shows that clustering algorithms combined with
the optimal algorithm can be applied to software development effort assessment issues and get
more precisely outcome.
2. LITERATURE REVIEW
2.1 COCOMO
COCOMO (Constructive Cost Model) is developed by Prof. Barry W. Boehm, who first
published non-linear model of assessment methods in 1981. COCOMO categorizes the projects
into three different levels of details, namely, Basic model, Intermediate model, and Detail model
[1]. It is mainly based on past experience of software projects to assess the software effort. It is
still widely used in software effort prediction and assessment of manpower.
The COCOMO model accords to the following equation to calculate the software effort/costs:
1 5
1
B
i
i
C A S x
=
= × ∏
(eq. 1) [1]
Where C is estimated software effort in man-month, A and B for the estimated parameters, S for
the software size in thousands of lines of executable code (KLOC), and then multiplied by the
product of 15 project complex factors, xi is adjustment factor value.

Computer Science & Information Technology ( CS & IT ) 383
2.2 Pearson product-moment correlation coefficient [3]
Pearson analysis can calculate the Pearson correlation index, a linear correlation value between the
two indicator variables. Its equation as:
(eq. 2) [3]
r is the correlation coefficient for the XY co-variance divided by the standard deviation of X and Y
standard deviation of the product, and its value range between 1 and -1, the average for the X, for
the Y's average.
2.3 One-Way ANOVA
One-Way ANOVA has also been applied in many fields, for example, Chen et al used it in genetic
engineering [4], Tang applied it to hotel staff job satisfaction analysis [5], Ropponen applied it to
software development risks [6]. In short, One-Way ANOVA compares two or more samples of the
population if they are the same statistical population. ANOVA values generally determine P<0.05
and F>4, where F value compares the number of differences in average volume between groups, P
value of confidence level, P<0.05 represents by 95 percent confidence level.
2.4 K-Means clustering algorithm [7]
K-Means clustering algorithm is a simple and efficient data clustering method. Compared to other
clustering algorithms, it is relatively simple and faster. K-Means clustering of the basic concept
of randomly generated in the initialization N group centers, and then calculate distance of each
group and the cluster center. If distance smaller than other cluster center that classified in to the
group, .
2.5 Particle Swarm Optimization algorithm
Particle swarm optimization is simulated biological behavior of the flock and genetic algorithms
[8, 9, 17, 18]. Comparing with other method of search for optimal solutions, such as differential
evolution algorithm [15], ant algorithms [16], Particle Swarm Optimization method is more
simple and rapid. It is less than genetic algorithm and differential evolution algorithm for mating,
mutation, natural selection and other complex simulation of biological evolution. It can more
efficiently find the optimal solution.

3. MULTIPLE FACTORS CLUSTERING MODEL
Figure 1. Multiple factors clustering model
We use the COCOMO database of 63 historical project data as evaluation data set. By Pearson
correlation analysis and One-Way ANOVA to the 15 factors in the data set for correlation analysis,
we choose N key factors. According to these N-factors, we apply the K-Means clustering
algorithm to clustering all software projects into some groups. For each group, selected a software
project data to be the prediction project, the others in the same group as training projects, using
PSO algorithm to find the optimization parameters. Do it loop to make each project to be the
prediction project. For example, suppose a group have five projects. In the first round, selected the
first project as prediction project, using other four projects as training projects; in the second
round, selected the second project as prediction project, using other four projects as training
projects, and so on. While each time the optimization parameters is found, let each project to
calculate the MMRE as evaluation criteria. This method is namely N-1 test method. The system
architecture is shown in Figure 1.
3.1 Effort adjustment factor and correlation analysis
Instead of using human experts to do the clustering, our approach uses the clustering algorithm to
do it automatically. However, there are 15 effort adjustment factors in the COCOMO data set. It
is too many to calculate for the automatic clustering algorithm. We should reduce it.
Pearson correlation analysis is used to analyze the close degree of the COCOMO 15 effort
adjustment factors to the software effort. We use Pearson correlation analysis by the eq. 2 to
calculate the correlation coefficient of 15 factors, we select according to the slope of the
correlation coefficient is positive (r> 0) and the high degree of relationship factors as K- means
clustering based. One-Way ANOVA analysis methods is also used in more than three population
COCOMO
Pearson
One-Way
ANOVA
K-Means
Project
Clustering
Group 1 Group 2
PSO combined with N-1 test
method to estimates of the
software effort

