Software reliability prediction using neural network

Abstract

Software engineering is incomplete without Software reliability prediction. For characterising any software product quality quantitatively during phase of testing, the most important factor is software reliability assessment. Many analytical models were being proposed over the years for assessing the reliability of a software system and for modeling the growth trends of software reliability with different capabilities of prediction at different testing phases. But it is needed for developing such a single model which can be applicable for a relatively better prediction in all conditions and situations. For this the Neural Network (NN) model approach is introduced. In this thesis report the applicability of the models based on NN for better reliability prediction in a real environment is described and a method of assessment of growth of software reliability using NN model is presented. Mainly two types of NNs are used here. One is feed forward neural network and another is recurrent neural network. For modeling both networks, back propagation learning algorithm is implemented and the related network architecture issues, data representation methods and some unreal assumptions associated with software reliability models are discussed. Different datasets containing software failures are applied to the proposed models. These datasets are obtained from several software projects. Then it is observed that the results obtained indicate a significant improvement in performance by using neural network models over conventional statistical models based on non homogeneous Poisson process.