You are on page 1of 5

Crow Search Optimization Based Approach for Parameter

Estimation of SRGMs
Ankur Chaudhary1, Arun Prakash Agarwal2, Arvind Rana3, Vijay Kumar4
1,2
Department of Computer Science and Engineering, Amity University, Noida, India.
3
Department of Computer Science and Engineering, Amity School of Engineering and Technology, Delhi, India.
4
Department of Applied Mathematics, Amity Institute of Applied Sciences, Amity University, Noida, India.
1
ankur.tomer@gmail.com; 2arun_pr_agrawal@yahoo.co.in,
3
arvind.rana1998@gmail.com; 4vijay_parashar@yahoo.com

Abstract: Software Reliability Growth Model (SRGM) is a LSE – “Least Square Estimation” and MLE – “Maximum
mathematical relation amidst the diverse attributes of testing Likelihood Estimation” have been proposed by researchers.
as it proceeds, and enables the estimation of the optimal time LSE method has been used widely for parameter estimation of
of release, stop time for testing and numerous findings various engineering problems. LSE uses a form of nonlinear
related to software faults, failures and release policies. regression [6] and works by following the below mentioned
However, parameter estimation of SRGMs is a challenging basic steps:
task and plays an important role in predicting failure
behavior. The parameter estimation for a SRGM needs to be x Initial inspection and pre-processing of sample data.
addressed with utmost care as traditional parameter
x Regression curve estimation.
estimation technique does not reflect consistency in
prediction. This paper presents an efficient approach for x Analysis of confidence interval for sample data.
parameter estimation of SRGMs using Crow Search
Optimization. Furthermore, we have compared the proposed MLE is also widely used parameter estimation approach with
approach with traditional and genetic algorithm based variance and unbiased estimators for a large sample size. It
approaches .The results have been validated on a real data chooses an estimator which is most likely to occur for all
set. possible data samples. MLE is slightly difficult to implement
as the likelihood equations used for MLE are solved
Keywords: Software Reliability; Testing; Parameter numerically [7].
estimation; SRGM; Crow Search Optimization.
The major issue for industry practitioners in applying MLE is
I. INTRODUCTION the complexity of its likelihood equations [8]. Wood et al.
[9]compared the relative performance of MLE and LSE using
Software reliability quantifies the failure-free behavior of an experiment with the assumption that errors of estimation are
software under specified environment for specified time normally distributed, and observed that MLE performs better
duration [1]. Estimation of software reliability during testing than LSE approach when sample size is large. However, both
phase helps managing various challenges in software of these are not the optimal choices for all types of problems
development phase viz. resource pool optimization, budget due to their own limitations such as: difficult implementation,
management, and other scheduling activities. SRGMs help us high biasedness in estimation particularly in case of MLE,
to predict future software failures so that precautionary sensitivity of initialization of model parameter values [5].
measures can be taken well in advance. SRGMs were broadly
classified into two types (i) Failure Rate model (FR) and (ii) Estimation of the unknown model parameters is an NP-hard
Non-Homogeneous Poisson Process model (NHPP). Literature problem and requires a reliable optimization approach. The
revealed that FR based SRGMs have an edge over NHPP software failures are stochastic in nature so the researchers aim
based SRGMs in few cases; however in others the latter were to estimate the optimal parameters for SRGMs so that failures
found to be more meaningful and important [1–4,21-25]. can be better mapped [5]. The aim of this study is to propose
an approach based on Crow Search Optimization for
NHPP models rely on factors such as fault content function, estimation of optimal model parameters so that the model can
testing time, and fault detection rate which express the provide thebest output.The article is designed as: Section II
stochastic behavior of occurrence of software failures. NHPP consists of Literature Review. Section III describes the Crow
based models predict faults at specific point of time [1]. search optimization. The results and discussion is augmented
Prediction ability of any model depends on its parameter in Section IV. Finally, in Section V the conclusion is drawn.
values and hence SRGMs require an optimal parameter
estimation technique. Parameter estimation techniques namely

