摘要：Predicting software faults is one of the most challenging issues in software engineering but has not been reachedyet to obtain satisfactory results. Among various methods ofsoftware fault prediction, non-homogeneous Poisson process(NHPP) based software reliability models (SRMs) and artificialneural network (ANN) are used vastly. But, the suitable modelselection with parameter estimation of SRMs and appropriatearchitecture selection of ANN complicate the task of softwarefault prediction. The purpose of this paper is to predict thelong-term software faults from the software fault count datausing a refined artificial neural network approach (RANN). Inorder to pre-process software fault count data, we have used fivedata transformation methods to transform Poisson count data toGaussian data. The long-term behavior of the software faults isthen predicted by means of point and interval predictions. Thepoint prediction of RANN is compared with the conventionalSRMs for the case of our newly made synthetic data and eightreal data in terms of predictive performance (average relativeerror). To ensure software reliability, we have constructedprediction intervals (PIs) of the predicted fault points usingour proposed simulation based method (PI simulation) andcompared them with those of existing delta method (PI delta)in terms of coverage rate and mean prediction interval width.The PI simulation method covers both the real and predictedfault point within narrower interval width than the PI deltamethod. Thus, the RANN can afford a cost effective predictiondevice from the viewpoint of predictability in the early phaseof software testing.
