Feature weighting for Case-based reasoning software project effort estimation

Background: Software effort estimation is one of the most important activities in software development process. Un-fortunately, estimates are often substantially wrong and specifically most projects encounter effort overruns. Numerous methods have been proposed including Case based reasoning (CBR). Existing research shows that feature subset selection (FSS) is an important aspect of CBR, however, searching for the optimal feature weights is a combinatorial problem and therefore NP-hard. Objective: To develop and evaluate efficient algorithms to generalise FSS into an effective feature weighting approach that can improve accuracy further, since not all features contribute equally to solving the problem. Method: Use various search algorithms e.g., forward sequential weighting (FSW) and random mutation hill climbing (RMHC) to assign weight to features in order to improve the estimation accuracy. We will extend an existing CBR java shell ArchANGEL1. We will perform experiments based on repeated measures design on real world datasets to evaluate these algorithms. Limitations of the proposed research: Dataset quality cannot be assured therefore our findings could be influenced by noisy data. Older datasets may be misrepresenting current software development approaches and technologies. CBR could be sensitive to the choice of distance metric; however, we will only use standardised Euclidean distance.