[rc5] Genetic Algoritms

Chris Arguin Chris.Arguin at unh.edu
Wed Nov 5 21:35:56 EST 1997


On Wed, 5 Nov 1997, Michael Driscoll wrote:

> >I'm not quite sure what you're proposing.  Treat distributed.net as one
> >large computer and release several programs that try to win by
> >multipying so much that it takes everything over?  Isn't this virus-like
> >behavior?  What is the benefit from this?
> >
> >I can see some benefits in learning about program interactions and
> >genetic like behavior in software but what is the advantage of doing
> >this on a distributed computer as opposed to a stand-alone?
> 
> No, that's not exactly what he's proposing.  corewar is a game popular
> among asm aficianados.  The object of the game is to write a snippet of
> asm code (in a made-up asm language called "redcode") which runs in
> another program which acts like a computer running on the redcode
> assembly.  The programs don't actually try to trash your core, instead
> they only battle in the virtual machine created by the simulation
> program.

Thank you. That last one is what I meant. I'm still working on how to best
distribute this. Right now I am thinking that the server sends out, say,
10 warriors to each client (each warrior is a couple of hundred bytes of
code). The clients then use their 10 warriors to create 1000 warriors via
some genetic algorithm. It then competes those 1000 warriors, 2 at a time.
Using the score from the competition, it returns the top 10 warriors to the
server, along with their scores. The server then sends the warriors out
again. Eventually (at a predetermined time?), it takes the best warrior
and submits it for competition.

There are some problems with this (figuring out what numbers work best
could be troublesome), but I think it makes a good starting point. Also,
note that the same 10 warriors sent to two different clients will not
produce the same pool of warriors, since most genetic algorithms have a
random factor or two. This means that if we have less warriors then
computers to process them, we can still get productive results by sending
out older warriors.

> It's a pretty nifty concept.  I'd suggest taking a look at
> http://www.koth.org for more info on corewar.

In fact, there is a paper at this site describing somebody's previous
attempts at a Genetic approach to corewars.

> Anyways, lemme just add my two thousand pesos and say that I think a
> corewar evolution project would be *extremely* cool, and even sounds
> plausible (apparently it's been done before, but to a much more limited
> degree).  This would also do a bit for recruitment since it would tell
> all corewar players about distributed.net.

Yep. There aren't all that many corewars programmers (sadly, assembly is
going out of style), but I'm sure we would get a few of them interested.
 
> Distributed genetic algorithms?  Wow, almost sounds like we can get
> government funding for this :-)

:) We'll claim it's a Massively-Parallel-Fuzzy-Genetic-Neural project. With
that many buzzwords, we're bound to get something.

--
Chris Arguin                 | "The duke had a mind that ticked like a  
Chris.Arguin at unh.edu         |  clock and, like a clock, it regularly 
                             +--------------+	went cuckoo." - Wyrd Sisters,
http://leonardo.sr.unh.edu/arguin/home.html |               Terry Pratchett




----
To unsubscribe, send email to majordomo at llamas.net with 'unsubscribe rc5' in the body.



More information about the rc5 mailing list