978-1-5386-9346-9/19/$31.00 ©2019 IEEE


II. LITERATURE REVIEW Step 3: Evaluate the fitness of each crow to achieve the
objective of missing unknown parameter.
Researchers have sought solutions to many complex
engineering optimization problems in various domains like Step 4: Repeat step 5, 6, 7 while itr<MI
software engineering, testing and reliability from several
natural phenomena. Nature has always been a great source of Step 5: Update the new position of each crow using following
inspiration to solve such complex problems resource condition.
scheduling, test effort estimation, test optimization and many
more [5, 10].Minohara et al. [11] utilized for the first time a
Genetic Algorithm(GA) based approach for parameter
estimation of SRGMs inspired from nature and found it more
stable than LSE based approach. Later on Sheta et al. [12] Step 5: check each updated crow position if updated position
investigated the use of particle swarm optimization (PSO) in cross the parameter boundary than persist previous position
solving parameter estimation problem of SRGMs. else updates the newly generated position to the last iteration
crow position.
Zhang et al. [13]further extended the work of Sheta et al. [12]
and proposed a PSO based parameter estimation approach with Step 6: Perform the memory updation for each crows follows:
slow convergence speed to search large solution space.
Aljahdali et al. [14] discussed a multi-objective GA for
parameter estimation of SRGMs. Hsu et al. [15] investigated
the use of Modified Genetic Algorithm (MGA) to estimate the
parameters of SRGMs. AL-Saati et al.[16] investigated the Step 7: Store the best position of crow in current iteration and
application of Cuckoo Search (CS) approach for parameter update itr=itr+1.
estimation.Kim et al. [6]used Real-Valued Genetic Algorithm
(RGA) for estimating parameters of SRGMs and found that
RGAperforms good than other variants of GA and
MGA.Choudhary et al.[5] proposed Gravitational search based
parameter estimation approach for SRGMs. Choudhary et
al.[17] Proposed approaches for parameter estimation of
SRGM using firefly optimization.

III. PROPOSED APPROACH


Crow search optimization (CSO) is best example of “Finding
Hidden Treasure”. We also utilized this intelligent behavior of
Crow to propose an effective parameter estimation approach
for SRGMs. various other intelligent behaviors evident from
the literature motivated us to utilized this metaheuristic[18].It
is evident from literature that CSO has various advantages over
traditional searching approaches [18].Some advantages are:

x More users friendly.


x Easy and effective implementation.
x Avoid premature convergence.
The proposed approach for parameter estimation of SRGM is
depicted in figure 1 and also given as follows:

Step 1: Formulation of fitness function and initialize the CSO


parameters such as problem dimension (pd), flock size, max
iteration (MI), length of flight (f_len) and probability of
awareness (PA).

Step 2: Randomly initialize the crow position and crow


memory for whole population in pd dimension. Here each
crow represents a possible solution set of unknown pd number Fig. 1. Parameter estimation of SRGM using CSO
of parameters of SRGMs.

584
IV. RESULTS AND ANALYSIS TABLE 4: Proposed Algorithm’s Parameters

This section focuses primarily on the experimental setup, Crow Search Optimization Parameter Value
results and their discussion. We designed a robust and stable
experiment, selecting Goel-Okumoto(GO) model [19] and Problem Dimension (No. of decision variables)
2
Delayed S-Shaped [20]model on seven open and closed (PD)
software failure datasets listed in Table1 and Table 2 Population (Flock) Size (N) 100
respectively. The selected datasets as well as the models are
benchmarked and have been widely used for performance Awareness Probability (AP) 0.1
analysis of various parameter estimation approaches. Flight Length (FL) 2
TABLE1: Adopted SRGMs Maximum Number of iterations (i-max) 500

Number of To compare the approaches under study, we performed 30 runs


SRGM Mean Value Function
Parameters of each algorithm on each data set and for each performance
parameter. Thesummarized results are shown in Table 5 and 6
G-O where a > 0, b
2 below.
Model[19] >0
TABLE 5: Comparison of MSE values for proposed approach,
Delayed S- , where
2 GA and GSA [5]
Shaped[20] a > 0, b > 0
Data Set Model GA GSA Proposed
TABLE 2: Data Sets Used
G-O
42.8292 20.68638 19.01404
Model[19]
Total FC2[21]
Total Number of Delayed S-
Data Set Numbers of 16.54767466 7.490235 8.9703
Failures Shaped[20]
Weeks
G-O
FC2-Dataset[21] 54 10 22.8292 11.4848 8.709723
Model[19]
FC6[21]
FC6-DataSet[21] 73 8 Delayed S-
33.54767466 14.03783 14.09724
Shaped[20]
FC40-Dataset[21] 101 50
G-O
217.8292 51.42722 30.04411
DS1-Dataset[22] 100 20 Model[19]
FC40[21]
Delayed S-
DS2-DataSet[22] 176 18 44.54767466 18.85571 17.04411
Shaped[20]
STS-Dataset[23] 231 38 G-O
65.8292 41.64973 25.66361
Model[19]
TS-Dataset[24] 86 22 DS1[22]
Delayed S-
60.54767466 59.26688 59.66361
Shaped[20]
Further, we have selected traditional LSE approach, GA and
GSA based approaches to compare withproposed approach . G-O
1570.8292 367.8824 332.3972
Mean Square Error (MSE) and Theil Statistic (TS) as depicted Model[19]
DS2[22]
in Table 3 below are utilized as performance metrics. Delayed S-
948.54767466 214.7282 202.3972
Shaped[20]
TABLE 3: Performance Parameters
G-O
83.8292 41.19227 23.76615
Model[19]
S. Tohma[24]
Criteria Formula
No Delayed S-
32.54767466 9.607635 13.76596
Shaped[20]
1 Mean Square Error(MSE)
G-O
301.8292 27.8974 21.02108
Model[19]
2 Theil Statistic(TS) Mishra[23]
Delayed S-
2885.54767466 206.5113 201.021074
Shaped[20]
The Crow Search Optimization parameters selected to perform the
experiments listed in Table 4 below:

