Alt-BEAM Archive

Message #13434



To: beam@sgiblab.sgi.com
From: Jean auBois aubois@trail.com
Date: Fri, 14 Apr 2000 22:19:11 -0600
Subject: And now for something completely different...


I've been thinking about what kind of bigger systems could use BEAM
technology as a basis. One architecture that comes to mind would look like:

BEAM Sensor Array --> Conventional Digital Logic As Neural Net -->
Latch --> Policy Logic --> Actuators.

One of the important points about the conventional digital logic is that
we'd be using the same 74... chips that we know and love as a sort of
neural net. Now, a "normal" neural net depends on interconnectivity and
weights applied to every input that comes into a neuron. When you are
using digital logic, you've got on or off and nothing in between, nothing
that could provide weighting. However, I have a sneaking suspicion that
this wouldn't make any difference whatsoever. Instead of levels, what
would be important would be differences in phase and frequency between the
inputs of the elements.

Just to give an example, anyone who has made a two motor photopopper can
attest to the fact that their are two inputs. The result (that is, which
way the photopopper moves) depends completely on the phase of these signals
-- either they are out of phase (essentially by 180 degrees, otherwise
known as inverse) or they are in phase. When they are out of phase -- that
is, one of the sensors receives more light than the other -- then the
photopopper (usually) turns towards the brighter direction. If both are
equally bright or dim, then the photopopper will go more-or-less straight
forward (presuming it has enough light.)

Now, that is an extremely simple example. If you wanted to scale up such a
robot, just replicating the light sensor to the solar engine to the motor
circuit doesn't work very well. On the other hand, a boolean combination
of a set of inputs might produce outputs that are recognizable, are useful
(a "it is bright to the right of me but I've got a touch sensor triggered
on that side and I'm low on power" kind of result.)

Now, at this point, there would be a simple latch. The latch would be
triggered by the overall clock for the device -- its heartbeat if you
will. As a result, the robot would be taking a snapshot of "recognized
situations" and would extend it, stretch it for a while. The clock might
be fairly fast...

The next section would be policy logic -- in other words, if the robots
recognizes a given situation, more-or-less what would you want its
actuators (motors or wheels or whatever) to do. Since the latch holds the
input to the policy logic steady for a while, the robot would get a chance
to actually complete an action before a new situation comes along.

I fear that this scheme might fall prey to the failing Tilden expressed
about computer controlled robots -- what do you do if a situation occurs
that you don't recognize or don't have a policy for (or even worse, your
policy screws things up). On the other hand, the rich processing that
might be available from the neural net section -- comparing and contrasting
signals for frequency and phase in real time (which continues in real time,
unaffected by the latch) -- might be quite powerful.


jab

p.s. there could be "true" feedback from the output of any of the sections
to the input of any of the others. The behavior chosen, for example, might
be factored into what the robot is "perceiving" (not unlike implex) or as
"memory" if it came back around to the latch.

Home