mutuality; imagination; contrasts

[Tom_BOURBON3]

From Tom Bourbon [931014.1040]

I will use two of Martin's posts in reply to Rick as points of entry into
the discussions on a few threads. Other work has kept me away from the
mail for a few days.

[Martin Taylor 931013 12:00]
(Rick Marken 931009.1500)

Martin, earlier:

Consider a second ECS (B), which has as its CEV-B some part of the variable
that disturbs CEV-A. ECS-B has no perception of ECS-A or CEV-A. It simply
controls CEV-B. But as a consequence of that control, the disturbance
to CEV-A is reduced, easing the control by ECS-A of its own perception.
Now suppose that CEV-A is part of the variable that disturbs CEV-B. Now
ECS-B has an easier job of control, which directly eases control exercised
by A. Each ECS in this figure-8 loop has a multiplicative effect on the
apparent gain of the other.

Rick:

In fact, this situation (as described and shown in Martin's diagram) is
virtually identical to the interactive control situations analyzed by
Tom Bourbon and the coordinated control situations analyzed by yours truly
(and described with icy clarity in "Mind Readings").

Martin:

Well, you may be right, but it looked (and looks) different to me. I'm
looking at why multicellular organisms exist, and why complementary
cooperative structures have evolved to be so prevalent. There is no
scintilla of "helping" in my diagram, since neither ECS can perceive any
aspect of the CEV that corresponds to the other's controlled perception.
BOTH act quite independently on controlling perceptions according to their
independent reference signals, and it is only a strange vagary of the
environment that lets an outside analyst see that there is some interaction
between the two CEVs and therefore between the two control loops.

I'm not sure what you mean, Martin, when you say my modeling of
interactions between various combinations of agents looks different from
what you proposed. True, I did not have as my specific goal, "looking at why
multicellular organisms exist, and why complementary cooperative structures
have evolved to be so prevalent." In that line of questioning, I wish you
well. Questions of that magnitude are far to ambitious for me! What I
have done is try to study interacting control agents. Some of them
are models that can be diagrammed without one scintilla of helping in them.
I want to see how various combinations of those agents use the same
environmental givens ("affordances?") to create interactions that people
identify by such labels as interferance, cooperation, control of one by the
other, helping or aiding, and so on. The examples I use are variations on
trasking tasks, and the participants are people and PCT models, but I
believe the results are general, so they would apply to the cells and
"simple" organisms that would inhabit your imagined early world.

In my demonstrations, each of the models acts to control its own perceptual
signal, nothing more. When a set of models imitates a set of people, or
when a model interacts on-line with a person, the result as seen by an
observer is one of the several possible varieties of coordinated action I
listed earlier. Sometimes the participants are also aware of the kind of
interaction. Sometimes unilateral or mutual interference can occur with one
or both of the particiants unaware that it even occurs. Sometimes control
of the actions of one agent by the other agent can occur without the
recipient of control being aware; so can "aiding" or "helping." Certainly
there is no such awareness in the interacting PCT models. But when people
cooperate, they know it. Even then, the models that duplicate the
cooperative results do not "know" what they are doing.

Martin, how are these demonstrations different from what you propose?

···

================================

Tom now:

I believe some parts of the following exchange between Martin and Rick raise
serious questions we should discuss on the net. As Martin indicates at one
point, I believe some strong residues of "other ways of thinking" still lurk
in the background and lend confusion to our discussions.

[Martin Taylor 931013 16:00]
(Rick Marken 931007.1600)

Martin earlier:

Sure, but just where IS the closure of the feedback loop before the
word is spoken. Remember, it is spoken once and then is gone.

Rick:

It occurs WHILE the word is spoken. Remember the Kelso, et al and
Abbs et al studies of compensation for disturbance to lip movement
WHILE a word is being uttered? This is what I'm thinking of; as the
person says "flunky" they are controlling (continuously, over time)
the progress of all the lower order perceptions that make up this
event (so nicely defined now by W. T. Powers).

Martin:
No problem here, except that I don't see how the on-line (real-world)
perceptual control that always happens at these levels can be used to
account for the different outputs of the "same" word when there are
different situational contexts. ....

