Program Repair With Repeated Learning

Abstract

A key challenge in generate-and-validate automated program repair is directing the search for fixes so that it can efficiently find those that are more likely to be correct. To this end, several techniques use machine learning to capture the features of programmer-written fixes. In existing approaches, fitting the model typically takes place <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">before</i> fix generation and is independent of it: the fix generation process uses the learned model as one of its inputs. However, the intermediate outcomes of an ongoing fix generation process often provide valuable information about which candidate fixes were “better”; this information could profitably be used to retrain the model, so that each new iteration of the fixing process would also learn from the outcome of previous ones. In this paper, we propose the <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Liana</small> technique for automated program repair, which is based on this idea of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">repeatedly</i> learning the features of generated fixes. To this end, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Liana</small> uses a fine-grained model that combines information about fix characteristics, their relations to the fixing context, and the results of test execution. The model is initially trained offline, and then repeatedly updated online as the fix generation process unravels; at any step, the most up-to-date model is used to guide the search for fixes—prioritizing those that are more likely to include the right ingredients. In an experimental evaluation on 732 real-world Java bugs from 3 popular benchmarks, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Liana</small> built correct fixes for 134 faults (83 ranked as first in its output)— improving over several other generate-and-validate program repair tools according to various measures.