Noninformative/Informative Bayesian Measures of Probability to Compare Predictive Accuracy of Estimation Methods in Software Reliability

Abstract


ARE: The Absolute Relative Error and SRE: The Squared Relative Error, as measurements of predictive accuracy of predictions of number of software failures for a target mission time for alternative reliability models is currently an important concern to all (Sahinoglu, 1992, 1997) The proposed way of quantifying by calculating the Bayesian probability of how much one method's prediction accuracy is better (less ARE and/or SRE) than the other alternative is far more realistic than a) deciding by comparing the mere expected values of ARE and SRE as conventionally done b)deciding by conducting statistical hypotheses tests of pairwise means, an approach more sensible than (a). We have first used noninformative and then informative priors by placing half-normal priors on the ARE(and SRE) r.v. as they are peaked around the ideal zero percent relative error. These details are discussed in Berger and Deely(1988) for the general problem of ranking normal means.