3 - Software reliability models

Abstract

This chapter presents different aspects related to software reliability models. The fundamental problem of estimating software reliability from records of failure and running time has also been discussed. A prediction system for doing this is divided into three parts—a model that expresses the distribution of random time to failure as a function of parameters, an inference procedure that derives parameter values from failure history, and a prediction procedure that combines the results of these to yield statements about future times to failure. Three methods of assessing the predictive accuracy of models on a given data set are described—u-plots, j-plots, and prequential likelihood. Criticisms of current models are stated and reasons for these problems are discussed, along with possible ways round them. It is pointed out that many models have as a parameter the “number of faults in the product”; indeed some investigators have regarded the estimation of this mysterious quantity as one of the main purposes of modeling.