- author = "Marta Smigielska and Aymeric Blot and Justyna Petke",
- title = "Uniform Edit Selection for Genetic Improvement: Empirical Analysis of Mutation Operator Efficacy",
- booktitle = "GI @ ICSE 2021",
- year = "2021",
- month = "30 " # may,
- editor = "Justyna Petke and Bobby R. Bruce and Yu Huang and Aymeric Blot and Westley Weimer and W. B. Langdon",
- publisher = "IEEE",
- address = "internet",
- pages = "1--8",
- keywords = "genetic algorithms, genetic programming, genetic improvement, APR, mutation operators, PyGGI, SrcML, Java",
- isbn13 = "978-1-6654-4466-8/21",
-
URL = "
https://geneticimprovementofsoftware.com/paper_pdfs/gi2021icse/smigielska_gi-icse_2021.pdf",
-
URL = "
https://github.com/SOLAR-group/gi2021artefact",
-
video_url = "https://www.youtube.com/watch?v=cpN9pLLKA7s&list=PLI8fiFpB7BoKDaxvS7SQp0iA7fN7rrvDD&index=3",
-
video_url = "https://www.youtube.com/watch?v=9nqrQ7YkHQM&list=PLI8fiFpB7BoKDaxvS7SQp0iA7fN7rrvDD&index=4",
-
video_url = "https://www.youtube.com/watch?v=e8uJf67hL0w&list=PLXTjhGKkSnI-se7uQneCX-pEDiDrQ7TIS&index=9",
-
DOI = "
doi:10.1109/GI52543.2021.00009",
- size = "8 pages",
- abstract = "Genetic improvement (GI) tools find improved program versions by modifying the initial program. These can be used for the purpose of automated program repair (APR). GI uses software transformations, called mutation operators, such as deletions, insertions, and replacements of code fragments. Current edit selection strategies, however, under-explore the search spaces of insertion and replacement operators. Therefore, we implement a uniform strategy based on the relative operator search space sizes. We evaluate it on the QuixBugs repair benchmark and find that the uniform strategy has the potential for improving APR tool performance. We also analyse the efficacy of the different mutation operators with regard to the type of code fragment they are applied to. We find that, for all operators, choosing expression statements as target statements is the most successful for finding program variants with improved or preserved fitness (50.03 percent, 33.12 percent and 23.85 percent for deletions, insertions and replacements, respectively), whereas choosing declaration statements is the least effective (3.16 percent, 10.82 percent and 3.14 percent for deletions, insertions and replacements).",
-
notes = "Typically GI deletions are over represented in the
space of all possible GI changes. So make all edits
equally probable.
QuixBugs small Java program. Parse using SrcML. Landscape analysis. Better than prior work on QuixBugs.
Fewer unique patches and much slower (slide 10), efficiency depends on nature of bug.
Future work (slide 15), expect difference between APR and non-function (eg speedup) GI.
Video e8uJf67hL0w Aymeric Blot
12:52 Discussion Chair: Bobby R. Bruce. A: Justyna Petke, Aymeric Blot.
13:13 Q: Westley Weimer, A: Justyna Petke
13:58 Q: Alexander Brownlee, why is delete mutation operator less successful? A: Aymeric Blot depends upon on actual bug fix, only if required fix is a deletion is deletion mutation operator successful.
14:40 Q: Bobby R. Bruce, other mutations may also lead to a deletion (A:yes) A: Aymeric Blot based on mutation type only. Tracking is complex. Reinforcement learning, dynamic mutation type selection or mutation targets (line number in source code).
17:45 Q: Stephanie Forrest, assembly code. A: Aymeric Blot may be no change. Q: Bobby R. Bruce, Syntax not so constraining? Assembly language gives GI more scope? A: Aymeric Blot, Possibly more need to learn which type of mutation to apply than in Java.
20:32 Q: Oliver Krauss only 1/4 of bugs due to deletion A: Aymeric Blot only if fix is deletion, having higher fraction of deletion mutations helps. A: Justyna Petke multiple changes, importance of location of mutations not sufficiently studied. A: Aymeric Blot may depend on APR or non-function (memory reduction) GI
part of \cite{Petke:2021:ICSEworkshop} http://geneticimprovementofsoftware.com/events/icse2021.html",
