imperfect control

CSGnet Archive CSG1995
1

[Martin Taylor 950705 17:00]

Bill Powers (950704.1650 MDT)

+Rick Marken (950705.0820)

I don't know why you persist in looking for these tricky paradoxical-
sounding generalizations, Martin.

And I don't know why you think I do that. Apparently both you and Rick
misunderstand my intent just as badly as I initially misunderstood Rick's.

You're actually much better at simple
straightforward PCT analysis and teaching.

Better than very bad is worth something, I suppose.

+Martin Taylor (950704 14:30) --

+> My point was and is that we all want to investigate the organism, and to do
+>so effectively one must concentrate on the details of how control fails, not
+>on the degree to which it succeeds. Rick's statement emphasises that point.

···

+
+I completely disagree with your point so I am disappointed if any statement
+of mine emphasizes it. Let me make a clear, de-emphasizing statement: The
+best information we have obtained about how organsisms work comes from
+studies where control (by the organism, of course) has been nearly perfect.

Neither of you even mention the issue about which I have been asking.

I'm very well aware that by looking in the environment (with our own PIFs)
we can approximately find CEVs that the control activities of a subject
are maintaining reasonably stable. I'm aware that from discovering these
stabilities (through the Test--which I'm aware takes more than just noting
a stability) one can infer the existence of perceptual functions that
imply controlled perceptions. I'm aware that the simplest description of
controlled perceptions is likely to involve units that correspond to
less complex perceptions that might also be controlled.... etc.

But where in all this is there any approach to the internal organization
of the control hierarchy? Where is there information about, say, the forms
of output functions, or the reference-setting linkages between higher and
lower control systems, or about the potential conflicts among control systems
at a level, or...? Where is there information about the changes that occur
in the (presumed) hierarchy during learning or forgetting?

All of these aspects of the internal organization are invisible to a simple
test for the nature of controlled perceptions. They depend on the dynamics
of the control, on the ways that the actual effects on the various controlled
CEVs (stabilized complexes in the environment) differ from ideal control.
I see nothing in the least paradoxical about this, nor do I think it either
trivial or premature to note it.

I'd be very happy to be shown where I'm wrong, and in what way studies of
nearly perfect control, that emphasize the perfection rather than the
imperfection, can illuminate the issues I mention. Rick, in particular,
is "very clear" that they can. Bill is a bit more circumspect, as usual,
but still I have a problem with:

Even with perfect control, we can still look at the conditions under
which the reference level changes, and that can tell us about higher
systems. We don't need detailed analysis to get the general picture
correctly.

From what observations do you infer that a reference level is changing?

And is that change not itself an aspect of controlling (perfectly or
otherwise) some other perceptual variable? The observations are still of
the same kind. One can infer WHAT is controlled, at least to some
approximation, and one can do so better the more perfect the control.
But the same issues still (seem to me to) arise.

To see
behavior keeping the controlled variable constant as "failure of
control" is rather strange. On the contrary, this would be evidence of
_ideal_ control, so accurate that we can't detect any change in the
controlled quantity. Your attempt to pique interest by saying that we
are interested mainly in failures of control carries all the wrong
messages.

Now who's making silly statements? Does ANY of that refer to something
I wrote? How on earth does keeping the controlled variable constant get
identified as "failure of control." And I didn't say that you were
interested mainly in failures of control. I suggested that perhaps
it might be a good thing to develop an interest in the ways control
deviates from perfection--the ways it fails.

It implies that the worse control is, the better we understand
it. In the limit, we would understand control the best of all if it
didn't even occur.

That's really ridiculous, and not implied by anything I have (consciously)
written. If you can't tell WHAT is controlled (which you can't, if control
is very poor), then how on earth would you expect to discover HOW it is (not)
being controlled. It may be a good rhetorical debating trick to impute
a nonsensical position which you can then destroy. But it doesn't help us to
advance our mutual understanding, which I take to be the point of posting
to CSG-L. Notice that I never did that to you; what I did was to note that
you did NOT subscribe to the views that your (or Rick's) words seemed to
suggest.

You said, for example, that you didn't want to confuse behaviorists by
talking about how control worked, but that they had first to see that
control is happening and that it explains more than they can explain--or
something like that. My comment was that to do so would be akin to asking
them to believe in magic (i.e. unexplained technology) when they would be
much more likely to trust something whose workings were explained.

Do I present things to people as magic?

No, never. But your words said that you prefer to do so.

The words might have said something odd, but past context made it quite clear
that here, as elsewhere, you could not have meant what they said. This is
quite different from what you seem to be doing with my statements. You take
something perfectly reasonable (at least Bill Leach seemed to understand)
and reinterpret them as total nonsense. Even if I had misspoken so badly,
you should know better than to treat what I said as if I could have intended
to be ridiculous.