Tom now:
Martin, are you saying that on-line perceptual control cannot account for the
necessarily variable actions ("outputs") that result in the same perceptions
in the presence of disturbances ("different situational contexts")? Or do I
just terribly misunderstand that sentence?

Martin earlier:

I'm saying that the closure is in imagination.

Rick:

Why? The fact is that we continuously hear what we are continuously
articulating; the loop is closed IF what we articulate also depends
on what we hear (and there is tons of evidence that it does -- eg.
the delayed speech studies).

Martin:
The closure of the perceptual control loop at levels near the actual
articulation is not in imagination. But there is a timing problem with
the dynamics of control at higher levels, if that control is through
the real world. The perception to be controlled is ultimately that
the listener acts in some manner or other. That is akin to a ballistic
control problem. In ballistic control, one shoots a bullet, and corrects
the aim for the NEXT shot according to the resulting error. But one is
much less likely to use a next shot if one has an accurate telescopic
sight. Likewise in using language, one can be misunderstood, as one can
determine from an unwanted perception of the listener's actions. It is
much better if one does not have to correct one's first linguistic aiming
point. That aiming point is constructed in imagination. The shooter of
the bullet controls the arm muscles on-line, but corrects the aiming error
off-line, between shots. So with the shooter of words.
\--------------------------------------------------------/

Tom now:

Martin, your reply to Rick raises severl questions for me. In it, and in
many other posts during our various rounds of discussions on speaking, I
have seen statememts that we control our outputs or that we control our
muscles. In what sense do you mean that here? I know that seeing statement
like those will always make me wonder if the writer and I are talking anobut
the same theory of behavior. I suspect Rick is reacting similarly, but more
often and more pointedly than I.

A few specific points on the material I just quoted. First, to me it looks
as though you dismiss or attach litle significance to the extensice
liteature on the effects of delayed speech. Why so? Second, I am a shooter
-- both of words and bullets. One of the first things I learned was that a
marksman (today, marksperson?) -- shooter should not try to "control" the
muscles. Most inexperienced shooters try to hold their muscles perfectly
still in the "correct" place. It doesn't work. Instead, the shooter
points the gun slightly above the target, then lowers it and looks
at and controls the trajectory of the sight as it drops slowly across the
target. The shot is squeezed off while the muzzle is descending across the
target. The shooter does not, indeed cannot, hold the muscles and gun
motionless. (Excluding shots fired from a bench or a mount; after sighting
in, those are all open loop.) I do not hunt; I shoot for the pleasure of
achieving that kind of control of perception -- control of perecption, not
muscles or outputs.

So with the shooter of words.

/------------------------------------\
Martin earlier:

If there is error in the
imagined percept of the category of the spoken word, the appropriate
changes in output are done in the normal way.

Rick:

Your introduction of this unnecessary imagination loop has created
on output generation (open loop) model. This is fine, but it will not
produce consistent results in variable contexts -- the very thing
you said my model (PCT) could not do (and which, in fact, it can).

Martin:
I didn't say it wouldn't. I asked you to show how it would, without
the arm-waving appeal to "All actions take place in an environment so
disturbed that only on-line control can account for accurate action."
In fact, lots of actions can, most of the time, produce "fire-and-forget"
accurate results that support the control of higher-level perceptions.
These are useful when the world is likely to provide substantially delayed
feedback.
\----------------------/

OK, Martin, but "fire-and-forget" works only in a disturbance-free world --
exactly Rick's point. Often it only works *qualitatively*, creating
results "sort of like" the intended or planned output. The technique fails
when there are very many "different situational contexts." In fact, the
technique usually fails quantitatively even if there are NO discernable
environmental disturbances. That is especially true when the intended
result is a particular "output." I'll describe a simple case.

I am finishing a paper that describes a few demonstrations I prepared for
"Models and their worlds," but Bill and I did not use them --the paper was
too already long. In the demonstrations, a person and a PCT model each "act
like" the two lineal models from "Worlds" -- a S-R model and a plan-driven
model. In both lineal models, behavior occurred as a controlled output. The
lineal plan-driven model simply made its momentary handle position match the
momentary value of the target function from an earlier experimental run; the
model had a "perfect memory" and a "perfect plan" for its actions. A
simplified plan-driven model like that cannot fail to reproduce the
remembered plan.