585
TABLE 7: Comparison of TS values for Proposed approach, GA
and GSA[5]

Data Set Model GA GSA Proposed


G-O
16.94935 12.35498 11.87616
Model[19]
FC2[21]
Delayed S-
14.81846 7.421363 7.7765
Shaped[20]
G-O
8.94935 6.332403 5.574428
Model[19]
FC6[21]
Delayed S-
9.81846 7.071327 7.084428
Shaped[20]
G-O
20.94935 10.34132 8.018943
Model[19]
FC40[21]
Delayed S-
9.81846 6.316877 8.018942
Shaped[20]
G-O
10.94935 8.452943 6.690614
Fig. 2. Comparative analysis of MSE values for Proposed Model[19]
approach, GA and GSA [5] DS1[22]
Delayed S-
9.81846 10.16031 10.46031
Shaped[20]
Table 5 shows the average of MSE values collected after 30
runs of each approach under study on seven datasets. The G-O
33.94935 15.88208 15.21154
Model[19]
results of MSE values as presented in the table and figure 2 DS2[22]
above reflect that the proposed approach performs well against Delayed S-
25.81846 12.20088 12.19154
GA and GSA based approach on all the datasets as Shaped[20]
summarized in Table 6. G-O
14.94935 10.40315 8.04195
Model[19]
Tohma[24]
TABLE 6: Summarized results of Proposed approach, GA and Delayed S-
GSA [5] 9.81846 5.079425 8.04192
Shaped[20]
G-O
Data Set Model Performed Well 38.94935 10.01629 8.316779
Model[19]
Mishra[23]
G-O Model[19] Proposed Approach Delayed S-
FC2[21] 34.81846 18.374149 18.31677
Shaped[20]
Delayed S-Shaped[20] GSA
G-O Model[19] Proposed Approach
FC6[21]
Delayed S-Shaped[20] Proposed Approach
G-O Model[19] Proposed Approach
FC40[21]
Delayed S-Shaped[20] Proposed Approach
G-O Model[19] Proposed Approach
DS1[22]
Delayed S-Shaped[20] Proposed Approach
G-O Model[19] Proposed Approach
DS2[22]
Delayed S-Shaped[20] Proposed Approach
G-O Model[19] Proposed Approach
Tohma[24]
Delayed S-Shaped[20] GSA
G-O Model[19] Proposed Approach
Mishra[23]
Delayed S-Shaped[20] Proposed Approach

Table7 shows the TS values on GA, GSA and proposed


approach. The figure 3 and table 7 reflect the superiority of Fig. 3. Comparative analysis of TS values for Proposed approach,
proposed approach over other implemented approaches. GA and GSA [5]

