Analysis of Errors in Software Reliability Prediction Systems and Application of Model Uncertainty Theory to Provide Better Predictions

Models are the medium by which we reflect and express our understanding of some aspect of reality, a particular unknown of interest. As it is virtually impossible to grasp any situation in its entire complexity, models are representations of reality that are always partial resulting in a state of uncertainty or error. However the question of model error from a pragmatic point of view is not one of accounting for the difference between models and reality at a fundamental level, as such difference always exists. Rather the question is whether the prediction or performance of the model is correct at some practically acceptable level, within the model's domain of application.Here lays the importance of assessing the impact of uncertainties about predictions of a model, modeling the error and trying to reduce the uncertainties associated as much as possible to provide better estimations.While the methods for assessing the impact of errors on the performance of a model and error modeling are well established in various scientific and engineering disciplines, to the best of our knowledge no substantial work has been done in the field of Software Reliability Modeling despite the fact that the inadequacy of the present state and techniques of software reliability estimation has been recognized by industry and government agencies. In summary, even though hundreds of software reliability models have been developed, the software reliability discipline is still struggling to establish a software reliability prediction framework.This work intends to improve the performance of software reliability models through error modeling. It analyzes the errors associated with a set of five software Reliability Prediction Systems (RePSs) and attempts to improve their prediction accuracy using a model uncertainty framework. In the process, this work also statistically validates the performances of the RePSs. It also provides a time and cost effective alternative to performing experiments that are required to assess the error form which is integral to the process of application of the model uncertainty framework.