Optimal selection of failure data for predicting failure counts

In the use of software reliability models it is not necessarily the case that all the failure data should be used to estimate model parameters and to predict failures. The reason for this is that old data may not be as representative of the current and future failure process as recent data. Therefore it may be possible to obtain more accurate predictions of future failures by excluding or giving lower weight to the earlier failure counts. While there are techniques that involve using trends in failure data or model predictions for selecting reliability data, one software reliability model that has a built-in method for optimally selecting a subset of the failure data for parameter estimation is the Schneidewind Non-Homogeneous Poisson Process (NHPP) software reliability model. We did not find use of this idea in the other models we surveyed. In order to use the concept of "data aging", there must be criterion for determining the optimal value of the starting failure count interval. Our research has identified the mean square error as a good criterion for selecting the starting interval of the failure data. We apply the criterion to select the optimal starting interval. We show that significantly improved reliability predictions can be obtained by using a subset of the failure data, based on applying the criterion, and using the Space Shuttle On-Board software as an example.