Discrete software reliability growth model based on maximum entropy principle with higher order polynominial moment contraints

Software managers depend on software reliability growth models (SRGM) during the testing phase in order to gauge when an application will be ready for release. However, many SRGM introduce an unnecessarily large number of parameters to fit the observed failure data more precisely, but in doing so compromise the ability of these models to accurately predict future failures. This paper presents a nonparametric software reliability growth model based on the Maximum Entropy Principle (MEP) with moment constraints, which provides an unbiased method to fit the observed data in a manner that is maximally noncommittal with regard to missing information. The MEP model is applied to a widely studied software failure data set from the historical literature. The results indicate that the MEP SRGM imposing first and second moment constraints achieves greater predictive accuracy than an earlier model based only on the first moment constraint. It is also shown that more complex MEP models utilizing third and fourth moment constraints actually produce worse predictions than the second order model. Thus, the proposed approach can avoid model over fitting by identifying the order of the MEP model that most accurately predicts future failures for a given data set.