[From Bill Powers (930129.1300)]
Rick Marken, Avery Andrews (930129) --
Response latency is measured in a lot of ways, as Rick says. Most
of them leave you with little idea of what the actual transport
lag through the nervous system is. I saw something recently in
which subjects indicated a response by saying "HA!" into a
microphone. Apparently the experimenter thought this was such a
simple thing to say that saying it required no time. I suspect
that any linguist would be laughing by now, thinking of what a
sonogram of "Ha!" looks like. Just imagine the diaphragm
beginning to tense, the pressure building up and starting to leak
through the throat, the hiss, and finally the "Ah" sound building
up to a measurable level. By the time that sound starts, the
moment that the nervous-system output actually began is fading
into the distant past.
Or consider indicating a response by pressing down on a key, or
releasing a key. Depending on the relative sensitivity of the
shoulder-muscle systems, the biceps and triceps, and the forearm
muscles that operate the fingers (not to mention the strength of
the spring under the key), the first tendency of the finger on
the key to move might be either in the right or the wrong
direction. If the biceps/triceps respond the most, initially, the
dynamics of the arm will make the finger press down harder
instead of releasing, or rise further off the key instead of
depressing. By adjusting parameters you can make the Little Man
model do these things quite clearly, to varying degrees. So an
unknown part of the "response latency" consists of the dynamics
of arm segment movements under angular acceleration.
And of course as Rick said, where do you place the threshold for
detecting presence of absence of a response? At the 10% point?
50%? 90%? Just above the noise level? The inflection point in the
movement? The point of maximum velocity?
When Bob Clark and I measured reaction times for mechanical
disturbances of human arms, we used an electromyograph so we
could pick up the moment when nerve impulses reached the muscles;
we ignored the actual mechanical movements occurring after that.
To my mind that is the only meaningful way to talk about a
reaction time.
Heck, I've seen experiments with rats in mazes in which "response
latency" was defined as the time it took the rat to get from the
photocell at the entrance to the maze to the photocell at the
entrance to the box where the cheese was.
Avery has it right: "Feedback is too slow" is a slogan, repeated
because someone else said it in an authoritative manner.
Rick's question is highly germane: too slow for what? If feedback
is too slow for control within 40 milliseconds, it's too slow,
even if it's present. This is true in a model and also in the
real system. The feedback pathways aren't surgically removed just
before a 40-millisecond disturbance occurs. They just don't do
much good. Look at the patellar reflex (knee-jerk). The rubber
hammer puts in an impulse stretch of the tendon. The control
system tries to correct the error, but it's much too late; the
disturance is gone. So the leg kicks upward, trying to oppose a
disturbance that's not there any more. Feedback was too slow. So
what? It was still a control system.
In the discussions of fast and slow movements, there seems to be
an idea floating around that the nature of the physical system
changes depending on the speed of the movement. This can't be
true. Either you have a system hooked up as a control system, or
you don't. If the control system isn't present for fast
movements, it isn't present for slow movements, either. If a
feedback connection is present for slow movements, it's still
there during fast movements even if it's responding too slowly to
do any good.
It doesn't matter whether you practice a movement a lot or are
doing it for the first time. The parameters of control may slowly
improve with practice, but you can never get to the point where
you can dispense with the feedback. In fact, the better you get,
the more important the feedback becomes in assuring an accurate
action despite the speed. No matter how good the control gets, on
the other hand, you can always present the system with a step-
change in reference signal, or a disturbance in the form of a
brief enough impulse, so the system is asked to perform beyond
its capacities. In that case there's no control during the
transition. But nothing about the control system has physically
or functionally changed. If you took away the feedback, the
movement wouldn't end up in the right place.
I hate to say it, but such statements about fast and slow or
practice and unpracticed movements simply show the lack of a
working model in the background.
···
--------------------------------------------------------------
Martin Taylor (930129.1230) --
Words don't have meanings so much as tendencies which coalesce
in the course of a dialogue (or monologue) to have some effect
on the listener (reader).
I'm afraid that that way of putting it gives too much comfort to
those who want to endow words with special properties. Words
don't have EITHER meanings OR tendencies. Saying it that way
ignores the active agent that is using the words for a purpose.
The receiving nervous system may treat words in certain ways,
interpret their conjunctions in certain ways, connect the
meanings they evoke in typical (for that nervous system) ways, or
employ them as pointers to experiences according to certain
strategies. The words have no ability to alter those actions by
the receiving nervous system. The way you're putting it is like
saying that knives and forks have a tendency to be on place-mats,
or that wet socks have a tendency to be taken off, or that
cigarettes have a tendency to be smoked. The sense of agency is
being put into the wrong thing. It's the person's brain that
gives words all the effects they can have, and those effects
depend entirely on what the brain does with the word-perceptions
once they're present. To give the words any causal properties is
to revert to a form of animism.
I agree that one of the biggest difficulties of language theory
is the hangup that the symbols that words (and letters, and
phrases) are have some definable, bounded, and mathematically
capturable relationship to one another.
Yes.
They do have tendencies that way ...
