Re: 'not an artist' article

[Erik <br0ke@math.smsu.edu>]

On 01-Jul-2000 Steve Baker wrote:
> Erik wrote:
> 
>> Here I thought I was a supergenious having
>> this idea, and it turns out everyone and their brother is thinking along the
>> same lines.
> 
> Genetic algorithms are becoming pretty commonplace these days.
>  
>> One problem with evolving systems in software is that it uses a lot of
>> technique that are primarily in the field of AI, and most programmers don't
>> have a strong AI background.
> 
> For what I propose (evolving artwork) this isn't an issue.
> 
> The original biomorphs program was very simple - no AI at all.
> 

I d'no if you can equate AI with complex :) artificial intellegence is
basically the computer equivelent of 'stage magic'... if ya see how it's done,
it loses it's magic and you sit there going "that's it?" Maybe I'd be better
off saying AL, since none of what we're talking about involves any kind of
emulated cognative ability

> The goal of the evolution is to please humans - hence a human has to say
> which mutant lives and which dies on the basis of which one comes closest
> to their idea of 'cute' or 'agressive' or whatever criteria is needed.
> 

I would think that this would severely limit the algorithms ability. But my
experience with them was primarily route finding in complex networks, so I'd be
grinding tens of thousnads of evolutions with many hundreds of nodes :) If you
implement a system that does that, I'd be interested in knowing how it turns
out. Like how many generations on average before an acceptable result, how the
models were generated, details of the algorithms input, etc. 

> To misquote Darwin: It's a matter of "Survival of the Cutest".
> 
> People who have tried to use automatic selection criteria have
> discovered that the critters that are produced have a tendancy
> to exploit 'holes' in your criteria of bugs in the program.
> 
> I forget the name of the guy - but there was a SigGraph paper
> a few years ago about this software package that evolved locomotion
> systems - with the hope that it would create critters that could
> walk, slither, crawl, whatever.
> 
> He set his selection criteria to be 'the creature that moved
> furthest in a given (fixed) amount of time.
> 
> He ran the genetics for a gazillion generations - then viewed
> the resulting population - expecting to see elegant running
> machines or something.
> 
> What actually happened was that his critters evolved into
> EXTREMELY tall, unstable structures that simply fell over.
> Since his 'distance travelled' criterion measured the distance
> travelled by the center of gravity of the critter, all they
> had to do was to fall over in order to beat out any other
> locomotion mechanisms.
> 
> The critters did what they were told - they produced an
> extremely simple, elegant (and unexpected) solution to
> the problem of moving the largest distance in 10 seconds.
> 
> He fixed that problem and found that critters would evolve
> to do things like exploit a roundoff error in his friction
> math to pick up energy from 'nowhere'.  Any bug or loophole
> in the simulated 'world' would result in creatures evolving
> to exploit it.
> 
> This is a hard-won lesson.
> 

Obviously you need to accurately describe the problem for the algorithm to work
:) The problems he described were probably solved very optimally... The time
involved in making the algorithm run well is why I concluded that it's not
feasable for small game worlds.

I'm reminded of an experiment a guy in canada did... The infamous hopping
tri-joint lamp was the output of a genetic algorithm. The thing produced two
possible locamotion methods, one was hopping, the other was dragging itself
like an inchworm or snake or something (the hopping one is the one that got
popular). It's just a matter of being able to describe the problem correctly :)

> Your economic model had better be pretty tightly coded or
> you'll end up with entire civilisations that evolve to
> exploit pennies that end up lodged down the sides of
> sofas.  :-)
> 

whatchew talkin' 'bout, willis? I'm makin' my millions from pennies stuck in my
trucks bench and in my couch :) From a mathematical or scientific view (which
is where my training is), economics is basically a few very simple formulas.
They seem to be relatively accurate, so I'd hope that a simulation would yeild
relatively beleivable results. (who knows, mebbe I'll make billions off of
business simulation software someday ;)

        -Erik <erik@smluc.org> [http://math.smsu.edu/~br0ke]

The opinions expressed by me are not necessarily opinions. In all
probability, they are random rambling, and to be ignored. Failure to ignore
may result in severe boredom or confusion. Shake well before opening. Keep
Refrigerated.
        

---------------------------------------------------------------------
To unsubscribe, e-mail: linuxgames-unsubscribe@sunsite.auc.dk
For additional commands, e-mail: linuxgames-help@sunsite.auc.dk