Mary's grump

CSGnet Archive CSG1993
1

[from Mary Powers 931117]

Hal.

Seems to me that a tetrahedronal anything is a solid figure with
four sides. I've never seen one that controls anything. Either
this is a metaphor (and one which does not particularly resonate
with me) or a hallucination. After your last post, I'm leaning
towards the latter conclusion.

When you fire off a blast onto who-knows-how-many-nets, you've
been asked not to include csg-l. This is because this is a very
busy net, and some participants have limited capacity in their
mailboxes. You have said we can just delete your stuff if we
don't want it. Why put the burden on us? Pick up your own socks.
If you have something to say to this net, relevant to this net,
fine. Otherwise, go fly your kite somewhere else.

Oded.

I don't understand your point about PCTers neglecting the
question "why does (a control system) work in the real world?"
You say this can't be answered analytically or by simulation in
(for instance) the case of making a cup of coffee. How can we
prove that a control system will always converge to a drink-
coffee state?

No one can prove that ANYTHING will always happen. That next shoe
you drop just might fall up, right? All we have going for us
(which works pretty well) is a hypothesis, a body of theory,
called physics, that says the shoe will always drop.

In the case of the coffee drinker, what we have going for us is a
hypothesis, a body of theory, called control theory. This says
that control systems have reference signals. In the real world,
most of those control systems are alive. In the living case, we
call reference signals intentions. The reason a control system
converges to the drink-coffee state is because it WANTS to have a
cup of coffee, and, given a fairly friendly environment (which
the coffee drinker has arranged to have, with the help of a lot
of other control systems) the control system does indeed get the
coffee. Every time. You don't have to solve this analytically or
simulate it - just watch it happen.

The question of "why" is embedded in the question "what is a
system controlling for". If a system is controlling for a certain
state of something, it's because the system wants it in that
state. And to do that it's got to be a control system, because
that's the only kind that has wants.

Hans.

You say: "feedforward" denotes a type of control in which no
"feedback" (perception of outcome) is required. Feedforward
normally refers to a process of computation of output based on
inputs only.

Inputs of what? As I understand control systems, their inputs ARE
perceptions of outcome. What other inputs might there be?

OK. Suppose the particular outcome can't be perceived. You are
saying that in that case the system has to compute outputs. Like
the linkage between the steering wheel and the wheels of a car.

This doesn't make sense. The steering wheel doesn't compute
anything, ever. The linkage is a passive mechanical device. If
the steering wheel is turned, the wheels turn. Gracing it with
the name feedforward makes more of it than it actually is,
doesn't it?

In the satellite example, when the signal is lost, a computation
has to be made to point the receiver where the operator expects
the satellite to be. This isn't computing output. Certainly the
computations lead to particular outputs, the movement of the
receiver to a certain position. But the consequence of the
computation will be PERCEPTIONS of a) the receiver being in a
particular position and (hopefully) b) signals coming in again
from the satellite. The output is the means, not the end: what
matters is the perception: is the receiver pointing where it
should be, is the signal now being received? You can output until
you're blue in the face and it doesn't mean a thing - you won't
know where the receiver is pointing or whether it's picking up a
signal unless you perceive those things. The problem is we're so
used to perceiving what's going on that we don't notice the
perceiving. We say we're looking at outputs. How are we doing
that? We've got to be perceiving them. There's no other way. You
seem to be able to see that to fine tune the receiver you need
feedback. But you need feedback to accomplish "feedforward" too:
the whole process of guessing and computing and pointing is the
working of your control systems, whether or not you really know
where the target is.

Martin.

"A control system is just a particular type of dynamical system".

Yes. But the point is that of all the many types of dynamical
systems, this is the only one that is of interest here. I'm sure
the general subject of dynamical systems is fascinating to many
people, but for our purposes all the other kinds there may be are
irrelevant. If you go to a supermarket for a gallon of milk, the
fact that there are several thousand other things on the shelves
doesn't really matter, does it? A dynamical system that doesn't
resist disturbances is a loaf of bread. We're here for the milk.

