[From Rick Marken (920826.1000)]
From Greg Williams (920826)
Thanks!!! Very helpful.
Avery Andrews (920826.1615)
Somehow, for some reason, that fact that you need feedback to point
at things or stand upright doesn't seem to get to AI-ers
The "Blind men" paper shows that feedback is THERE whether you need it
or not. Organisms are ipso facto locked in a loop and behaviorally it
is clear that the sense of the feedback in this loop is negative. Who
cares what ai-ers think? I'm not much interested in the Pope's ideas
about population growth either.
The right examples, I think, have
a capacity to burn holes in people's minds that mere arguments don't
We've been demoing this stuff for YEARS. Obviously using the WRONG examples.
I'm pretty convinced that there are no right examples (any more than there
are the right stimuli to make people do stuff). The only thing that can be
right is the willingness to learn -- and I haven't seen much of that amongst
ai types. In fact, the usual approach is to explain to us why our demos
and models are not (or don't need to be) doing what they are doing
-- controlling perception.
Well, could you build a system that could look at a video screen &
process what it sees there well enough to play the game?
Sure. Believe me, we can design perceptual systems with the best of them;
and if we can't, we are happy to learn how they do it and incorporate
those results into our models, when necessary. I have always said that
the main substantive problem for PCT behavior modelling will be learning
how to build functions that perceive some of the more complex variables
that we obviously control. The whole idea of PCT is that behavior IS
perception in action -- controlled perception. It's all about perception.
So if anybody is building great perceptual models, that is not an argument
agaist PCT and for the approach to behavior of the perceptual modeller --
it just shows that there is information about perception that we can use.
Suppose, for example, that we discover that "honesty" is something people
control. We certainly can't build an "honesty detector" now (or in the fore-
seeable future). Would such a situation make PCT less plausible? NOT.
Suppose the advocates of some other theory could build an honesty detector.
Would that make the other theory more plausible than PCT. NOT. But it
would certainly suggest that we should look at their honesty detector algorithm
for use in our PCT model.
PCT modellers are well aware of the problems of modelling perceptual
functions and we TRY to keep up with the state of the art. So I think we
are at least at the level of a Chapman and Agre, unquestionably at the level
of a Beer or Brooks.
Martin Taylor (920826 11:00)
Rick is controlling for the
perception that dynamics is irrelevant to the operation of a control system,
whereas I believe that the control system is defined by its dynamic behaviour.
Dynamics are NOT irrelevant to the operation of control systems; if the
dynamics are wrong, there is NO control. But I'm not particularly interested
in modelling the dynamics, per se. Once the system works (so that it behaves
(dynamically) just like the real system) then I'm happy. I'm MUCH more
interested in WHAT is controlled (what perceptions), WHY its controlled
(higher level goals?) and HOW its controlled (use of lower level perecptions?).
Dynamics "fall out" of the operation of the control model. They certainly
have to be there (and appropriately -- witness the discovery of the need for
a lag to account for the marken effect) but it sounds like you want to model
the dynamics itself. That's fine with me (they have been doing that in human
factors for years -- and missing completely the MAIN point, that these
systems are controlling perceptual variables relative to internally specified
references for those variables).
Not everything can be relegated to the result at infinite time, as the
stability equations often trotted out assume.
I agree. But once you know that there is a dynamically stable solution to the
control equations, then you can present the algebraic results to clarify
the main points of control -- control of perception, resistance to disturbance.
You can summarize the main properties of a lever without giving the differ-
ential equations for it, can't you??
Why is there a conceptual problem with understanding the importance of the
differing numbers of degrees of freedom in different parts of the hierarchy?
Because I don't know what phenomenon this relates to. I am interested in
modelling human behavior; that's what I do well. I can't do all these fancy
mathematical analyses of properties of the mdoel itself. If you want to
do that then go ahead. You may find out something important about the
model and that would be great. But I'm more interested in the phenomenon
of control. So if your degrees of freedom analyses of the model predict
something about the phenomenon of control, then let me know and I'll test it.
Dynamics is central to the importance of PCT as a psychological theory. As
a static theory it only presents a trivial truth that has some amusing
consequences. You'll never persuade the world of its beauty by repetitive
ranting based on that trivial truth.
I don't think I could possibly disagree with you more. I presume the
trivial truths revealed by the static equations are
p = r and o = -kd
Perception (p) is made to match an internally specified reference (r) and
this is accomplished by producing outputs (o) that cancel the effects of
disturbances (d) on p.
OK. There's the trivail truth.[If it's so trivial, why don't psychologists
TAKE IT FOR GRANTED AND ACT AS THOUGH THEY UNDERSTOOD IT (like by not doing
research in the stupid way it is currently done)]. Now, please explain the
non-trivial truth that is revealed by dynamic analysis.
Forgive me, Bill, for I care not what I do.
Your resistance to disturbances leads the outside observer to doubt the
truth of this claim.
That's the fact, Jack.
Regards
Rick
ยทยทยท
**************************************************************
Richard S. Marken USMail: 10459 Holman Ave
The Aerospace Corporation Los Angeles, CA 90024
E-mail: marken@aero.org
(310) 336-6214 (day)
(310) 474-0313 (evening)