The Study Of Software Reliability Early Prediction Based On Support Vector Machine

With the rapid development of computer technology, we depend more and more on computer systems. Software is the primary factor among all that result in computer system errors. So how to secure the quality of software as well as designing and developing reliable software are our urgent task. As the early prediction of software reliability enables the developers and testers to get general ideas about the software reliability before testing it, it is important for further development, testing and quality control of software.Support Vector Machine (SVM) is a new learning machine based on VC (Vapnik-Chervonenkis) dimension and SRM (Structural Risk Minimization) principal of Statistic Learning Theory (STL). It finds the best balance between model complexity and learning capacity with limited samples, so that it has the best generalization ability. Compared with traditional model building methods, SVM has its obvious advantage. It can resolve the problems remaining in Artificial Neural Network (ANN), such as difficult to confirm the structure, fit-less or over-fit, easy to get into part minimization and so on. It also has excellent performance in small samples condition.This paper summarizes the status of software reliability early prediction methods in both home and aborad, puts emphasis on two classic modeling methods: multivariate linear regression and ANN, and points out their each faults. Thereout, we introduce SVM theory into the field of software reliability early prediction, and advance the software reliability early prediction model based on SVM with the help of many mathematic tools such as clustering analysis, Primary Components Analyse (PCA), discriminant of strange samples, data scaleing, cross validation, non-enlighten Depth-First Search (DFS) and so on. Simulation shows that compared with classic models, the new model has a better prediction precision, a better generalization ability and a lower dependence on the number of samples. At last, we design a model system based on the new model to help the automation of software reliability early prediction.