I seem to be somewhat grumpy this morning :frowning:

(later)
What I see going on here, and it's been going on for months, is
people who think PCT is interesting and has its points, but who
see Bill and Tom and Rick as dancing on a high wire without a
safety net. How can they possibly have such reliance on control
theory that they don't need the security and comfort - and wisdom
and right-thinking - of (for instance) feedforward or information
theory or dynamic systems theory or whatever.

Most of this endless quibbling is between people who have lots of
training and professional experience with control systems of the
artificial variety, or who have approached the same issues PCT
addresses, but in a different way - between those people and the
three dancers on the wire. Those people have their feet on the
wire, but they've got to have that net, and a safety harness, and
they aren't letting go of the platform, either.

Meanwhile Bill and Tom and Rick are saying "what's with the
harness and the net?" Because for years they've been out there on
that "wire", and it isn't a wire at all, but solid ground. A lot
of what PCT "leaves out" or "doesn't explain" - a lot of the
really valuable, important stuff that PCT dismisses so arrogantly
and unwisely - is harnesses and nets. Stuff that's really
superfluous (and even a barrier) to understanding.

I think I'm becoming disillusioned with the net. I can look over
at the bookcase and see 6 computer-paper boxes full of printouts,
a sickeningly large proportion of which is quibbling of the sort
that came through this morning, and Bill endlessly and patiently
explaining, and explaining again, and explaining yet again - and
nobody seems to be able to stop and think MY GOD, how is he able
to keep these discussions going with psychologists and
information theorists and control engineers and roboticists and
AI types and linguists and sociologists and educators and
therapists and organizational development people and bicycle
designers and human factors mavens - and the answer staring all
you klutzes in the face is not that he's some kind of frigging
genius - pretty smart, sure - but that his ace in the hole, the
ground on which he takes his stand, the source of his insights
and analyses and ability to talk to people in all these
"separate" fields, is PERCEPTUAL CONTROL THEORY. Simply that, no
frills, none of this baggage you keep dragging in. Just a simple
little unified theory of living systems which enables him to
handle the complexities of all these fields. Doesn't anyone else
want to try it too? The only catch is: you'll never get anywhere
with PCT if you keep looking at it from where you are. You need
to look at where you are from the point of view of PCT. Then
maybe Bill could quit teaching kindergarten over and over again,
and have some time to get some development done, and have some
company doing it.

I guess I'm still grumpy :slight_smile:
Hi, Rick.

Mary P.

2

From Tom Bourbon [931118.1115]

[from Mary Powers 931117]

. . .

(later)

. . .

Meanwhile Bill and Tom and Rick are saying "what's with the
harness and the net?" Because for years they've been out there on
that "wire", and it isn't a wire at all, but solid ground. A lot
of what PCT "leaves out" or "doesn't explain" - a lot of the
really valuable, important stuff that PCT dismisses so arrogantly
and unwisely - is harnesses and nets. Stuff that's really
superfluous (and even a barrier) to understanding.

Mary, stay grumpy! :-))

The ground out here is solid indeed! When I read that section of your
post, I was immediately reminded of the Indiana Jones movie, in which Indy
stands at the brink of a yawning crevasse, torn between his new insight that
there *must* be an invisible bridge to the other side, and his powerful fear
of stepping into what looks like empty space. Finally, he raises one foot
and plants it forcefully -- on the invisible bridge. Even while he and his
associates make their escape across the bridge, there remains the dizzying
fact that when they look down, there is only the yawning emptiness.

