دانلود رایگان مقاله انگلیسی پیش بینی قابلیت اطمینان نرم افزار با استفاده از تکنیک های یادگیری ماشین - اشپرینگر 2018

Abstract

Software Reliability is indispensable part of software quality and is one amongst the most inevitable aspect for evaluating quality of a software product. Software industry endures various challenges in developing highly reliable software. Application of machine learning (ML) techniques for software reliability prediction has shown meticulous and remarkable results. In this paper, we propose the use of ML techniques for software reliability prediction and evaluate them based on selected performance criteria. We have applied ML techniques including adaptive neuro fuzzy inference system (ANFIS), feed forward back propagation neural network, general regression neural network, support vector machines, multilayer perceptron, Bagging, cascading forward back propagation neural network, instance based learning, linear regression, M5P, reduced error pruning tree, M5Rules to predict the software reliability on various datasets being chosen from industrial software. Based on the experiments conducted, it was observed that ANFIS yields better results in all the cases and thus can be used for predicting software reliability since it predicts the reliability more accurately and precisely as compared to all other above mentioned techniques. In this study, we also made comparative analysis between cumulative failure data and inter failure time’s data and found that cumulative failure data gives better and more promising results as compared to inter failure time’s data.

8 Conclusions

Software reliability is one amongst the important facet of the software quality. Presence of faults/failures makes the software unreliable. Software Reliability is dynamic and stochastic in nature so we may say that reliability is a probabilistic measure that assumes that the occurrence of failure of software is a random phenomenon (Quyoum et al. 2010). In this paper, we have applied machine learning techniques namely ANFIS, FFBPNN, GRNN, SVM, MLP, Bagging, CFBPNN, IBK, Lin Reg, M5P, RepTree, M5Rules for predicting software reliability based on past failures of software products. The performance of above mentioned Machine Learning techniques have been evaluated using five different types of data sets being extracted from industrial data to predict the failure intensity of the software’s in use. We have empirically proved that ANFIS outperformed the model predicted using other mentioned ML techniques for all the datasets being taken into consideration in predicting reliability in terms of the various statistical efficacy measures applied. For each of the five datasets taken into consideration, results show that the correlation coefficient is above 0.99 in most of the predictions for ANFIS which signifies that the actual and the predicted values are very close. Also we found that the MARE is between the ranges of 0.025–1.5 in most of the predictions for ANFIS and is quite lowest in comparison to the MARE’s is being calculated by other mentioned techniques. The results also depicts that the MSE ranges between 0.5 and 16.0 in most of the predictions for ANFIS and MRE is quite lowest in comparison to the other mentioned techniques. Apart from this, GRNN shows very appreciating and encouraging results and can also be used for reliability prediction. We may also infer that GRNN and MLP follow ANFIS in predicting reliability. Also we observed that cumulative data produces always better results as compared to inter failure time’s data. The result shows that the cumulative failures yield high correlation coefficients within the ranges of 0.74–0.99 in comparison to the correlation coefficient’s being calculated for interfailure time’s data which signifies that the actual and the predicted values are very close. The results also depicts that the cumulative failures yields low MARE within the ranges of 0.04–1.38 in comparison to the MARE’s being calculated for interfailure time’s data. This is the reason that cumulative failure data is always chosen for failure prediction experiments.