SDTEST3 Progress

[From Bruce Abbott (950318.1305 EST)]

I've been working on another version of the analysis program for the SDTEST3
data. Although it is not yet finished, I thought there might be some
interest in the outcome of one attempt to automate finding the best k and
delay values for the current one-level model.

The procedure seems reasonable: find the best k for the model while
excluding from the analysis the transition data immediately following the
switch in cursor color (and active target), then use this k-value while
seeking the best switching delay. Both phases would use an iterative binary
search strategy to find the best value.

It turns out that there are at least two problems with this method. First,
the fits between model handle and actual handle are not as good as those
achieved by the previously used hand method of selecting a delay, fitting k,
and iterating these two steps. Second, the procedure failed to produce
reasonable results for the case in which the data were collected absent an
active disturbance to the cursor. In that case there was no need to keep
adjusting the mouse during the non-transitional portions of the run, so the
k-fitting procedure had little to go on. Yet the old procedure works with
or without a disturbance being applied to the cursor, as the switch itself
provides a disturbance to which a measurable response occurs, providing a
basis to compute k.

As a consequence, I'm rewriting the analysis program to automate the
previously used hand fitting of the delay. I plan to measure the actual
switching lags and then use the median lag as the starting point for the
iterative procedure. In hand fitting the data, I've noticed that the median
lag is usually slightly higher than the optimal delay, as the integration
lag (inversely proportional to k) of the model absorbs part of the observed
switching lag.

The participant's responses to the switching transitions can be viewed
partly as a reflection of the operation of the same lower-level control
sytem that handles responses to the disturbance during the time that a given
target remains selected, so it may be more appropriate to compute k while
including the transition data than while excluding them. In fact, for data
collected while the disturbance was acting, the k-values computed with and
without transition data are not all that different once the optimal lag has
been entered, as would be expected if the post-lag transition reflects only
the activity of the lower-level control system.

Meanwhile, I've run two more participants through one trial each of the
STTEST3 task. The first was a student who dropped by my office to ask about
what research I am currently involved in. I though that the easiest way to
answer that quesiton was to demonstrate, so I had her complete a run and
then showed her the analysis of her own data. The other participant was my
daughter Cyndi, who is 21. Here are the results of the runs:

Run lag k rms r
016 (student) 27 0.0830 24.20 0.979
017 (Cyndi) 22 0.3467 19.62 0.987

Both participants initially showed a marked tendency to overshoot the target
following a switch but quickly recovered. What is most remarkable about
these runs is how good they were on what was, after all, the participant's
first attempt. Good, of course, is a relative term, and by it I mean good
compared to the early performances of us old folks. Cyndi has agreed to go
through a series of sessions so that we can get some idea of what the
terminal performance of a well-practiced perticipant looks like.

Rick Marken has offered to develop a model of SDTEST3 performance in a week
or so that will incorporate a second-level "relationship" control system
(although for some reason he wants to do it on the Mac instead of a real
computer); I plan to do the same when I get the time, but for now I'm going
to have to attend to other business. I've got some other ideas to try as
well with respect to data collection, but I'll hold off discussing those
until I have had an opportunity to experiment a bit.

Regards,

Bruce