Can AI help to improve debugging substantially? debugging experiences with value-based models

Abstract

Finding and fixing faults in programs is usually an expensive and tedious task. Consequently the development of intelligent debugging tools that aid the programmer in this task is a topic of major industrial interest. This work describes two representations for applying model-based diagnosis to Java programs, a technique that permits locating (and partly correcting) faults without requiring a formal specification of the desired program behavior, since interaction can be limited to test cases and observations of variable correctness. One of the models uses a special transformation to provide more accurate diagnoses on programs with loops and this is borne out by the experiments. The presented results on actual debugging performance show clearly superior accuracy to classical debugging techniques, and better discrimination than dependency-based programs models. We discuss the results in terms of the properties of the two models and the various example programs and present avenues for further improvement.