[From Bill Powers (931120.0850 MST)]
Ed Ford (931129.1410)
Interesting example of controlling perceptions: to the dog
following the rabbit trail, smell is the perception that matters;
to the human being who saw the trail only after a snow, vision is
what counts.
···
---------------------------------------------------------------
Rick Marken (931129.1500) --
I took Martin's post as a description of the world as it looks
through the eyes of a dynamicist (including physicists and
biologists and others who are trying to form a Big World Picture
using the methods of physical sciences). You're right, of course,
about the Test, but from the dynamicist's point of view the
quantitative difference in stability between a controlled
variable and a marble in a bowl doesn't suggest anything
qualitatively different between active and passive stabilization.
Martin's diagram yesterday showed the qualitative difference,
where he indicated the power source and sink associated with the
output function. Active stabilization works by drawing on energy
from a large supply, and expending it to create stability that is
greater than what any passive (equilibrium) system can achieve.
This is a process incompletely appreciated in most Big Picture
approaches, and the one at the heart of PCT.
---------------------------------------------------------------
Bob Clark (931129.1747) --
RE: keys-and-lights learning experiment
And, as you note, we found subjects occasionally demonstrating
a "negative" reaction time. This can be explained in three
different ways: 1) guessing; 2) "reorganization of a learned p
erformance;" 3) "anticipation" from a remembered sequence. I
seem to recall that subjects, when questioned, were often able
to recite the sequence.
Remember the technician Roy X? He used to be very good at
following instructions, even if they contained fatal errors. He
gave up before finding the "anticipation" solution. And Dr.
Branfonbrenner, who was extremely bright, stated the solution as
soon as he heard the instructions, yet exhibited all the plateaus
when we insisted he actually perform the experiment.
The only way to beat the machine was to press the correct key
_before_ the next light came on, which required having learned
the sequence of lights. So the critical perceptual variable to be
controlled was a temporal relationship -- at least that's how I
interpreted the results.
It seems to me that the conditions of the experiment would have
been unlikely to produce an Intrinsic Error.
I seem to recall some pretty earnest sweating over this task by
many subjects, and even Branfonbrenner had a good laugh over it
when he finally learned the task. Dick Robertson, who tested many
subjects with a computer-based version of the experiment, might
have some additional comments about evidence of intrinsic error
among his subjects.
My recollection of those results includes finding a bi-modal
response for level 2. This did not seem to make much sense
(theoretically), so we re-examined the set-up and found it did
not adequately distinguish among the signals presented. We
found that it was in some ways too difficult, sometimes
yielding a level 3 response, and it was also possible for the
subject to respond at level 1.
Our problem was in finding tasks that a subject couldn't perform
by using a lower level of information than what we thought we
were presenting. The bimodal distribution occurred when we tried
to measure response time to a sequence. A dot on an oscilloscope
would move right-left-right or right-left-left, with the
difference in sequence being indicated by the last jump so we
could measure a clear reaction time. But some subjects simply
waited until the third jump and responded to "light" or "no
light" at a particular position on the screen, and thus showed a
lower-level reaction time without paying attention to the
sequence. The distribution was bimodal over the population, but
not within an individual. I forget how we solved this, but I
remember that we did, and that we got the same peaks but no
ambiguities.
I would not use "Feed-Forward" in this [keys and lights?]
situation. Rather: "Using imagination to Predict (Anticipate)
future events by means of remembered (recordings of) similar
previous events."
I agree. Feedforward is an especially bad term in this context
because it implies that somehow the system can look forward in
time, which it can't. Your description recognizes that however
the phenomenon works, it uses present-time information only.
A prediction is always a present-time calculation.
For me, memory/imagination provides relatively simple and
familiar descriptions for some situations.
For me, too.
Have you ever seen the book on memory by Yates, mentioned in BCP
on p. 215-216? Was the method you learned similar to the "method
of loci" discussed in BCP?
---------------------------------------------------------------
Avery Andrews (931130.1249) --
"Model-assisted control" it is. Brilliant, and exactly right.
(931130.1342) --
Suppose you're on an exposed position on a small boat,
controlling for staying on the boat, and you see a large wave
coming.
I would say that you perform a present-time extrapolation of the
position of the wave to produce a present-time imagined encounter
with it, and you take action if that imagined encounter differs
from what you would like it to be. As you act, you continuously
or repeatedly re-calculate the extrapolation, changing your
actions as the extrapolated situation changes, until you are
experiencing in imagination a result that suits your goals. As
the wave nears, your extrapolations become more and more accurate
and your corrective actions more and more appropriate, so when
the wave actually arrives you have done everything required to
resist its effects -- except for resisting effects you failed to
imagine, like a piece of driftwood you didn't see surfing into
the side of your boat.
This is the same principle that pilots use in landing airplanes,
and that is used in automated systems for hands-off landing. I
believe that the customary term is "predictive control." The
perception you are controlling is a _present-time_ perception
derived by an iterative prediction from other _present-time_
perceptions. We never experience any actual perceptions of the
future.
I think this is a much more informative way to describe what
happens than using the metaphorical term "feedforward."
Predictive control is a form of model-assisted control, but not
quite the kind Hans Blom has been talking about. In Hans'
approach, the model runs concurrently with the real environment.
In iterative predictive control, the model runs faster than real
time, over and over, so the outcome predicted by the model is
compared with the reference level again and again. The actions
based on the error alter the perceived present-time environment,
which alters the outcome the next time the internal model is run,
which alters the error, which alters the action, and so on. This
kind of control is needed when one must begin acting
substantially in advance of the desired outcome at some fixed
time in the future.
Because predictive control of this kind is iterative, the method
of extrapolation is not critical; even if the prediction is in
error for long times into the future, as the critical event
nears, the extrapolations become shorter and shorter, so errors
of prediction make less and less difference. After the final
prediction, control merges into normal present-time control.
Modeling this kind of control for human behavior would require a
lot of stipulations about perceptual functions that we don't know
how to model.
---------------------------------------------------------------
Best to all,
Bill P.