feedforward test

[HANS_BLOM1]

[Hans Blom, 931111]

(Rick Marken (931110.0900))

This leads me to a proposal for a "test for feedforward behavior".
The first part of the test is to identify some result that is
produced repeatedly and consistently when there are no disturbances.
The second part of the test is to apply disturbances. There should
be NO RESISTENCE to these disturbances. If there is resistence then
the result might be controlled and the test can proceed to identify
the control system involved ( the sensor and output components).
So "the test for feedforward behavior" assumes that a feedforward
behavior is a REPEATABLE, CONSISTENT result that is NOT under control.
I think it must also be established that this feedforward behavior
is directly the result of neural outputs -- and NOT a side effect of
control of other variables.

It seems to me that the ball is now in the court of the advocates
of feedforward behavior; show me an example of a feedforward behavior
(as established by "the test for feedforward behavior") that can only
be explained as a computed output of the human nervous system. That is,
show me an example of a consistently produced behavioral result (such
as reaching for the bed sheets in the dark) that is NOT controlled
(can be disturbed) and THAT CAN BE TRACED DIRECTLY TO THE OUTPUTS
OF THE NERVOUS SYSTEM (ie. is not a side effect of control of
other perceptual variables).

Fair enough. Would the following cursor tracking test do?

Present your subject with a predictable cursor to track, say a sine wave
or a triangle. Let the subject track for one minute, say, at least until
the quality of tracking is adequate. Then, slowly, start to withhold feed-
back information, i.e. do not show parts of the curve in order to force
your subject into feedforward mode. The subject will see something like a
dotted line at first. Now, slowly by slowly, decrease the dot density --
space the dots ever farther apart. To make the task a little easier, you
might want to display short line segments rather than dots. Continue to
decrease the information displayed until whole periods or more are in-
visible. See how well the subject performs. Repeat this with a number of
different cursor movements, say sines, triangles and block waves of dif-
ferent amplitudes and different periods.

That is the easy part. I'm sure you must have done things like this in the
past. Now comes the difficult part. Design a controller (feedback, feed-
forward or combination of both), CSG-style or not, that performs equally
well when whole periods or more of the signal are missing. And not only
equally well on ONE sine wave signal with a fixed period and a fixed
amplitude, but on ANY periodic signal that a human subject can easily
handle after his one minute's training.
Inputs to the controller are two perceptions: a binary signal that indi-
cates whether the cursor is visible or not, and a number that specifies
the cursor position when it is visible and is zero (or random) elsewhere.

When you evaluate the quality of your subject's "tracking" performance
when the cursor is invisible, a simple correlation might not provide the
best measure. The reason is that our internal clocks are noisy and show
some drift. Amplitude noise is, I think, evaluated fairly by a correlation
function, phase noise is not. By some "time warping" of the time scale a
much better fit is possible. For the moment, however, a visual comparison
of the "tracking"/"prediction" performances of subject and control system
will do nicely.

Greetings,

Hans

[Tom_BOURBON3]

From Tom Bourbon [931111.0901]

[Hans Blom, 931111]

(Rick Marken (931110.0900))

This leads me to a proposal for a "test for feedforward behavior".

Me:
Which I will not repeat here.

Hans:

Fair enough. Would the following cursor tracking test do?

Present your subject with a predictable cursor to track, say a sine wave
or a triangle. Let the subject track for one minute, say, at least until
the quality of tracking is adequate. Then, slowly, start to withhold feed-
back information, i.e. do not show parts of the curve in order to force
your subject into feedforward mode. The subject will see something like a
dotted line at first. Now, slowly by slowly, decrease the dot density --
space the dots ever farther apart. To make the task a little easier, you
might want to display short line segments rather than dots. Continue to
decrease the information displayed until whole periods or more are in-
visible. See how well the subject performs. Repeat this with a number of
different cursor movements, say sines, triangles and block waves of dif-
ferent amplitudes and different periods.

That is the easy part. I'm sure you must have done things like this in the
past. Now comes the difficult part. Design a controller (feedback, feed-
forward or combination of both), CSG-style or not, that performs equally
well when whole periods or more of the signal are missing. And not only
equally well on ONE sine wave signal with a fixed period and a fixed
amplitude, but on ANY periodic signal that a human subject can easily
handle after his one minute's training.
Inputs to the controller are two perceptions: a binary signal that indi-
cates whether the cursor is visible or not, and a number that specifies
the cursor position when it is visible and is zero (or random) elsewhere.

When you evaluate the quality of your subject's "tracking" performance
when the cursor is invisible, a simple correlation might not provide the
best measure. The reason is that our internal clocks are noisy and show
some drift. Amplitude noise is, I think, evaluated fairly by a correlation
function, phase noise is not. By some "time warping" of the time scale a
much better fit is possible. For the moment, however, a visual comparison
of the "tracking"/"prediction" performances of subject and control system
will do nicely.

Me:
Hans, I cannot tell from those paragraphs whether you have read the several
posts during the past few days in which Bill and Mary Powers and I have
discussed the facts that people have other senses than vision, that when
people control visual perceptions they also (unavoidably, necessarily)
control peceptions from those other senses, and that higher-level
perceptions arise form perceptual functions that sum, integrate, or
otherwise pool perceptual signals from many senses.

I have done some modeling that addressess at least some of the points you
raise. I do not offer this description as a complete answer to your
questions.

An ECS with a reference signal that is a time-indexed array of values can
control the momentarily sensed position of its simulated hand, relative to
the present value of the reference. When it does, the ECS is a controller
of its sensed actions. It no longer controls the position of the
cursor, which position is now an uncontrolled side effect when the ECS
controls its hand position. The ECS no longer cancels the effects of
disturbances to the position of the cursor. A person does the same thing.

Prior to the elimination of the visual display, the person controlled the
position of the cursor relative to that of the target. A simple PCT model
can do that, as well.

The next part of your "test" is a little more complex.

The challenge I see concerns how to model a person's transition from control
of the presently seen position of the cursor relative to the target, to
control of the presently felt position of the hand relative to the present
value of a remembered or imagined pattern of positions. One solution I can
imagine is to rely on a single PCT loop, in which case the challenge is to
construct a sufficiently complex input function that includes both visual
and "somatic" inputs and creates a perceptual signal that is an analogue of
the states of many external variables all at the same time. Another
solution, which seems as though it would be easier -- at least I can begin
to imagine how I might do it -- would use at least two independent
lower-level ECSs, one for the visual display and one for all of the
"somatic" perceptions. Both loops, of course, would use the same
lowest-level output function to act on the simulated handle. When the
visual display is present, the system tracks in standard PCT fashion,
*using* but not *controlling* the position of the simulated handle. For the
regular target functions you described, the model would store the average
created by several periods of sampling the target and use those as the
reference signal for the somatic loop when the visual display is missing.

I would try this approach before any open-loop model.

Until later,

Tom