A Genetic Programming Approach to Automated Software Repair.

http://www.cs.virginia.edu/~weimer/p/weimer-gecco2009-preprint.pdf

by mrlebowski • 17 years ago

22 7 17 years ago

cs.virginia.edu

mrlebowski• 17 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&...

unignorant2• 17 years ago

He is at Berkeley tomorrow:

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

Attend if you can!

jacquesm• 17 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• 17 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.

unignorant2• 17 years ago

You might try something like CUTE:

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

extension• 17 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• 17 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.

A Genetic Programming Approach to Automated Software Repair.

Related Stories