differences. We set P<0.05 and F>4 as selection standards, and select high F value of factors as
base of project clustering.
3.2 K-Means Software Project Clustering
Selected key adjustment factors, using Pearson analysis and One-way ANOVA methods, is used
as the coordinates of multiple dimensions, In short, the two dimensions of the data used for
clustering, coordinates can be used as two-dimensional X, Y coordinates to do clustering.
3.3 Parameter Optimization using PSO
Using PSO to optimize the parameters of COCOMO, each particle contains two dimensions X
and Y coordinates, X and Y are A and B parameters in the COCOMO model equation in eq. 2,
respectively. Using PSO for optimal characteristics finds the best value of A and B parameters as
the prediction project parameters. First, X and Y coordinates are randomly generated 40 particles
of range between 0 and 1 in a two-dimensional space, and give each particles random initial
speed. Then the X and Y coordinates of particles as a predictor parameters, and use MMRE as
fitness Value. Each particle must be recorded optimal solution that on their through path, the
solution called the local optimal solution (Pbest). Each particle must also have social behavior
that each particle to find the optimal solution in the current search, called a global optimal
solution (Gbest).
)(())(() 21 idgdidididid xpRandcxpRandcVwV −××+−××+×=
(eq. 3) [18]
ididid Vxx +=
(eq. 4) [18]
When the particles are getting Pbest and Gbest, using the equation (4) update each particle's flight
speed. The next particle coordinates for the current location coordinates add updated flight speed
as the new coordinates (eq. 4). In the eq. 3, which Pid is Pbest, Pgd is Gbest, C1, C2 is learning
constant，in general, learning constant can be set to 2, and w is inertia weight, usually set to
decimal that between 0 and 1, Rand() is between 0 and 1 random numbers [8]. This study tested
the weight of inertia to 0.05 as the more appropriate inertia weight. The algorithm steps are as
follows:
Step 1. Randomly generated 40 particles and give initial velocity in
two-dimensional space.
Step 2. Calculate the fitness of each particle.
Step 3. Update Pbest of each particles.
Step 4. Update Gbest of each particles .
Step 5. Update velocity of each particles using eq. 3.
Step 6. Update position of each particles using eq. 4.

Step 7. Repeat Step 2 to Step 6 until the stop condition.
3.4 Performance indicators
It’s usually using Mean of MRE (MMRE) and Prediction level (Pred) as accuracy reference value
In the research of software effort estimation. In this study, using the Pred and MMRE as accuracy
reference value.
• MMRE
•
Software effort estimation in the assessment of evaluation criteria commonly used Mean
Magnitude of Relative Error (MMRE), the formula as equation (6) below.
1
_ _1
_
i N
i
actual effort predicated effort
MMRE
N actual effort
=
=
−
= ∑
(eq. 5) [11]
In this study, MMRE for the PSO algorithm as the effort estimates of fitness value and
evaluation criteria. MMRE value is the smaller that the prediction effort closer the actual
effort. Which actual_effort is actual effort , predicated_effort is prediction effort ， N is
number of project，i is NO. i project [10, 11].
• Pred(x)
Prediction level is used in software project effort assessment. The accuracy of the assessed
value of less than set percentage of the total assessed value accounting. The Pred(x)
equation is as follows:
( ) , # ( , )i
k
PRED x k i MRE x
N
= = ∀ ≤
(eq. 6) [11]
In this study, we will use MMRE as the main evaluation criteria. Pred (x) value is the higher the
better. x is a percentage value, k is the representative of the assessed value of MRE is less than or
equal to x. N is the number of all the projects [11, 12].
4. EXPERIMENTAL RESULTS
4.1 Pearson correlation analyze COCOMO adjustment factor
First, we combined all three of the original COCOMO model project type, including simple
model, Semidetached Mode, embedded mode, total of 63 historical project data. 15 factors and
software effort for Pearson correlation analysis(equation(3)), and generate the correlation
coefficient r. The result of each factors and effort by Pearson correlation analysis, we take the
adjustment factors that correlation coefficient is larger than 0.2 as clustering factors, The results
(Figure 2) shows the database size (DATA) and the effort is positive correlation, followed by
Modern Programming Practices (MODP) next to Required Software Reliability (RELY) and
Computer Turnaround Time (TURN). Pearson analysis phase, we selected four adjustment

