[From Bruce Abbott (950314.1230 EST)]
Bill Powers (950313.1540 MST) --
Not to gang up on you, but ...
An open-loop perceptual "translation" from color to selected target
would be very efficient of resources and speedier.This is one of the myths I have been battling for years: that open-loop
reactions are more efficient, simpler, and faster than closed-loop
control. This is simply not true. The closed-loop system is always
simpler and faster, as well as being more reliable (less sensitive to
changes in system properties or external disturbances).
Bill, your argument on the advantages of closed-loop versus open-loop is a
strong one, but I'm not sure it necessarily applies specifically to the
problem I was trying to address, which involves establishing a relationship
between PERCEPTIONS, i.e., if green then left target else right target. It
seems to me that this is an associative process (i.e., involving long-term
memory) in which a particular perception (e.g., green cursor) triggers
another perception (memory of the associated target shape, position, etc.).
I was suggesting that the formation of this association may require a
supervisory perceptual control system that checks the retrieved perception
for accuracy (did I remember the correct target?). The activity of this
high-level supervisory control system may be analogous to what we call
"attention." My initial thought was that this checking may become
unnecessary once the associative connection has become reliable and may even
interfere with the smooth functioning of the lower-level system that employs
the associative mechanism in its perceptual input function.
When we model the participant's performance on the SD task, we simply assume
that the perceptual input function exists without asking how it came to be;
in such an analysis the learning mechanism I propose becomes almost
irrelevant. But then we are left with no answer to the question of why the
switching lag keeps getting smaller with practice, other than to state
vaguely that reorganization is going on.
Now I'm perfectly willing to entertain the notion that what we call
"association" is itself the product of the operation of a control system,
one who'se smooth operation gives the appearance of S-R (or, more accurately
in this case, S-S). We really know very little about what is going on when
someone asks us "What's 3 times 4?" and "12" pops into consciousness.
Viewed from the outside, it LOOKS like the input "triggers" the
recollection. Current neural network modeling may hint at a solution; at
least it illustrates one way the process COULD work, although serious
questions can be raised about whether today's models have anything to say
about how the brain actually accomplishes this prodigous feat. In these
systems (as I understand them) the output is compared with a "training"
pattern and adjustments are made to system parameters (e.g., via "back
propogation") until the the "correct" input-output relationships are
established (if possible). Isn't this effectively a control system, with
the output being the system parameters and these being altered by the error
between perceptions of actual versus desired pattern?
As should be apparent by now, I'm not really trying to "go ballistic" on the
issue of associative mechanisms, but merely thinking about possibilities to
discuss and evaluate.
MORE SDTEST3 DATA
Bill Powers sent me a modification to SDTEST3 that "flashes" a tiny
rectangle, located immediately to the right of the cursor, whenever the
cursor color changes. The flash is highly discriminable, easily noticed
even when one is concentrating on minimizing the target-cursor distance. I
tried it out yesterday evening and again this morning (with several
variations as I will explain presently). Last night's runs produced the
following results (optimal lags shown):
Run lag k rms r
010 16 0.0635 18.89 0.987
011 26 0.0947 19.76 0.988
Individual lags:
Run Mean Median
010 38 23 18 27 14 14 23 17 14 20.6 18.0
011 42 28 30 22 26 27 31 24 25 21 27 24 27.4 26.5
(Bill: Note corrected run numbers.) Bill had reported that the addition of
the flash cut his delay in half; I obtained the same result in run 010 but
the delay in run 011, while under my non-flash typical performance of 30,
was not cut nearly as much. The effect may have been due to fatigue; run
011 was conducted about 2 hours after run 010.
This morning I tried it again, with the following results:
Run lag k rms r Comments
012 16 0.1025 18.88 0.989 flash mouse sens = 0.33 var isi
013 17 0.1299 18.82 0.987 flash mouse sens = 1.00 var isi
014 15 0.1729 20.81 0.983 no flash mouse sens = 1.00 var isi
015 13 0.1318 17.69 0.987 no flash mouse sens = 1.00 fixed isi
Individual lags:
Run Mean Median
012 24 15 23 15 24 18 12 16 15 18.0 16.0
013 21 26 20 13 14 17 16 21 20 15 18.1 18.5
014 21 13 13 14 15 18 20 14 17 16 15 16 16.0 15.5
015 16 13 16 13 14 14 14.3 14.0
Bill's modifications to SDTEST3 included reducing mouse sensitivity by
dividing mouse output by 3 (sensitivity = 1/3 or 0.33) and using randomly
varying inter-switch intervals. I wondered about the contribution of these
changes to the reduction in lag times. Run 013 put the mouse sensitivity
back to 1.0; this change had no discernable effect on the switching lag.
Run 014 removed the flash while keeping the code in place that produces the
flash (the flash "color" was changed from white to invisible black). I was
astonished to find that removing the flash did not affect my performance; in
fact, the lag was even shorter. Run 015 used the original SDTEST3 program,
which formerly had yielded lags on the order of 30. The inter-switch times
were fairly long and constant; the lags produced were the shortest yet.
Apparently we're still seeing improvements with practice. Perhaps this is
making it easier to attend to the target-cursor distance and cursor colors
at the same time, so that the flash becomes less needed to cue the change in
cursor color. I'll be interested to see whether this phenomenon occurs with
others.
Regards,
Bruce