Newsgroups: comp.ai.genetic,comp.ai.alife,comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!MathWorks.Com!europa.eng.gtefsd.com!howland.reston.ans.net!news.moneng.mei.com!uwm.edu!lll-winken.llnl.gov!trib.apple.com!amd!netcomsv!netcomsv!netcom.com!park
From: park@netcom.com (Bill Park)
Subject: Re: Make a BattleMech Stand Up!
Message-ID: <parkCxn2KA.1F0@netcom.com>
Keywords: Sims Koza genetic programming evolutionary programming artificial life
Cc: stephane@dcs.gla.ac.uk
Organization: Netcom Online Communications Services (408-241-9760 login: guest)
References: <none> <781773199snz@mondas.demon.co.uk>
Date: Fri, 14 Oct 1994 01:24:10 GMT
Lines: 126
Xref: glinda.oz.cs.cmu.edu comp.ai.genetic:4022 comp.ai.alife:1134 comp.ai.neural-nets:19447

-- Oy, oy, oy!  You lot in the UK: Your excellent New Scientist
magazine has a thorough article on hexapod locomotion, including gait
selection ...

Jim Collins and Ian Stewart, "The Mathematical Springs in Insect
     Steps,", New Scientist, 8 Oct 1994, Vol. 144, No. 1946,
     pp. 36-40, ISSN 0262-4079.

------------------

In article <781773199snz@mondas.demon.co.uk> ward@mondas.demon.co.uk writes:

> You need to see the movie by Karl Sims. Try ftp to ftp.think.com
> and look in (?) /users/karl/.. for an MPEG file. Amazing stuff...
> -- 
> 
> Pete Ward                   I know it's irrational but at times
> Mondas IT Ltd               of stress I take great comfort from
>                             the belief that somewhere out there
>                             there really is a free lunch.

Another way to see Karl Sims' remarkable critters walking, swimming,
and flying is to find a tape or rebroadcast of the Scientific American
television program that was shown on a Public Broadcasting Station
here in the San Francisco Bay Area last week.  The host/presenter is
Alan Alda of M*A*S*H fame.

In that same show, Alda also visits SRI International's robotics lab
in Menlo Park, California, and tries out Flakey, a voice-controlled
mobile robot.  Flakey's just a wheeled cart, no legs, sorry, but with
very sophisticated control software and stereo vision.

See also the lead article, by Sims, in the proceedings of the July
1994 SIGGRAPH conference, available from the Association of Computing
Machinery (ACM) in the U.S.

Sims' simulated his robots using computer graphics (does he have an
appropriate name, or what?).  He developed them in a series of
artificial life experiments using competitive evolution, by killing
off less "fit" individuals from a population of 300 robots, and
allowing the remainder to crossbreed to produce new variations.  The
robots' artificial chromosomes (the genotype) described not only how
many body parts (rectangular solids) comprised a robot, but what sort
of joints connected the parts, what sort of actuators operated the
joints, what sort of sensors the robots possessed, and how the robot's
"nervous system" connected its sensors to its actuators (in other
words, the phenotype).

I don't think the nervous systems learned; that is, I think Sims
worked with hard-wired networks that were completely specified in the
chromosomes.  Others (e.g., Bergman at Interval Research, Palo Alto,
CA) have, of course, used both techniques -- breeding a population of
neural networks, training each one separately, culling the ones that
didn't learn their lessons very well, then letting the smarter
networks go out on dates with one another to produce a new generation
of offspring networks.

Sims' test of fitness was how far an individual robot was able to
walk, or swim, or (I think) how long it could stay up by flying,
depending on what mode of locomotion he was trying to obtain in an
experiment.  But, to find out how well a robot could get along, Sims
had to simulate the second-order dynamics of each robot, including
collisions, coefficients of friction on the ground, and viscous drag
of the water or air.  This was nontrivial programming!

As often happens when we let evolution design something, Sims'
programs came up with surprising solutions that would "never" have
occured to a human designer.  For example, one walking robot had no
legs, just a big base and a jointed stalk that stuck up into the air.
It waved this stalk back and forth to rock its base in such a way as
to move along the way you might shift a tall, heavy piece of
furniture.  Another pulled its "body" along with two "arms," sort of
like a person with no legs.  More obvious designs included some
snake-like swimmers, and one robot that "rowed" itself along very
gracefully with four "fins."

It took about 15 generations of breeding to produce robots with decent
locomotion abilities.  The amount of computation was so great that he
used a megabuck-scale Thinking Machine parallel processor.

Obtaining this expensive hardware was not a problem, since Sims worked
at Thinking Machines.  At least until it went bankrupt a few weeks
ago.  He might be looking for a job about now, if some
location-based-entertainment company were hiring.

You could probably pick up a Thinking Machine for him to work on
pretty cheap these days, too.

You probably wouldn't need a Thinking Machine, though: You already
know what shape a BattleMech is, so you would just have to evolve a
control system for it.  A much simpler problem.  A good high-end
workstation ought to suffice -- with a MIPS, ALPHA, PowerPC, or
Pentium chip in it.  You want general numerical computation power, not
special-purpose graphics rendering power, for the evolution and
testing.  See Professor John Koza's two-volume set, "Genetic
Programming," for some ideas on various approaches you could use to
"breed" a smart BattleMech.

But why stop with just breeding one that can get up again after it
falls down?  You could breed a really mean, clever, deadly sneaky
S.O.B. that could homogenize your more masochistic customers until
their credit cards maxed out.  Heck, if you have the budget, you can
breed pack robots that co-operate to hunt down your screaming patrons.

Maybe you'll need parallel processors again if you are that ambitious.
But take note of how much computing the computer graphics guys on the
Babylon 5 TV show get done each week.  They use a bunch of piddly
little NuTek Toasters (Amigas on boards, another defunct computer
company!). They grind away on a network, sharing the computation
burden, which is their case IS mostly rendering.  You want to see hot
graphics? Check out the Toaster-generated alien in the B5 episode
"Grail" (out on video in a month or two, I hear).

Bill Park
=========

Koza, John R.  Genetic programming : on the programming of
     computers by means of natural selection /  Cambridge, Mass. : MIT
     Press, c1992.  xiv, 819 p. : ill. ; 27 cm.
     LC CALL NUMBER: QA76.6 .K695 1992
 
Koza, John R.  Genetic programming II : automatic discovery of
     reusable programs /  Cambridge, Mass. : MIT Press, c1994.  xx, 
     746 p. : ill. ; 26 cm.
     LC CALL NUMBER: QA76.6 .K693 1994
-- 
Grandpaw Bill's High Technology Consulting & Live Bait, Inc.