factors, followed by use of One-Way ANOVA selected N factors as a basis for K-means
clustering.
4.2 One-Way ANOVA Analyzing COCOMO Parameters
In the One-Way ANOVA analysis phase, we using each adjustment factors and software effort to
analysis, we decided to P value <0.05, F value > 4 as the selection criteria. After analysis, all
Figure 2. Pearson correlation analysis COCOMO factors
of P values are greater than 0.05. By the One-Way ANOVA analysis results (Figure 3), we will
select Analyst capability (ACAP), Applications Experience (AEXP), Programmer Capability
(PCAP), Programming Language Experience (LEXP).
Figure 3. One-Way ANOVA analysis COCOMO factors
The Pearson correlation analysis and One-Way ANOVA analysis of the eight selected adjustment
factor as clustering. The eight of factors that are Database size(DATA) Required Software
Reliability(RELY), Analyst capability (ACAP) , Programmer Capability (PCAP), Applications
Experience (AEXP),Computer Turnaround Time (TURN), Programming Language Experience
(LEXP), Programming Language Experience (MODP) .

4.3 Software Project Clustering
Figure 4. Data distribution after log conversion
The eight of selected key adjustment factor and effort are total of nine variables. We use K-
Means clustering algorithm to do project clustering. However, the value of project effort is much
higher than other eight key adjustment factor value in COCOMO data set. If we use the values of
project effort directly, the outcomes of clustering is easily impacted by this outstanding value. It
results that clustering is very poor. In order to make clustering more desirable, so we take the
effort value to do logarithm conversion. After conversion, indeed, the clustering results are
improved. It is shown in Figure 4.
Finally, each group could be assigned to the data points about 10 projects, this study tested three,
four, five as fixed group of numbers. Divided into five groups led to some group’s data less than
five projects, so, we divided into three groups and four groups as fixed number of group for
software project clustering. Clustering results of detail in Table1 and Table 2 .
Table 1. Clustering of results in three groups
Group Number of Projects
Group1 34
Group2 19
Group3 10

Table 2. Clustering of results in four groups
Group Number of Projects
Group1 24
Group2 8
Group3 13
Group4 18
4.4 Parameter Optimization and Performance indicators
According to the results of clustering software project, each group has been identified by particle
swarm optimization to optimize COCOMO parameters and use N-1 test method to estimate
effort. In this study, particle swarm optimization to 5000 generations as the stop condition, inertia
weight through experimental tests to 0.05 as the most appropriate inertia weight, C1 and C2 are set
to 2. The test results presented in Table 3 and Table 4.
Table 3. Estimation software effort in three groups
Group MMRE Pred(0.25)
Group1 0.2267 0.5588
Group2 0.2358 0.5789
Group3 0.2284 0.5000
Mean 0.2303 0.5459
Table 4. Estimation software effort in four groups
Group MMRE Pred(0.25)
Group1 0.202123804 0.6666
Group2 0.213253173 0.625
Group3 0.289618454 0.461
Group4 0.165956676 0.7222
Mean 0.217738027 0.6187

4.5 Comparative result
Prediction of effort has been much research in COCOMO, Huang and Chiu [11] using fuzzy
neural network on software effort estimation, Koch and Mitlöhner using voting rules on software
effort estimate [12], Ahmed et al using fuzzy logic- based framework prediction software
effort[13]. Azzeh et al using Fuzzy grey relational analysis for effort prediction in multiple
databases [14]. However, most research estimation software effort base on three project type of
COCOMO. It is little related research that using project database of the effort adjustment factor
and software effort to clustering project and estimation software effort. So, this study will be
compare with based on the COCOMO three project type to estimation software effort and using
statistical and intelligent computing on project clustering and estimation software effort. Related
research to compare the data shown in Table 5.
Table 5. Related research data comparison table
Compare Category Mean of MMRE Mean of Pred(0.25)
This research with three group 0.2303 0.5459
This research with four group 0.2177 0.6187
COCOMO 0.26 0.54
ANN 0.37 0.40
FNN 0.22 0.75
FGRA 0.232 0.667
5. CONCLUSIONS
This study aimed to apply computing intelligence techniques to estimate software effort. This
approach is different from the past method in that human experts categorize the software projects
into various groups. We first do the factors correlation analysis, and then cluster the projects
using automatic clustering algorithm. The last is parameter optimization and estimate the
software effort. The experimental data shows that our approach can accurately estimate the
software effort than by human experts.
ACKNOWLEDGEMENTS
This research was supported by the National Science Council, Taiwan, R.O.C. (grant NSC-100-
2221-E-036-026-) and Tatung University, Taiwan, R.O.C. (grant B100-I02-031).

