Control:Inside and Out -- So What?

CSGnet Archive CSG1996
1

[From Rick Marken (960622.1215)]

Regarding my earlier post [Rick Marken (960622.1040)] I forgot
to mention the point of the whole thing.

The point is that, even though a higher oder system can perceive
and control the "informativeness" of the relationship between
a perceptual signal and other variables in a control loop such
as the control systems own outputs and possible disturbance
variables, this has nothing to do with how the control system
itself controls. "Informativeness" is something that can be perceived
_about_ the behavior of a control system (when the system is looked
at from a higher level); but it is of interest only to the system
doing the observing; it has nothing to to with how the way the observed
control system actually operates. For example, when we observe a
control system in a tracking task, we can see that its outputs mirror
variations in the disturbance to to the cursor. We can see this mirror
relationship but the system doing the tracking cannot. The mirror
relatinship that we see is an _irrelevant side effect_ of controlling.
Saying that the information in perception is essential for control is
like saying a mirror image relationship between disturbance and output
is essential for control. It's just a mistake -- but one that can be
costly because it can (as it has for the last 100 years) point research
on living systems in exactly the wrong direction.

That's why it matters.

Best

Rick

2

[From Bruce Gregory 960624.1100 EDT)]

(Rick Marken 960622.1215)

Regarding my earlier post [Rick Marken (960622.1040)] I forgot
to mention the point of the whole thing.

The point is that, even though a higher oder system can perceive
and control the "informativeness" of the relationship between
a perceptual signal and other variables in a control loop such
as the control systems own outputs and possible disturbance
variables, this has nothing to do with how the control system
itself controls. "Informativeness" is something that can be perceived
_about_ the behavior of a control system (when the system is looked
at from a higher level); but it is of interest only to the system
doing the observing; it has nothing to to with how the way the observed
control system actually operates. For example, when we observe a
control system in a tracking task, we can see that its outputs mirror
variations in the disturbance to to the cursor. We can see this mirror
relationship but the system doing the tracking cannot. The mirror
relatinship that we see is an _irrelevant side effect_ of controlling.
Saying that the information in perception is essential for control is
like saying a mirror image relationship between disturbance and output
is essential for control. It's just a mistake -- but one that can be
costly because it can (as it has for the last 100 years) point research
on living systems in exactly the wrong direction.

That's why it matters.

Thanks. This and your earlier post make clear what seemed to me
to be errant nonsense. The entire exchange involving you, me,
Bill, Martin, and Hans is a vivid example of why communication
is so damned difficult and why the principle of charity (the
notion that an interlocutor is making _some_ sense, no matter
how much it may not be obvious)is one that it is well to fall
back on when all else fails!

Regards,

Bruce

3

[From Bruce Gregory 960624.1340)]

(Rick Marken 960622.1215)
(Rick Marken 960622.1040)

Let me try my version of what I understand you to be saying. We
can use as the thermostat as an example. _All_ the thermostat
_knows_ is the difference between the temperature recorded by its
thermometer and its reference level. The thermostat knows nothing
about termperature perturbations. Nor does it know anything about
its actions as far as turning the furnace on or off. (It turns
the furnace on and off, but receives, and needs, no signal
verifying what it has done, or whether, in fact, it has done
anything.) We, on the other hand, can observe the temperature
fluctuations and the hear the furnace turning on and off, (we
have information) but we are not part of the system controlling
for the heat of the house (except insofar as we set the reference
level for the thermostat). O.K.?

Regards,

Bruce

4

[From Rick Marken (960624.1200)]

Bruce Gregory (960624.1340) --

Let me try my version of what I understand you to be saying. We
can use as the thermostat as an example. _All_ the thermostat
_knows_ is the difference between the temperature recorded by its
thermometer and its reference level.

Precisely!

The thermostat knows nothing about termperature perturbations.

Well, it knows nothing about the _cause_ of the perturbations. It knows the
perturbations as changes in the percetual signal.

Nor does it know anything about its actions as far as turning the furnace on
or off.

Right!

We, on the other hand, can observe the temperature fluctuations and the
hear the furnace turning on and off, (we have information) but we are not
part of the system controlling for the heat of the house (except insofar as
we set the reference level for the thermostat). O.K.?

By George, I think you've got it!

Actually, we (observers) can use instruments to measure (perceive)
fluctuations in the thermostat's perceptual signal, p(t), and heater output,
o(t). We can also derive (logic level perception) the existence of a
relationship between p(t), o(t) and the room temperature, d(t): the
relationship is

(1) p(t) = f(d(t)+ o(t))

If we know the form of f(), we can determine the value of d(t) from our
measurements of p(t) and o(t). We can then say (as Martin says) that we are
able to do this because there is "information about d(t) in p(t)". This is
obviously just a weird way of saying that, if we know that p(t) = f(d(t)+
o(t)) and we know o(t), p(t) and f(), then we can solve for d(t).

