Abstract
This thesis presents the results of an investigation of the application of neural network to software reliability assessment. Recently neural networks have been applied for software reliability because of its characteristic which can perform as human brains. But in general, software reliability analysis requires specifications of parametric distributions and certain assumptions that are difficult to validate at times. In this thesis, we first present the methodology of the neural network on the software reliability growth model. We try to derive the neural network approach into mathematics expressions while most researchers think that neural network is a black-box method. Furthermore we compare the neural network model with the conventional parametric models. We also demonstrate how to apply the neural network method on the traditional parametric models. Furthermore, we show how to make the models efficiently by using neural networks to achieve combinational models. At last, experimental examples using the real software reliability failure data sets are given to evaluate the performances of the proposed model. We compare the performances of models from three aspects: goodness-of-fit, short-term predictions, and long-term predictions. From the numerical results, we can conclude that the neural network model has better performances than the traditional models.