586
V. CONCLUSION Genetic Algorithms. Int Symp Softw Reliab Eng 324–329 . doi:
10.1109/ISSRE.1995.497673
Prediction of software reliability requires estimating the [12] Sheta A (2006) Reliability growth modeling for software fault
parameters of the software reliability growth models. detection using particle swarm optimization. Evol Comput 2006
Parameters of SRGMs generally follow non-linearity of failure CEC 2006 IEEE 3071–3078
data. This in itself is a complex problem and hence requires [13] Ma J, Ting TO, Man KL, et al (2013) Parameter estimation of
utilizing metaheuristic approaches. In this paper, we proposed photovoltaic models via cuckoo search. J Appl Math 2013:10–
an effective crow search optimization based parameter 12 . doi: 10.1155/2013/362619
estimation approach and compared it with three other [14] Aljahdali SH, El-telbany ME (2008) Genetic Algorithms for
approaches based on GA and GSA. Results obtained Optimizing Ensemble of Models in Software Reliability
wereencouraging and hencemotivate us to continue this study Prediction. Learning 8:5–13
with some more datasets and performance parameters to [15] Hsu CJ, Huang CY, Chen TY (2008) A Modified Genetic
Algorithm for Parameter Estimation of Software Reliability
confirm the validity of results.
Growth Models. In: 2008 19th International Symposium on
Software Reliability Engineering (ISSRE). pp 281–282
REFERENCES
[16] AL-Saati D, Akram N, Abd-AlKareem M (2013) The Use of
[1] Kapoor P, Pham H, Gupta A, Jha P (2011) Software reliability Cuckoo Search in Estimating the Parameters of Software
assessment with OR applications Reliability Growth Models. arXiv Prepr arXiv13076023 11:
[2] Williams DRP (2007) Study of the warranty cost model for [17] Choudhary A, Baghel AS, Sangwan OP (2018) Data
software reliability with an imperfect debugging phenomenon. Engineering and Intelligent Computing. 542: . doi:
Turkish J Electr Eng Comput Sci 15:369–381 10.1007/978-981-10-3223-3
[3] Bai CGG, Hu QPP, Xie M, Ng SHH (2005) Software failure [18] Askarzadeh A (2016) A novel metaheuristic method for solving
prediction based on a Markov Bayesian network model. J Syst constrained engineering optimization problems: crow search
Softw 74:275–282 . doi: 10.1016/j.jss.2004.02.028 algorithm. Comput Struct
[4] Gokhale SS, Trivedi KS (2006) Analytical models for [19] Goel AL, Okumoto K (1979) Time-Dependent Error-Detection
architecture-based software reliability prediction: A unification Rate Model for Software Reliability and Other Performance
framework. IEEE Trans Reliab 55:578–590 . doi: Measures. IEEE Trans Reliab R-28:206–211 . doi:
10.1109/TR.2006.884587 10.1109/TR.1979.5220566
[5] Choudhary A, Baghel AS, Sangwan OP (2017) An efficient [20] Yamada S, Ohba M, Osaki S (1984) s-Shaped Software
parameter estimation of software reliability growth models Reliability Growth Models and Their Applications. IEEE Trans
using gravitational search algorithm. Int J Syst Assur Eng Reliab R-33:289–292 . doi: 10.1109/TR.1984.5221826.
Manag 8:79–88 . doi: 10.1007/s13198-016-0541-0 [21] Kapur, P.K., Chanda, U. and Kumar, V., 2010. Dynamic
[6] Kim T, Lee K, Baik J (2015) An effective approach to allocation of testing effort when testing and debugging are done
estimating the parameters of software reliability growth models concurrently communication in dependability and quality
using a real-valued genetic algorithm. J Syst Softw 102:134– management. An International Journal Serbia, 13(3), pp.14-28.
144 . doi: 10.1016/j.jss.2015.01.001 [22] Kapur, P.K., Pham, H., Chanda, U. and Kumar, V., 2013.
[7] Hsu C-JJ, Huang C-YY, Chen T-YY (2008) A Modified Optimal allocation of testing effort during testing and
Genetic Algorithm for Parameter Estimation of Software debugging phases: a control theoretic approach. International
Reliability Growth Models. In: 2008 19th International Journal of Systems Science, 44(9), pp.1639-1650.
Symposium on Software Reliability Engineering (ISSRE). pp [23] Kumar, V., Khatri, S., Dua, H., Sharma, M., & Mathur, P.,
281–282 2014. An assessment of testing cost with effort dependent FDP
[8] Genschel U, Meeker W (2010) A comparison of maximum and FCP under learning effect: A Genetic Algorithm Approach;
likelihood and median-rank regression for Weibull estimation. InternationalJournal of Reliability, Quality and Safety
Qual Eng Engineering, 21(6), 145002.
[9] Wood A (1996) Predicting software reliability. Computer (Long [24] Kumar, V. and Sahni, R., 2016. An effort allocation model
Beach Calif) 29:69–77 . doi: 10.1109/2.544240 considering different budgetary constraint on fault detection
[10] Arora D, Baghel A (2015) Application of genetic algorithm and process and fault correction process. Decision Science Letters,
particle swarm optimization in software testing. IOSR J Comput 5(1), pp.143-156.
Eng [25] Kumar, V., Mathur, P., Sahni, R. and Anand, M., 2016. Two-
[11] Minohara T, Tohma Y (1995) Parameter Estimation of Hyper- dimensional multi-release software reliability modeling for fault
Geometric Distribution Software Reliability Growth Model by detection and fault correction processes. International Journal
of Reliability, Quality and Safety Engineering, 23(03),
p.1640002.

587

You might also like