[From Bill Powers (931117.1500 MST)]
Hans Blom (931117) --
As I've been saying to Martin Taylor, I'll agree that feedforward
can enhance control behavior. In the light of your post, I'll
even agree that engineers have been able to make good use of it
in designing artificial control systems. But please try to
understand what both Rick and I have been saying: there is no
point in OUR using feedforward in a model of real human behavior
until we come across an example of behavior in our
experimentation that leaves a large part of the behavior
unaccounted for, when only a simple control system is used as a
model.
Now that Martin Taylor is looking at some real experimental
results using our approach, I think he will agree that accounting
for the differences that remain between the model and the real
person is not as easy as adding a feedforward connection. There
*are* details of behavior that are not accounted for by the
model, but they are rather small, and they consist mainly of
high-frequency oscillations showing no obvious connection to
disturbances or actions (I assume that Martin is seeing the same
things -- my own performance has a few extra wobbles in it,
alas!).
You mention the fact that the performance changes when the
disturbances become more predictable. That fact interests me
greatly, and one day I hope to extend the model by another level
to try to account for that difference. Perhaps feedforward will
then prove to hold the solution. Or perhaps some other kind of
higher-level control process will be required: we won't know
until we are looking at the data. Remember, we are not designing
a control system, but trying to figure out the design of a system
that already exists. It does help to have in mind some
alternatives to the basic model, so I don't consider such
explorations wasted. But if we start trying to force a model onto
the behavior just because it has proven useful in engineering
designs, we don't be able to see the data in an unprejudiced way.
···
-------------------------------------------------------------
Your description of your proposed tracking experiment is clear
now. What is the advantage of scrolling the display to show past
positions of the target? Our current experiments, which are done
in graphics mode, by the way, simply show the target moving from
side to side, and the cursor that is operated by the handle
attempting to follow it. Nothing else is shown on the screen. The
only difference from your experiment is that in ours, the subject
sees no indication of past positions of either cursor or target.
Do you have some reason for showing them?
-------------------------------------------------------------
Obviously you have done experimental work with tracking. I
apologize for my nasty suspicious mind. You say
As soon as the target was more or less predictable, the
operator soon learned to make use of the regularities. This
resulted in ridiculous results like zero or even negative delay
times.
Why is that a ridiculous result? It's what you observed, isn't
it? Did you give up on the modeling at that point, or did you try
to find a model that would behave the same way?
By the way, while you were doing these experiments, did you ever
realize that it was a perception that the person was trying to
control?
-------------------------------------------------------------
I know that open-loop phase errors usually do not average to
zero. In long-duration cave experiments, the "diurnal clock"
starts to tick for 25-hour days on average, with extremes
ranging from 12 to 36 hours.
I was thinking more in terms of 12 to 36 seconds. Tom Bourbon has
already mentioned his experiments in which he did many
repetitions of experiments where he tracked a simple triangular
target, and then shut off the feedback to see how long he could
keep the cursor moving accurately. The errors became significant,
as I recall, long before the end of a one-minute run.
--------------------------------------------------------------
The basic problem is, however, how to measure errors. What is
your criterium? Do you just pick a "standard" test -- and if
so, which one?
The criterion is very simple: it's the distance between the
cursor's position and the position where it should be to produce
zero error. To evaluate a subject's control error over a one-
minute run in which the display is refreshed 60 times per second,
we take the RMS value of these differences over the 3600 data
points (in the latest implementation).
The same criterion is used in comparing the model's behavior with
that of the real subject. A real handle position is recorded 60
times per second during an experimental run. The model computes
and records handle positions at the same rate (it's given the
same table of disturbance values that the real subject
experiences). The RMS difference between model and real handle
positions over the 3600 data points is the measure of the model's
prediction error. We simply substitute the model for the person
in the same experiment, and compare the handle movements.
We express error as these RMS measures divided by the peak-to-
peak disturbance, which is essentially the same as the peak-to-
peak excursion of the handle measure. This normalizes the results
to a constant disturbance amplitude.
I'm curious -- is there another criterion that would be more
appropriate?
---------------------------------------------------------------
The fact is that the [feedforward] models are inadequate
for good control over any prolonged time.
That depends -- see my anaesthetics example above. If sensors
are not available or if sensors fail, feedforward is all you
can do, even over long time periods.
Most of human behavior involves control that is far more precise
than your 80-120 percent criterion. I don't buy the argument that
if poor control is all you can achieve, you have good control.
You don't. You have poor control.
I just got up out of my chair, walked around the computer table
to the stairs, climbed the stairs, turned into the kitchen, got a
cup of coffee, and retraced my steps downstairs again and am now
seated where I was within a couple of inches. I didn't spill a
drop of coffee on the rug, either. That is what I think of as
normal human control behavior. Even now, as I type, my fingers
are remaining within about 1% of their total positional range,
falling on the keys I want to hit perhaps 98 or 99 percent of the
time. That is normal human control, and many people are much
better at it than I am. Practically everything we do shows
extraordinatily accurate control, in the 1 or 2 percent error
range. That is the sort of control behavior I'm trying to model.
Perhaps I am misreading this sentence, but to me this implies
that attention is required to make a feedback signal operative
in
a control system, and I disagree with that.
If, in your tracking experiments, your subject is not paying
attention to the position of the target but to the fly that
crawls across the screen, his tracking performance will
deteriorate or tracking will cease. More was not meant.
True, and this is a subject that cries out for experimental
investigation. But what about all the other control systems that
are working at the same time, of which the subject is unware or
only dimly aware? There is, after all, an arm holding a handle,
and it is operated by a position control system too. The feedback
signals reaching the spinal, brainstem, and midbrain control
systems are certainly present and participating in these control
loops, even though the participant is attending primarily to the
visual controlled variable. And the participant is maintaining an
erect posture in the chair, as well as tracking the target with
the eyes. Most of the feedback signals involved in this control
process are unawared, at least until attention is called to them.
------------------------------------------------------------
When *I* look at the behavior of an organism, I DO look around
for events in its environment and in its history, to explain
what I see now that I did not see before....
I was just trying to emphasize that I am concerned mainly with
how the system works right now, not how it became able to work
that way. I wasn't trying to tell you what to be interested in.
--------------------------------------------------------------
... for me the central question is: what is behavior and how
do skills improve over time? Isn't that a question worth
answering?
Yes, it's worth answering once you can say exactly what a "skill"
is in terms of a model that explains it. You can't even say what
a skill is, in any lasting terms, until you can say what the
person was intending to do. Is a clown stumbling and falling
across the arena "unskillful?"
-------------------------------------------------------------
Best to all,
Bill P.