That thought was followed with recollections of a topic discussed at the
latest CSG meeting -- the "visual cliff" experiments of Eleanor J. Gibson
(and subequently a host of others). The "visual cliff" is a large box that
from the side looks like:

sheet of heavy glass
               >
           > \|/ |
           >____________|
           > >\\\\\|
space-------> |\\\\<-----solid
           >______|\\\\\|
            /|\
             >
             >
          floor

The surfaces of the solid section and of the floor are covered with a
distinctive red and white checkerboard pattern. Visually, the pattern
"affords" an unambiguous perception of an elevated solid surface and a
surface that drops away steeply to a solid floor, some distance below -- a
visual cliff. In fact, the sheet of heavy plate glass provides a continuous
solid surface above the entire cliff.

On the title page of Gibson's _Principles of Perceptual Learning and
Development_ (Appleton-Century-Crofts, 1969) is a famous photograph of "a
kitten looking down a visual cliff." Not shown is the tiny band of PCT
modelers, dancing on the glass above the deep pit.

Come on, gang! Plant your feet in midair.

Tom

3

[Martin Taylor 931118 11:30]
(Mary Powers 931117)

Always nice to get Mary's views. A refreshing note.

Martin.

"A control system is just a particular type of dynamical system".

Yes. But the point is that of all the many types of dynamical
systems, this is the only one that is of interest here.

A point I tried to make in posting to Gavin. The essence of this
particular sentence is to make the bridge to those who say that
dynamical analysis is wrong or irrelevant. It is neither, even within
the context of control. It is just that once one knows that we are
dealing with control, specialized analyses and experiments based on
that fact are much simpler to do and to use than are analyses based on
general dynamical theory. There's no conflict, but there is good reason
for not worrying too much about the general dynamics, so long as we
are considering a particular system that is actually controlling.

...Bill endlessly and patiently
explaining, and explaining again, and explaining yet again - and
nobody seems to be able to stop and think MY GOD, how is he able
to keep these discussions going with psychologists and
information theorists and control engineers and roboticists and
AI types and linguists and sociologists and educators and...

One may well wonder, indeed. My guess, and it is only a guess, is that
he is looking for two perceptions: that more people come to understand
fully the force and beauty of PCT, and that he might gain some new
insights into how PCT actually works in living systems.

Anyone who participates in this discussion (which is really very tight
technically and open-minded socially compared to other discussion groups
on the Internet), must be assumed to have some interest in perceptual
control concepts, if only to find out whether they might apply in their
own areas of interest. Some might actually have useful ideas, coming as
they do from different backgrounds. But those useful ideas can't be
made useful until the people concerned DO understand where they might
fit into PCT.

Hypothesis: Bill is always right.

Test: Challenge Bill's ideas.

Result: ??

I have a different hypothesis. Bill is fundamentally right, and is
usually right when he makes a bald statement. But he isn't always right.
And in lots of places Bill states flat out that he is providing a best
estimate that could be quite wrong. He is intellectually honest, and
that makes all the interactions worthwhile. It really makes sense to
challenge Bill's notions, because if they do (very occasionally) turn
out to be wrong, he says so, and the whole structure of PCT is stronger
for the challenge (not for the result of the challenge). If we all just
assumed Bill to be right always, because he usually is, then PCT would
wind up as a very fragile flower indeed: one man's fancy.

the source of his insights
and analyses and ability to talk to people in all these
"separate" fields, is PERCEPTUAL CONTROL THEORY. Simply that, no
frills, none of this baggage you keep dragging in. Just a simple
little unified theory of living systems which enables him to
handle the complexities of all these fields. Doesn't anyone else
want to try it too? The only catch is: you'll never get anywhere
with PCT if you keep looking at it from where you are. You need
to look at where you are from the point of view of PCT.

So, the beauty of a landscape may be seen only from within one picture
frame? I like to see the drops of water fall into the pool, as well
as to look from the cliff over the treetops. One doesn't invalidate
the other. I would add "as well" at the end of your passage.

