Empirical evaluation of an entropy‐based approach to estimation variation of software development effort

摘要

As effort estimation has gained increasing attention, most of the techniques proposed have focused on the accuracy of effort estimates. Yet no clear conclusions have been drawn on which techniques perform best in all contexts. We propose an entropy-based approach to effort estimate variation caused by measurement and model error sources whatever the effort estimation technique used. The proposed approach was empirically evaluated by exploring three entropy formulae, four interpolation methods, and two analogy-based effort estimation approaches (crisp and fuzzy analogy) over seven datasets using the Jackknife evaluation method. The obtained results show that the three entropy formulae have in general the same positive influence on the performance of the entropy-based approach measured in terms of absolute error of effort deviation. In addition, the spline interpolation outperformed all other interpolation methods, using any of the entropy formulae. Moreover, achievement percentages of the best variants of our approach closely approximated those of the Gaussian distribution confirming that the Gaussian distribution is useful for characterizing effort estimate variation.