Anyway, back to the question: would one of you care to discuss the issues
of how to study the internal organization of the (presumed) hierarchy, or
should we let them drop for now? I'd be happy to let them drop, since I'm
shortly going to take a couple of weeks of holiday and there's lots I need
to finish up before I do. I'm sure the same questions will come up again.

Martin

2

[From Bill Powers (950706.1545 MDT)]

Martin Taylor (950705.1700) --
RE: imperfect control

You're actually much better at simple
straightforward PCT analysis and teaching.

     Better than very bad is worth something, I suppose.

It comes out to "very good at PCT," so perhaps it doesn't start at "very
bad."

···

------------------------------------------------
     I'm aware that the simplest description of controlled perceptions
     is likely to involve units that correspond to less complex
     perceptions that might also be controlled.... etc.

     But where in all this is there any approach to the internal
     organization of the control hierarchy? Where is there information
     about, say, the forms of output functions, or the reference-setting
     linkages between higher and lower control systems, or about the
     potential conflicts among control systems at a level, or...? Where
     is there information about the changes that occur in the (presumed)
     hierarchy during learning or forgetting?

Before we can make an approach to the internal organization of the
control hierarchy, we have to know what we are trying to model. In some
cases, as for the arm model, our tracking experiments, some results from
operant conditioning experiments, and a handful of demonstration
experiments, we have some idea of how control behavior appears and at
least a few of the mechanisms involved, so we are able to construct
provisional working models. But even as we build these models and use
them to interpret observations, we are aware of many details we are
leaving out -- necessarily, because we don't have data. Constructing a
picture of the hierarchy is a matter of building a piece here and a
piece there, seeing what is missing, and trying to fill in the gaps with
new experiments.

I have little hope that we can anticipate much about the final model.
Building up complex conceptual structures built out of repeating
identical elements, I think, is a waste of time; there is no reason to
think that the elements of one level will resemble those of another
level beyond the overall resemblance that is common to all control
systems. The input computations may vary drastically from one level to
the next; there may be modes of control involving manipulation of
parameters of lower-level systems; we may find control systems that
switch their input connections around from place to place in the lower-
level world, applying a single control process to different sub-sets of
lower-level perceptions. At some levels there may be model-based control
systems like those that Hans Blom describes, if not in that exact form
then in some equivalent form.

There's nothing wrong with exploring possible structures, if only to get
an idea of the kinds of modeling problems we may be up against. But
without some specific type of observable phenomenon to focus such
efforts, the "possible" structures are very unlikely to have any
relevance to the real system.

Beyond that, I don't see much value in generalizations such as yours
about studying "failures" of control rather than "successful" control.
That is a dichotomy that doesn't exist; all working control systems are
"failures" if you look at them closely enough, and they are all
"successes" if you consider them in the larger context of what they are
supposed to accomplish. The only valid point that you're trying to make
is that we must measure variations in the controlled quantity to get a
picture of the dynamics and other details of the control system. But my
answer to that is "what else?" Haven't we been doing that all along?
Isn't that the standard engineering approach to control system analysis?

The point you're missing is that this level of analysis is not essential
to building up a general picture of the organization of the hierarchy.
We can get a good picture of how the kinesthetic arm control systems
work without doing a ten-year research project. And having seen what
perceptual signals these systems control, we can then try to discern
larger control loops in which control of the lowest-level variables is
the means of controlling higher-level variables, as we use joint-angle
control systems to control arm configuration, and arm configuration
control systems to control visual relationships. For purposes of
understanding organization, we can treat the lower systems as if they
were ideal; there is no practical effect of this approximation on the
way the higher systems work. Questions of stability and dynamics are
really details that make no difference in the big picture (especially
since we know we are dealing with a stable system), and we are still at
the stage where the big picture is the main concern.

     From what observations do you infer that a reference level is
     changing? And is that change not itself an aspect of controlling
     (perfectly or otherwise) some other perceptual variable? The
     observations are still of the same kind. One can infer WHAT is
     controlled, at least to some approximation, and one can do so
     better the more perfect the control. But the same issues still
     (seem to me to) arise.

Yes, the same issues arise, and we will no doubt make mistakes. But we
can make very reasonable guesses about higher systems by assuming that
we understand what variables lower systems are controlling, and putting
those controlled variables together with other observations to
understand what a higher system is controlling.

Obviously we infer that a reference level has changed when we see that a
well-understood control process has suddenly switched to seeking a
different goal -- from keeping the cursor on the target to keeping the
cursor beside the target, for example. But down the road there will be
much more sophisticated approaches possible, where we don't require that
the reference signal simply change from one fixed value to another. The
overall problem is one of solving a set of simultaneous differential
equations, looking for steady-state solutions in terms of hypothesized
controlled variables or several types and at several levels. We don't
know -- at least I don't know -- how to do this right now. I suspect
that we will always need to be able to arrange for fixed reference
levels at _some_ level in order to study lower systems, but this has not
proven to be an obstacle yet. The problem is like that of analyzing the
behavior of any complex system.

