Skeptical about information

[From Rick Marken (931115.1100)]

Martin Taylor (931115 10:40) --

I said:

So your suggestion to Bill P. that he be "skeptical" seems a bit gratuitous,
especially since you have not suggested a similar degree of skepticism from
those (like Hans) who are advocating feedforward models.

Martin replies:

There is an intrinsic difference between being sceptical about claims of
"some, occasionally" and claims of "all, always." "Some, occasionally" is
itself a statement of scepticism, which Hans has been making.

Does "skepticism" just mean NOT using words like "all" or "always"?
What about observation and TESTING? Isn't this what skepticism is
about. It seems to me that what you call "skepticism" is what I would
call "sophistry".

Be that as it may, I seem to recall Hans saying something about
feedforward being involved in all behavior? Ah, here it is: Hans said
"feedforward is everywhere". Well, OK, he didn't say it's "always"
everywhere. Does that make Hans skeptical about the prevalence of
feedforward processes in behavior?

You still don't seem to have realized that testing for whether the variable
is controlled is not a test for whether there is a feedforward component
in the action mechanism.

What makes you think that? I never said that feedforward models are ruled
out by establishing that a variable is under control. What I said (and, of
course, Bill P. independently said exactly the same thing) is that
feedforward models are not NEEDED until you make observations that
suggest that they ARE needed. Observations of control behaviors that
CANNOT be handled by a pure feedback model are the ones that might make
one want to consider adding feedforward to the model.

I said:

The Lang-Ham model explains nothing that cannot already be quite
precisely handled by the existing model.

Martin replies:

A big claim, indeed. Bill and I have been having a little dialogue with
simulations on this question over the weekend (no answers yet).

The Lang-Ham system may confer some interesting performance charateristics
on a control system. But Lang-Ham is not YET needed as a component of
a MODEL OF HUMAN BEHAVIOR -- unless you know of some data that CANNOT
be handled by a pure feedback model and that MIGHT be handled if the
Lang-Ham connection were added.

I see that you still don't understand the basic concept of information.

I guess not. If it's THAT subtle, maybe my little brain will just have
to do without it.

The entire functioning of the control
loop depends on not "whether" but on "how much of"

THAT kind of "information"
about the disturbance exists in the perceptual signal

Well, then, how come I have been able to build functioning
control systems that behave almost exactly like people -- without
having ANY idea 1) what information about the disturbance is or
2) how much of it is in the perceptual signal?

How about a SIMPLE explanation (one that even I could understand) of
WHAT information about the disturbance is, HOW it is measured and
WHY the functioning of the control loop depends on how much of it
there is? Also, how about a quick lesson for modellers on how to
vary the amount of information about the disturbance that
is in the perceptual signal? If I could vary this information using
a computer program I think I could understand you point a lot better.

As I said to Tom last week, it seems fruitless to continue this discussion
here.

I bet it would be fruitless anywhere. But it would at least be more
fun here.

It gets so wound up in misunderstandings of the basic notions of
probability and information that nothing useful is communicated.

There's got to be SOME way to get this across to people of average
intelligence (like myself). The net gives you a chance to practice
communicating this important idea to those who are stochastically
and informationally challenged (like me).

I have
been encouraged by a private posting to continue my "Information,
Perception, and Control" paper, and maybe I will do so.

That's great -- and let me add my voice to the chorus of encouragement.
But I think it's worth it to keep trying to make your point in net
discussions. If you keep trying different ways of presenting your ideas
to me on the net, you'll be able to see where my misconceptions and
failures of understanding lie and you can try various ways to get me
past them.

I'm REALLY trying to understand what in the world you mean when you
say things like "...the entire functioning of the control loop depends
... on "how much of" THAT kind of "information" about the disturbance
exists in the perceptual signal". Help me out. Or, if you can
explain it to Bill (who is a LOT smarter than I) then maybe he can
exlain it to me (Bill has been able to explain many difficult
concepts to me -- I'm sure he can help me understand what you mean
by the sentence above -- once you get Bill to understand it himself).

Best

Rick

[Martin Taylor 931115 18:45]
(Rick Marken 931115.1100)

I'm not going to take up the challenge of getting back into the information
discussion that so bored most of the audience last time. Nor am I
going to continue with my interpretations of what Hans has been saying,
since you are as likely to be right as I am. Hans can fend for
himself. But:

feedforward models are not NEEDED until you make observations that
suggest that they ARE needed. Observations of control behaviors that
CANNOT be handled by a pure feedback model are the ones that might make
one want to consider adding feedforward to the model.

