Year: 2024
Type: dissertation
Abstract: Automated program repair (APR) enhances software quality by fixing bugs automatically, but it faces challenges due to software complexity. The vast number of possible patches makes exhaustive search impractical, and identifying correct patches is difficult since tools may generate incorrect fixes that overfit test cases. Mitigating these challenges involves leveraging the structure of code, which consists of a formal channel (execution semantics) and a natural language channel (comments, variable names). Machine learning excels at interpreting the natural language channel using large datasets but struggles with generating semantically correct patches. Conversely, program analysis provides detailed insights into program semantics. Combining program analysis with machine learning can address these challenges, using program analysis for execution specifics and machine learning for natural code aspects, like identifiers and comments. This thesis consists of four different works: • Chapter 3 advances semantic repair by synthesizing patches with side effects, employing symbolic execution with state merging and effective patch prioritization to repair bugs in open-source projects. • Chapter 4 reduces the search space by utilizing neural networks to learn variable information, ranking variables and patch templates to improve accuracy and reduce test overfitting. This enhances existing approaches, allowing them to repair previously unfixable bugs by leveraging program namespace information. • Chapter 5 leverages abstract interpretation and fuzzing to probabilistically approximate reachable program states, focusing on high-probability states PSP boosts performance in abstract interpretation, symbolic execution, and patch prioritization, benefiting strategies discussed in Chapter 3 and Chapter 4. • Chapter 6 addresses the challenge of identifying the most relevant facts, such as test errors and angelic values, for constructing effective prompts for LLM based APR. Extracted through program analysis, these facts build prompts whose effectiveness varies across different bugs. We develop a strategy to select facts tailored to each specific bug, significantly enhancing the effectiveness of LLMs in APR.
Author: Parasaram, Nikhil
Institution: University College London
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7062315239