Handle the complexities of all these fields? Do Newton's beautiful
graviational laws handle all the complexities of multibody dynamics?
Sure they do. Can you use them for such prediction? Not easily. There
are better ways, averaging huge numbers of bodies and calling them
galaxies, for example. Likewise there are better ways of handling
issues in all these difference fields, recognizing that PCT can be
used _when the situation doesn't get too complex_. It's a long way
from using a 3-level 6-wide control hierarchy model to determining the way
people interact when apparently rich people drive through a ghetto after
a particular jury decision involving neither. Sure, we have faith that
PCT applies there. Can you use it to handle the complexities? We don't
even know whether simple feedforward at one control level is ever useful.

Faith, it's wonderful!

I guess I'm still grumpy :slight_smile:

Don't grump at Bill, please.

Martin

4

[From: Oded Maler 931118]

* [Mary Powers 931117]

* I don't understand your point about PCTers neglecting the
* question "why does (a control system) work in the real world?"
* You say this can't be answered analytically or by simulation in
* (for instance) the case of making a cup of coffee. How can we
* prove that a control system will always converge to a drink-
* coffee state?

···

*
* No one can prove that ANYTHING will always happen. That next shoe
* you drop just might fall up, right? All we have going for us
* (which works pretty well) is a hypothesis, a body of theory,
* called physics, that says the shoe will always drop.

Physics is orders of magnitudes simpler than what we are talking about
in human behavior. The simple Newtonian mechanics that tells me that a
shoe (maybe a ball is a better example) will fall, is a theory about
idealized mass points, location, and velocities. A rule of physics
will say that always unsupported masses will fall down under
gravitation. The relvance of this theory to my actual concrete ball
dpends of course on the correspondence between the hat and the
theoretical entity via a hypothetical measurement device. As for
coffee making, the analog of "a ball will always fall" (under such and
such condition) is "O. will always succeed in making coffeee" (under
such and such condition). The former is expressible in "objective"
coordinates: for every (x,y,z) position with z>0, there will be some
time t where the ball will be at (x',y',z') with z'=0. How can you
*express* (in objective coordinates) the fact that for every kitchen
configuration, the agent O. will eventually succeed in pouring into
his mouth a cup of coffee? You don't even have a description of "a cup
of coffee" in objective terms (otherwise the object recognition
problem would have been solved). Now you and me know the answer:
somehow the perceptual hierarchy of the agent has developed in such a
way that whenever some coffee-willing intention signal start blinking,
a process is triggered such that under a *wide variety* of
circumstances, will lead to what we usually observe as drinking
coffee, just by the fact that various control systems in all levels
try to match their respective references. But what in the rules of the
world (including the perceptual machinary) makes this particular
high-dimensional dynamical system converge to a state (more precisely
a region of the state-space) where the signal matches its reference.
This the *why* I referred to.

*
* In the case of the coffee drinker, what we have going for us is a
* hypothesis, a body of theory, called control theory. This says
* that control systems have reference signals. In the real world,
* most of those control systems are alive. In the living case, we
* call reference signals intentions. The reason a control system
* converges to the drink-coffee state is because it WANTS to have a
* cup of coffee, and, given a fairly friendly environment (which
* the coffee drinker has arranged to have, with the help of a lot
* of other control systems) the control system does indeed get the
* coffee. Every time. You don't have to solve this analytically or
* simulate it - just watch it happen.
*
* The question of "why" is embedded in the question "what is a
* system controlling for". If a system is controlling for a certain
* state of something, it's because the system wants it in that
* state. And to do that it's got to be a control system, because
* that's the only kind that has wants.

I hope I have clarified a bit what I intended to say.

As for the meta-level criticism concerning the bandwidth wasted by
educated-but-non-dedicated PCTers, I hope I will be able to respond
to it tomorrow (if I had Bill P.'s work ethics, everything would have
looked differently :slight_smile:

Best regards

--Oded

--

Oded Maler, VERIMAG, Miniparc ZIRST, 38330 Montbonnot, France
Phone: 76909635 Fax: 76413620 e-mail: Oded.Maler@imag.fr