open-loop tests of control systems

[From Bill Powers (941002.0830 MDT)]

Bill Leach (941001.1851 EDT) reminds me of a comment I had meant to make
on

[Martin Taylor 940929 10:30]

When an electronic engineer wants to examine an unknown control system,
one way is to break the feedback connection and examine the S-R
behaviour of the broken loop.

It's very difficult to break a control loop that a person normally uses,
without disrupting the control system. As soon as the loop is broken,
the person loses control and furthermore knows that control is lost, for
actions no longer affect perceptions. The normal consequence of breaking
the external part of a control loop is for the person to switch to some
alternate method of control as soon as loss of control is detected, or
to give up controlling completely until the outer part of the loop is
restored.

The electronic engineer doesn't have this problem because the system
being investigated is usually a one-level system, or at least has no
ability to adapt to loss of control. When the external feedback path is
opened, the test signals applied to the input have to be made much
smaller (by a factor equivalent to the output error sensitivity) if the
result is not to be simply driving the output from limit to limit as
fast as possible with every excursion of the test signal. If that does
NOT happen when the external path is broken, you can be pretty sure that
you aren't looking at the same control system -- something inside has
reduced the output gain to prevent such wild output fluctuations, or has
turned off the former control system entirely.

Investigations of passive perception usually require that the person
perform an action which tells the experimenter something about the
subject's experience of the perception but doesn't affect the perception
directly. So there is no control loop to be broken -- not one that
involves the perception in question as a controlled variable. You can't
assume in that case that the response to the test stimulus is the same
response that would exist if the perception were under control.

The nearest you can get to seeing open-loop behavior of a living control
system is when the external path is very briefly interrupted and then
restored before the loss of control is detected by a higher-level
system. This requires special quantitative experimental techniques and
real-time instrumentation. Rick Marken did something like that in his
"reversals" experiment.

Investigations of open-loop passive perception are peculiar in that they
make the test subject a partner, and even a substitute, for the nominal
experimenter. If the test subject is not an acute and eloquent observer
of subjective phenomena, the experimenter may be unable to understand
what the test subject is reporting, or to know the meaning of a report,
or a response, in terms of the perception actually being experienced.
Worst of all, the experimenter may assume that the test subject is
seeing and reporting on the "objective stimulus," meaning the stimulus
as the experimenter sees it in his own subjective world of perception.
What is generally done, as I understand it, is for the experimenter to
tell the test subject to make some standard indication when some feature
of the perceptual field is experienced -- but there is no way to know
what that feature is, in the test subject's world. Thus perceptual
experiments tend to be cast as S-R experiments, with occurrance of a
response being assumed to indicate occurance of the perceptual situation
that the experimenter has in mind. In very, very simple (low-order)
situations this may not be a bad guess, but it is always a guess, and it
will become progressively worse as more complex perceptions are
investigated.

In any case in which the experimenter can independently determine the
stimulus inputs on which perceptions are based, the situation is by
definition open-loop. If any control is involved, it cannot be control
of the same variable that the experimenter is manipulating. This is why
I insist, against considerable resistance from some directions, that the
standard PCT diagram include an explicit disturbing variable separate
from the controlled variable. Effects on the variable actually under
control are achieved not by arbitrarily altering that variable (which
would break the loop) but by altering some other physical variable, the
state of which _contributes to_ but does not _determine_ the state of
the controlled variable. The action of the control system must be able
to produce independent effects on the controlled variable regardless of
the state of the disturbing variable. If the experimenter can
arbitrarily set the state of the variable being sensed and controlled,
then while that arbitrary state exists the loop is no longer closed, for
the system's action can have no effect on the variable.

Even in cases where the experimenter-applied manipulation of the
controlled variable is very brief, such as an impulse disturbance, the
response that is observed is not the open-loop response. It is true that
for the first delay-time the loop is effectively open, but the ensuing
changes in action are feeding back to affect the input in every
subsequent instant; the observed response curve is the closed-loop
impulse response, not the open-loop response. If the loop were truly
open, it would be seen that the output response develops over a period
equivalent to many delay-times, is very much larger, and has a form very
different from (and more protracted than) the form of the closed-loop
response.

In summary, "breaking the loop" to observe open-loop responses to
arbitrary changes in the input is not a practical method for
characterizing human control systems. Nor are tests of passive
perception that do not involve control likely to tell us about the
characteristics of control systems.

···

----------------------------------------------------------------------
Best,

Bill P.

[Martin Taylor 941003]

Bill Powers (941002.0830 MDT)

[Martin Taylor 940929 10:30]

When an electronic engineer wants to examine an unknown control system,
one way is to break the feedback connection and examine the S-R
behaviour of the broken loop.

It's very difficult to break a control loop that a person normally uses,
without disrupting the control system. ... The normal consequence of breaking
the external part of a control loop is for the person to switch to some
alternate method of control as soon as loss of control is detected, or
to give up controlling completely until the outer part of the loop is
restored.

...

Investigations of passive perception usually require that the person
perform an action which tells the experimenter something about the
subject's experience of the perception but doesn't affect the perception
directly. So there is no control loop to be broken -- not one that
involves the perception in question as a controlled variable. You can't
assume in that case that the response to the test stimulus is the same
response that would exist if the perception were under control.

...

Investigations of open-loop passive perception are peculiar in that they
make the test subject a partner, and even a substitute, for the nominal
experimenter.