I disagree. Words have no tendencies at all. They are perceived
or they are not perceived. That is all they can do. The rest is
up to the receiving system.
---------------------------------------------------------------
Oded Maler (930129) --
The answer to the question "what is feed-back slow for" you
must invoke some "objective" performance criterion independent
of the internal perceptual coordinates of the acting
individual.
Feedback is too slow to permit a person to catch a passing
bullet. So what? Arbitrary objective criteria don't have anything
to do with the interior design of the behaving system, which
behaves to achieve its own aims, not those of an external
observer. The control systems in the human body are exactly fast
enough to permit the kind of behavior that human beings are able
to accomplish. They are not fast enough to accomplish ends that
human beings can't accomplish, even though an artificial control
system with faster feedback might be able to do them.
You must assume that "playing a piano trill correctly at some
speed" or "knocking out a boxing champion" has some more or
less agreed-upon meaning.
This won't help to answer the question of whether feedback is too
slow to explain any particular instance of behavior. I certainly
can't play a piano trill "correctly" in terms of objective
criteria that might apply to my friend Sam Randlett, who teaches
concert pianists. Yet my feedback is exactly fast enough to
account for the speed with which I CAN play a trill (Sam would
fall asleep waiting for the next note). I can't knock out a
boxing champion, or a drunk paraplegic, yet my feedback control
systems are just exactly fast enough to make my fist move as fast
as it actually moves.
Then you can build a mathemtical model of that act and the
component involved (muscles, nerves and their reaction time)
and show WITHOUT USING THE CONTROL MODEL that it is impossible
for information to travel and affect the muscle at the time
scale between the intitiation of the action and its outcome.
Again this misses the point. With the Little Man arm model, I can
find the maximum speed of movement at which it is impossible for
feedback signals to make any change in the movement between its
initiation and its outcome. That speed is one in which a movement
occurs in response to a step-change in the reference signal, and
its time constant is about 40 milliseconds, just like Mohammed
Ali. The initiation of the act consists of an instantaneous
change in the reference signal. The control system begins with
maximum possible error, which then reduces asymptotically to zero
with a 40-millisecond time constant. This is simply the fastest
speed at which the error can be reduced.
The determination of the delays is quite independent of the
model; it's due to neural transport lags and to the leaky-
integrator form of the muscle response. The feedback model
incorporates these factors, and when the parameters are adjusted
for the nicest and fastest error correction possible under those
conditions, you get a 40-millisecond time constant of error
correction. The feedback model is acting as fast as physically
possible.
If you didn't have the feedback present, the model wouldn't be
able to execute the same movement anywhere nearly as fast. The
feedback model starts, immediately after the step-change in
reference signal, with an error signal that would move the arm
something like 10 times as far as the actual distance to the
reference or intended end-point. This produces a tremendous
acceleration, on the order of 20 gravities in linear terms. But
as soon as the arm begins to move, the error signal begins to
drop (about 9 milliseconds later, actually). Now the effective
target location is not so far beyond the intended end-point. This
process continues, the effective end-point coming back inward
toward the intended end point while the arm moves outward toward
the intended end-point. Before the arm reaches the endpoint, the
rate feedback actually moves the effective target point to the
negative side of the intended end-point, decelerating the arm.
All this happens automatically with no particular computational
difficulties, and the arm comes to rest at exactly the right
position, 100 or 150 milliseconds after the initiation.
If you wanted an open=loop system to bring an arm to an endpoint
after a step-change in the initiating signal, you would have to
use a much smaller signal, one tenth as large, to avoid
overshooting the intended end-point. As a result, the initial
acceleration would be much less, and the movement would take far
longer.
Actually, the problems would be much worse than that. You
couldn't actually use a step-change in the driving signal,
because without the feedback you'd have a mass on a spring with
very inadequate damping. You would need a full blown motor
program that would apply a complex waveform to the muscle, to
prevent oscillations. Of course you could then use a larger
driving signal, and in fact supply the same driving signal that
would be observed during the operation of the feedback system.
You could then achieve equal speed -- but at what cost!
By showing that an ideal pianist or boxer with ideal "objective
sensors" and "objective effectors" cannot achieve something
because of timing constraints you show a-forteriori (?) the a
realistic (i.e. PCT-based) pianist/boxer with the same timing
constraints cannot do it either.
This is more like my point. In fact, real boxer/pianists come
very close to achieving the theoretical limits of performance.
Models, of course, can always work better than real people,
because we can give them more favorable properties. But if we
match masses, delays, and time-constants, we can then adjust the
model to reproduce the human behavior reasonably well. The Little
Man, given very strong muscles, can in fact move the fingertip
from one point to another with a 40-millisecond time constant,
about the same as Ali or Joe Louis. But not a lot faster, given
the delays and time-constants. Trying to push for still faster
movements takes the system to the edge of instability. I presume
that's why the real system doesn't go any faster, even for the
most practiced practitioners.
The emphasis of the important PCT insight that within the
individual "it's all perception" in contrast to the naive
objectivism of, say, cognitive psychology, should not be
exaggated into a solipsist neglect of the external environment.
