A greedy reliability estimator for usage-based statistical testing

Markov chain usage models have been a basis for statistical testing of software intensive systems for more than two decades. During this time, several reliability estimators have been formulated and used in testing. This paper presents an improvement on the arc-based Bayesian estimator distributed with Version 4.5 of the JUMBL (J Usage Model Builder Library) [1]. The arc-based Bayesian estimator is conservative, and especially so for samples that are small relative to the entropy in the model. We call the new model the “greedy estimator” because it combines the specific information from testing with the inference attributed to the total population. The greedy estimator is shown analytically and experimentally to give more accurate estimates than its predecessor on concrete models, although they converge in the long run. The results of using the greedy estimator are demonstrated for a set of testing data for a tape drive controller.