Analysis of error processes in computer software

Abstract

A non-homogeneous Poisson process is used to model the occurrence of errors detected during functional testing of command and control software. The parameters of the detection process are estimated by using a combination of maximum likelihood and weighted least squares methods. Once parameter estimates are obtained, forecasts can be made of cumulative number of detected errors. Forecasting equations of cumulative corrected errors, errors detected but not corrected, and the time required to detect or correct a specified number of errors, are derived from the detected error function. The various forecasts provide decision aids for managing software testing activities. Naval Tactical Data System software error data are used to evaluate several variations of the forecasting methodology and to test the accuracy of the forecasting equations. Because of changes which take place in the actual detected error process, it was found that recent error observations are more representative of future error occurrences than are early observations. Based on a limited test of the model, acceptable accuracy was obtained when using the preferred forecasting method.