Software Debugging through Dynamic Analysis of Program Structures Submitted by

Software covers every corner of our lives, so do software faults. Currently, a popular approach in fault localization compares program feature spectra in passed execution and failed executions, and aims at predicting program elements that are close to the faults, by locating program elements whose exercising have strong correlation with program failures. In this thesis, we focus on the so-called statistical fault-localization techniques, investigate four topics, and present the results of our investigations to tackle four related problems. First, we address the problem that strong correlations may not necessarily be the root cause of the observed failures. We model the propagation traffic through every edge and proportionally apportion the probability of a block being faulty to its directly connected blocks, resulting in a set of homogenous equations in which the probabilities of individual blocks being faulty after propagation of infected program states are the unknown and can be solved. Empirical studies on real-life medium-size programs show that this technique is more effective than state-of-the-art techniques