In order to use the "information in p(t) about d(t)" you have to know o(t),
p(t) and the functional relationship between p(t), d(t) and o(t). Clever
observers of the thermostat can, indeed, find d(t) given p(t), o(t) and
equation (1). They can "extract the information from p(t) about d(t)". I
prefer to say that they can solve for d(t) given p(t) and o(t).

Of course, the thermostat itself can't do this "information extraction"
process -- and it doesn't need to because control doesn't require that the
control system know about o(t) or the relationship between o(t), p(t) and
d(t). All a control system has to be able to do in order to control is
continuously generate output based on error: o(t) = f(r(t)-p(t)). If f() is
the right function to do the job (a function acquired through learning) then
o(t) will have a negative feedback effect on p(t); error (r(t)-p(t)) will
continuously cause outputs that reduce error. There is no "extraction of
information from p(t) about d(t)" nor any solution of equations to determine
d(t) given p(t) and o(t).

The only reason I can imagine why people might want to believe that control
systems work by "extracting" or "using" or "consuming" information in p(t) is
if they want to maintain the idea that input _drives_ or _guides_ behavior.
That is, the idea of information in p(t) about d(t) is consistent with an
input- output view of behavior, a view of behavior that is completely
inconsistent with PCT. According to the input-output view, p(t) is a _cause_
of behavior; according to PCT, p(t) is _controlled_ by behavior.

Best

Rick

5

[From Bruce Gregory (960624.1555 EDT)]

(Rick Marken 960624.1200)

Actually, we (observers) can use instruments to measure (perceive)
fluctuations in the thermostat's perceptual signal, p(t), and heater output,
o(t). We can also derive (logic level perception) the existence of a
relationship between p(t), o(t) and the room temperature, d(t): the
relationship is

(1) p(t) = f(d(t)+ o(t))

If we know the form of f(), we can determine the value of d(t) from our
measurements of p(t) and o(t). We can then say (as Martin says) that we are
able to do this because there is "information about d(t) in p(t)". This is
obviously just a weird way of saying that, if we know that p(t) = f(d(t)+
o(t)) and we know o(t), p(t) and f(), then we can solve for d(t).

I'm not sure why you find this wording "wierd", but I'm pleased
that you have room for Martin's idiosyncracies,

The only reason I can imagine why people might want to believe that control
systems work by "extracting" or "using" or "consuming" information in p(t) is
if they want to maintain the idea that input _drives_ or _guides_ behavior.
That is, the idea of information in p(t) about d(t) is consistent with an
input- output view of behavior, a view of behavior that is completely
inconsistent with PCT. According to the input-output view, p(t) is a _cause_
of behavior; according to PCT, p(t) is _controlled_ by behavior.

I don't perceive this motive in Martin's writings, but he can
plead guilty or innocent without my help.

Thanks again for your time and illuminating response.

Regards,

Bruce

6

[Martin Taylor 960625 14:00]

Rick Marken (960624.1200)

Referring to the recently disinterred demonstration that the disturbance
waveform can be recovered from the perceptual signal, given knowledge of
the reference signal and the output/feedback function, Rick (1996) says:

If we know the form of f(), we can determine the value of d(t) from our
measurements of p(t) and o(t). We can then say (as Martin says) that we are
able to do this because there is "information about d(t) in p(t)". This is
obviously just a weird way of saying that, if we know that p(t) = f(d(t)+
o(t)) and we know o(t), p(t) and f(), then we can solve for d(t).

In order to use the "information in p(t) about d(t)" you have to know o(t),
p(t) and the functional relationship between p(t), d(t) and o(t).

Rick Marken (1993) at the time of the experiment, before he knew the results:
+RIck Marken (930331.2100)

+Your method does not use knowledge of o to extract
+the information about d from p -- so I consider it quite a fair method.

Which is correct? Marken puzzling over what would happen when the simulation
was not yet run, or Marken reconstructing from memory what _must have_
happened, because he knows that the reconstruction is possible given o.

+It does use information about the output function (that transforms error
+into the variable that affects the perception). But I think you imagine
+that it is the output function that extracts (recovers) the information
+from p. In fact it doesn't -- but that's what the simulations will show.

In fact, it did, and that's what the simulations showed.

In 1993, Rick knew that knowledge of the output waveform would be required,
and that _because we were not going to use it_, therefore our simulation
would fail and we would see that the perceptual signal carried no information
about the disturbing influence. He was happy about that.

In 1996, Rick asserts that we did use knowledge of the output waveform,
which was why we were able to do the reconstruction successfully.

Actually, we used three things: (1) f(), The unvarying output-to-CEV function
(which I sometimes label the output/feedback function), (2) The reference
waveform r(t), and (3) the perceptual signal waveform p(t).

They suffice to reconstruct the disturbing influence, to within the accuracy
of control.

Selective memory is funny, isn't it:-)

Martin