Newsgroups: comp.robotics
Path: brunix!uunet!europa.eng.gtefsd.com!howland.reston.ans.net!sol.ctr.columbia.edu!The-Star.honeywell.com!cs0.dasd.honeywell.com!burro.dasd.honeywell.com!astro.dasd.honeywell.com!skp
From: skp@astro.dasd.honeywell.com
Subject: Re: Micro-coding Robots?
Message-ID: <1993Jul26.104242.1@astro.dasd.honeywell.com>
Lines: 95
Sender: news@burro.dasd.honeywell.com (USENET News System)
Organization: Honeywell Inc. DASD
References: <22ooc0INNe10@uwm.edu>
Distribution: world,local
Date: Mon, 26 Jul 1993 17:42:42 GMT

In article <22ooc0INNe10@uwm.edu>, rick@ee.uwm.edu (Rick Miller) writes:
> The notion of 'microcoding' robotic behavior has been knocking around in
> my head since I began fidgeting with neural networks... even though those
> two topics have almost nothing to do with each other.
> 
> The idea is that instead of putting a microprocessor on-board a robot, I'd
> simply put a *memory* chip in it.  A little support circuitry (a clock to
> latch the address lines, maybe some R-C delays, power MOSFETs for motors)
> would be sufficient, the memory chip should probably be an EEPROM.
> 
> The 'address' lines would be used as inputs and the 'data' lines as outputs.
> Inputs would necessarily be binary, like micro-switch closures or maybe
> threshold-crossings or the differentials (change) of analog sensors, and
> some outputs could be fed back to inputs to implement counters, branching,
> and subsumption.  You could divide up the input/output lines any way you
> like, maybe even making some into a program counter or signifying the
> "state" of the machine...
> 
> I originally concocted this idea to implement neural-net behavior.  Since
> digital neural-net computations are quite intensive, I figured I'd train
> the network simulation, then feed it all possible inputs while storing the
> outputs as data on the memory chip.  Thus, the chip would *behave* exactly
> as the neural-net... but instantaneously, without run-time computation.
> 
> It then occurred to me that I didn't *have* to use a neural-net algorithm
> to "program" such a device.  I could do it with *any* algorithm which gives
> digital outputs for digital inputs.  The point is that the *computations*
> are already done, the results at run-time are practically instantaneous in
> comparison.
> 
> The difficulties I see are:
>          .
>          .
>          .
> 
> As per my original notion, this type of robot would be an *ideal* test-bed
> for neural-net robotics...  no run-time computation, little analog circuitry,
> practically instantaneous action!  All that laborious computation would be
> done before the 'bot ever powers-up.
> 
> Any thoughts or comments?  Anyone?
> 
> RICK MILLER            <rick@ee.uwm.edu> Voice: +1 414 221 3403 FAX: -4744
> 16203 WOODS            Send me a postcard, and I'll return another to you!
> 53150-8615 USA         Sendu al mi bildkarton, kaj mi redonos alian al vi!

Rick-

	Last (school) year I was working on a similar project in which I was
going to use an Artificial Neural Network for a simple robotic guidence system,
using microswitches and possibly a hard wired neural network.  As the project
proceeded I talked to a good friend here at Honeywell (who works with EEPROMS)
about my project and he said that you could use EEPROMS to do the actuall
computation's of the neural network.. So we sat down and looked at many
possibilities, and found out that inorder to enable the EEPROM to do the
actuall processing, you would need many EEPROMS and it would be a reall complex 
problem getting the floating point accuracy needed for the neural net matrix,
so then we explored the option of runing the neural network and getting the
results and then just storing them on the EEPROM.. Well this turned out to be
very simple.  The major problem I found was that we would work out all the
possible solutions to the problem in our heads without using the neural network, and
then I would go to the neural network and try to train it so that it would give
me the same responses...  I concluded that the way in which I was using the
neural network in this particular application was totally useless... But I
coulld see possible problems in which this technique would be extreamly
usefull.  But the only problem I see with it is that it is a totally static
system, unless you go ahead and design the EEPPROM to actually change the
neural matrix values, and then do all the computation to presend situations to
the neural network...  A LOT OF WORK...  

	The EEPROM programer that we were going to use ran under Windows 3.x
and the programer it's self was a little box that was attached via either the
serial or parallel port...  I'm not sure exactally how much it costs, but if
you would like more information about this system, just let me know...

That's all folks...
Sean Patrick  

   _____________________________________________________________________
  /\                         ___    _   _   _____                       \
  \_|                          _|    | / /    __ |                       |
    |                         |_     |/ /      __|                       |
    |                        _  |       \     |                          |
    |                        ___|   __^__\  __|                          |
    |                                                                    |
    | Kenneth Sean Patrick          e-mail  : skp@cs0.dasd.honeywell.com |
    | HONEYWELL Inc.                tel     : (505) 828-6533             |
    | Defense Avionics Systems Div. fax     : (505) 828-5500             |
    | 9201 San Mateo N.E.           location: Sim Lab                    |
    | Albuquerque, New Mexico                                            |
    |   _________________________________________________________________|___
     \_/____________________________________________________________________/
  
              

