From Greg Williams (920108 - 2)
Rick Marken (930108.0900)
My point is that you have not reached
that goal in predicting cursor position.
After reading this I got out my HyperCard conflict stack and
did a tracking run and model run (400 data points each) with a
low conflict and got a correlation between subject and model
mouse movements of .996831 -- the correlation with cursor
movements would be lower but at least we seem to have part of
a true science. I repeated this with a slightly higher level
of conflict and got a sublject/model correlation of .993398. Still
in the "true science" range. Higher conflicts will take us
well below .99 (to .98 maybe?) suggesting that there is something
to be learned there.
If "true science" is your aim, why not compute the subject/model correlation
for the INTEGRAL of the handle movement? That should get you within spittin'
distance of 1.0! But if you want to be sobered a bit, use your model to
compute the correlation between the modeled derivative of handle position and
the derivative of the subject's handle position. That might even be lower than
the cursor-prediction correlation.
How about H = K * integral(C - T)? (K is a constant to be adjusted for best
fit to the data by running the simulation with the model in it). That's
just a first cut, of course, since it doesn't predict the cursor position
well enough for "true science."
I think the idea, now, is to use your S-R model in a real tracking
situation. As I understand the challenge, you are to derive an S-R
model (like your equation above) from your observation of the
relationship between S (cursor) and H (handle movement). Bill
apparently sent you that data. I'd throw in the disturbance too -- I
don't think it's an unfair advantage for you at all -- in many experiments
you CAN see the disturbance (or the cause thereof) even if the subject can't.
So I would suggest that Bill give you D, C and H from a tracking task. Based
on that data, you come up with an S-R model that generates H based on what
the subject can see (C and T).
We're back to what is and is not an "S-R" model. If I can't fit parameters
with the model taking the place of the subject with the loop closed, I won't
be able to get reasonable values for K. If fitting parameters with the loop
closed makes the model above into a PCT model, rather than an S-R model, why
is that? As required, the model has the form H = f(C,T). Input-output. But if
I try to adjust K to make H follow the data, GIVEN THE FIXED C data, my "best"
K will not be best when the loop is closed. By adjusting K with C NOT fixed,
and the (given; it is just the difference between the other two givens)
disturbance operative, the K WILL be best for OTHER disturbances, too.
Your model must then be tested by seeing if it can do what the subject
does -- control the cursor in a new situation. So your model must be
"run" (this could be cone analytically but it's easier with a computer
simulation) with a new disturbance -- to see if it generates the H that
controls the cursor (as the subject would).
I would go beyond that, and try to change the FORM of the model to predict C
better, too. Maybe even get the subject-model cursor correlation up to the
"True Science" range... of course, that might not be possible, given noise in
the subject. My challenge to you and Bill would then be to make an underlying
generative model for the noise which results in a better s-m C correlation
than is possible with behaviorist "models" which contain no underlying
hypotheticals.
As ever,
Greg