Engineering topics
Publications and source records attributed to Littlewood, B..
A detailed look is given to software reliability techniques. A conceptual model of the failure process is examined, and some software reliability growth models are discussed. Problems for which no current solutions exist are addressed, emphasizing the very difficult problem of safety-critical systems for which the reliability requirements can be enormously demanding.
Different software reliability models can produce very different answers when called upon to predict future reliability in a reliability growth context. Users need to know which, if any, of the competing predictions are trustworthy. Some techniques are presented which form the basis of a partial solution to this problem. Rather than attempting to decide which model is generally best, the approach adopted here allows a user to decide upon the most appropriate model for each application.
The Jelinski-Moranda (JM) model for software reliability was examined. It is suggested that a major reason for the poor results given by this model is the poor performance of the maximum likelihood method (ML) of parameter estimation. A reparameterization and Bayesian analysis, involving a slight modelling change, are proposed. It is shown that this new Bayesian-Jelinski-Moranda model (BJM) is mathematically quite tractable, and several metrics of interest to practitioners are obtained. The BJM and JM models are compared by using several sets of real software failure data collected and in all cases the BJM model gives superior reliability predictions. A change in the assumption which underlay both models to present the debugging process more accurately is discussed.
Maximum likelihood estimation procedures for the Jelinski-Moranda software reliability model often give misleading answers. A reparameterization and a Bayesian analysis eliminate some of the problems incurred by MLE methods and often give better predictions on sets of real and simulated data. Practical difficulties in estimating the initial number of errors N and the failure rate of each error phi by the method of maximum likelihood are: N, the MLE of N, is occasionally infinite (i.e., the routines for calculating N and phi do not converge). It is shown that N is finite sub i only if the regression line of the interevent times t sub i vs. i has positive slope. A serious problem is that often N approximates n, the sample size, and sometimes N = n. Thus the MLE predicts that the program is perfect even when it is far from being so. Only when almost all failures have been removed can N and phi be trusted near the end of debugging.