[From Bill Powers (960402.0700 MST)]
Bruce Gregory 960401.16:45 --
The Test only tells us if a system is successfully controlling, not
if it is trying unsuccessfully to control or ignoring the variable
altogether.
Modeling the control behavior can tell us more than that. While learning
is in progress, the parameters of the model that best fit the data will
be changing. The Test itself is just an outline for making qualitative
decisions about what is controlled. But a simulation that mimics the
behavior using a postulated control model gives us a quantitative
picture of control, including how good the control is and how the
quality of control changes with time. We have more choices than
"successful" and "unsuccessful" control. There is a continuum between
those extremes.
···
-----------------------------------------------------------------------
Rick Marken (960401.1600) --
I think there was a strong suspicion in PCT circles that rats in
operant conditioning experiments were controlling food delivery
rate; this suspicion was based on observation of conventional data
-- data that was collected without the introduction of disturbances
to possible controlled variables. I have always argued -- against
mild to strenuous objection from virtually everyone, including Bill
Powers -- that conventional data is virtually useless for the study
of control. I think Bruce Abbott's rat experiment illustrates my
concerns rather nicely.
I still disagree with you that these data are "useless." That's an
exaggeration. What Bruce found was that a variable (pressing rate)
assumed to be a measure of the rat's compensatory behavior was not, in
fact, varying in the experiments in question. But in the same analysis,
I found that changing the output variable to "meal size" and treating
rate of pressing and meal spacing as disturbances produced a pretty good
control model (correlation with actual meal sizes over 0.95). Now, of
course, Bruce is taking really good data with every press and every
reinforcement recorded to the nearest 1/30 second, as well as a video
record showing where the rat is and what it is doing with the same time
resolution. He's recording exactly how much food is ingested and how
much is spilled, in the experimental sessions and between them, and he's
weighing the rats every day. Of course you could say that this removes
Bruce's data from the realms of "conventional data."
I do agree with you in the sense that I think these experiments are much
too complex, and involve artificial constraints not found in a normal
environment (the on-off nature of keypresses being a major constraint).
But I think it's a worthwhile challenge to try to understand what the
rats are controlling and how they're doing it even in this complex
situation. It's been a relevation to me to see how many different
behaviors are discernible even in this "simple" fixed-ratio-1
experiment. To call this "lever-pressing behavior" is a gross
oversimplification of what is actually going on. In fact, Bruce is
developing a notation for ethologically recording behaviors from the
video tapes, and for the first five minutes of the first tape being
analysed, has had to use 11 categories of behavior, four of the
categories occuring 60 or more times, with more categories known to
occur later on in the tape.
The only disturbances involved so far are naturally-ocurring ones
created by variations in the rat's own behaviors. But after 25 days of
running the first three rats, we're still waiting for them to reach
asymptotic weight on a FR-1 schedule! Another week might be needed.
We'll think up systematic disturbances after the rats settle down to
steady-state body weight.
It's becoming clear to me, as Bruce patiently continues the experiments,
that there is no way in hell that previous experimental data could have
been taken with enough care to support all the hypotheses and
conclusions that have been reported in the literature during the last 50
or 60 years. We could easily spend the entire lifetimes of these rats
just figuring out what happens on two or three fixed-ratio schedules.
Fortunately, Bruce's attitude is that we'll take as much time as
necessary to make sure the experiments are done as well as possible,
with no target date for finishing or publishing. If it takes two years,
it takes two years, and if we don't get beyond FR-4, we don't.
-----------------------------------------------------------------------
Best to all,
Bill P.