On Using Chained Neural Networks For Software Reliability Prediction

To predict the reliability of a software system based on inter-failure time series, neural networks can be used to learn about the software behaviour. The paper considers the DLS (discounted least square) principle for constructing a learning algorithm. Various circular back-propagation networks are experimented with a large number of inter-failure time series available for some projects. In order to deal also with multiple steps time series prediction, the networks can be chained in different ways. The experiments present the performances obtained with different chained architectures, including aspects concerning the convergence speed.