ACM Comm 2010 May Automatic Program Repair with Evolutionary Computation (Notes)

"Beyond Total Capture: A Constructive Critique of Lifelogging" CACM May 2010

Automatic Program Repair with Evolutionary Computation by Westley Weimer, Stephanie Forrest, Claire Le Goues, and ThanhVu Nguyen, p.109-116

Automatic Program Repair with Evolutionary Computation

Ideas

1. Genetic programming (GP)[1]
2. Fix the failure.
3. Pass all the test cases.
4. Produce output identical to known-good test results.
5. Do not introduce new errors.
6. Dead Code might be introduced. Or calls to irrelevant functions.
7. Generation count limit to stop an infinite loop. Here that limit is 20.
8. Some defects can't be repaired this way.
   1. Many simultaneous edits.
   2. Changes that cannot be made directly at the statement level. (e.g., \( \textbf{matmul}(b, a) \) should be \( \textbf{matmul}(a, b)\) ).

References

1. Field Guide to Genetic Programming by by Riccardo Poli, William B. Langdon, and Nicholas Freitag McPhee. 2008[2]

Internal Links

Parent Article: Reading Notes