Re: Semi-Automated Programming

Brian Atkins (brian@posthuman.com)
Wed, 14 Jan 1998 11:29:35 -0500


Ahh yes, brute force and tweaking- some of my favorite things :-)
Brings back memories of waiting for fractals to render on my
Amiga 500...

Anyone seen any GA stuff done in Perl?

Eugene Leitl wrote:
>
> On Tue, 13 Jan 1998, Brian Atkins wrote:
>
> > Yeah, everyone should start coding in Perl :-) It rocks!
> > Seriously, it would seem to require some kind of AI to get
> > any farther than what we have now ?
>
> The Genetic Algorithms Archive
> http://www.aic.nrl.navy.mil/galist/
>
> John Koza has a variant of GA he called GP patented (strictly speaking, I
> do not see any difference in evolving code and data, for code is just data
> interpreted by an abstract entity). He initiated a school of evolving
> code trees of a Lisp subset, with sometimes interesting results.
>
> The good things about GA/GP is that you need just to formally specify the
> constraints, i.e. the results which you want to get. The method, in
> theory, would generate systems with progressively higher fitness, which
> ideally would met your specs fully. This might be useless for
> cryptosystems, but great for VLSI design and real-world control problems.
> Another cool things are GA evolved IFS for image compression, synthesis
> and artwork.
>
> I think what cripples current system is that the GA/GP system's metalevel
> (obviously only the fitness function should be sacred, orelse we float off
> into Never-Never land) is not open to GA optimization, which makes the
> resulting system limited, and brittle. A good GP must adjust at least the
> mutation rate and the coding to guarantee a certain kind of optimal
> fitness landscape (roughly, a small ball in genespace encompassing maximal
> variety in phenotypespace, while genespace being permeated by lots of
> neutral filaments (multiple nonlethal migration pathways out of the local
> minimum trap)).
>
> I think such systems are easiest to design as CAs, which coincides nicely
> with the expected (it'd better) future evolution of computer hardware.
> Random defect hit density for deep-submicron WSI considered, evolvable
> hardware must obviously favour tiniest dies possible -- CA cells. (Just
> for this time, I won't plug molecular dot CAMs here).
>
> ciao,
> 'gene

-- 
The future has arrived; it's just not evenly distributed.
                                                       -William Gibson
______________________________________________________________________
Visit Hypermart at http://www.hypermart.net for free business hosting!