Fault Localization With Data Flow Information and an Artificial Neural Network

Abstract

Fault localization is a technique for identifying the exact source code line with faults. It typically requires a lot of time and cost because, to locate the fault, a developer must track the execution of the failed program line by line. To reduce the fault localization efforts, many methods have been proposed. However, their localized suspicious code range is wide, and their fault localization effect is not high. To cope with this limitation, this paper computes the degree of fault suspiciousness of statements by using an artificial neural network and information of the executed test case, such as statement coverage, execution result, and definition-use pair. Compared to the approach that uses only statement coverage as input data for training an artificial neural network, the experiment results show higher accuracy in 15 types of faults out of 29 real fault types in the approach that the definition-use pair included.