A PCT model can easily duplicate the performance of the lineal model. All
that is needed is a reference signal that is a time-indexed file of
intended hand positions. At each moment, the model compares its perceived
hand position against the momentary reference position and the rest you
know -- standard PCT control.

The story is different for a person. At least it is for this person. I
have been tracking the same triangular target function for over nine years.
I have performed thousands of replications, under various experimental
conditions and while developing different programs. I should know the
pattern by now. The pattern is a triangular wave, with the target moving
up and down on the screen at uniform velocity -- the same one for nine
long years! The target is two short horizontal lines moving vertically

                /|\
t c t |
--- --- --- \|/

with a cursor in between. The positions plotted against time would look
like this:

  /\ /\ /\
/ \/ \/ \. The task runs for one minute with a new target

position every 1/30 second, for a total of 1800 positions vs time. There
are 11 upper peaks on the complete function -- I know this baby well!

For the paper, I practiced the task 22 times, with my hand covered by a
piece of cardboard. I concentrated on the feelings as I succesfully made
the cursor track the target. My intention was to perform the pattern of
movements during runs of the task when the screen went blank as soon as the
run began. (A slight case of reversible "deafferentation.") My
*impression* is that I could do this task in my sleep. That is a mistake.

For data collection, I ran 16 replications with my hand covered and the
screen blanked after the first few seconds of the run-in period on each
replication. For each run, I correlated the positions of my handle with
the positions in the target -- 1800 pairs of points in each of the
16 replications. The result? The mean of the 16 correlations was -.003.
The slightest error in my handle velocity, or in the point at which I
reversed the direction of the handle, threw the actual and ideal functions
out of alignment, and there were many such "lapses" on my part during each
run, in spite of my trying hard enough to end the evening with a mild
headache.

I have done this procedure more informally many times over the years, and
have asked other experienced trackers to try it. The results are always the
same. *Qualitatively*, the actual handle function *always* "looks sort of
like" the ideal one. *Quantitatively*, pre-planned "fire-and-forget" action
does not work. If experienced people cannot move a tracking handle to
accurately and reliably reproduce a heavily-practiced triangular waveform, I
wonder how well they would do at producing discourse?

Enough. My reactions to the exchange between Martin and Rick are
about (1) an impression (mine) that some people are saying when we speak we
control our output, and (2) the idea that pre-planned, canned, plan-driven,
programmed, command-driven actions are *quantitatively* accurate and
reliable.

Until later,

Tom Bourbon

[_Martin_Taylor]

[Martin Taylor 931014 14:50]
(Tom Bourbon 931014.1040)

A thoughtful post, requiring a thoughtful answer.

Two short answers:

Martin, are you saying that on-line perceptual control cannot account for the
necessarily variable actions ("outputs") that result in the same perceptions
in the presence of disturbances ("different situational contexts")?

No, I would never say something that intentionally suggested that. I might
very easily say something that could lead you to read that into what I say.

Martin, your reply to Rick raises severl questions for me. In it, and in
many other posts during our various rounds of discussions on speaking, I
have seen statememts that we control our outputs or that we control our
muscles.

Have I actually ever said either of those things? I often catch myself
in reading my writings before posting them in something similar, but what
I catch is statements that we control CEVs. I usually change them to
statements that we control the perceptions corresponding to the CEVs, but
I'm sure I don't catch all of those cases. My mind no longer consciously
even conceives of controlling outputs, as it might have done two or three
years ago, and you will have seen quite a few postings from me commenting
on statements by other people that we do control output.

I grant that I said "the shooter ... controls the arm muscles." It should
be said that "the shooter controls the perceptions derived from the
kinaesthetic sensations from the arm muscles" but this convoluted clause
would detract from the main point.

