Software reliability modeling based on SVM and virtual sample

Abstract

Software reliability prediction models, which receive the most attention in software reliability engineering, use the failure data collected in testing phases to predict the failure occurrence in the operational environment. Currently, as the requirement for reliable software is increasing, ways to predict and estimate the reliability of software systems which require high reliability with small size test data is more problematic. What's more, there exists a difficult problem in software reliability modeling that the prediction capability of a model varies with failure data change. This inconsistency problem limits the promotion and application of software reliability techniques; mainly because the assumptions the models are based on may not be suitable for most cases. In addition, there exists a contradiction in traditional software reliability prediction methods. Prediction accuracy is low due to a lack of comprehensive consideration of factors affecting reliability. However, when more factors are taken into account, it is difficult to establish a statistical model and solve multivariate likelihood equations. For these reasons, software reliability prediction modeling method based on machine leaning techniques for small sample size is studied in this paper. Firstly, gene expression programming algorithm is used to analyze the small size sample by symbolic regression. The symbolic regression function acquired can then be viewed as the priori information of the data and used to generate a virtual sample. Then, with the virtual sample a regression model based on a Support Vector Machine (SVM) can be established, with which the software reliability can be predicted. Finally, a case study based on real failure data-sets is presented verifying the effectiveness of the method.