REFERENCES
[1] B.W. Boehm, “Software Engineering Economics” ,Englewood Cliffs ,NJ：Prentice-Hall,1981.
[2] The Standish Group Internaction “EXTREME CHAOS Report 2001” ,2001.
[3] Joseph Lee Rodgers and W. Alan Nicewander. “Thirteen Ways to Look at the Correlation
Coefficient” The American Statistician, Vol. 42, No. 1 (Feb., 1988), pp. 59-66,1988.
[4] Dechang Chen，Dong Hua，Jaques Reifman，Xiuzhen Cheng “Gene Selection for Multi-Class
Prediction of Microarray Data” IEEE Proceedings of the Computational Systems Bioinformatics
(CSB’03), 2003.
[5] Xiuli Tang，Yingkang Gu “ Research on Hotel Staff Job Satisfaction: the Case of Shanghai” IEEE
Information Management and Engineering (ICIME), pp.118-120, April 2010.
[6] Janne Ropponen and Kalle Lyytinen “Components of Software Development Risk : How to Address
Them? A Project Manager Survey” IEEE Transactions On Software Engineering,Vol.26, NO.2 ,2000.
[7] J. B. MacQueen “Some Methods for classification and Analysis of Multivariate Observations,
Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability”, Berkeley,
University of California Press, 1:281-29,1967.
[8] Kennedy, J. and Eberhart, R. C. Particle swarm optimization. Proc. IEEE Int’l. Conf. on Neural
Networks, IV, 1942–1948. Piscataway, NJ: IEEE Service Center ,1995.
[9] Shi, Y. Eberhart, R. “A modified particle swarm optimizer” IEEE Evolutionary Computation
Proceedings 1998 ,1998.
[10] Shan, Y. ,McKay, R.I. ,Lokan, C.J. ,Essam, D.L. “Software Project Effort Estimation Using Genetic
Programming” IEEE Communications, Circuits and Systems and West Sino Expositions vol.2,
pp.1108-1112, 2002.
[11] Sun-Jen Huang , Nan-Hsing Chiu “Applying fuzzy neural network to estimate software development
effort” Springer Science+Business Media, LLC,pp.73-83, 2007.
[12] Stefan Koch , Johann Mitlöhner “Software project effort estimation with voting rules” Decision
Support Systems Volume 46, Issue 4, March 2009, pp. 895-901, 2009.
[13] Moataz A. Ahmed, Moshood Omolade Saliu , Jarallah AlGhamdi “Adaptive fuzzy logic-based
framework for software development effort prediction” Information and Software Technology
Vol.47, Issue 1, 1 January 2005, pp. 31-48, 2005.
[14] Mohammad Azzeh , Daniel Neagu , Peter I. Cowling “Fuzzy grey relational analysis for software
effort estimation” Springer Science + Business Media, LLC 2009, 2009.
[15] Yucheng Kao, Jin-Cherng Lin , Jian-Kuan Wu “A Differential Evolution Approach for Machine Cell
Formation” IEEE Industrial Engineering and Engineering Management, 2008.,pp. 772-775,2008.
[16] Marco Dorigo, Mauro Birattari, Thomas St¨utzle “Ant colony optimization” Computational
Intelligence Magazine, IEEE,pp.28-39,2006.
[17] Yahya Rahmat-Samii “Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) in
Engineering Eelectromagnetics” IEEE Applied Electromagnetics and Communications, 2003.
ICECom 2003, 2003.
[18] Shu Wang ,Ling Chen “A PSO Algorithm Based on Orthogonal Test Design” Natural Computation,
2009. ICNC '09. Fifth International Conference,pp.190-194, 2009.

Authors
Jin-Cherng Lin received the Ph.D. degree in Information Engineering and Computer Science from
National Chiao-Tung University, Taiwan, in 1989. He is currently an associate professor in the
Department of Computer Science and Engineering at Tatung University, Taiwan. His research interests
include software testing and validation, software quality assurance, computer network management, and
computer network security. His email is jclin@ttu.edu.tw.
Han-Yuan Tzeng received his M.S. degree in Information Engineering and Computer Science from
Tatung University, Taiwan, in 2010. He is currently an R&D engineer. His research interests include
computing intelligence techniques, software engineering, and system design.
Yueh-Ting Lin received his M.S. degree in Mechanical and Electro-Mechanical Engineering from The
National Sun Yat-Sen University, Taiwan, in 2011. He is currently an R&D engineer. His research
interests include computing intelligence techniques, intelligent transformation system, and digital signal
process.

Automatically Estimating Software Effort and Cost using Computing Intelligence Technique

More Related Content

What's hot (18)

Similar to Automatically Estimating Software Effort and Cost using Computing Intelligence Technique (20)

More from cscpconf (20)

Recently uploaded (20)

Automatically Estimating Software Effort and Cost using Computing Intelligence Technique