Study of software reliability prediction based on GR neural network

The failures of safety-critical software may result in the serious loss of property and life, thus software reliability has become very demanding. As an important quantitative approach for estimating and predicting software reliability, software reliability prediction technique is significantly useful for improving and ensuring software quality and testing efficiency. A novel software reliability prediction method based on general regression neural network (GRNN) is proposed, which makes it feasible that without constructing a statistical model like classic software reliability models and having difficulties of solving multivariate likelihood equations, this method can be used to predict software failures. It also incorporates test coverage which has increased prediction accuracy. By using probability plot technique and the least square fitting, the probability distribution functions of the original failure data can be determined. And large amount of data can be simulated to make the reliable prediction, which provides a way for solving the inaccuracy problem caused by small size sample of test failure data. A case study has also been done in a real failure data set. The results show that the proposed method can reflect the relationships among the time, test coverage and number of the faults. And it can improve the prediction accuracy.