Estimation of software project effort using nonlinear regression models

Accurate estimation of software project effort plays an important role in software project management and is one of the most difficult empirical modeling tasks in software engineering. This paper introduces a piecewise linear regression model with a change point [9, 10] which includes the well-known COCOMO model as a special case. Our model is used to estimate software project effort using a NASA software project dataset [2] and a small software project effort dataset from Global Software GroupFlorida (GSG-FL) of a local company. For comparison, we also study three other models: Simple linear regression, COCOMO model [4] and the model based on radial basis functions [14]. Among these four models, we conclude from the empirical results that a piecewise linear model is the best one in modeling the NASA software project effort and GSG-FL dataset. The empirical results show that one has to consider different models due to a structure change in effort for projects from medium then to large size. We also found that the coding methodology is not a significant factor in predicting the effort, after accounting project size and Development Lines (DL), but it is a significant factor, if only accounting project size, not DL.