Models and demos -- what sells

[From Rick Marken (920808)]

The recent post by penni sibun got me to thinking about the
problem of making impressive demos of control phenomena.

penni sibun (920807) says:

well, i've seen the crowd demo (thanks to avery) and i've seen sonja
(pengi's daughter). quite frankly, you know, they're both just demos,
and one can look at them, or even play with them, and assume they're
video games that aren't trying to demonstrate any particular theory of

(don't get me wrong--i think both deontrol models and demos is the same

I believe that one of the problems confronted by those of us who are
trying to "sell" PCT with models and demos is the same as the problem
we confront when trying to point out to psychologists that there is
a phenomenon (called "control") that is going on in front of their
eyes that they have not taken into consideration in their attempts
to understand mind and behavior. The problem is that the disturbances,
constraints and calibration problems that make control necessary
and obvious are simply invisible. When you point your finger at a
target, the pointing just seems to happen; the fact that you can
repeat this pointing with great precision seems completely unimpressive.
You just point at the target again and again. Disturbances (such as
changes in your orientation with repect to gravity), constraints
(such as the fixed length of the segments of the arm) and calibration
problems (like the fact that a neural signal never produces exactly
the same amount of muscle tension) go completely unnoticed. When
disturbances are visible (such as movements of the target) they look
like stimuli guiding the response. It is, thus, very easy for those
who want to, to ignore control.

The "little man" demo is a very impressive piece of graphics
programming. But what is really amazing about the little dude is
that he (or she?) CONTROLS. But this aspect of the demo is very hard
to see. In fact, there are no disturbances in this demo other than target
movements. It might help if you could "tug" on the arm and see that
this had little effect on the accuracy of pointing. But basically,
the problem with demonstrating models of control is the same as the
problem of seeing control in normally occuring behavior; what is most
amazing about control is what you can't see. And you can't see the
amazing aspect of control (disturbance resistance, constraint
satisfaction, and calibration compensation) because control itself
prevents these things from having any noticable effect. So it
is the fault of control itself that the process of control is invisible.

In order to see control, you must be the agent of disturbance; you
must be able do something that you know should have an effect on
a variable if it were not under control. If you think a person is
controlling the poition of a limb then you can literally "push" on
the limb to see if the push has the expected effect (movement of
the limb). This "test" must be done carefully -- not too much
disturbance (control systems have limits to the amount of output
they can produce) and an appreciation that control of some variables
occurs more slowly than others (so the disturbance may seem to have
an an effect but will be slowly cancelled if there is control).
My "mind reading" demo is an attempt to show that it is difficult
to see control (the movements of all 5 numbers around a computer
screen appear to be behaviors of the subject) but that it is
absolutely necessary to know what is being controlled in order to know
what a person is doing.

I don't know if there is any real dramatic way to show control;
we keep trying but we obviously haven't found a real "grabber"
that would get psychologists to throw up their hands en masse an
cry "oy vay, I've been missing the point for my whole career; people
don't respond to stimuli or generate outputs -- they control. Now I
have to abandon all my work and start studying control. Damn, how did
I miss that -- I guess that guy Powers was wasn't just a stubborn,
contrarian radical outsider after all."

Ah. Someday.

penni goes on:

is this [S-R modeling (rm)] what the
alife people claim they are building? it is emphatically not what
agre, chapman, brooks and co claim they are building. for these folks
and others in ai, for gibsonian psychologists, for
ethnomethodologists, etc., agents and their worlds are *not
separable*. how can you have a respond to b's stimulus when a and b
are the same thing? that's actually not a terribly useful
simplification of the argument; try this one: how can you individuate
a's-stimulus and b's-response when you can't individuate a and b?

