Imagination construction!

[From: Chris Love (920914.1300)]
[To: Bill Powers (920910) and Rick Marken (920910)]

  Thanks for the encouraging posts!
Bill, I have already started working on this *new* model. And
I don't know either when it will be complete, but I can't see
it taking too long (< 1 month) since all the *control objects*
have already been created, such as linking objects, connection
routines, the basic ECS object, etc.. But in development work,
as I'm sure your aware this is never a sure thing!

Perhaps what you feel is missing is the pairing up. In fact, it's

still there...

I see Bill. This really does appear that, for the purpose of that
n-level ECS, it is imagining without even knowing it. Kind of like
drifting in the middle of a lecture - it's nor until you bring
yourself back to attention (reality) that you realize you ever left!
I say this because that n-level ECS is being provided all information
in a regular manner as if nothing unusual was happening to it,
although the information (percepts) are really imagined ones.

  The problem: By allowing those n-1 level ECS' to provide the
imagined percepts to your (focus) n-level ECS you are getting
(undesired?) influence from neighboring n-level ECS'?

How? This is occuring because those n-1 level ECS' are also
receiving outputs from the neighboring n-level ECS' in the
process of generating the net input refernce sum.

  Do you want this? If you wish to evaluate the effectiveness of
that n-level ECS, I think not since you now have external
uncontrollable influence from neighboring n-level ECS' through
an indirect path. If this was your intention then I do not
understand the reason, or possibly overlooking something. Please
explain.

  In my proposal, you short circuit the ECS that you want in
imagination mode. No influence from neighboring n-level ECS'
occurs. I know it does in the *global* sense through the
environment **eventually** as recieved refernces, but this is
far enough removed, I feel, to be of little effect. When that
n-level ECS imagines you want it to try to perceive doing well,
that's the objective, right? So, in achieving, or attempting to
achieve this objective we must vary it's direct weights in relation
to it's imagined percepts such that it reduces its error. THis
results, of course, in a better controller. THis is my hypothesis.

  Bill, if you removed the influence of those other n-level ECS'
contributions to those other n-1 level ECS refernce sums then I
think we have agreement on the imagination objectives.

I'll be interested in your implementation of an imagination loop

that can adjust perceptual weights...

  Me too! I'm open to any suggestions. My baseline proposal is to
use some form of Hebbian stlye learning (also known as the Widrow-
hoff rule). As applied to
PCT it goes something as follows:
  Delta weight = learning rate* connection wt. * error signal
  New wt. = Old wt. + Delta wt.

  Comments:
   The learning rate is typically set to a small value between
0.1 and 0.3.
   The connection weight is the exisiting weight (wt.) of the
connection of interest.
   The error signal is that signal eminating from the comparator.

  What does this do?
   Well, as I see it it correlates particular references with
particular percept signals. In wording the Hebbian rule says
something like, "... if two connecting units are correlated
then strengthen that connection." THislowers the threshold
necessary to *fire* that receiving unit in the future under
similar circumstances.

  Is this the right thing to do?
That depends on wether their is correlation between the refernces
an the percepts? And whether this correlation is desireable to
learn, i.e., will it make controlling better?

  If anyone has any comments, improvements, either to prove or
disprove the above proposal or anything else - great! Let's have
at 'em!

···

------------------------------------------------------------
Rick Marken-

Whatever throws the switch must have reason for doing it...
the reason is to control something.

... it would be nice to have a system that alters the imagination

connections through the perceptual and output system...

the imagination swithch is probably the output of a control system..

Sounds good to me too Rick.
  I also thought that another possibility would be to allow the
ECS to detemine that it needs to imagine. THis would happen, like
me, when I work too hard and get (temporarily) *brain dead* and
my mind automatically swithches to imagination mode to try to
solve *the problem*, I guess subconciously.

When I say "allow the ECS to determine", I should specify that
the *swithch* is internal rather than coming externally from
another ECS. I would like to avoid this path of exploration
first if i can since it requires, i feel, a dedication to
function, i.e., that ECS' objective is to *only* know when to
send a *swithch to imagination* signal to other ECS'. I, as
a designer, now have to worry about how to make such an ECS do
just that. I would prefer to develop, if i can, a *almighty*
ECS that is VERY GENERIC in the sense that everything that must
happen in an ECS or to an ECS can be handled *EXCLUSIVELY* by
that ECS. It does not require outside influence to make decisions
on what its actions should be.

  As an engineer, if it turns out that your proposal works ok,
great I'll implement it tommorow. I'm in no way biased to either
approach - just one that works and is reliable, reproduceable,
and expalainable (is that a word?)!!! I think at this point,
however, that the almight ECS may be the way to go. What do you
think about this?

>figure out how to do experiments that would guide the modelling
of imagination...

  I can't imagine these expts. now either but you can't always see
the end of the tunnel; you just have to beleive that it will appear
before you drop!

Thanks,
Chris. (DCIEM)