An Artificial Neural-Network Approach to Software Reliability Growth Modeling

Abstract

Abstract Software reliability growth models (SRGM) are statistical interpolation of software failure data by mathematical functions. The functions are used to estimate future failure rates and reliability or the number of residual defects in the software. The SRGM facilitates reliability engineers to decide when to stop testing. Although more than 200 traditional SRGMs have been proposed to estimate failure occurrence times, the research is still continuing to develop more robust models. Inherently the SRGMs are based on assumptions. In order to increase the estimation accuracy of the models we propose the SRGM based on Feed-Forward Neural Network (FFNN) approach. It seems to have significant advantages over the traditional SRGMs. Traditional parameter estimation of SRGMs need estimation ranges of parameter beforehand. The proposed artificial neural network (ANN) model does not have this requirement and hence the parameter estimation gives consistent results without any assumptions. In this paper a new neural network combination model based on the dynamically evaluated weights is proposed in order to improve the goodness of fit of already proposed traditional SRGMs and ANN based combination models. The performance comparison from practical software failure data sets seems to confirm that, the goodness of fit of proposed model is better than that of traditional SRGMs, both independent and ANN based models.