Code Naturalness to Assist Search Space Exploration in Search-Based Program Repair Methods

摘要

Automated Program Repair (APR) is a research field that has recently gained attention due to its advances in proposing methods to fix buggy programs without human intervention. Search-Based Program Repair methods have difficulties to traverse the search space, mainly, because it is challenging and costly to evaluate each variant. Therefore, aiming to improve each program's variant evaluation through providing more information to the fitness function, we propose the combination of two techniques, Doc2vec and LSTM, to capture high-level differences among variants and to capture the dependence between source code statements in the fault localization region. The experiments performed with the IntroClass benchmark show that our approach captures differences between variants according to the level of changes they received, and the resulting information is useful to balance the search between the exploration and exploitation steps. Besides, the proposal might be promising to filter program variants that are adequate to the suspicious portion of the code.