Research And Application Of Robust Regression Algorithm In Software Cost Estimation

Along with the unceasing expansion and the complex degree daily enlarging of the software system scale, it appears the typical software crisis such as the massive software project progress extension, the budget overspending and the quality flaw from the 20th century 60's last stages. We can obtain the correlated software organization or the software product certain historical data when we carry on the estimation of the software project in the usual situation, and it is helpful to take advantage of these data to forecast the future software projects. The robust regression analysis (RRA) is such one kind quite commonly used and the effective data actuation method. This thesis based on some retrospective studies of RRA, focuses on some problems of the software cost estimation data with ordinary methods and tries to propose the RRA methods to analysis of the software development cost estimation data and effectively solutes to the problem of masking effects when the outliers exist. Firstly, it demonstrated the outlier conception and discussed the OLS applied to the outliers. Secondly, based on robust statistics theory, some robust regression techniques were reviewed and discussed further and discussed in detail the distinguish of the one-dismession and two-dismession outliers, the RRA of the total location parameters and scale parameters, and then put up the algorithmic implementation steps. The procedure could be applied to the cases of a large sample size where the usual algorithms require an unfeasible; It might be incorporated in any robust multivariate calibration method and may improve efficiency. The better results suggest that the procedure is a prospective robust regression method in the outlier research. Finally, further research was made on the basis of the above methods. It proposed the robust problem in the software cost data, and carried on the computer simulation in the Bureau of Civil Administration Funeral Rites ERP system by RRA. From the results, we found that this method could solve effectively the masking effects by outliers and obtained better results.