Research On Software Reliabilityprediction Techniques

Along with software system becomes larger in scale and more complicated increasingly, the problem of software reliability becomes more outstanding. So how to secure the reliability of software as well as design and develop reliable software are helpful to allocate the resource correctly to reduce the development cost, shorten the development cycle and improve the quality of software.In the thesis, based on the analysis of current software reliability prediction, a frame of software reliability prediction based on classifiers ensemble is proposed. Then, software metrics attribute selection and classifiers ensemble are studied. The main contributions of the thesis are as follows.1. A frame of software reliability prediction based on classifiers ensemble is proposed. This frame consists of three parts. In the former stage, attribute selection of software metrics is adopted in data preprocessing. In the kernel stage, classifiers optimization and classifiers ensemble are studied. In the back-end, evaluation and comparison of algorithms are studied.2. A method of attribute selection of software metrics based on information gain and adaptive genetic algorithm (IG-AGA) is proposed. First, information is used to filter redundant attribute of software metrics, then the method takes the classification accuracy of KNN and size of attribute subset as the evaluation index and uses adaptive genetic algorithm to search randomly the software metrics attribute which are mostly contributing to the software reliability.The experiment result on the software metrics data set of NASA shows that the selected metrics subset can achieve good classification accuracy.3. A novel method of software reliability prediction based on support vector machine ensemble is proposed. It designs the support vector machine ensemble from two aspects: individual classifier optimization and classifiers ensemble. In the aspect of individual classifier optimization, it applies adapative genetic algorithm(AGA) to choose the best parameters of support vector machines. In the aspect of classifier ensemble, according to the ensembling mind of'many could be better than all', revised fuzzy C-means clustering algorithm based selective ensemble(RFCMSE) is proposed. The prediction result on the selected attribute subset of software metrics shows that it has better prediction accuracy and generalization performance.4. Software reliability prediction prototype system based on support vector machine ensemble is designed and realized from the overall architecture, process and key modules.Finally, the work of this paper is concluded. The further research topics about software reliability technique are outlined, and the research directions in future are discussed.