A Genetic Programming Approach to Automated Software Repair.

A Genetic Programming Approach to Automated Software Repair. (cs.virginia.edu)

22 points by mrlebowski 16 years ago | 7 comments

jacquesm 16 years ago |

Very interesting stuff. I did notice that this bug was due to a broken test procedure, a simple test loop during the initial write would have brought this problem out instantly.

If your code has an integer range of several 10's of thousands of input possibilities and it costs you next to nothing to run it why not exhaustively test that function?

That way your confidence goes up tremendously, just a couple of edge cases vs the whole range, I know which one I'd pick.

The run output would simply be two columns of integers, very easy to scan for errors. (the output should be equal or monotonous increase from day by day, and should not hang...).

Of course that's after the fact, but there is really no routine so trivial that you can get away without really testing if it does what you intended.

scott_s 16 years ago | |

That's an excellent point - if you have a finite range, why not test all of it? - and I know software testing researchers who look into just that. The difficulty to keep in mind is that in a large code base, there are going to be many small pieces that have a different finite range.

unignorant 16 years ago | | |

You might try something like CUTE:

http://osl.cs.uiuc.edu/~ksen/cute/

mrlebowski 16 years ago |

Westley Weimer has a presentation on the same @ Stanford University today at 2pm (1 hr from now) in Gates 104

Map - http://maps.google.com/maps?f=q&source=s_q&hl=en&...

unignorant 16 years ago | |

He is at Berkeley tomorrow:

http://events.berkeley.edu/index.php/calendar/sn/eecs.html?e...

Attend if you can!

extension 16 years ago |

Don't get too excited, this requires unit tests to detect the bug. In other words, if you already have a pretty good understanding of the bug, this thing will evolve a fix for you instead of having to write it yourself.

If you can write an algorithm that finds bugs, then you've really got something.

wlievens 16 years ago | |

A bug is a defect. A defect is a deviation from the intended meaning of the program. To automatically detect that, you will need a formal specification of the meaning of the program.

It's the automated programming paradox. You can't automate programming because you have to tell the automaton what to program... which is called programming.