As a matter of principle, of course you are right, provided you add
the caveat "of comparable complexity" after "pure feedback model."
Any Lang-Ham or other structure that combines feedforward with feedback
is more complex than the same structure without the feedforward component,
if the other components are unchanged. But if it takes two pure
feedback ECSs to accomplish what one forward-and-back structure will
do, the choice is less obvious. It is obvious that ANYTHING that a
feedforward system can do can also be done at least as well by a pure
feedback system, because the pure feedback system can be transformed
identically into a feedforward one by setting the PIF to a zero-gain
element. So you will NEVER find any observation of control behaviour
that cannot be handled by a pure feedback model, but that can be handled
by a feedforward model. The right question is different.

Also, how about a quick lesson for modellers on how to
vary the amount of information about the disturbance that
is in the perceptual signal? If I could vary this information using
a computer program I think I could understand you point a lot better.

OK. Three ways right off the bat.

(1) change the bandwidth of the perceptual function.

(2) change the resolution of the perceptual function (for example, by
quantizing its output and altering the height of a quantization step).

(3) Add noise to the output of the perceptual function before the
comparator (equivalent to (2), but works continuously). Make sure the
noise is at least as wide-band as the perceptual function itself.

As I said to Tom last week, it seems fruitless to continue this discussion
here.

I bet it would be fruitless anywhere. But it would at least be more
fun here.

One disagreement, one agreement (Guess which is which). But I would
prefer to spend the excessive time I devote to CSG-L on more productive
endeavours. The information issue has proved so difficult to get across
in this medium that I think the more formal paper should be written and
used as a basis for later argument.

Or, if you can
explain it to Bill (who is a LOT smarter than I) then maybe he can
exlain it to me (Bill has been able to explain many difficult
concepts to me -- I'm sure he can help me understand what you mean
by the sentence above -- once you get Bill to understand it himself).

No fair. Bill is smarter than all of us, and his skill at explaining
stuff is unmatched on this net. Despite that, it is MY lack of clarity
that is the problem that has to be addressed, and I don't think that the
relatively freewheeling and necessarily short communications here are
suited to resolving it. I hope that a considered paper, if I ever can
get to it, will be much better. Besides, what I want isn't just an
analysis of what happens in a control hierarchy. I want to find an
argument that will satisfy Tom that stability in an informational sense
implies control. My original argument that I thought was satisfactory
had too much intuition in it, and Tom rightly called me on it.

Till later.

Martin

[From Oded Maler 931116 09:30 old continent]

* [Rick Marken (931115.1100)]

ยทยทยท

*
*
* I'm REALLY trying to understand what in the world you mean when you
* say things like "...the entire functioning of the control loop depends
* ... on "how much of" THAT kind of "information" about the disturbance
* exists in the perceptual signal". Help me out. Or, if you can
* explain it to Bill (who is a LOT smarter than I) then maybe he can
* exlain it to me (Bill has been able to explain many difficult
* concepts to me -- I'm sure he can help me understand what you mean
* by the sentence above -- once you get Bill to understand it himself).
*

There is a crucial point the PCTers like yourself neglect, and it is
the central point for other people interested in control (engineers,
mathematicians, etc.) You emphasize the question "what is a system
doing", namely what is it controlling for. They ask the question
"*why* does it work (in the real world)?" If you think about it, you
will see that you don't adress this question seriously. Saying thar
reorganization takes care of it, does not answer the question. This
question cannot be answered analytically (by building a model of a
dynamical system for, say, all possible
person-in-the-room-making-coffee situations, in terms of elementary
sensations, and proving that a control system will always converge to
a drink-coffee state), neither by simulation (building a numerical
model of the same system and trying a small fraction of the
uncountable number of possible situations). Hierarchization is
apparently a good idea, and the intuiuition that, I think, Martin has,
about *why* it works in the world, is related to gross properties of
systems in general, to information. That was mu guess.

Other comments: due to my recent professional adventures in elementary
dynamical systems, DE, etc., I think I track is a misconception about
what "linear" means. When someone speaks about non-linear dynamics
he means the the differential equation is non-linear, i.e.,
d(x1,..,xn)/dt - f(x1,..,xn) where f is a non-linear function.
This means, the interdependence between the various state-variables
is much more intricate than when f is a linear combination. So when people
say that PCT cannot cope with non-linear dynamics, it does not mean that
PCT models cannot control for a non-linear perceptual function.

I've been reading recently about so-called intelligent/autonomous
control (engineering point of view). It is essentially a question of
how a discrete controller can (is required in order to) control a
continuous plant. If you translate it to HPCT terminology, it
addresses the issue of levels, the passage from the continuous lower
levels (which everybody knows how to control in some operation
regions) to higher discrete levels, that detect qualitative changes,
and control them by discrete actions.

--

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