Let's digress a bit, and ask why it is so difficult to get PCT ideas across.
In part, it is because ordinary language is cast not in terms of erceptions,
but in terms of CEVs ("Complex", not "Controlled", Environmental Variables),
which include the actions of other people. We can see and talk publicly
about the actions of a third party, easily. We cannot detect, let alone
easily discuss, the perceptions that third party is controlling. It is
so much quicker to say "He controls the position of the cursor," which is
wrong, than to say "He appears to be controlling for his perception of
the position of the cursor to be at a certain level," particularly when
the clause is only part of a sentence of which the main point is elsewhere.
Using sloppy language leads to all sorts of possible misunderstanding,
perhaps even internally to the talker, and it is good to be called on it.
(But not too often).

Using long-winded language also leads to misunderstanding, unless you
are an expert writer (as is Bill P.).

···

==================
To the two main points:

I do not know all of your studies on two-CT systems, though I would like to.
You have thought about these things much longer than I have. I just had
not seen any reference to a system in which ECS-A was REDUCING the
disturbance to CEV-B at the same time that CEV-B was REDUCING the
disturbance to CEV-A. In their comments, both Rick and Bill mentioned
that all the actions of A would be effective as disturbances to CEV-B,
which is true. What they didn't say (or perhaps didn't see) was that
DIST-A in the absence of control by A would be providing much larger
disturbances to B. Much of their argument would have been valid if
DIST-A had been zero in the absence of ECS-A. But it isn't, under
the proposed situation.

Have you done studies in which the CEV of a control system is subject to
the influence of a disturbing variable, and then added another control
system to perceive and control the state of that disturbing variable?
If you have, is it not true that the first system's control is improved
when the second reduces the range of the disturbance to its CEV?

Let's propose an example to make it concrete. Person A is doing an
ordinary tracking task, moving a handle to control the perception of
a difference in height between a cursor and a target that moves up and
down. The target velocity is incremented periodically by a random number
drawn from a range -N to +N that is added to the current velocity. Now
Person B is introduced. B sits in another room, looking at a number written
on the screen. The random number that goes to disturb A's cursor is
subtracted from the number of B's screen, and B moves a handle to restore
the visible number to its reference value. B's handle movement is used
to generate another number that is added to the random number before it
is used to affect A's target. If the sign of the addition is right,
the velocity changes of A's target should now be appreciably less violent.

That's the non-looped condition. Unless there is some survival benefit
to the new controller of doing so, such a situation will not endure in
nature. Now make the situation reciprocal, so that the first ECS's CEV
influences the disturbance to that of the second. Forgetting for the
moment Bill's analysis that shows the situation to have positive feedback
(I'll come to that), will not control by the first system reduce the
disturbance to the second, thus enhancing the second's control? If it
does, each ECS benefits by the relationship, and the survival probability
of each is enhanced. (One assumes that there is some survival benefit to
the control of the individual perceptions in the first place).

Does it still sound like what you have done?

What I have done is try to study interacting control agents. Some of them
are models that can be diagrammed without one scintilla of helping in them.
I want to see how various combinations of those agents use the same
environmental givens ("affordances?") to create interactions that people
identify by such labels as interferance, cooperation, control of one by the
other, helping or aiding, and so on.

And what I have seen of it has all been quite inspiring to me. While I
was away last month, I handwrote a taxonomy of "helping systems" that I
had intended to submit to you for comment before posting it, but it sounds
as though you have already done the same. I'd love to see it. Maybe you
posted it before I came on the net, or before I was aware of its importance.
(Incidentally, my library has been unable to get me a copy of the PCT issue
of ABS.).

OK, now let's consider Bill's analysis that the situation leads inherently
to positive feedback, and therefore can never be useful. Even if each
ECS controls its CEV so that in the absence of the loop the effect on
the corresponding DIST is reduced (CEV-A on DIST-B and vice-versa),
nevertheless if the loop introduces positive feedback, I accept that mutuality
cannot be helpful.

Here's the picture again.

                     > >
                     > ref-A | ref-B
           ----------O--------- ---------O----------
          > > > >
        PIF-A OUT-A PIF-B OUT-B
          > > > >
   =======^====================V======^===================V=======
          > > > >
           <-----CEV-A<-------- <------CEV-B-------
               > > > >
               V DIST-A<-------------------V DIST-B
               > ^
                ----------------------------------| This loop in the
                                                     environment is unperceived
                                                     by either ECS

Let's trace out the signs of an effect that starts with some external
(positive) influence on DIST-A.

