Control and the Test;Evolution; cause-effect

[From Bill Powers (960215.1200 MST)]

Martin Taylor 960215 11:00 --

     If you have an observation of a variable that is clearly being
     disturbed by some unobservable cause, but which nevertheless
     persists in returning to or close to its original position, how can
     you tell whether the variable is or is related to the CEV of some
     control system?

In replying to Rick Marken's comments on the Test, you saw his reply as
diametrically opposite to mine:

     I don't know that I agree with either of you, but you do seem to
     represent opposite ends of the spectrum of possible answers.

In fact, our answers were completely consistent.

If all you see is the variable varying, you have no idea why it is
varying. If you see it return to its original state, you have no idea
why it did that. The tides return to their original state twice a day
(if you measure from mean sea level). The earth returns to the same
sidereal position once a year. A traffic light changes to red, and then
back to green again. Merely watching a variable change does not tell us
anything except that it changes.

We are still mired in the problem of what "disturbance" means. You have
always questioned my insistence on defining a disturbing variable that
is separate from the CEV; your own examples of disturbances always seem
to start with a change in the CEV (this goes clear back to the original
arguments about information, in which your idea of a disturbance was an
instantaneous step-change in the CEV). Hans Blom clearly defines a
disturbance in the same way you do: to disturb the position of his
Brownian particle, he seizes it and forces it into a new position, then
releases it. This is not how disturbances are applied in the Test.

My reason for wanting a separate disturbing variable is simply that a
change in the disturbing variable does not necessarily produce any
change in the CEV. It may or it may not, depending on whether control is
present and on the behavior of the reference signal. If you define a
disturbance as a change in the CEV, you are either breaking the loop or
giving up the ability to manipulate the CEV.

When we do the Test, we are not interested in the behavior of the
putative CEV _per se_. What we are interested in is the _relationship_
between a disturbing variable and the state of the CEV. In the absence
of control, varying an independent disturbing variable will result in
some amount of change of the CEV. When control is present, _the same
variation_ in the independent disturbing variable will result in a far
smaller change in the CEV. Naturally, to do the Test you must know the
behavior of the disturbing variable, independently of the behavior of
the CEV.

If you think of a disturbance as a change in the CEV, then the idea of
the _same_ disturbance having a _smaller_ effect on the CEV is nonsense.
If a disturbance is a change in the CEV, then the same disturbance is
the _same_ change in the CEV -- it has to be, under this definition of
disturbance. So you can't do the Test by arbitrarily changing the CEV.

To apply a disturbance when doing the Test, you have to vary something
in the environment other than the CEV. For example, to apply a
disturbance to the angle between the vertical axis of a standing person
and the horizon, you would have to vary a _force_ applied to the person
-- not the angle. The reason is that the force is only an influence on
the angle, and it may or may not have effects on the angle, depending on
what other forces are present and how they vary. If you vary the angle
directly as a way of applying a disturbance, you make it impossible for
other forces to be influencing the angle at the same time: you break the
loop, if one exists.

A good part of our failure to connect on this subject, I think, is the
difference in our fundamental orientations toward physical processes. I
think of them first as continuous analog processes, with discrete
changes being only an extreme limiting case, never actually reached in
real macroscopic systems. You (and apparently Hans as well) start from
the other side: you speak of physical processes as fundamentally
discrete in nature, with the continuous case being reached only under
exceptional conditions.

As a result, you and I think of "applying a disturbance" very
differently. From your point of view, as I interpret it, one would apply
a step-force to the standing person, and this force would instantly make
the angle start changing, before any control system could react to the
change. As a result, you don't see any important difference between
defining the disturbance in terms of the applied force and in terms of
the resulting change in angle. As I see the world, applying a force
disturbance would entail (for example) bringing my hand in contact with
the person, and gradually raising the force from zero to some specific
amount over a finite period of time, with the applied force starting at
zero and rising smoothly to its maximum value. I am prepared to find
that as the force increases, a counterforce develops, preventing any
significant change in the angle.

And of course, since I am trying to see whether the angle is under
control, I would be sure to increase the force slowly enough so that if
the standing person was going to resist it, the resistance could develop
right along with the rise in disturbing force, with only an
insignificant lag. If I pushed sharply and strongly, any control that
existed would be overcome and the angle would change radically.

I am not interested in sending the person sprawling to the floor. If
there is control, I want to see it working, not to thwart it. The point
of the Test is to give the person a chance to oppose the effects of the
disturbance, if this is going to happen at all. The control system, if
there is one, should be able to continue working normally all during the
Test.

