Intelligent software debugging: A reinforcement learning approach for detecting the shortest crashing scenarios

Durmaz E., TÜMER M. B.

Expert Systems with Applications, cilt.198, 2022 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 198
  • Basım Tarihi: 2022
  • Doi Numarası: 10.1016/j.eswa.2022.116722
  • Dergi Adı: Expert Systems with Applications
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, PASCAL, Aerospace Database, Applied Science & Technology Source, Communication Abstracts, Compendex, Computer & Applied Sciences, INSPEC, Metadex, Public Affairs Index, Civil Engineering Abstracts
  • Anahtar Kelimeler: Reinforcement learning in automated bug, detection, Exploring crashes by SARSA, Exploring crashes by prioritized sweeping, Delta debugging, Detected goal catalyst, REPRODUCTION
  • Marmara Üniversitesi Adresli: Evet

Özet

© 2022 Elsevier LtdThe Quality Assurance (QA) team verifies software for months before its release decisions. Nevertheless, some crucial bugs remain undetected in manual testing. These bugs would make the system unusable on field, thus merchant loses money then manufacturer loses its customers. Thus, automatic software testing methods have become inevitable to catch more bugs. To locate and repair bugs with an emphasis on the crash scenarios, we present in this work a reinforcement learning (RL) approach for finding and simplifying the input sequence(s) leading to a system crash or blocking, which represents the goal state of the RL problem. We aim at obtaining the shortest input sequence for the same bug so that developers would analyze agent's actions causing crashes or freeze. We first simplify the given crash scenario using Recursive Delta Debugging (RDD), then we apply RL algorithms to explore a possibly shorter crashing sequence. We approach the exploration of crash scenarios as a RL problem where the agent first attains the goal state of crash/blocking by executing inputs, then shortens the input sequence with the help of the rewarding mechanism. We apply both model-free on-policy and model-based planning-capable RL agents to our problem. Furthermore, we present a novel RL approach, involving Detected Goal Catalyst (DGC), which reduces the time complexity by avoiding grappling with convergence via stopping learning at a small variance and attaining the shortest crash sequence with an algorithm that recursively removes the unrelated actions. Experiments show DGC significantly improves the learning performance of both SARSA and Prioritized Sweeping algorithms on obtaining the shortest path.