Reinforcement learning for mutation operator selection in automated program repair

Area of Science:

• Software Engineering
• Artificial Intelligence

Background:

• Automated program repair (APR) uses heuristic search to fix software bugs by mutating code.
• Current methods often randomly select mutation operators, generating many non-compiling or incorrect program variants, wasting resources.

Purpose of the Study:

• Investigate reinforcement learning (RL) for selecting mutation operators in heuristic-based APR.
• Aim to reduce the generation of invalid program variants and improve repair efficiency.

Main Methods:

• Developed an RL-based approach for mutation operator selection, agnostic to programming language, granularity, and search strategy.
• Conducted extensive empirical evaluation on 353 real-world bugs from Defects4J, using 30,080 repair attempts.

Main Results:

• The RL approach yielded a higher number of test-passing variants compared to the baseline random selection.
• No significant improvement was observed in the number of successfully patched bugs.

Conclusions:

• While RL enhances the generation of valid variants in APR, it has not yet demonstrated a clear advantage in improving bug-fixing rates over random selection.
• Further research is needed to fully leverage RL's potential in optimizing heuristic-based program repair search strategies.