# Skepticism, S-R and Control, PCT Research

[From Rick Marken (950524.1000)]

My experiments with Hans' model-based control program have shown me once
again the importance of being skeptical; it is especially important to be
able to doubt what you _want_ to believe most. It is important, that is, if
you want to avoid wasting time trying to explain things that don't exist.

I proposed one such explanation myself (950523.1130) when I said:

The control that appears to exist when ft = .1 and pff = .01 is, apparently,
an artifact; it probably results from the oscillation in u is that is set up
by the the Kalman filter based control of x; with ft = .1 and pff = .01, the
oscillations may just happen to be the opposite of the disturbance that
"comes with the program".

It turns out that the "control that appears to exist" doesn't; it's not an
artifact because it's not anything; Hans' program exhibits no "model- based"
control at all. So my explanation of this "phenomenon" in terms of
coincidental oscillations that "just happen to be the opposite of the
disturbance" is completely irrelevant. It's rather like an explanation of
the phenomenon of "reinforcement" -- a make-believe explanation of a make-
believe phenomenon.

Failure to be skeptical about "model-based control" led others besides myself
to some claims and explanations that can now be seen to be irrelevant because
they are based on the presumed existance of a phenomenon (model-based
control) that does not exist. For example, here is part of Hans Blom's
(950522) explanation of model- based control:

Because gamma and delta are set to values greater than zero, forgetting
forces continuous relearning. And the relearning attempts to subsume the
sine wave into the model. Therefore the parameter estimates become rather
meaningless. But control is fine! ... By the way, "forgetting" in my demo,
including the new one, may be far too drastic and/or done in an
inappropriate way.

This is an explanation of how a model does something that it doesn't actually
do -- control.

Martin Taylor (950523 16:40) falls victim to the same temptation (to explain
something before we know it exists):

if you model the world as introducing the kind of noise effect that the hot
air from the fire would introduce, then your perception can be stabilized in
part by ignoring the noise effects... The practiced hunter is controlling
for the perception that the cursor (foresight) is aligned with where the
target (animal) is.

Martin uses a model that doesn't work (model-based control) to explain a
phenomenon that doesn't exist (controlling for alighnment of the
foresight with "where the animal is" -- the equivalent of controlling xt --
when the perception of that reality is corrupted by noise).

Even Bill Leach (950523.22:22) gets caught up in the excitement:

It seems to me that this thing could control depending upon what the
output of the Kalman filter is capable of doing to the model.

Such a "control loop" seems appropriate to me for systems with poor
quality "perception" but very constrained "xt".

In fact, this thing (Hans' model) cannot control at all so it's makes no
sense to apply it to any task where control is involved.

Seeing these speculations (including my own) about a non-existant phenomenon
helped me understand more deeply what has been going on in the behavioral
sciences for the last hundred years or so. What we have is a lot of
authoritative and important sounding explanations of phenomena that don't
actually exist; they are just misinterpretations of the phenomenon of
control. Attempts to explain, for example, reinforcement (the apparent
selective effects of stimuli on behavior) are equivalent to my attempts to
explain the apprent "control" found in model-based control demo.

Bruce Abbott (950524.1000 EST) --

Bruce:

The signal to stop flying comes from sensors on the foot pads.

Me:

Signalus-response?

Bruce:

Not necessarily, but I don't seem to have a fly/bee wiring diagram here in
front of me at the moment, so it's a little hard to tell.

You won't find evidence against the stimulus-response view in the fly/bee
wiring diagram. You have to look at the whole loop. It may be true that what
is felt at the foot pad affects flying; but then it is also true (from
observation) that flying affects what is felt at the foot pads. So the S-R
relationship between foot pad sensor and flying is part of a closed loop; if
this is a stable, negative feedback (ie. a control) loop, then the perceptual
variable in this loop is under control; what is felt at the foot pads (the
perception) is not a "signal": it is a controlled variable.

By the way, I'm finding a treasure-trove of control system research in this
area. Did you know about it?