     Anyway, back to the question: would one of you care to discuss the
     issues of how to study the internal organization of the (presumed)
     hierarchy, or should we let them drop for now?

I was under the impression that this is what we are doing, bit by bit.
We think of phenomena to set up as experiments; we do the experiments
and see what happens; we try to interpret what we find as some specific
control process; and we try to devise working models that will reproduce
the phenomena. These models constitute guesses about how parts of the
internal organization work. When we have enough part-models, we can try
some tentative hierarchical arrangements of the parts to see if they
will work, too. Along the way, there may be some useful evidence from
neurology that will help us converge on a physiologically more correct
model. I don't see any other way to approach the problem that is likely
to remain relevant to the real system.
-----------------------------------------------------------------------
Best,

Bill P.

3

[Martin Taylor 950707 14:20]

Bill Powers (950706.1545 MDT)

We have differences of degree, I think, not of kind.

we try to devise working models that will reproduce
the phenomena. These models constitute guesses about how parts of the
internal organization work. When we have enough part-models, we can try
some tentative hierarchical arrangements of the parts to see if they
will work, too. Along the way, there may be some useful evidence from
neurology that will help us converge on a physiologically more correct
model. I don't see any other way to approach the problem that is likely
to remain relevant to the real system.

I'm happy to leave it at that, for now. But I hazard a guess that we will
have future disagreements in this area, only partly because of problems of
wording. There's nothing in your posting to which I would object
qualitatively, and with most of it I would unhesitatingly agree. But
there remain differences of personal style of enquiry, and I think those
differences will remain.

For once I was on the other side of the disagreement; usually you dislike
my being too speculative, and introducing possibilities for which you can
see neither proof nor disproof. This time I was enquiring into the details
of how one might acquire data to look at the structural organization, rather
than speculating on what that organization might possibly be.

Building up complex conceptual structures built out of repeating
identical elements, I think, is a waste of time;

It is at least a more parsimonious approach than to assume _a priori_ that

The input computations may vary drastically from one level to
the next; there may be modes of control involving manipulation of
parameters of lower-level systems; we may find control systems that
switch their input connections around from place to place in the lower-
level world, applying a single control process to different sub-sets of
lower-level perceptions.

All of these statements may well be true. How would one go about looking
for instances of them? And are any of them incompatible with the notion
of building from complex structures of repeated elements? It is demonstrable
that such structures can produce wildly different kinds of computations, and
it may be the case that control systems with drastically different input
functions are nevertheless constructed with similar elemental units.

As I now conceive the Layered Protocol Theory of communication, each protocol
node is formed from a complex of ordinary ECUs, independently controlling
different scalar perceptions. Likewise, even in the most straightforward
approach to a control hierarchy, any perception at a high level is
inevitably the result of a complex of perceptual computations performed
at lower levels as well as internally to the ECU in question. I don't
see a dichotomy between having a complex function and having a function
built of simple repeating elements. After all, is it not true that the
whole of algebra can be built from the "neither-nor" operator, used often
enough in a logical structure?

without some specific type of observable phenomenon to focus such
efforts, the "possible" structures are very unlikely to have any
relevance to the real system.

Right. And different presuppositions lead to different observations. There's
no such thing as an a-theoretic observable. Observations and theories
tend to form a saturating positive feedback system (more than one loop),
which is one reason why it is so difficult to get people to shift from one
scientific worldview to another, and why the shift is so dramatic when they
do (the "epiphany" experience).

···

----------------
I don't expect to contribute much before around July 26. Have a good
meeting, scientifically and socially productive.

Talking of the meeting, I'd aprpeciate it if the authors for the special
issue could get together and discuss the relationships among their
contributions. I had hoped to send a draft of mine (using PCT and LP
in designing interfaces to computers and consumer electronics), but
with four authors the editing takes longer than I had hoped.

Hi, Rick. Sorry miss the opportunity for long chats. Maybe next year.
Will you have a draft paper for the issue to show in Durango?

Martin

4

[From Bill Powers (950707.1600 MDT)]

Martin Taylor (950707.1420)--

     For once I was on the other side of the disagreement; usually you
     dislike my being too speculative, and introducing possibilities for
     which you can see neither proof nor disproof. This time I was
     enquiring into the details of how one might acquire data to look at
     the structural organization, rather than speculating on what that
     organization might possibly be.... All of these statements may well
     be true. How would one go about looking for instances of them?

