Alt-BEAM Archive

Message #02028



To: beam@corp.sgi.com
From: Justin jaf60@student.canterbury.ac.nz
Date: Mon, 05 Apr 1999 14:07:24 +1200
Subject: [alt-beam] Re: Socer bots/ Hive behaviour


> Very interesting. I would like to see (even the beginnings of your
> Biocore Idea) and try to experiment with your circuit Ideas.


> Now as far as the ants go (I am not an entomologist) are there not
> specific ants that have specific functions?

Yep, but ants of the same caste can do things like build a bridge out of
themselves. Maybe I should just get an ant-farm rather than ponder the
impossibility of building anything that even remotely compares... :-)

> Also the Pheromone thing
> (at least I think) is a substitute for memory. The scent is left to for
> example say warning and the ants know this not because they remember it
> but because they are programmed to.

Yes, that's an interesting point. And an interesting question - could
leaving a trail be considered a memory system, or an alternative system
that performs the same function?

> I do believe that your ideas will
> result in some very interesting dynamic behaviour but I am really
> interested in "moving to the next level" so that we can hard wire some
> stuff program some stuff and let the robot program some stuff on its
> own. I'm not looking to create a DATA but if one bot could communicate
> to another bot like BEES do this would be a very interesting and magical
> leap forward in robotics. I do not want you to think I'm slaging BEAM,
> I'm not but I believe it is just the beginning and not the ending.

Actually, I was more concerned that you seemed to have overlooked what
to me, is the really cool part: that two or three extremely simple rules
can create extremely ordered high-level behaviour (the boids example is
a good one).
Actually _building_ BEAM soccer bots would be take a year, and within a
few hours of completion they would probably be boring to watch, being of
fixed nature.

What would be _really_ cool (and very difficult), would be to build some
bots that are basically just motorised sensor platforms, with a
reprogramable CPU.
Build an environment for them. These things have to be build to
incredibly high standards of precision (to virtually eliminate chaotic
anomalies such as sensor reflection, bounce, etc).
Then set up the same system inside a computer to simulate as closely as
possible sensor results, speed, etc. Stick a simple program in the
simulated CPU's, and run a breed&select program on it. For simplicity,
the game is tag. Run the software to evolve the robots for thousands and
thousands of generations (ie, until they're actually quite good and
surprisingly clever for 100 lines of code or whatever the limit is), and
then stick the program in the physical bots and see how well it
translates.

Then you can do all sorts of experiements: as well as the bots giving a
cool tactical display, and as well as the creepy feeling of "Erm... I
have _no_ idea how this thing thought of doing that!", you can also test
earlier breeds against newer ones to find out whether the difference in
tactics nullifies the evolutionary difference (ie they've become too
specialised), try to breed for versitility, etc etc.

Then the extra cool thing: you build another identical bot, and replace
the CPU with a radio control unit, and discover whether 100 lines of
utterly bizzare evolved code can beat the crap out of you - the
super-intelligent creator with no code limits.

It would also make an awesome interactive display for one of those
science museums. Extremely difficult to do, but probably well worth it.

With regards to human vs robot, a speed adjustment makes a nice skill
setting - ramp it up a bit if you do tend to outhink the bot, as its
brain-to-wheels reaction will be faster, or turn it down if you have
difficulty with the controls, or a young child is playing. (Hmmm, make
that "a young child before video games were invented" perhaps :-)

------------------------------------------------------------------------
eGroup home: http://www.eGroups.com/list/alt-beam
Free Web-based e-mail groups by eGroups.com

Home