A systematic approach to software reliability modeling

We propose a systematic approach for distinguishing between software reliability models and parameter estimation procedures using predictive performance. We consider several composite modeling procedures consisting of a variety of elementary modeling procedures based on some NHPP models together with either maximum likelihood or least squares parameter estimation. We also consider a number of selection criteria for choosing an elementary modeling procedure with which to predict the future. We use several predictive performance measures to assess these composite modeling procedures, including a new measure based on predicting failure intensity at the end of the observation interval. We demonstrate these techniques by analyzing a specific software reliability data set