+ DIST-A influences CEV-A (by definition)
+ Perception A
+ Error A
- Output A
- CEV-A (of course the opposes the original effect of the disturbance, so don't
         conclude I'm doing a discrete sequential analysis here. I'm trying
         to get at the overall sign of the gain of the loop between the two
         control systems)

? DIST-B (we cannot know this function, so there is no telling whether a
          positive change in CEV-A influences DIST-B positively or negatively)

And so on. It seems to me that provided the signs of the effects of CEV-A on
DIST-B and of CEV-B on DIST-A are appropriate, the mutuality feedback loop
could be positive or negative.

This analysis is not really adequate. There are loop transport delays within
the mutuality loop as well as within each ECS, so that the analysis gets
quite complex even in linear systems. When they get non-linear, I could
well imagine that oscillatory interactions are quite probable if the
gains of the ECSs are high and the systems are tightly coupled. It was
to avoid the problems of tight coupling that I used the term "influenced"
in respect of the relation between CEV-A and DIST-B.

I hope that this makes it clearer what I am talking about in respect of
mutuality. If it still seems to be something on which you have worked,
I'd love to see the results.

=============

First, to me it looks
as though you dismiss or attach litle significance to the extensice
liteature on the effects of delayed speech. Why so?

Because these effects occur on the time-scale of a syllable, which is
not the time-scale over which one chooses one's words. When one is
speaking a syllable, one is speaking it to accomplish the perception
of having spoken a word already specified in a reference level from
a higher level. I'm interested in where that reference level came
from, and no amount of disturbance on the 200 msec time-scale will
change it before actions to speak the word are initiated. The successful
speaking of the word is what is affected by delayed auditory feedback,
and if I may requote what you quoted:

The closure of the perceptual control loop at levels near the actual
articulation is not in imagination.

----------

[on learning to shoot accurately]...
One of the first things I learned was that a
marksman (today, marksperson?) -- shooter should not try to "control" the
muscles. Most inexperienced shooters try to hold their muscles perfectly
still in the "correct" place. It doesn't work.

No, neither would it in articulating words.

To continue the shooting analogy:

the shooter
points the gun slightly above the target, then lowers it and looks
at and controls the trajectory of the sight as it drops slowly across the
target. The shot is squeezed off while the muzzle is descending across the
target. The shooter does not, indeed cannot, hold the muscles and gun
motionless.

Now, what does the shooter do if the shot turned out to have hit a metre
to the right of the target? Isn't the aim for the next shot perhaps taken
a little to the left of the target (not, perhaps, a whole metre, but some).
Isn't that kind of control like correcting the word that has been
misunderstood?

OK, Martin, but "fire-and-forget" works only in a disturbance-free world --
exactly Rick's point. Often it only works *qualitatively*, creating
results "sort of like" the intended or planned output.

We agree on that. What we might not agree on is the amount of disturbance
that exists in this particular piece of the world, remembering that the
"world" for the language-level perceptual control (whatever that might be)
includes the perceptual control functions of the levels that execute
language. The the world for any individual ECS includes all the ECSs in
the hierarchy below it, so to a large extent higher-level ECSs do see
a fairly stable world. But that's not really relevant to the issue at
hand.

Let's continue the analogy with shooting. Language is highly redundant.
That means that there are many possible language actions that will result
in the same language perception. It is as if the bullseye for the shooter
were to be made several metres wide. Then a one-metre error in the
perception of where the bullet hit, compared to the aiming point, would
not matter to the score. So with language. Results "sort of like" are
often good enough.

But...

The technique [fire-and-forget; MMT] fails when there are very many
"different situational contexts."

I don't think this is correct. The technique fails when there are
disturbances, which by their nature are not perceptible to the ECS that
is controlling against them. But as was clarified by Bill P during the
discussion on "information in perception," a higher-level system can
perceive a function of both the disturbance and any signal in the lower
ECS (we usually say that only the perceptual signals go up, except for
the output return through the imagination loop, but that's not an absolute
restriction). A higher ECS's output can alter the reference levels
for a lower, providing it with situational context dependence (of which
it, of course, knows nothing.

