The CSR approach to software reliability prediction

A plethora of software reliability models have been developed over the years but, in spite of extravagant claims for their efficacy, none can be trusted to give accurate results in all circumstances. This presents potential users with a serious difficulty, and has probably been the main reason why there has been such a poor take-up of these techniques. Recent work in CSR has largely overcome this problem, and it is now possible in most cases to obtain reliability measures and predictions that can be trusted. This position arises not because of the creation of better models, but of techniques for analysing predictive accuracy. Essentially, it is now possible to apply many models to a particular source of software failure data, and decide which (if any) is giving results of acceptable accuracy. The techniques which have brought about this new perspective on software reliability modelling also bring with them a bonus: they enable even more accurate predictions to be made from the available models by a process of recalibration. In this paper we show how these techniques work when applied to some real data sets.