Software quality is directly correlated with the number of defects in software systems. As the
complexity of software increases, manual inspection of software becomes prohibitively expensive.
Thus, defect prediction is of paramount importance to project managers in allocating the limited
resources effectively as well as providing many advantages such as the accurate estimation of
project costs and schedules. This thesis addresses the issues of defect prediction and learning in
the geometric framework using statistical quality control and genetic algorithms.
A software defect prediction model using the geometric concept of operating characteristic
curves is proposed. The main idea behind this predictor is to use geometric insight in helping
construct an efficient prediction method to reliably predict the cumulative number of defects during
the software development process. The performance of the proposed approach is validated on real
data from actual software projects, and the experimental results demonstrate a much improved
performance of the proposed statistical method in predicting defects.
In the same vein, two defect learning predictors based on evolutionary algorithms are also
proposed. These predictors use genetic programming as feature constructor method. The first
predictor constructs new features based primarily on the geometrical characteristics of the original
data. Then, an independent classifier is applied and the performance of feature selection method
is measured. The second predictor uses a built-in classifier which automatically gets tuned for
the constructed features. Experimental results on a NASA static metric dataset demonstrate the
feasibility of the proposed genetic programming based approaches.