Refining Fitness Functions for Search-Based Program Repair

Created by W.Langdon from gp-bibliography.bib Revision:1.8120

@InProceedings{bian:2021:apr_icse,

• author = "Zhiqiang Bian and Justyna Petke and Aymeric Blot",
• title = "Refining Fitness Functions for Search-Based Program Repair",
• booktitle = "APR @ ICSE 2021",
• year = "2021",
• month = "1 " # jun,
• editor = "Sergey Mechtaev and Shin Hwei Tan and Martin Monperrus and Lingming Zhang",
• publisher = "IEEE",
• address = "internet",
• pages = "1--8",
• keywords = "genetic algorithms, genetic programming, genetic improvement, APR, Software Engineering, SBSE, Program Repair, Fitness Function, GenProg, ARJAe, 2Phase, Gin, EvoSuite, QuixBugs",
• isbn13 = "978-1-6654-4473-6",
• code_url = "https://github.com/SOLAR-group/apr2021artefact",
• URL = "http://www.cs.ucl.ac.uk/staff/a.blot/files/bian_apr-icse_2021.pdf",
• video_url = "https://www.youtube.com/watch?v=2CbHQMhkMTU",
• DOI = "doi:10.1109/APR52552.2021.00008",
• size = "8 pages",
• abstract = "Debugging is a time-consuming task for software engineers. Automated Program Repair (APR) has proved successful in automatically fixing bugs for many real-world applications. Search-based APR generates program variants that are then evaluated on the test suite of the original program, using a fitness function. In the vast majority of search based APR work only the Boolean test case result is taken into account when evaluating the fitness of a program variant. We pose that more fine-grained fitness functions could lead to a more diverse fitness landscape,and thus provide better guidance for the APR search algorithms.

  We thus present 2Phase, a fitness function that also incorporates the output of test case failures, and compare it with ARJAe,that shares the same principles, and the standard fitness, that only takes the Boolean test case result into consideration. We conduct the comparison on 16 buggy programs from the QuixBugs benchmark using the Gin genetic improvement framework. The results show no significant difference in the performance of all three fitness functions considered. However, Gin was able to find 8 correct fixes, more than any of the APR tools in the recent QuixBugs study.",

• notes = "'no significant difference between effectiveness and efficiency of the three fitness functions on our dataset'

  cf video 2CbHQMhkMTU 11:17 A: due to QuixBugs, may be different results on different benchmarks. A2 (red) Future work more benchmarks and especially bigger faulty software, with possibly larger search spaces. More formal analysis of search space, cf search landscape of SBSE testing.

  Q2: fitness diversity. A: different patches (of same example bug) should have different (numerical) fitness values. Low diversity means different patches have the same fitness. Without fitness diversity it is hard to guide the search for better patches.

  Q3: would other types of diversity be better? A1(off screen, voice only): tried new operators. a2: Future work may be need new measures of diversity. A3(grey): We concentrate upon fitness diversity between patches to guide search. a2: Smooth the landscape so EC algorithm is better guided (to acceptable bug repair)

  'test-suite adequate patches for 11 programs, with 8 being correct fixes (more than any of the APR tools in the recent QuixBugs study)'

  checkpoints around numerical values in the source code as part of fitness distance calculation. Internal test oracles? But not used?

  patch over fitting. EvoSuite hold out tests.

  'The mutation operator either appends a random new edit to a patch or removes one from a non-empty patch. The crossover operator combines two parent solutions by concatenating, in both orders, the two sequences of edits before each edit is removed with 0.5 probability to create the two children patches.'

  University College London, London, United Kingdom

  https://program-repair.org/workshop-2021/",

}

Genetic Programming entries for Zhiqiang Bian Justyna Petke Aymeric Blot

Citations