[Martin Taylor 921201 16:00]
(Bill Powers 921130.1300)
Martin Taylor (921130.1340) --\
Some very well-made points about planning. If you can plan, you don't
need to control. I would add that at any level, there are uncontrolled
perceptions that can introduce disturbances, so that even if all
controlled perceptions at a given level are perfectly controlled, the
higher levels still work with partially-uncontrolled inputs.
I'm afraid I don't follow. If ECS A provides as its output reference signals
for ECSs X, Y, Z, ..., and they provide the desired percept (perfectly,
by hypothesis), how can the percept of A be uncontrolled? The hypothetical
extreme situation that I was proposing was that A would be assured that
the percepts it demanded of X, Y, Z, ... would be delivered. Any disturbance
applied by the world would be eliminated by the perfect control at the
lower level, wouldn't it? If the sensory values applied to the PIF of X
don't change, then the output shouldn't change. (I should think this would
apply even when we are at time-based levels such as sequence or program,
but my intuition gets a bit dicey under these conditions).
Or are you referring to the degrees-of-freedom problem, that says that
not all perceptions are controllable simultaneously, but that any perceptual
signal is potentially an input to any higher ECS? If so, I agree, but it
raises an interesting point about imagination. I have thought of imagination
as a "what if" kind of exercise, issuing virtual output signals that would
bring the percetual signal colser to the reference if they were real. Now
we are talking about a "what if I did nothing" exercise, imagining what the
perceptual signal would be, and comparing it with what is to get the new
information. The information rate is presumably somewhat higher for
uncontrolled perceptual signals than for controlled, but it is still
less than a naive description of the temporal pattern of incoming sensory
data would require.
It [passing the error up vice the perceptual signal] has some interesting
possibilities, including the ability to continue controlling with
intermittent input data. Such a system is in the imagination mode ALL
THE TIME, but there is provision for continually comparing the model's
behavior with real perceptions and using the difference to modify the
model. This requires error signals from lower level systems to be passed
upward, not perceptual signals.
I didn't want to move the input to the PIF from the lower-level perceptual
signal to the lower-level error signal, because to do so seems to me inherently
less accurate than using the perceptual signal directly. My point was that
the imagination mode provides the subjective probability distribution in
the ECS that determines the information being supplied by the perceptual
signal, and if the lower levels are doing what they are asked to do, then
they are providing only a very little information to the higher ECS. I
would hazard a guess that in a well organized system in good control, the
information rate is reduced by a factor of G/N at each level, where G is
the average (geometric, I suppose) gain of the lower level system, and N
is the number of lower-level systems contributing sensory input to the
higher-level system. G/N then would be an appropriate inter-level slowing
factor for simulations. There would have to be a modification of this
for uncontrolled inputs.
Note that the low information rate can be achieved by connections that
provide information at a high rate intermittently, just as readily as by
systems that provide it continuously but slowly. There's nothing in this
that argues one way or another for whether the sensory data comes from
lower-level perceptual signals or error signals, but it does demand that
the higher-level ECS be in imagination mode all the time.
Martin