I think I'm safe from that accusation. I probably spend 90
percent of the time required to produce a model like the Little
Man in constructing a realistic physical model of the environment
with which the system interacts. The actual control-system model
is trivially simple in comparison. Even in our simple tracking
models, we include more in the environment than most other
modelers do -- for example, disturbances that directly affect the
outcome, and often nonlinearities and changes in parameters.
The question of how and under what conditions people can
achieve certain "objective" performance, in other words, what
guarantees that a system organized in a certain way survives
("objectively") in a given environment, deserves more attention
and better answers than "otherwise, reorganization will
continue".
It certainly does, and we have never used reorganization to make
up for lack of specificity in a model. We match models to
performance by adjusting parameters, thus answering the question
as to what guarantees success (equal to that of the human) in the
task. "Survival" isn't so much of a concern; so far all of our
experimental subjects have survived. Most of our talk about
reorganization is by way of speculating about how we would go
about including it in a model, when we finally get around to
doing that.
Of course when we're just doing shirtsleeve conjecturing or Big
Picture fantasizing, anything goes. It's all reorganization.
Heck, maybe it's all chocolate syrup.
... but still some neurophysiological evidence about basic
properties of nerves and muscles can replace this data and
prove the uselessness of feedback for certain kinds of actions.
I suspect that it would also prove the uselessness of supposing
that such actions can actually occur. Given such data, which I
certainly used a good deal of in designing the Little Man, we can
find out how good a control system can be built around those
properties, and that is in fact what I did. I learned enough to
know that I am not going to waste my time trying to devise a
central pattern generator that could reproduce the same
performance within even a factor of ten worse. Anyone else is
welcome to try. I wish someone would. Then we could drop this
whole stupid subject of open-loop behavior.
------------------------------------------------------------
Avery Andrews (930130.0420) --
My perception is that the next issues to look at are (a)
pattern-generators (b) the idea that the output of the
effectors is what is controlled (as in the Jack Adams quote,
and Abbs & Winstein's `technical' definiton of feedback. This
is sort of like what you call the `objectification blunder' in
QAPR, but I seem to want to call it the `output blunder'.
Yes. I really appreciate all the time, research, and thought
you've given to PCT in your off-season, so while I agree that
these are important subjects I don't mean to imply that you're
expected to go on doing all the work. Linguistics calls, I know.
Pattern generators aren't a big problem, actually. It's true that
a pattern generator can always be devised to duplicate the
performance of a control system in a specific situation, at least
over a brief period of time. But when you consider the data that
must be available to the generator, and the accuracy with which
calculations must be carried out, and the fact that physiological
machinery like nerves and muscles change with use, and that there
are independent and unpredictable disturbances that can act
directly on the outcome and that can't be sensed or anticipated,
and that behaviors can last for hours and days, each move
beginning where the last one left off -- the whole idea begins to
look highly impractical. The simplest behavior requires a
supercomputer to carry it out.
The pattern generator can be part of a control system, as you've
noted, but that's not the usual idea. The normal open-loop
pattern generator is actually ruled out because the output
blunder must be ruled out. Pattern generators rely on output
devices that behave uniformly and with infinite precision, and on
total absence of disturbances that enter the output chain after
the effectors. The only way that I know of to deal with effectors
of variable properties and independent disturbances downstream
from the effector is through feedback control. There's just
nothing else that will work. So the output blunder is really the
most important one.
This blunder isn't quite so easy to find in the literature, not
because it isn't there but because it's like a black hole in the
middle of an explanation: invisible unless you realize that
something ought to be there. Consider the idea of an "orienting
response." This is a response that orients the body or head
toward a stimulus like a sound. But what kind of response could
make the angle of the head end up in exactly the right direction
regardless of the orientation of the body or the location of the
stimulus? On one occasion the head might have to turn 2 degrees
on the neck; on another, 90 degrees; on still another, 40 degrees
the other way. How can a response be so specifically what is
objectively required, despite large differences in initial
conditions? The whole problem is wrapped up and concealed in that
little word, "toward."
There are lots of cases in which responses are simply named after
the effect that the organism's actions produce, thus skipping the
entire question of how the nervous system can cause the same
remote consequence to occur over and over under varying
conditions. Just by calling all outcomes of actions "responses"
you can leap right over the black hole without giving it a
glance. The shoelace-tying response. The problem-solving
response. The balancing response. The verbal response. The
tracking-the-target response. The putting-it-on-the-Visa-card
response. The car-steering response. You can take it for granted
that any time the word "response" is used, it's really referring
to an outcome, not the action that happens to be required this
time to create that outcome.
The output blunder is hard to detect because hardly anyone even
realizes that there's a problem here. It's not discussed one way
or the other. What's wrong with saying that you respond to a
question by stating the answer?
---------------------------------------------------------------
Bob Clark (930129) --
Let me whisper so nobody else will hear: you wouldn't have a
spare reprint of that "integrated data collecting" paper, would
you? I don't have it.
--------------------------------------------------------------