[A discussion of trying to reproduce the sawtooth motion of an unseen
tracking target whose track has been well practiced over the years]
...
I have done this procedure more informally many times over the years, and
have asked other experienced trackers to try it. The results are always the
same. *Qualitatively*, the actual handle function *always* "looks sort of
like" the ideal one. *Quantitatively*, pre-planned "fire-and-forget" action
does not work. If experienced people cannot move a tracking handle to
accurately and reliably reproduce a heavily-practiced triangular waveform, I
wonder how well they would do at producing discourse?

One of the things that is most evident about language use is how different
it is depending on whether you see and hear your partner face-to-face.
In face-to-face communication you depend heavily on feedback at various
levels of abstraction. If the partner is separated in time or (absent
telephones) space, you can't do that. Written language is much more
highly structured than is spoken conversation, just to make the discourse
more redundant (the bullseye more metres wide). But even written language
depends eventually on feedback. The issue is in the temporal gap between
the feedback you get on your own productions--your perception of your
acoustic productions or your words, for example--and the feedback that
goes through the control systems of your partner and results in actions
you can sense.

In using language, you are trying to control perceptions that depend in
part on sensing actions performed by the partner. There is a time lag
in any such feedback loop that is substantially greater than the lags
that are involved in controlling the perceptions of correct articulation,
especially if, as here, the language is written.

One CAN correct a word or phrase that is mis-spoken. In conversation, this
is so frequently done that one notices it only when one's attention is drawn
to it. One cannot correct a word or an argument that is misunderstood,
until some action of the partner indicates that the misunderstanding has
happened. For effective communication, one needs to act so that there
are few misunderstandings. If there are many, the loop gain between the
partners can easily go positive, so that the entire traffic between them
consists of clearing up misunderstandings about clearing up misunderstandings
about clearing up.... (something like CSG-L, sometimes).

This is presumably why language is built on categories rather than on
continuum levels of perception. (I mean "language" not the execution of
linguistic actions). In a category, the target is metres wide, and small
deviations of the result from the intended aiming point change the score
only if the aiming point was near the edge in the first place. So, your
correct point that it is wrong to say:

"pre-planned, canned, plan-driven,
programmed, command-driven actions are *quantitatively* accurate and
reliable"

is a foundation for why language is as it is. You must be
able to "fire-and-forget" successfully most of the time when you use
language, and you can't be quantitatively accurate in doing so. But
in a category you can be qualitatively correct, especially if you keep
your aim away from the category boundary near other categories that
in this context might be confused with your target category. (You can,
of course, be quantitatively correct in the perceptions you control
that relate to the execution of language; you have to fire-and-forget
in respect to the desired perceptions that relate to the partner).

I hope that this posting is sufficiently redundant to allow you to get
my meaning, and to correct it if it is wrong, rather than to misunderstand
me.

Martin

[Tom_BOURBON3]

From Tom Bourbon [931014.1754]

[Martin Taylor 931014 14:50]
(Tom Bourbon 931014.1040)

A thoughtful post, requiring a thoughtful answer.

Martin, this is jus a brief remark about one small point in your beautifully
redundant reply.

The point in question is this.

I grant that I said "the shooter ... controls the arm muscles." It should
be said that "the shooter controls the perceptions derived from the
kinaesthetic sensations from the arm muscles" but this convoluted clause
would detract from the main point.

Whet prompted my reaction to your statement, about the shooter controlling
the arm muscles, was not that you did not add the other details, but that
the statement can be taken as implying that effective control is the control
of actions. As I tried to suggest, but not with sufficient clarity, at the
beginning of the post to which you replied, I was using the exchange between
Rick and you as a place to jump back into the discussion on language. I
have been away from the net for a few days. Your example of the shooter
reminds me of a number of other examples, many from other people, not you,
in which there is also a stronger than average possibility of a reader
concluding that to speak is to control actions. I think we agree that
shooters and speakers control their own perceptions of the consequences of
their actions, not their actions.

More later, on the remainder of your reply.

Regards,

Tom