The Test; Open-loop measurements

CSGnet Archive CSG1995
1

[Martin Taylor 951107 16:40]

Bill Powers (951107.1330 MST)

You are still taking the experimenter's view rather than the cat's.

Martin Taylor 951107 11:15 --

    I'm only considering the case in which everything the cat does
    purposefully is actually irrelevant to what the God-experimenter
    knows to be effective.

Are you saying that you could devise an experiment in which this would
occur, and in which a person "versed in the art" could not discover what
the cat is actually controlling?

I don't know whether an experimenter could determine what the cat is
controlling during any specific escape sequence. I don't care. The
question is about what the cat might be doing, not the experimenter.
I know I started by asking about the application of the Test, but the
Test leads into the organism tested, doesn't it?

There are two levels of control being
considered here: one, the perception that the cat is in fact
controlling, and two, the indirect effect of controlling that variable
on a second variable that the cat is also controlling.

The "second variable" is, I presume, related to the escape and the subsequent
access to food. The first variable exists, but keeps changing during any
one escape sequence.

It should always be possible to discover what the cat is actually
controlling by applying disturbances and seeing if they are resisted,
and of course then going on to complete the Test.

During rapid reorganization?

In that case, if I moved the stick so that the cat no longer brushed
against it, the cat would do nothing to restore contact and I would
reject contact as part of the controlled variable.

    But the cat WOULD do something to restore contact, or it wouldn't
    get out of the box.

But the test goes on, if it isn't failed at that stage. We then have to
establish that preventing the cat from sensing contact (even when there
is contact) will destroy control.

By hypothesis throught this interchange of messages, the cat doesn't perceive
the contact. So, again by hypothesis, preventing it from sensing contact
will not destroy control--the cat escapes nevertheless.

In that case, we will be able to show that the cat is not controlling
the stick position or contact with it. If we move the stick so that
contact no longer occurs, the cat will not restore contact, and it won't
escape from the box, either. What you're trying to imagine is, I think,
an impossible situation: where the cat's behavior just happens reliably
to restore contact with a stick moved in any arbitrary direction or to
any arbitrary position, but where there is no control of contact.

Two answers: (1) you are asserting that no matter what variation there
is in the set of feedback functions made available in the environment,
control will be maintainted IF the truly effective perception is what
is being controlled, and (2) you are asserting that if the truly effective
perception is non-existent, then control will not occur as the side effect
of other control, no matter how restricted might be the universe of
available feedback functions.

No, Thorndike's cat would have been extremely unlikely to have escaped
even once, if the box were 2 miles square. And No again, the cat would
almost certainly escape very quickly if the box were so small it could
hardly avoid the stick. The actual situation is somewhere inbetween.
As a cat explores (or at least that's what it looks as if it is doing),
it moves around and brushes things. There's a reasonable chance that
its exploration may involve it brushing the (possibly only virtual) stick.
And at that point it escapes, as we presume it wants to do.

By analogy, in the world in which we live--society and all--if we do certain
things, other things happen (usually). We need never know why, so long
as God doesn't disturb the environment as you propose your experimenter to
do. So long as the feedback paths remain unchanged, if we discover a way
to influence our perceptions, we can control them. It is inconsequential
that some element of the feedback path is a pure side-effect of controlling
another perception for which the main output sets the reference levels, so
long as the environmental feedback paths don't change. But when they do,
those side-effects may be quite different.

So with your disturbances to the cat. When you change the cat's environment,
it may well go back to "random" behaviour, changing its actions until some
different side-effect of some different perceptual control system happens
to have the right influence on the main controlled perception--being out
of the box and at the food.

In
other words, you are saying that there can be a reliable coincidence
that will fool us. Considering how unlikely such coincidences are, is
this really a serious problem?

In view of the above, do you think that it is either unlikely or a problem?

The action that the cat learns from
one success will not work on the next trial because the disturbance will
be different. Yet if the cat has learned to control the right variable,
it will succeed on every trial even though it must produce a different
amount or direction of action each time.

Yes, quite. And I presume it is through attaining control and retaining it
as the environmental feedback functions change that the cat (and we) learn
what is really the effective perception to control in order to control the
higher-level one. Once the feedback path ceases to depend on side-effects
that change with the environment, it becomes reliable and control consistent.

