Generalized linear models in software reliability: parametric and semi-parametric approaches

The penalized likelihood method is used for a new semi-parametric software reliability model. This new model is a nonparametric generalization of all parametric models where the failure intensity function depends only on the number of observed failures, viz. number-of-failures models (NF). Experimental results show that the semi-parametric model generally fits better and has better 1-step predictive quality than parametric NF. Using generalized linear models, this paper presents new parametric models (polynomial models) that have performances (deviance and predictive-qualities) approaching those of the semi-parametric model. Graphical and statistical techniques are used to choose the appropriate polynomial model for each data-set. The polynomial models are a very good compromise between the nonvalidity of the simple assumptions of classical NF, and the complexity of use and interpretation of the semi-parametric model. The latter represents a reference model that we approach by choosing adequate link and regression functions for the polynomial models