I've been suggesting for a very long time that the place to look at the
structural organization of the brain is _in the environment_. The
environment that you perceive and control is the one that reveals the
brain's structure. If you see sensations, the brain perceives
sensations. If you see configurations, the brain perceives
configurations -- or at least _your_ brain does. The same goes for all
the other levels I've tried to define. They're right out there in the
world you experience. There is no need to guess at what goes on inside
that blank volume in your world where your head seems to be, although
invisibly. The entire world of controlled perceptions is sitting right
there in front of us. We have the front-row seat.

When I reach out to touch a key on the keyboard, I am looking at
perceptions and relationships among perceptions. My hand that carries
the finger is a perception, the distance from the finger to the key is a
perception, the pressure I feel on and from the key is a perception, the
movement of the key is a perception, "downward" is a perception. It's
all right there. All we have to do is sort out all these perceptions
into some kind of order. I know that if I shut my eyes, all the other
perceptions go away. So all those perceptions must be built out of what
enters my eyes when I open them. I know that to type "the" I have to
bring one or more fingers into a certain relationship with three keys on
the keyboard; therefore relationship is a perception that I can control.
I know that these relationships must be brought about in one and only
one sequence to get "the": therefore I know that I can perceive and
control sequence. And so it goes. Everything we need to know to build a
model of the whole hierarchy (if it is a hierarchy) is woven into the
world we experience.

This whole idea is so simple that I despair of ever getting it across.
When I say "perception," people automatically think of what is going on
behind their eyes, not what is going on in front. What we're trying to
model is the part behind the eyes, but the phenomena we're trying to
account for is in the part right here before us.

So it's really simple to find examples of controlled variables at all
levels of organization. The problem is to see them for what they are. As
we experience them, they aren't sorted out into layers with boxes and
arrows showing what depends on what. We have to work that out for
ourselves, by experimenting with the world and trying to get a sense of
the dependencies. It's just like being a baby again, only this time a
baby who is aware of the nature of the game.

     And are any of them incompatible with the notion of building from
     complex structures of repeated elements?

That remains to be seen. It is difficult for me to understand how the
same structure that creates a sensation-perception (a simple weighted
summation, we presume) can also create the perception of sequentiality,
or the perception of rotation rate, or the perception of a syllogism. It
would, of course, make life simpler for the theoretician if there were
one simple building-block of brain organization, repeated again and
again with only changes in the parameters and interconnections. But I
can see no evidence to support that idea. The evidence against it is
strong, even in the neuroanatomy of the brain. Different layers of the
brain contain neurons and networks with very different properties; there
is no simple repeating structure that extends from bottom to top.

Judging from the kinds of perceptual classes I think I have identified
in my environment, it would be extremely surprising to me if the
perceptual functions at one level worked anything like perceptual
functions at the next higher or the next lower level. This system
evolved, it wasn't designed. It probably evolved layer by layer, each
new layer being added to deal with conflicts of new types. Comparative
HPCT will be a fascinating subject once someone invents it.

···

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

Bill P.

5

[Martin Taylor 950710 13:15]

Bill Powers (950707.1600 MDT)

I'm gone in an hour or so, for a couple of weeks. But first...

The entire world of controlled perceptions is sitting right
there in front of us. We have the front-row seat.

When I reach out to touch a key on the keyboard, I am looking at
perceptions and relationships among perceptions. My hand that carries
the finger is a perception, the distance from the finger to the key is a
perception, the pressure I feel on and from the key is a perception, the
movement of the key is a perception, "downward" is a perception. It's
all right there. All we have to do is sort out all these perceptions
into some kind of order.
...

So it's really simple to find examples of controlled variables at all
levels of organization. The problem is to see them for what they are. As
we experience them, they aren't sorted out into layers with boxes and
arrows showing what depends on what. We have to work that out for
ourselves, by experimenting with the world and trying to get a sense of
the dependencies. It's just like being a baby again, only this time a
baby who is aware of the nature of the game.

Unless I totally misread you (a distinct possibility), you are asserting
that subjective experience of one's own hierarchy--perception OF perception
rather than simple perception of the environment--is our main access to
the organization of the control hierarchy. This may be all well and good
for one's own experience, but where does that leave you when someone else
claims to have a different experience that suggests a different kind of
organization?

Even when all our perceptions are our own, including our perceptions of
what other people say, nevertheless if we believe in "science" we have to
have some kind of observations that "ought" to correspond to observations
other people will agree that they have made. We need to work with
perceptions of the environment, rather than with perceptions of perceptions
of the environment.

What we're trying to
model is the part behind the eyes, but the phenomena we're trying to
account for is in the part right here before us.

And it is those phenomena that we have to try to select in such a way that
various people can agree to their existence.

I think I MUST be misreading you. Maybe by the time I return, I'll be able
to see another interpretation of what you wrote.

Have a good meeting.

Martin