    I am indeed trying to deal with the situation in which it is the
    side-effects of control that do the intended job.

This can be the case only if you always use the same disturbance that
has the same relationship to a side-effect.

I hope you see now that it is precisely the variation in the "effective
side-effects" that is at the heart of what interests me in this situation.

Must go. Possibly more on general reorganization and "open-loop"
experiments tomorrow. Perhaps not.

Martin

2

[From Bill Powers (951107.1330 MST)]

Martin Taylor 951107 11:15 --

     I'm only considering the case in which everything the cat does
     purposefully is actually irrelevant to what the God-experimenter
     knows to be effective.

Are you saying that you could devise an experiment in which this would
occur, and in which a person "versed in the art" could not discover what
the cat is actually controlling? There are two levels of control being
considered here: one, the perception that the cat is in fact
controlling, and two, the indirect effect of controlling that variable
on a second variable that the cat is also controlling.

It should always be possible to discover what the cat is actually
controlling by applying disturbances and seeing if they are resisted,
and of course then going on to complete the Test. Once that is done, the
secondary controlled variable could be discovered in the same way, and
the physical link between controlling the first variable and controlling
the second could be found by inspectin the environment.

In that case, if I moved the stick so that the cat no longer brushed
against it, the cat would do nothing to restore contact and I would
reject contact as part of the controlled variable.

     But the cat WOULD do something to restore contact, or it wouldn't
     get out of the box.

But the test goes on, if it isn't failed at that stage. We then have to
establish that preventing the cat from sensing contact (even when there
is contact) will destroy control. This is how we rule out unintended
side-effects.

     I am assuming that the cat completely ignores the stick and any
     related sensations, so this isn't a better example. It's an
     example of something quite different.

In that case, we will be able to show that the cat is not controlling
the stick position or contact with it. If we move the stick so that
contact no longer occurs, the cat will not restore contact, and it won't
escape from the box, either. What you're trying to imagine is, I think,
an impossible situation: where the cat's behavior just happens reliably
to restore contact with a stick moved in any arbitrary direction or to
any arbitrary position, but where there is no control of contact. In
other words, you are saying that there can be a reliable coincidence
that will fool us. Considering how unlikely such coincidences are, is
this really a serious problem?

     Initially, the cat is doing a lot of what we call "random" things.
     It has an unsatisfied reference to see itself out of the box, at
     the food. When there is no pre-organized control structure that
     affects that perception, given the currently available
     environmental feedback paths, reorganization begins. Things get
     "tried out." And at some point, the desired perception comes to
     pass and reorganization stops. If that new organization isn't
     altered in the interim, then when the cat is put into the box at a
     time that it has the same reference level, it will "do" the same
     things and get out of the box.

This is why we use disturbances; so "doing the same thing" will NOT
reliably have the same effect unless there is actual control.

     Now assume that in all this the cat is topically anaesthetized, so
     that it can't feel the stick. Maybe the stick is even not a real
     stick, but a virtual one in an experimenter interface, so that the
     cat's "brushing" it is an event detectable only to the
     experimenter. The cat nevertheless gets out on one occasion by
     "marking the box and lying down on its left facing the clock," and
     will do so on future occasions SO LONG AS THE ENVIRONMENTAL
     FEEDBACK FUNCTIONS REMAIN THE SAME.

Again, that's why we use disturbances. If the condition is visible to
the experimenter, the experimenter can vary the nominal position in
which the cat must be to open the box. If the cat continues with its
former behavior, the box will not open.

In the Test, we NEVER stop applying disturbances. So either the cat
learns to control the right variable, or it continues to fail most of
the time, and keeps on reorganizing. The action that the cat learns from
one success will not work on the next trial because the disturbance will
be different. Yet if the cat has learned to control the right variable,
it will succeed on every trial even though it must produce a different
amount or direction of action each time.

     I am indeed trying to deal with the situation in which it is the
     side-effects of control that do the intended job.

This can be the case only if you always use the same disturbance that
has the same relationship to a side-effect.

