[From Rick Marken (930316.2000)]
Cliff Joslyn --
In the end, I suspect that waht's at stake is a philosophical
dispute about the nature of CAUSAL vs. INFORMATIONAL processes.
Clearly D is CAUSALLY linked to O, but the INFORMATIONAL link is
hardly (!) clear.
Nicely put, Cliff. There is no question that d(t) is causally
linked to p(t); I'm arguing that there is no informational link.
Martin's and Allan's (where is Allan?) chagrin is understandable
if they thought I was denying the causal link from d(t) to p(t).
Maybe that's why Martin thought that denying the informational
link from d(t) to p(t) made PCT a mathematical abstraction; maybe
he thought that my use of the term "information" was synonymous
with "causal". Is that right, Martin?
Please don't stop, people. Even if it's "just" an argument about
semantics, those tend to be far more important than they're given
credit for.
I'm willing to go on until I'm satisfied that I understand what Martin
and Allan are talking about; but I don't know if I can get anyone to
come out and play.
Martin Taylor (930316 16:45) --
What I want to talk about is a stage leading to that--
an experimental study of PCT and grammar.
Before this "information in perception" debate broke out I was planning
to submit a research proposal to the net also; I don't have time to do
much research (and post to the net too -- he he; well, actually, I'm still
trying to write -- I'm even slower at it than Gary Cziko). But I think
this is a good idea; there may be some graduate students looking in
(or graduate advisors) who might be able to run with some of these
research suggestions.
The objective of the study is to see whether, and to what degree, the
control system can learn to keep its perceptual signals near zero.
I really didn't understand your description of the study. But after
reading this far I am getting the distinct impression that the
control systems you plan to study are simulations. Why not do a
study with real control systems -- like people? If you are intending
to study people then I need some clarification of the vector
representation of the letters and all that. I sure hope you are planning
to do this with living control systems. What we really need in PCT
is good data; who cares how a particular implementation of an algorithm
works? Unless your working in AI or Artificial Life, of course, in which
case you only care about the behavior of the model itself.
Research Proposal
My proposal is to study the human ability to control sequences (I
use letters, just like you do, Martin). The studies would be an
extension of the perceptual studies described in my "Hierarchical
behavior of perception" paper. I think there are at least two very
good reasons for doing these studies; 1) they would demonstrate the
applicability of PCT to a complex behavior-- "sequence produciton" and
2) they would show that limitations in what appear to be the ability
to generate "outputs" may actually be limitations in the ability to
perceive the intended perceptual consequences of those outputs.
I have already set up and done the basic experiment and it works like
a charm -- but there are many variations and some modelling work to
do, and that's where the PhD thesis student would come in.
The basic task is for the subject to control a sequence of letters. The
letters appear at alternating positions on the screen like so (with
time moving from left to right:
A C E B D
B D A C ....
The subject is to keep the ABCDE sequece occuring on the screen. He or
she does by pressing the mouse button when necessary -- a button press
becomes necessary when the sequence is "disturbed" which means that
a new transition rule between letters comes into effect such as
ABCDEADBECADBEC...
^
The new transition rule occurred at the ^; note that the new sequence
has the same letters but none of the transitions are the same as in
the "reference" sequence. Pressing the button returns the "reference"
transitions. If the button is pressed when the reference transitions are
in effect the "disturbance" transition sequence is implemented. So, in
order to control the sequence, you have to push the button at the right
time.
Some quick preliminary (but very highly reliable) results are as
follows: When you speed up the rate at which the letters in the
sequence occur, control (measured as number of correct transitions
in a run of 100 or so letters) decreases. I would argue that this is
because you can't perceive the sequence when it is going too fast; but
it could also be argued that this is a reaction time limit. So I
did the same experiment but changed the disturbance; instead of a
new transition rule using the same letters I just put in a new set
of letters as the disturbance so:
ABCDEABUYTSWUYTSW...
Now the subject doesn't have to perceive the sequence in order to
control the ABCDE pattern; the disturbance can be detected at the
configuration level. And sure enough, the ABCDE pattern can be
controlled at about twice the rate relative to the maximum rate for
sequence control. SO the limitation on sequence control was not due
to a limitation in how quickly the subject can push the mouse button
when the disturbance is detected; at the fast rates the disturbance
to teh sequence is not even detected -- because you can't detect the
sequence (this is subjectively obvious, by the way).
There are many variations on this kind of experiment; but the goal is
to show 1) that complex variables (like sequences) can be controlled;
2) that control of these variables requires that they be perceived and
3) that what seems like an output limit (the speed with which you can
move your fingers to type letters, for example) may not be what
is limiting the ability to control higher order perceptual variables
like producing typed sequences.
I have not even tried to model this situation yet; it would sure be
nice to have a grad student or three.
Best
Rick