(There are related posts by Rick Marken and Bill Leach, but this one will
cover their comments, too).

The problem of "breaking a control loop" and finding that there is an
enormous tranisent output is not at issue. The key is in the last sentence
I included in the quote from Bill P.'s comment. The subject in ANY
experiment (at least any ethical one that involves "informed consent")
is a partner to the experimenter, and the subject's primary controlled
perception is to perceive the experimenter satisfied with the subject's
participation. The "open loop" is open because the subject has turned
to zero any gain that might normally exist in the loop of the tested
perception. The situation in many of these experiments is such that the
subject couldn't control the perception anyway.

The question of whether the perceptual function itself works the same
way when it is part of a functioning control loop is quite another issue.
The supposition I have always been working with is that the control effects
are the action of a fixed mechanism, the components of which can be
described. If they change, it is because of reorganization or adaptation,
not because the loop is or is not now functioning with non-zero loop gain.

In any case in which the experimenter can independently determine the
stimulus inputs on which perceptions are based, the situation is by
definition open-loop. If any control is involved, it cannot be control
of the same variable that the experimenter is manipulating.

Yes, that's the central point.

Effects on the variable actually under
control are achieved not by arbitrarily altering that variable (which
would break the loop) but by altering some other physical variable, the
state of which _contributes to_ but does not _determine_ the state of
the controlled variable. The action of the control system must be able
to produce independent effects on the controlled variable regardless of
the state of the disturbing variable.

Right.

Anyway, I think that it is worthwhile to re-post my analysis from last
November of what happens in a psychophysical experiment, to see whether
Bill P. thinks as well of it as he did at that time, especially in view
of

Nor are tests of passive
perception that do not involve control likely to tell us about the
characteristics of control systems.

Martin
=====================(old posting follows)=====================

[Martin Taylor 931130 19:00]

A long time ago, I asked how to look at psychophysical experiments within
a PCT framework. The replies I got were not very helpful, amounting more
or less to "it's all controlled perception." As I now have an answer that
satisfies me, I'd like to put it out for criticism. Maybe I will be repeating
what I was told so long ago, but I don't think so. Anyway, it might help
others who are struggling with the issue that bugged me for so long. (The
reason I bring it up now is that one of the participants in my small PCT
seminar raised the issue today. It is a live problem.)

In the kind of psychophysical experiment I am considering, a subject
looks at some kind of display during a series of "trials." On each trial
the display is different, and the subject makes some kind of appropriate
"response." The response is supposed to relate to what was displayed on the
last trial (the beep was in the second interval; the third line is longest).
Whatever the response might be, it has no effect on the next display.

The whole thing sounds like a perfect exemplar of S-R. But of course it
isn't, and that was very hard for me to understand early in my acquaintance
with PCT.

Here's my current analysis. For some reason (satisfying a perception of
pleasing the experimenter, for example) the subject has set a reference
perception, not for the display as such, but for a perceived relationship
between the "response" and the display. For example, if the response is
categorical, its category should be in some defined way the same as the
category of the display. The controlled perception is of a relationship
between two perceptual signals, one uncontrolled (the display) and the other
controlled (the perceptual effect of the response).

At the time of the display, the response has not yet been made. Neither
is there any intrinsic (S-R) reason why one should be made. But there
is an error signal in the control system concerned with the display-response
relationship. The perceived display is "X" but the perceived response is
null. In the relation-perception control system there is an output that
serves to reduce or eliminate this error, setting a reference for perceiving
the response to have been "X." When that reference comes to have an equivalent
perception, so does the relational reference, and the error goes to zero.
(Sometimes the subject makes a "response error," and knows it, like the
synchronized swimming judge in the Barcelona Olympics, and tries to
correct it. The display-response perceived relation fails to match its
reference of equivalence, and the error persists. In the case of the
Barcelona judge, the response correction has only recently been completed,
so the output can persist for quite a long time!)

The situation is parallel to that of getting to bed (whether the lights
are on or not). One does not control the location of the bed as such
(as a rule), but one does perceive it, and one does control one's own
location as part of satisfying the reference for the perceptual relation
between one's own location and that of the bed.

Any PIF (except perhaps at the lowest level) has more than one input
component (if it didn't, it would be at a lower level in the hierarchy).
For any PIF, not all of the input components are necessarily controlled
perceptions. Some may be uncontrolled at that moment, and must be taken
as given. Control of the perceptual signal (the output of the PIF) is
performed by varying those input components that are currently controlled
lower-level perceptions. In the case of psychophysics, the experimenter's
display is an uncontrolled input component of the perception of the
relation between display and "response," but that relation is a controlled
perception, control being effected by variation in the response.

I hope this helps somebody who is in the position I was a while back.
(Also, I hope it is right).

Martin

<[Bill Leach 941003.17:13 EST(EDT)]

[Bill Powers (941002.0830 MDT)]

Bill Leach (941001.1851 EDT) reminds me of a comment I had meant to make
on

[Martin Taylor 940929 10:30]

As usual, you point out issues at a level far more significant than most
of the rest of us.

The electronic engineer doesn't have this problem because the system
being investigated is usually a one-level system, or at least has no
ability to adapt to loss of control. When the external feedback path ...

I could not but help thinking of an encounter with an item of military
equipment that was "so sophisticated" that it could loose several
feedback loop and still function to within spec. or at least so close to
spec. that one could not be sure that there was actually anything wrong.
And of course, again, this was only an engineered control system!

-bill