     The thesis, which I am beginning the think is what you have been
     saying about reorganization all along, is that the ENTIRE
     perceptual control hierarchy depends on the consistency of the
     side-effects, in the environment, of intrinsic control. Intrinsic
     control is what is really happening. There are lots of different
     perceptual controls that influence variations in the values of the
     intrinsic variables, as pure side effects that are more or less
     consistent. Some reorganizations build one hierarchy, others build
     quite different hierarchies, but in all survivable cases the side
     effects influence the intrinsic variables (and probably a lot
     else).

Yes, that is what I have at least been trying to say all along -- one's
words are not always guaranteed to mean the same things to everyone, and
this is not what you would call a simple idea. Your last sentence is the
key: "Some reorganizations build one hierarchy, others build quite
different hierarchies, but in all survivable cases the side effects
influence the intrinsic variables (and probably a lot else)." That's it.
All that counts is control of the intrinsic variables; the organization
of behavior that accomplishes this is not the point. Reorganization
stops when intrinsic error goes to zero, not when behavior is organized
in any particular way. There might be lots of objective conditions that
determine whether a given behavioral organization will increase or
decrease intrinsic error, but unless we're specifically interested in
what they are, they don't matter. The reorganizing system neither knows
nor cares what the objective conditions are.

One way out of some of the problems this seems to create is to increase
the scope of what you are willing to call intrinsic variables. Perhaps
we need truth and beauty, too. When you add constraints like those,
perhaps some of the problems go away.

     What this analysis highlights is that the reorganized PERCEPTUAL
     hierarchy is highly likely to contain a great deal of conflict.
     Since it is the side-effects of perceptual control that really
     matter to survival, there is no reason why conflict itself should
     not, in some instances, ensure that those side effects continue to
     occur.

One argument against this is to recognize the consequences of conflicts:
loss of range of control by both systems involved in the conflict. If
each system was acquired because it had a beneficial effect on intrinsic
error, then when the two systems come into conflict they are no longer
able to control against the same range of disturbances as before. Maybe
that is the main reason that conflicts tend to get resolved.

     Food for thought?

Definitely useful.

···

---------------------------
Rick asks how you can be sure that part of a system is operating open
loop. You reply:

     The experimenter predetermines (or uses a predetermined algorithm
     to select) what "stimulus" is to be presented on trial 37. Nothing
     the subject does on trials 1-36 has any effect on this choice. But
     what is presented on trial 37 does influence the subject's
     "response" on trial 37 -- or if it does not, the experimenter
     becomes dissatisfied with the subject, which is something that the
     subject may be able to perceive and for which, by assumption, the
     subject is controlling.

How do you determine that the experimenter's manipulation changes a
quantity that would ordinarily be changed by the subject's own behavior?
As you state the situation, there is no way to tell whether the
manipulated variable is a CEV for a control system, or a disturbance of
a CEV. In either case there would be a response, but if the manipulated
variable is actually only a disturbance, the response would be very
different from what it would be if the CEV is a variable that is
normally under control.

An example of a true open-loop measurement of a control system is
testing the iris response to light intensity using a beam of light that
is much smaller in diameter than the aperture of the iris. Here you're
varying the intensity of the light and affecting a variable that is
normally affected by the iris, but in a way that the iris can't affect.

It's hard to measure open-loop response in a human control system,
because breaking the loop destroys control, and other systems inside the
same organism will immediately start changing reference levels,
sensitivities, whatever. Suppose you sat a subject down in front of a
tracking experiment to measure open-loop response to cursor position. To
do this you have to break the connection between the handle and the
cursor, and move the cursor yourself. But the moment the subject
discovers that the handle no longer has an effect, you won't be
measuring the same system any more: the subject will ask if you're
kidding and will probably turn that control system off. By being sneaky
you can get a brief look at the system open-loop, but only if you
generate cursor movements that are recordings of the same subject's
results with the same disturbance in a previous run, so the subject is
fooled into thinking he still has control. If you put in a test-jog of
the cursor, you will get a look at the open-loop response for a fraction
of a second, but you will have given the game away. The best way to do
this would be to display the cursor position being generated by a
previously-determined closed-loop model of the subject, so when you
apply a sudden jog in the disturbance, the subject and the model will
react more or less the same. But it's really hard to keep up the
deception for very long.

I think that open-loop measurements are quite possible, as in your
lighthouse example, but that they measure perceptions that are not
normally under control.
----------------------------------------------------------------------
Best,

Bill P.