Adaptive estimation framework for software defect fix effort using neural networks

This dissertation describes research on a solution to the general current problem of limitations, in using existing software defect fix effort estimation techniques, that arise when using scarce, incomplete, inaccurate, and sparse historical data for quality required in software development. The goal of this dissertation is to provide a framework and a methodology for estimating the software defect fix effort for a specific problem, when historical data with certain features missed, by using a nonparametric technique, applying dissimilarity matrices and self-organizing neural network for software defect clustering. This Adaptive Software Defect Fix Effort Estimation framework and methodology is intelligent and self-adaptive, with the capability of dealing with various categorized input data, and uncertainty, when needing to accurately estimate defect fix effort. The framework includes three stages. Stage I uses feature abstraction of input variables for the next step estimation. Stage II uses prediction tasks that are capable of estimating the software defect fix effort. Stage III uses probabilistic measurement. A Self-organizing Neural Network is used in my estimation techniques.; To evaluate the improved capabilities of my proposed estimation framework, the validation of estimation performance of my new methodology is addressed in the second half of the dissertation. Two experiments with different categorized incomplete software historical defect data inputs are applied for the performance evaluation. The estimation framework provides significant performance improvement over a current, widely used general defect effort estimation model with an average of MRE accuracy from experimentation with NASA Metrics Data Program (MDP). By using this new framework, the defect fix effort can be estimated using different incomplete historical data and different input domains. The Adaptive Software Defect Fix Effort Estimation framework extends traditional size-based estimation to multi-dimensional input metrics, combining numerical, nominal, ordinal, binary and other data types.; My research found that most defect fix efforts are performed during the software testing phase. Therefore, defect fix effort can be reduced further by applying some reuse techniques like product line testing in the software testing phase. I propose an economical effort model for software product line testing with corresponding discussions on Return of Investment (ROI) at the end of the dissertation. (Abstract shortened by UMI.)...