Abstract
Over the last 35 years many statistical models have been proposed for the quantitative evaluation of software reliability. All the existing reliability models show a wide variability in predictive validity across data sets. In this thesis, we discuss several Non-Homogenous Poisson Process (NHPP) models comprehensively. A theoretical review of these models is provided. For the two types of data, count and time data, the likelihood equations are obtained. We then generalize a very popular existing Software Reliability Growth Model (SRGM), the Littlewood model. The mathematical expressions of some important software reliability measures for the resulting general model and also the general Weibull model are derived. This theoretical analysis enabled us to develop easily configurable software tools to perform empirical studies of a number of SRGMs. For these we used two of published data sets. We used three techniques to analyze the predictive validity of the several special cases of the above two general models. The results of these evaluations emphasize the problem of the inexistence of one standard model that can be used accurately for all applications. Two models refinement approaches, recalibration and model combination, were then explored. For these, we used two published data sets, and also included two additional data sets from recent large-scale development projects in the consumer electronics industry. Evaluations of predictive validity showed that when used individually; neither approach was universally effective across our data sets. However, applying recalibration, then model combination did provide significant improvements in predictive validity. Finally, several prediction problems associated with using this conventional type of modeling and some corresponding solutions are summarized.