| Software cost and quality estimation has become an increasingly important field due to the increasingly pervasive role of software in today's world. In spite of the existence of about a dozen software estimation models, the field continues to remain not-too-well-understood, causing growing concerns in the software-engineering community.; In this dissertation, the existing techniques that are used for building software estimation models are discussed with a focus on empirical calibration of the models. It is noted that traditional calibration approaches (especially the popular multiple-regression approach) can have serious difficulties when used on software engineering data that is usually scarce, incomplete, and imprecisely collected. To alleviate these problems, a composite technique for building software models based on a mix of data and expert judgement is discussed. This technique is based on the well understood and widely accepted Bayes' theorem that has been successfully applied in other engineering domains including to some extent in the software-reliability engineering domain. But, the Bayesian approach has not been effectively exploited for building more robust software estimation models that use a variance-balanced mix of project data and expert judgement.; The focus of this dissertation is to show the improvement in accuracy of the cost estimation model (COCOMO II) when the Bayesian approach is employed versus the multiple regression approach. When the Bayesian model calibrated using a dataset of 83 datapoints is validated on a dataset of 161 datapoints (all datapoints are actual completed software projects collected from Commercial, Aerospace, Government and non-profit organizations), it yields a prediction accuracy of PRED(.30) = 66% (i.e. 106 or 66% of the 161 datapoints are estimated within 30% of the actuals). Whereas the pure-regression based model calibrated using 83 datapoints when validated on the same 161 project dataset yields a poorer accuracy of PRED(.30) = 44%.; A quality model extension of the COCOMO II model, namely COQUALMO, is also discussed. The development of COQUALMO from its onset enables one to understand how a comprehensive modeling methodology can be used to build effective software estimation models using the Bayesian framework elaborated in this dissertation. |