Only a bit. We heard about some studies of beavers (reported by G. Cziko)
that showed that the beaver apparently builds dams to control an auditory
perception of "rushing water", keeping that perception at zero.

Next time you're in an academic library, take a look at _The Journal of
Experimental Biology_. It would appear that biologists, at least, are
having little trouble getting control system studies published

There are control system studies and then there are PERCEPTUAL control
system studies. It's the latter that have been tough to publish. There are a
bazillion published studies of manual control in psychology, for example.
Many of these are fine but they don't get at the basic PCT question -- what
perception(s) is the organism controlling. I suspect that many of the
control system studies in biology are equivalent to the "manual control" type
studies in psychology. I would bet that studies done to determine the
perceptual variables that organisms control are as rare in the biological
as they are in psychological literature. My experience is that studies like
the bee study described by Avery and the "baseball catching" study in Science
are the rare exception rather than the rule-- in biology and in psychology.
It may be that there are, in fact, more PCT-like studies in the biological
literature. That would be a pleasant surprise, especially of there were a
fair number of them.

Best

Rick

[Martin Taylor 950524 14:00]

Rick Marken (950524.1000)

if you model the world as introducing the kind of noise effect that the hot
air from the fire would introduce, then your perception can be stabilized in
part by ignoring the noise effects... The practiced hunter is controlling
for the perception that the cursor (foresight) is aligned with where the
target (animal) is.

Martin uses a model that doesn't work (model-based control) to explain a
phenomenon that doesn't exist (controlling for alighnment of the
foresight with "where the animal is" -- the equivalent of controlling xt --
when the perception of that reality is corrupted by noise).

Rick might note that I did not use ANY model in my discussion. I merely
suggested to him that even if Hans' model worked, it would in no way
detract from the notion that what is controlled is perception. Hans
did not provide a counter-example to PCT, which is what Rick assumed

I don't hunt, but I have grave doubts as to whether a real hunter
would try to change the aim as the apparent position of the target moves
with the heat shimmy. I still claim that the hunter is controlling his
own perception.

If you want to take the "out" that the heat waver is faster the hunter can
move the gun, think of another scenario, a spear-fisherman fishing from the
bank of a stream. If he throws the spear at the position where the fish
seems to be, he will die of hunger before catching a fish. The refraction
at the water surface changes the apparent depth of the fish, and any slow
waves on the surface may make it seem to move back and forth.

The skilled hunter catches fish. I claim that in doing so he is controlling
his perceptions. If I understood Rick's earlier comment on Hans' posting,
Rick would say that if the hunter hits the fish with the spear, he is
demonstrably NOT controlling his perceptions, but is instead controlling
an outer-world variable. Success in controlling one's perceptions by means
of actions in the outer world is not incompatible with PCT.

Martin

<[Bill Leach 950525.21:37 U.S. Eastern Time Zone]

[From Rick Marken (950524.1000)]

Rick, I can not deal specifically with Han's model. Model based control
does work.

A very easy example is satellite tracking antennas. The antenna is, of
course, controlled to a position with a very good closed loop negative
feedback control system. The reference for this antenna position however
is generated by a model and in some systems, automatic correction to the
model takes place.

Obviously, should some unanticpated disturbance to the satellite's orbit
occur, the model's ability to track would either be completely destroyed
or seriously impaired depending upon the effectiveness of the feedback
provided by the "correction" system.

Such almost completely open loop control methods probably do not exist at
all in living systems. However, we do "control for" things for which we
have limited immediate perception. In most (but maybe not all) cases, it
is likely that we model as set of discrete immediate perceptions that if
achieved _may_ achieve the goal (I presume that at the lower or at least
lowest levels of control ALL perceptions are controlled with negative
feedback).

Maybe another way of saying what I am trying to express is that when I
drive over to someone's house to see them, I am controlling perceptions
using closed loop negative feedback control to full time but the
perception of meeting someone that is the ultimate control reference for
the whole series is "open loop" and model based.

-bill