Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Bojan Cukic.


As software systems are increasingly deployed in mission critical applications, accurate quality and reliability predictions are becoming a necessity. Most accurate prediction models require extensive testing effort, implying increased cost and slowing down the development life cycle. We developed two novel statistical models based on Dempster-Shafer theory, which provide accurate predictions from relatively small data sets of direct and indirect software reliability and quality predictors. The models are flexible enough to incorporate information generated throughout the development life-cycle to improve the prediction accuracy.;Our first contribution is an original algorithm for building Dempster-Shafer Belief Networks using prediction logic. This model has been applied to software quality prediction. We demonstrated that the prediction accuracy of Dempster-Shafer Belief Networks is higher than that achieved by logistic regression, discriminant analysis, random forests, as well as the algorithms in two machine learning software packages, See5 and WEKA. The difference in the performance of the Dempster-Shafer Belief Networks over the other methods is statistically significant.;Our second contribution is also based on a practical extension of Dempster-Shafer theory. The major limitation of the Dempsters rule and other known rules of evidence combination is the inability to handle information coming from correlated sources. Motivated by inherently high correlations between early life-cycle predictors of software reliability, we extended Murphy's rule of combination to account for these correlations. When used as a part of the methodology that fuses various software reliability prediction systems, this rule provided more accurate predictions than previously reported methods. In addition, we proposed an algorithm, which defines the upper and lower bounds of the belief function of the combination results. To demonstrate its generality, we successfully applied it in the design of the Online Safety Monitor, which fuses multiple correlated time varying estimations of convergence of neural network learning in an intelligent flight control system.