It is important to have a good cost estimate in order to budget a new project. Unfortunately, software effort estimation methods are often inaccurate. Molokken and Jorgensen report that 60-80% of the time a software project will overrun its estimate by an average of 30% [98]. Furthermore, most estimates do not describe the uncertainty of the estimate [52, 54, 118]. In addition, each source of uncertainty, as described by Kitchenham and Linkman [68], has yet to be represented.

In this thesis, the design principles of an effort estimation tool called 2CEE are discussed. This tool is currently being deployed at NASA's Jet Propulsion Laboratory, and feedback regarding the methodology improvements to industry is reported. This tool represents an approach to effort estimation that provides greater interaction of the cost analyst with the estimation model. The approach places an emphasis on representing estimation uncertainty which, among other benefits, allows estimates with increasing confidence throughout the software lifecycle.

Previously, Jorgenson has argued that most effort estimation is done manually [49, 52]. However, manual methods have difficulty sampling the space of uncertainty [54]. This thesis describes how 11 of Jorgensen's 12 expert judgment best practices may be automated in a model, 7 of which are demonstrated in 2CEE. Thus, the distinction between manual and automated methods for effort estimation is questioned. Instead, an alternate ideology is proposed where neither manual nor automatic estimation methods dominate, but rather each augments the other.

In addition, the techniques of feature subset selection, bagging, and boosting are explored for the COCOMO software effort estimation model. These methods are evaluated using nonparametric techniques to compensate for the non-Gaussian error distributions [95]. Improved estimation accuracy is reported. Finally, a startling discovery regarding the stability of the COCOMO software effort estimation model is reported.