As is still largely the case with hardware, much of current software reliability prediction is based upon established models which are largely black-box in nature. The structure of these models reflect supposed scientific observation of the physical processes of software failure, together with the conventional desire for modelling brevity. It is, unfortunately, the case that such models, if automatically applied, may neglect important but unknown structure in the failure data, and may squeeze out of the data much of its information content. For this reason such model validation as may subsequently be used, may not be effective.
In this paper the author introduces the nature of exploratory data analysis (e.d.a.), and discusses its use for software reliability assessment. Emphasis is on searching all the available data, in its greatest generality, for pattern to exploit in its analysis. As well as more basic techniques, the use of time series analysis and proportional hazards modelling in this context will be discussed. Examples will be presented of the types of results obtained from the application of e.d.a. to software reliability data and, in particular, of the discovery of structure ignored by established models.