|
Sat Jul 28 17:04:09 EDT 2007
Good article on GAs
--
This article as New Scientist caught my eye because if gives a
pretty good overview of evolutionary algorithms applied to prob-
lems of design. Their premise, that GA's are surpassing the abil-
ities of human designers is not new nor is it noteworthy in my
opinion. GA's do not exhibit intelligence, but they do allow an
efficient way to search a very large search space of viable com-
binations of properties.
http://xrl.us/3jy8
From my point of view, the most difficult part of writing a ge-
netic algorithm is determining how to formulate what you are
evolving. In the article, they talk about designing things like
ship hulls. In engineering design, there is often a balance be-
tween like the amount of material, stength, flexibility, etc. In
otherwords, the algorithms try combinations and designs that a
human engineer never would. In doing so, it is not unusual for
superior designs to be found. It is even more common that these
designs are like nothing a person would have tried.
Now, when I say formulate the problem, I am not talking about the
fitness function - the thing that tests if an individual is "fit"
enough to be mated with another fit individual. I am talking
about deciding what parts are the "genes" - or in otherwords, how
are the pieces and parts that are changed designed?
One interesting aspect of this for me would be to apply it to al-
gorithm design. My problem here is, how do you break up an algo-
rithm? It seems to me that on some level you have to design some
constituent parts that fit together - like Lego blocks. It is
these blocks that are then put together by the GA, tested for fa-
vorable characteristics. The combinations of these characteris-
tics are then propagated through the evolutionary process.
Here is a thought exercise: how do you evolved a code? It seems
that you can't simply use the syntax of the language as the genes
- it would be impressive just to get a piece of code that could
compile. So, you have to have chunks of code that you know are
valid already - chunks that could then be placed into some sort
of predefined template. After thinking about this for sometime,
it seems like you'd have to design your process to the point
where you are really just selecting the combination of a certain
set of routines that you know, when put together, do what you
want them to.
Maybe one day I'll sit down and write a program that evolves an
"algorithm," but for now it seems to me that applications for
GA's requires the problem to be broken up into pieces for which
the evolutions simply the combinations of routines. Is there a
way to evolve chuncks of code in an interleaving fashion such
that you can control what they do such that you have a good
chance of having "individuals" that pass your fitness function?
--
Powered by vee
Copyright © 2006-2008
|
|