Applying Machine Learners to GUI Specifications in Formulating Early Life Cycle Project Estimations

Producing accurate and reliable early life cycle project estimates remains an open issue in the software engineering discipline. One reason for the difficulty is the perceived lack of detailed information early in the software life cycle. Most early life cycle estimation models (e.g. COCOMO II, Function Point Analysis) use either the requirements document or a size estimate as the foundation in formulating polynomial equation models. This paper explores an alternative approach using machine learners, in particular neural networks, for creating a predictive effort estimation model. GUI-specifications captured early in the software life cycle serve as the basis for constructing these machine learners. This paper conducts a set of machine learning experiments with software cost estimation empirical data gathered from a “real world” eCommerce organization. The alternative approach is assessed at the program unit level, project subsystem level, and project level. Project level models produce 83 percent average accuracy, pred (25), for the client-side subsystems.