Theoretical Analysis of GP-Evolved Risk Evaluation Formulas for Spectrum Based Fault Localisation

Risk evaluation formulas convert program spectrum data from test executions into suspiciousness score, according to which statements are ranked to aid debugging activities. Designing such formulas remained largely a manual task until Genetic Programming has been recently applied: resulting formulas showed promising performance in empirical evaluation. We investigate the GP-evolved formulas theoretically and prove that GP has produced four maximal formulas that had not been known before. More interestingly, some of the newly found maximal formulas show characteristics that may seem inconsistent with human intuition. This is the first SBSE result with provable human competitiveness. Department of Computer Science University College London Gower Street London WC1E 6BT, UK Theoretical Analysis of GP-Evolved Risk Evaluation Formulas for SBFL Xie et al.