The Study On Software Reliability Model Based On Neural Network

The study on software reliability model has become one of the core and hot topics of software reliability engineering field. Until now, more than hundreds of software reliability models have been proposed. However, the application range of models is usually limited in one certain sub-space, which leads to lower generalization. Nevertheless, neural network have excellent ability of nonlinear calculation, which causes the research of software reliability model based on neural network becoming more and more important. The most urgent problem is to establish a model with higher prediction precision and generalization using neural network technology. In this paper, we do some research to resolve this problem, which is as follows.The author proposes a new cascade software reliability model, which combine feedforward neural network model and several classical software reliability models to establish a new cascade software reliability model. This model uses the output of four classical software reliability models as the input of feedfoward neural network, which is confirmed well than individual classical software reliability model through the results of experiments.The author proposes a new software reliability model, which we called the mixture of expert software reliability model, using the central thought of the mixture of expert model in neural network field. This model uses some classical software reliability models for the low level experts to establish a top model to control the output of expert model. We believe that this model combine the advantages of some classical software reliability models and have higher prediction precision, which is confirmed through the experiments using the real failure data.The author proposes a hierarchical mixture of software reliability model through the improvement of the mixture of expert software reliability model. This model increases in number of the low level experts and the layer of the gating network and uses Expectation Maximization algorithm to verify the convergence of training. We hope that this new hierarchical mixture of software reliability model can has better performance than individual low level expert models and can solve the problem of choosing expert model automatically. The results of experiments confirm that the best low level expert model can be chosen automatically by the hierarchical mixture of software reliability model based on the real data, which make the hierarchical mixture of software reliability model become more general than other classical models and have higher prediction precision.