From: Eugene Leitl (eugene.leitl@lrz.uni-muenchen.de)
Date: Thu Sep 14 2000 - 15:00:42 MDT
Emlyn O'Regan writes:
> Well, I'm not a VB programmer, but I'm still seriously underqualified to do
> more than rant vaguely about genetic programming. Eugene has a passing
> familiarity with this stuff; got a status report on state-of-the-art for us?
The (pathetic) state of the art is stuff being done by John Koza. You
can read him up on the web, and in several (by now three) fat MIT
Press bibles, all titled Genetic Programming (I, II and now III if not
IV, I believe), where he struts his (some of them actually pretty
nifty) achievements.
The field has huge potential, but is right now stuck imo slightly
before cyanobacteria stage, for several reasons. Most of them is that
mature (those having lived through a few iterations) evolutionary
algorithms are very subtle, not all details of natural instances of
them are known, and the few Comp. Sci. evolutionary algorithm
practitioners ignore even what little we already know, standing smack
middle in da Nile. Sounds a lot like AI, huh?
The difficulty resides in: having enough crunch (Koza has a 512 CPU
Beowulf in St.Anford, which is a good start), a large enough
population (depends on the problem, may be dynamic), a suitable
mutable substrate (machine (Java VM) instructions or even SEXPRs are
that not) and a flexible enough framework (which could compensate for
less than perfect substrate, and do further industrial-strength
magick). The latter is not at all addressed (and in fact is
surprisingly difficult to attack), for instance the importance of the
mutation function itself having to be mutable (most assuredly true for
wet evolutionary systems, it sure as hell doesn't equal-probability
shotgun-mutate, especially something as brittle as computer languages)
is not recognized. (At least Koza told me that he doesn't see the
point of it beyond the obvious crude hand-coded optimizations, making
moi very distraught indeed).
Another reason is that GP looks like programming straight from the
Salpetriere to most Comp. Sci. people, unless they are martians. Clash
of paradigms, very much so.
I'm not expecting Koza to make a breakthrough (partly, because of
above reasons), and I'm not aware of any practitioners with a real
budget and an open-minded approach, so near future prognosis is less
than good. But, yeah, eventually somebody somewhen will cut it, we go
into the Cambrian explosion regime and there will be much boisterous
frolicking and merriment in the streets (with the exception of
computer securety people, these who have not yet shot themselves will
be looking for new jobs).
> I'm a bit afraid that the business plan would look something like:
> 1 - Get lots o money
> 2 - Spend lots o money doing heaps of cool research into genetic
> programming, play around with oh so excellent equipment, discover some
> excellent stuff, have a bloody fantastic time
> 3 - Run out of money with no products/market to show for it,
> 4 - get a slapping (eek) from the VC(s)
> 5 - Reputation in ruins, become a tramp, scavenge in bins, waiting for
> the singularity
I dunno about 5 (participating in a .com burnout will raise your job
market value), but 1-4 is imo exactly how the shit can be expected to
go down. </old curmugeon>
> Emlyn, die hard optimist
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:30:59 MST