New ways to get accurate reliability measures (software)

Abstract

Two techniques that analyze prediction accuracy and enhance predictive power of a software reliability model are presented. The u-plot technique detects systematic differences between predicted and observed failure behavior, allowing the recalibration of a software reliability model to obtain more accurate predictions. The perpetual likelihood ratio (PLR) technique compares two models' abilities to predict a particular data source so that the one that has been most accurate over a sequence of predictions can be selected. The application of these techniques is illustrated using three sets of real failure data.<>