[Martin Taylor 991022 15:48]
[From Bob Christensen (991022.1020 PT)]
Bill Powers (991021.0710 MDT)
A perception is not necessarily an error signal just because it represents
something that we don't like.How is it that we know a perception is one we don't like or want?
Followed by a discussion of why we should consider that error values
should ordinarily be propagated up the hierarchy as inputs to the
perceptual input functions of higher-level units. It's worth contemplating
this idea, quite apart from whether it violates the standard HPCT model.
There are two ways to consider the question (at least):
(1) The inputs to higher-level systems may be derived from both the
perceptual signals and the error signals of the lower-level systems;
(2) The inputs to higher-level systems may be derived ONLY from the
error signals of lower-level systems.
In analysing the implications of these two possibilities, we have to
think also of the low-level reference signal, because e = r-p. If we
allow any two of the three signals to form part of the input to the
higher levels, the third is implicit, since the higher-level perceptual
function might be constructed to form it from the two that are available.
Therefore proposal (1) is formally identical to allowing the higher-level
perceptual input function to see both the lower-level perceptual signal
and the lower-level reference signal. Proposal (2) is different.
We can probably dismiss proposal (2), because there is no value in allowing
only the error signal to propagate upward. Here's why. Any elementary
control unit that functions properly will act until its error signal is
zero. If there is another unit trying to control the error signal to
a different value, there is inherent conflict between the two (it's a
different kind of conflict than the one we normally consider, but it's
conflict nevertheless). So nothing at any higher level can do more than
act to assist its own input to become zero--in other words, the higher
level system can only duplicate the function of the lower, and the
higher-level systems CANNOT have varying reference values for their
inputs. Their reference values MUST be zero, or they induce conflict at
the lower levels.
We cannot dismiss proposal (1) so readily. Indeed, with the addition of
a switch or some other non-linear combination scheme, it is identical
to the hypothesized "imagination connection." In the imagination
connection, the reference value sent down to a lower-level control
unit is substituted for its perceptual signal, as if the unit had
successfully controlled its perception, but without requiring the
lower unit to act on its environment.
Since the reference value for the lower-level control unit can
be derived from the values of its perceptual signal and its error
signal, the imagination connection _could_ work by propagating the
error signal back up the hierarchy. But it would be awkward, since
using such an error-based imagination signal would require a
switch to prevent the lower-level system from acting in addition to
the switches already inherent in the mechanism. It's possible, but
less plausible than the "standard" connection.
How, then does error get perceived? I think Bob hints at that:
For instance one cannot immediately make a
dent in one's car go away, but one might start imagining the dent not being
there. This is imagining the reference, isn't it? Now I can toggle backand forth between Perception (dent), and reference (no dent) leading to a
sequence of the two perceptions stored in memory. Now I can perceive the
sequence simultaineously like when perceiving both the A and the B at the
same time in AB. Now I can perceive that the two perceptions (regular
perception, and replayed reference) are different, just like I can tellthat A is different from B.
This sounds very like the way the imagination connection is assumed to
work, except that the imagination connection _also_ allows the higher
systems to determine the results of lower-level action if the lower-
level controls were to work properly. It could well be, as Bob says,
that a discrepancy here between current perception and the downgoing
reference signal that is returned as an imagined perception could
indicate "what we don't like."
The gist of what I'm trying to say is that the "standard" imagination
perception gives the same formal results as does upward propagation of
error signals, and is simpler in an engineering sense.
ยทยทยท
---------------------------
BUT...
As Mark Lazare (991020 1145 am) pointed out, there IS a body of research
in which the error signals themselves are the subject of study. Tom
Bourbon studied a kind of adaptation in which one control system used
the error signal of another as input to its Perceptual Input Function.
The output of this error-perceiving control unit was not, however, to
the reference input of the other control unit. It affected (if I remember
correcty) the gain of the other unit. It was a local tuning system that
attempted to optimize the performance of the other system. Mark called
it a reorganizing control, but I think that this is an improper name
for it.
I mention this mainly to show that it _is_ reasonable to consider the
place within the HPCT mechanism for units that sense the error signals
of other units. But it is less reasonable to place the error-sensing
units directly within the main perceptual control hierarchy. If Bourbon
is right, they tune the momentary operations of the main hierarchy. It
is tempting to equate them also with attention control. But I won't
pursue that here.
Martin