[From Bruce Abbott (950730.1300 EST)]
Rick Marken (950729.2020) --
I think I now understand what Bruce was up to in the post I took
to be a satire. Bruce was showing that the reinforcement rates
observed in ratio experiments (ratio requirement < 64 presses)
could be produced by a rat pressing at a constant rate after a
fixed delay (the consumption period, c). This is, indeed an
interesting possibility and it is why Bruce claimed that reinforcement
rate is not a controlled variable; the reinforcemnt rate could
be a side effect of the actions of an automaton that emits a
fixed delay (c) after a reinforcer and responds at a regular rate (p)
thereafter, until the next reinforcemnet. It is also why Bruce
suggested that response rate (p) and/or consumption period (c)
might be controlled variables.
By Jove, I think he's got it! (:->
Of course, response rate and
consumption period are consequences of action; but these would
be rather perculiar variables to control since doing so would mean
relinquishing control of nourishment. Moreover, there is no evidence
that they are controlled anyway, other than their computed stability . . .
But CYCLIC2 seems to show that response rate may not be as stable as it
seems, once you remove a constant collection time from the analysis;
moreover, Bill's control model seems to do an excellent job of fitting these
data under the assumption that reinforcement rate IS under control. So
perhaps things are not so strange as they seem.
I think the control model of operant conditioning (control of nourishment
in the context of changing feedback functions) keeps getting better as
we improve our understanding of the environmental constraints that exist
in operant conditioniong research. I think the fit of the control model with
reinforcement rate as the controlled perceptual variable makes a pretty
convincing case that reinforcement rate is, indeed, the variable under
An' you ain't seen nothin', yet!
But I also think this episode shows that conventional operant
conditioning experiments are not the way to go about studying control.
Studies like Staddon's baroque "cyclic ratio" cantata show how difficult
it is to make sense of data that was collected without any notion that
variables are being controlled by the subjects in these experiments.
But Rick, as I never tire of repeating to you, this is just preparation, to
gain some insight into the variables at work and their relationships.
Moreover, Staddon's cyclic ratio study was conducted to do exactly what you
ask for: to evaluate a control model. Staddon assumed that rate of food
intake was being controlled. Both you and Bill approached the data just as
Staddon did, assuming that the imposition of higher ratios acted as a
disturbance to food rate which was countered by means of increasing the rate
of lever pressing. So I really don't understand the criticism.
It's fun to try to make sense of data like that from Staddon's study
to the extent that we can -- but I still think it would be best to
just start doing some simple control experiments the right way. The
ambiguity about what is being controlled in operant experiments could
be solved with a few weeks of tests aimed at determining which
variables are, indeed, systematically protected from disturbance --
Rick, analyzing the Ettinger and Staddon (1982) data has been more than fun;
it has been something of an eye-opener. Having conducted this analysis, we
are now in possession of information which will help to improve the design
of our own control studies. For one thing, we now have a better idea as to
what data we need to collect. For another thing, we know that a naive model
which neglects collection time is inadequate.
Furthermore, we aren't done yet, as there are other aspects of the situation
we need to be thinking about and debating (when I have a bit more time).
We're having to think about some important issues that haven't really been
dealt with as yet within PCT, such as the allocation of behavior. In my
judgement, learning how to take such factors into account within a
control-system model will represent an important advance for PCT.
This isn't an either-or thing. We can do basic control studies and we can
anlayze extant data to learn what we can from them which will help up to
develop more adequate control models that apply more generally to real-world