Cases, Predictions, and Accuracy Learning and Its Application to Effort Estimation

Estimation by analogy EBA (effort estimation by analogy) is one of the proven methods for effort prediction in software engineering; in AI this would be called Case-Based Reasoning. In this paper we consider effort predictions using the EBA () method AQUA and pay attention to two aspects: (i) The influence of the set of analogs on the quality of prediction. The set of analogs is determined by a learning process incorporating the number of nearest neighbors and the threshold of the similarity measure used, (ii) Analyzing and understanding the conditions under which the prediction can be expected to be the most or the least accurate. We study two types of learning: One for finding the “best” set of analogs, and one for finding out factors for reliability. While both questions are relevant for different areas and disciplines, the focus of the paper is on estimation of effort in software engineering. For EBA method AQUA, the cases can be features or past projects characterized by attributes of various type. Classical estimation approaches just investigate the overall estimated quality of a system. However, in that case information is missing if and why estimation was performing the way it did. Bad estimates are often due to external influences. Therefore it is valuable for to find out under which conditions the estimates are more or less reliable.