Alt-BEAM Archive

Message #03688



To: beam@corp.sgi.com
From: Terry Newton wtnewton@nc5.infi.net
Date: Tue, 25 May 1999 11:17:17
Subject: [alt-beam] Re: diode


At 11:05 AM 5/25/99 +0200, Wouter Brok wrote:
>Hello Terry,

Hello Wouter

>What I like is that you program neural nets in the microprocessors you use
>for your robots.

Well... not exactly. The networks I usually use are just simple memory
arrays, the senses index the array and return the memory. Confidence
is kept for each location ranging from 0-3, if a move works (results
in the feelers not touching) the confidence increases, otherwise it
decreases. When confidence drops to 0 the memory is replaced by a
random memory. This is not my idea, but straight from David Heiserman's
"How to build your own self-programming robot", a book I found at a
yard sale for 50 cents. Dated 1979.

The problem with "real" neural nets (real as in back-prop and all that
stuff) is the algorithms are too complicated for tiny processors and
don't really work all that well, often requiring hundreds of training
passes to learn. Most discouraging, back-prop nets often forget what
they learn every time they learn something new! I reject. The array
idea is simple and can be programmed on anything, even a stamp,
learns quickly and doesn't forget unless the memory is ineffective.

> I have been thinking about doing this as well (before I
>knew BEAM), but I lacked (and still lack) the knowledge about neural nets.

Don't worry you haven't missed much

>In september I'll attend a course on neural nets at uni and I hope that
>after that I can implement a microprocessor (the AT90S1200) in a
>BEAM-walker (5 motor, I've the frame ready, but didn't find the time for
>the circuit yet), such that the neural net processes 'higher order' sensory
>information. At the moment this sounds really ambitious to me ... I don't
>know if it is possible ... I'll see.

It is very possible! Micros can easily control nervous nets with nothing
more than a resistor to interface, simulating a variable resistance to
ground (using pwm) to affect the timing of the neurons. You should have
better luck with a 5 motor walker than I did with my 2 motors, being able
to turn and all that...

Terry Newton


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

eGroups.com home: http://www.egroups.com/group/alt-beam
http://www.egroups.com - Simplifying group communications



Home