Subject:
|
Genetic programming (was "Re: IR camera?")
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Fri, 12 Feb 1999 20:13:35 GMT
|
Reply-To:
|
BRYAN.BEATTY@AUTODESK.ihatespamCOM
|
Viewed:
|
1712 times
|
| |
|
|
Mike Moran wrote:
>
> lego-robotics@crynwr.com (Pete Hardie) writes:
> > Frex, let's say you have your factory, and each legobot carries its
> > building instructions, and each
> > is programmed to run a maze. You have the factory at the end, and it
> > closes its doors after
> > 10 legobots have entered. You will get fast maze running robots from
> > this.
> Don't be too sure. Unless you constrain it a lot you could get one or
> two robots which travel fast enough to jam the door and the rest just meander
> on through becuase it's jammed open. This is perhaps a bit contrived but I'd
> be wary of making too specific statements of what a GA evolved system will do;
> too many people have got caught on this before. An example is one of Karl Sims
> creatures which was taking part in a GA which had maximim speed as the main
> metric: it's strategy was just to grow tall and then fall over hence achieving
> a maximum speed whilst falling over; I doubt if he would have been able to
> predict this given the seemingly reasonable constraints he placed on it.
Genetic programming in general often exhibits unintended results,
namely:
1. Unless you're very careful, it ends up optimizing for the special
cases rather than the general class of problem you want to solve. In
the above example, you're likely to end up with robots that can run that
particular maze, but would be utterly lost in a different configuration.
2. Genetically programmed critters will find and exploit unintended
weaknesses in your problem space-- they "cheat." The example you gave
of tall critters falling over was an excellent one; they were exploiting
a loophole in the rules for judging results. Another result that Karl
Sims got was in an early version of his physical simulator that had a
subtle bug in the collision-detection routines. His critters found and
exploited that bug: they evolved a behavior in which by "spanking"
themselves, they could propel themselves across his virtual universe at
practically lightspeed.
|
|
Message has 1 Reply:
Message is in Reply To:
 | | Re: IR camera?
|
| (...) performs in real life. (...) the first place? (...) design and try to combine (...) this is called Genetic Algorithms). (...) Don't be too sure. Unless you constrain it a lot you could get one or two robots which travel fast enough to jam the (...) (25 years ago, 12-Feb-99, to lugnet.robotics)
|
17 Messages in This Thread:
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
This Message and its Replies on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
