Control loop delay, Perception study

[From Rick Marken (940203.1000)]

Bill Leach (02 Feb 1994 22:59:13) --

I wonder if you guys have played around much with "control loop delay"
considerations?

There are two types of control loop delay. I think you are asking about
what we call "transpost lag" which is the time it takes for the effect of
a transient disturbance to move from one the input to the output of a
control system). The other delay we call the "slowing factor", which is
the time it takes for the effect of a transient disturbance to build to
it's full effect at one point in the system). We always incorporate a
slowing factor into our models; if we don't the models don't work. I rarely
put a transport lag into my models because it turns out that, in the
continuous control tasks I've studied, the transportlag parameter hardly
makes any difference in the success of the model (which typically picks up
over 99% of the variance in the variables of interest).

There was one interesting case where transport lag did make a BIG
difference. I did a compensatory tracking experiment where the disturbance
to the cursor was actually the output of a computer simulation of a
low gain control system that was controlling the same cursor. So there
was a low level conflict between the subject and the computer simulated
control system (CSCS). The gain of the CSCS was vary low so the subject
had no trouble controlling the cursor; the outputs of the CSCS were just
a disturbance added to the cursor. In order to show that this was the case
(that the outputs of the CSCS were just like an "inanimate" distrubance,
like the wind) I stored the outputs of the CSCS obtained during a tracking
run and used them as the disturbance during a second tracking run. The
subject's tracking perfmance was always MUCH WORSE when working against the
"replayed" distrubance even though it was EXACTLY THE SAME TIME WAVEFORM
that had been present during the previous tracking trial. The only
difference was that, in the first case, the waveform was being generated
in real time as the outputs of a CSCS; in the second case it was being
played back from a stored table.

This phenomenon (dubbed "the Marken effect" -- my kids were so proud)
was a real problem for PCT because, when you substituted a PCT model
for the human, you found no difference in tracking performance with
the "live" versus stored disturbance. Bill Powers solved the puzzle
because he somehow realized that the reason for the difference was
a "transpost lag" in the subject. When you put such a transport lag in
the PCT model (I think the equivalent of a 160msec lag gave the best
results) the model behaved just like the subject -- ie. the model
produced much better performance with the live (CSCS generated) disturbance
than it did with the identical "stored table" distrubance. The low gain
"live" opposition from the CSCS apparently acts like a "spring" that damps
out what are ordinarily VERY small instabilities in the control loop that
result from the existence of the transfort lag. Bill P. can probably explain
this; obviously I can't.

Bill P. even thought of a practical application for this idea -- which I
plan to sell to Microsoft for $10,000,000 (a pittance but a reasonable
start for the Living Systems Institute). The CSCS "spring" is just a teensey
piece of software that can be part of every mouse-based operating system,
allowing improved mouse operation with no hardware changes. Oops. Now the
secret is out. Maybe I'll ask for $100 instead; no Living Systems
Institute but a nice steak & lobster dinner for two at the marina.

Martin Taylor (940202 16:45) --

A colleague is interested in the perception of objects tilted out of the
frontal plane. What aspect of the object determines its perceived shape
(and tilt)?

I'd be inclined to ask "what perceptual variable is controlled when a person
controls the shape and tilt of an object?" But your project sounds VERY
interesting.

In the discussion, we designed an experiment using pursuit tracking.

The pursuit tracking part is OK. What I think you need in order to test
hypotheses about perceptual variables is a distrubance to your "planar
2-D object" that influences the different hypothesized perceptual
variables differently or that must be compensated for differently
depending on what the person is actually controlling. This is what I
did in the "size" control experiment. It was basically a compensatory
tracking task where the subject's output determine the x dimension of
the quadrilateral and the computer disturbance determined the y dimension.
If S were controlling x*y (area) then the subject's output, x, would have to
be proportional to 1/y. If the subject was controlling x+y (perimeter),
output, x, would have to be proportional to -y. So you can tell which
perceptual variable is controlled by looking at the relationship between
distrubance (y) and output (x).

Let's say tilt could be perceived on the basis of x/y or some funciton
of the time changes in x and y (dx/dy). To test this, all you have to
do is allow the subject to influence one aspect of the presumed variable
(say x) and see if this influence has the expected relationship to y. Or,
you could monitor the presumed tilt variable itself and see how well it is
controlled (using the stability factor, perhaps).

It would be very helpful (and interesting) if you could post the
alternative mathematical models of the perceptual function that produces
the "tilt" perception.

Best

Rick

<Bill Leach 04 Feb 1994 19:05:58

Rick Marken (940203.1000)

More control loop delay.

Well, I'm still musing but it seems to me that there are absolute
maximum response speeds for human activity. For things involving
movement, the responses are rather slow as compared to most computing
hardware.

Visual recognition OTOH probably has a fixed maximum response speed but
when compared to computer visual recognition... well it's embarrassing so
lets go back to "mechanical".

I am having a little trouble with your example. Are you saying that when
presenting a person with exactly the same sequences of changes and asking
them to follow that identical set of changes (pursuit?), that if you
introduce a real-time derived disturbance you get one set of results but
if instead, you record the real-time disturbance and "play it back" while
repeating the experiment, you get a different result?

Don't worry... what ever it is that you are/have/will develop, MicroSloth
is in the process of releasing it RSN (and their lawyers will contact you
about YOUR infringement) :slight_smile:

-bill