Resistance to disturbance is not seen just in deviations that are
subsequently corrected. It is seen even more clearly in the steady
state. A person standing erect in a steady wind is not deviating from
vertical and then returning to vertical; what we see is a steady small
angle of deviation from the vertical that is just enough to offset the
wind force. If we look only for deviations followed by corrections, what
we will be seeing is mainly _lack_ of control -- we will be observing
the effects of transient disturbances that have bandwidths outside the
bandwidth of control. That can give us useful information about the
dynamics of control, but it is not the most direct way to establish the
existence of control. The existence of control is best found through
steady-state relationships.

I don't know whether we will ever see eye-to-eye on this matter. Our
basic orientations are just too different.

···

-----------------------------------------------------------------------
Remi Cote 960215.1300 est --

     It is not pessimistic to explain why human are human and
     distinguish themself so dramaticaly from the rest of living
     species. It is the essence of what I get on this list.

One of the messages I hope we get across on this list is that ALL
organisms are control systems, from the amoeba on up.
-----------------------------------------------------------------------
Chris Cherpas (960215.1002 PT) --

     Controlling for a concept of linear (IV-DV) causality is often
     cited here as an impediment to seeing circular causality. The
     situation is perhaps analogous to the perceptual experiment in
     which a picture looked at from one perspective appears to be a
     young woman ("wife") but as an older woman ("mother-in-law") when
     "viewed differently." Both views are "correct," but cannot be held
     at the same time.

I disagree: the two ideas are as different as they can be. In lineal
causality, we have

(1) effect = f(cause)

In circular casuality, we have

(2) effect = f(cause,effect)

This is not just a matter of "perspective." There is a fundamental
difference between systems that can be described in these two very
different ways. A system that can be described correctly by (1) can't be
described correctly by (2), and vice versa. The feedback loop (or its
absence) changes the basic properties of the system.

Just to make this plain: if you have a system which is correctly
described by (2), then seeing it as a system of type (1) is not an
option; analysing it as if it were a type (1) system would simply be
incorrect.
-----------------------------------------------------------------------
Best to all,

Bill P.

[Martin Taylor 960215 18:00]

Bill Powers (960215.1200 MST)

We are still mired in the problem of what "disturbance" means.

Nothing in your posting leads me to believe that there is any difference
between us about what "disturbance" means.

You have
always questioned my insistence on defining a disturbing variable that
is separate from the CEV;

Never have.

your own examples of disturbances always seem
to start with a change in the CEV (this goes clear back to the original
arguments about information, in which your idea of a disturbance was an
instantaneous step-change in the CEV).

A breathtaking misreading!! And after all these years!!!

I always take for granted that there is a separate disturbing variable, or
many of them, but that since it is invisible to the control system's
perceptual function, the only aspect of interest is the disturbing influence
on the CEV. That's "d" in the usual equations about which we often talk.
Why bring in the red-herring of what causes "d" (unless it is a Tester)?

Hans Blom clearly defines a
disturbance in the same way you do: to disturb the position of his
Brownian particle, he seizes it and forces it into a new position, then
releases it.

That's one of his proposed mechanisms. Not a definition. He also proposed
making the particle magnetic and subjecting it to a varying magnetic force.
So did I, and I also suggested blowing on it, etc...

I'm glad you followed with a long dissertation on how you see disturbing
variables and disturbing influences. It may help for the archives. But I
see nothing in it that I could not have written myself, had I the linguistic
facility.

For example:

For example, to apply a
disturbance to the angle between the vertical axis of a standing person
and the horizon, you would have to vary a _force_ applied to the person
-- not the angle.

Of course.

If you vary the angle
directly as a way of applying a disturbance, you make it impossible for
other forces to be influencing the angle at the same time: you break the
loop, if one exists.

Right. Another way of seeing this is to use the analogy Bruce (I think) did
a couple of weeks back, by looking at zero and infinite impedance sources
and loads. If you are going to examine the parameters of a real system,
you can't load it with zeros and infinities.

How do you get the idea I think differently?

A good part of our failure to connect on this subject, I think, is the
difference in our fundamental orientations toward physical processes.

If there is such a difference, I do not see it in your writing. Or in mine.
Perhaps you think that talking about Nyquist rates is equivalent to deciding
that the underlying process is discrete. It isn't. The Nyquist rate is a
rate at which sampling a continuous waveform allows an exact reconstruction
of the continuous waveform. To work with a set of samples at or above
the Nyquist rate is to lose no information from the continuous waveform,
and in practical situations is often easier. It's certainly easier in
simulation to treat the sampled waveform than the continuous;-)

I
think of them first as continuous analog processes, with discrete
changes being only an extreme limiting case, never actually reached in
real macroscopic systems. You (and apparently Hans as well) start from
the other side: you speak of physical processes as fundamentally
discrete in nature, with the continuous case being reached only under
exceptional conditions.