My comments about people building control systems and calling them
SR devices were based on an article I read in a computer graphics
magazine where little "bugs" moved around the 2-D surface of the
screen, avoiding obstacles and, maybe, chasing other bugs. They
were described as SR devices because they were -- then sensed
some input (S) than was converted into a force (R) than moved
the bug. So the bug is SR but it was in a closed loop SITUATION;
the output continuously influenced the input. So there was an RS
connection too. The integration that produced movement provided
dynamic stability and the effect of force on input was such that
increases in force led to a reduced tendency of the input to generate
output -- that is, the sense of the feedback from output to input
was negative. So these bugs were control systems. But the authors
showed NO EVIDENCE of knowing this -- and they carried on about the
wonders of their SR model. It was wonderful -- but it was a control
system. If they had known that, their research could have taken off in
an exciting direction. I'm afraid instead they will end up just trying to
develop more and more elaborate methods of generating outputs in order
to make the behavior of the bugs even more complex.

I know that many scientists are aware of the fact that there is a
relationship between output and input (ie feedback). They just
don't know what this means. I maintain that, whatever they
call their models, they are thinking of them in SR terms. If they
were not, then they would have discovered already that closed loop
systems control and that this means that little of significance can be
learned about these systems by studying input-output relationships
(the traditional approach in psychology).

In 1976, just as I was starting to try to understand PCT, I ran into
a book by Neisser (called Cognition and Reality) where he talked
all about the fact that there is this "perceptual cycle" -- perception
leads to cognition which leads to action which leads to new
perception. It sounded great (at the time) but, it turns out, it
was just words. Without a quantitative anaysis Neisser had no
idea what his loop really meant -- it meant that perceptions were
controlled relative to fixed or variable internal references.
I gave a talk recently were someone cited Neisser's book as evidence
that psychologists already knew what I was talking about. It's at
times like that when I get very pissed at the penchant among
psychologists to treat a turn-of-phrase as understanding.

Trust me. Whatever psychologists call their models, they are SR
models pure and simple and they are treated this way.
When psychology (and the other sciences of life) really catch on
to PCT, it will be obvious, not from their words (they might call
control theory something else) but from their actions; they won't
do research in the way it is currently done. The IV-DV approach will
disappear and you will start seeing studies of controlled variables
using some version of "The test". Statistics will become unnecessary
(unless the controlled variable is itself statistical) and experiments
will be repeated regularly -- because they will always work (just like
physics experiments). When you see that, then you will know that
psychology has become the science of living control systems (as it
should be). I don't imagine that we'll see that for some quite
some time, unfortunately. But we'll hear a lot of people making up
words that sound like control theory ("situated behavior" seems to
be the latest contender). Remember, don't listen to what they say;
watch what they do. When they start testing for controlled variables,
then you know they've got it.

Best regards




Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
(310) 336-6214 (day)
(310) 474-0313 (evening)

From Oded Maler 920811]

[ Rick Marken (920808) about SR <--> Control]

I think I start to understand the reason for some ongoing
misunderstanding. You probably agree that some parts of the control
loop when, seen in isolation are S-R, that is, some input goes in and
some output goes out which depends on the input. And there is a
mechanism responsible for it, it takes reference and perception, say,
and produces force. This is *all* what roboticists mean by S-R. So
many of their systems as you admit are control systems (maybe of
a simple kind) although they might use terminology such as S-R,
reinforcement etc.

On the other hand, in psychology, these terms are much more loaded,
and in particular, they come with a "methodology" based on
statitistical tests etc., which are supposed to reveal the underlying
mechanisms from the outside without opening the black box. Your
insight on control suggests that this is useless, and probably, as
someone educated in psychology, this was one of the great revelations
of your life. But for engineers (unlike reverse-engineers) this
problem does not exist, because they build the boxes, they know the
I-O ("S-R") characteristics of (some of) the components they put in
the loop, and all your protests against S-R, which might be valid in
the context of psychological experiments, seem like fighting
non-existent windmills. They don't know and don't care about ID-IV
tests and the only arguments you may have with them is whether it is
useful or not to use the language of hierarchical servo loops,
reference signals etc. in order to achieve interesting performance (I
refrain from using the word "behavior"..).

Best regards