Using an RBF Neural Network to Locate Program Bugs

We propose an RBF (radial basis function) neural network-based fault localization method to help programmers locate bugs in a more effective way. An RBF neural network with a three-layer feed-forward structure is employed to learn the relationship between the statement coverage of a test case and its corresponding execution result. The trained network is then given as input a set of virtual test cases, each covering only a single statement. The output of the network for each test case is considered to be the suspiciousness of the corresponding statement; a statement with a higher suspiciousness has a higher likelihood of containing a bug. The set of statements ranked in descending order by their suspiciousness are then examined by programmers one by one until a bug is located. Three case studies on different programs (space, grep and make) were conducted with each faulty version having exactly one bug. An additional program gcc was also used to demonstrate the concept of extending the proposed method to programs with multiple bugs. Our experimental data suggest that an RBF neural network-based fault localization method is more effective in locating a program bug (by examining less code before the first faulty statement containing the bug is identified) than another popular method, Tarantula, which also uses the coverage and execution results to compute the suspiciousness of each statement.