Not unless we (at least I speak for myself) are dealing with quantum-level
physics. At that level, physical processes may well be discrete, but maybe
not. At the levels where we usually talk, there's no problem in talking
as if the processes are fundamentally analogue. And that's what we do.
Always, unless otherwise specified.

As a result, you and I think of "applying a disturbance" very
differently. From your point of view, as I interpret it, one would apply
a step-force to the standing person, and this force would instantly make
the angle start changing, before any control system could react to the
change.

If you want to look at the loop delay in the system, or are specifically
interested in the temporal dynamics, then an impulse disturbance is indeed
what you want. If you have reason to believe the system is non-linear, you
might do it, but only as one of a suite of tests. And since living systems
are unlikely to be linear, you wouldn't put too much faith in the results
of such a sudden push in predicting the effects of other disturbance
waveforms.

As a result, you don't see any important difference between
defining the disturbance in terms of the applied force and in terms of
the resulting change in angle.

Whether one deals with the dimension of the output variable or with the
dimension of the sensory variable depends on which one is of most
interest. The CEV can be written as Y=y(x1, x2, x3...), as you have
pointed out.In this example, Y is an angle, xn are forces. If x1 = 0,
the "disturbance" can be seen in either dimension. So what? I agree that
language is sometimes unclear, but I doubt the concepts are.

And of course, since I am trying to see whether the angle is under
control, I would be sure to increase the force slowly enough so that if
the standing person was going to resist it, the resistance could develop
right along with the rise in disturbing force, with only an
insignificant lag.

A long time ago, I proposed a test that involved moving a wastebasket to
see whether another person was controlling a perception of its position.
In that case, I would indeed displace it to a different part of the room.
I would not produce a gradually increasing force on it, nor would I move it
by a millimeter, and then another millimeter. But in general, I'm with you.

If you are interested in measuring the "insignificant" lag, you would find
this technique imprecise. If you already have determined that the angle is
under control, this technique is uninformative. But if you measure the
angular offset and the force of your push, you can at least determine the
effective gain _at the rate of your force increase_. (If there's an integrator
involved in the output function, the effective gain will be a function of
the rate). You'd want to try it at a variety of rates, if you really want
to understand the control system. And that variety would include rates
faster than the control system can handle, if you really want to know its
limitations. And then you'd have to test out various superositions, etc., etc.

I don't know whether we will ever see eye-to-eye on this matter. Our
basic orientations are just too different.

You can't see eye to eye with someone on whom you turn your back, which is
the case when orientations are the same, unless you are side-by-side.
Everything in your posting argues that our orientations are identical. Are
we side by side or are you turning your back?

Now back to the apparent disagreement between you and Rick, as I understood
it.

Hans and I had asked about the results of applying the Test's first phase
(in both discrete and continuous form) by influencing the particle's position.
Rick said, in effect, do the Test and you'll see that there is no control.
You said, in effect, the Test can't be done in this situation. That sounds
like a contradiction. Neither Hans nor I (to whom you were both replying)
had suggested simply looking at the particle, which is what you imply by:

If all you see is the variable varying, you have no idea why it is
varying. If you see it return to its original state, you have no idea
why it did that.
...
Merely watching a variable change does not tell us
anything except that it changes.

This is position for which I have argued in writing for at least 30 years.
I wrote about it as a probable reason for the difference between tactile
and haptic perception (Carterette and Friedman's Handbook of Perception,
Vol III, 1972). I now kick myself for not having discovered PCT then (even
if it was 20 years after Powers), since we used a 3-level control system
of touch behaviour whose intent was to do just as you suggest--probing the
world rather than just letting it happen

A quote from that chapter, after a discussion of the findings that the
sensation of touch is replaced by a perception of object when the subject
chooses to move the hand freely:

   "Although it does not lead to a "sensation", the purposive nature of the
   active touching process is important...Formally the touch process in the
   real world is a feedback process. This formal statement is fundamentally
   important to an understanding of touch."

We actually used the term "Control Loop" for a 3-level hierarchy labelled
"Behaviour Control," "Movement Control," and "Motor Control." The perceptions
and references are organized exactly as in the PCT hierarchy, but with
an added bypass for sensory input to each of the three levels of perceptual
input function. I had all the elements of PCT, but not the genius
to see its generality and importance. Maybe if I had, I too could have been
rejected as a psychologist:-(

···

---------------

As I now understand you, you are agreeing with Rick (and with me) that
the Test can be performed in the Hans Blom Brownian motion situation.
Maybe you would like to provide your own answers to the 5 questions
that I have twice asked Rick. No collusion, now!

Martin