[From: C. Love (920916.13:30)]
[To: Whomever is following these proceedings]
I have come up with a simplistic approach to imagination
based primarily on the way Bill proposes imagination to
occur, i.e., the three levels of interaction of ECS'. Now
this says nothing of how or why imagination is triggered by
that level (n+1) ECS, but it does say what to do when it
happens.
Let's set up a scenario...
OK, the level (n+1) ECS, we'll call ECS top.
Next there are three lower level (n) ECS', which we'll call
ECS M1, ECS M2, and ECS M3, (M=middle).
Lastly ther are also three even lower level ECS', which we'll
call ECS B1, ECS B2, and ECS B3, (B=bottom level).
So the ECS (B1 to B3) are the ones that are short circuited
according to Bill's model, and the ECS' (M1 to M3) are the
ones imagining. (For the moment let's assume that those other
level (n+1) ECS', which did not request imagination to occur
are simply shut off during this process for simplicity's sake.
That's the scenario. Now what happens...
One last point, all weights are continuous value and fully
interconnected between layers and upon initialization have
small random weighted values between {-1,+1}.
Let's start be assuming the the error of ECS top is >0
and error is calculated by the rule E=R-P, where
R=net weighted sum of references
P=net weighted sum of incoming percepts.
ECS top>0 ==> R-P>0 ==> R>P (==> = implies)
the necessary action here is therefore to
increase P and decrease R.
Which is better to do? What are the implications? I'm
not sure at this point. We must remember WHY we are doing
this - it is because we haven't been able to control our
percepts entering into ECS top.
The course of action I'm proposing to alleviate this
mismatch is the following:
To increase Net Percept of ECS top do the following -
1) Consider an arbitrary percept weight, Wn, entering
into ECS top along with thhe percept signal, Pn.
2) If Wn*Pn>0
If Wn>0 and Pn>0
Wn=Wn+Delta
Else
Wn=Wn-Delta
Else
If Wn>0 and Pn<0
Wn=Wn-Delta
Else
Wn=Wn+Delta
3) What's Delta?
Delta is, I feel, something related to a portion
of the error.
I propose the following:
Delta = L*ABS(E)
L= proportion factor, say 0.2
E= error of ECS top.
This same sort of process would be applied to
the outputs of ECS top also.
I think we would have stability in finding the
equilibrium due to the porportion of the error, L,
being less than one.
Any comments?
(Keep in mind this is more or less a little
idea that popped into my head,but I thought
it was worth sharing anyways.)
Chris.