Killeen the control theorist

[From Bruce Abbott (960113.1410 EST)]

From Bill Powers (960113.0530 MST) --

Sorry to have co-opted your example. Killeen's system, I agree, is not
conceived of as a control system, although the outer loop, as you say,
turns out to be a control system.

Maybe now Rick will accept that.

What you seem to be doing is looking for a justification for picking
incentive delivery as the prime mover in the loop, as if you really
agree with Killeen's choice of that prime mover.

. . .

Just stop a minute and look at what PCT would say is the prime mover in
this loop: it is the reference signal specifying how much incentive is
to be ingested, or the rate at which it is to be ingested.

I would say that it is the error signal that is the prime mover in this
loop, not the reference, but perhaps we have different definitions of "prime
mover." There is no output (behavior) until there is error, regardless of
the value of the reference signal. In that sense the system doesn't "move"
(act) unless there is error. Yet such a system will follow its reference
(if it can); in that sense the reference is what "moves" the cv about, and
this is the sense in which you mean the term, I would think. There is no
real disagreement between us here, just a difference in definition.

My identification of incentive delivery as the "prime mover" in Killeen's
system has nothing to do with any agreememt or disagreement on my part with
Killeen on this choice. I'm simply describing _Killeen's_ choice, and a bit
of the evidence he invokes to support that choice. I want to know what the
implications of that choice are in Killeen's system, and in particular how
those implications differ from those derived from the different choice
implied by a control model designed to account for the same performance
Killeen's system is designed to account for.

Bruce, remember the difference in sign, even within Killeen's own
deprivation model. He proposes that behavior increases with deprivation.
But deprivation _decreases_ with R. His arousal hypothesis requires that
behavior _increase_ with R.

I don't find anything particularly strange about that: we are, after all,
talking about two separate loops, with different variables being affected as
a function of R and in turn affecting R. You make it sound as if these
changes are mutually inconsistent, but they are not, as demonstrated by the
simulation.

The whole reason he needs all this stuff
about coupling and zeta and the rest is the fact that it seems that
behavior sometimes increases with R and sometimes decreases with R. So
he needs some way to switch between the two relationships, with the
switching being dependent on the ratio through a set of assumed indirect
relationships adjusted to work right.

In the ratio case we have been using to illustrate Killeen's system, R =
B/n, so B _must_ increase with R. I think you're talking about n here, not R.

I can't fault Killeen for trying to identify a set of variables and
relationships that will allow the model to work properly (i.e., fit the
data). These additions to the model must be evaluated according to certain
criteria. Are they reasonable (i.e., do they make sense, do they fit what
we know)? Do the parameter values needed to fit the data seem reasonable?
Do parameters that ought to remain constant across different experimental
settings (e.g., subject-related constants) do so, or must they be changed to
make the model fit the data in each case? Can they be independently
measured, outside the experimental context in which they are applied? Is
there independent evidence for the processes invoked? Is the model as a
whole consistently applied across the entire spectrum of experimental
conditions to which it presumably applies, or are bits and pieces included
or excluded as needed to make the theory fit the data? Perhaps most
importantly, do these additional parameters and functions render the model
capable of fitting virtually any set of data, making it incapable of disproof?

Several of the reviewers of Killeen's _BBS_ paper indicated their concern
about several of these issues, and they are of concern to me as well. But
at this point my understanding of certain details of Killeen's model is
still incomplete, so I am not in a position yet to offer an informed
opinion. I am certainly not arguing in favor of Killeen's model; I am just
trying to understand it well enough to give it a fair evaluation.

Since he has assumed that R is the
prime mover, he has to introduce effects by other routes that counteract
the "basic" effect of R, and then set up equations in which the
parameters can be ajusted to bring these other effects into play just at
the point where they're needed. In doing this he has to postulate a lot
of very fanciful unobservable processes, and assign specific properties
to them, the only criterion being to make the result match the data.
Where I come from, these properties are called fudge factors.

They are where I come from, too. But that's all part of theory development.
Galileo's parabola really doesn't describe the motion of cannon balls
through the atmosphere; to do the job accurately you need to add a fudge
factor that takes air-friction into account. The need for such a fudge
factor in order to make the model fit the data can mean that the basic model
is wrong, or that there is some additional process (like friction) that
needs to be included in the model. Of course, once you devise such a
process and add it to the model, it is incumbent on you to demonstrate the
reality of the process; otherwise you are just saving the theory.

The evidence you cite is no evidence at all. All the experimenter did
was make the food accessible. If the pigeons had not moved to it, picked
it up, and eaten it, there would have been no "delivery" of incentives.
The delivery, as usual, depended on the behavior. And this experiment
does not support the idea that a single incentive can cause a series of
pecks which, in turn, provide an incentive. All it shows is that once
the pigeons find that food is accessible, they will start looking for it
anywhere in the environment, until it is clear that it is no longer
accessible. In Killeen's experiment, the positions of the pigeons were
recorded in measuring "activity," but not what the pigeons were doing.
I'll bet they were pecking all over the place.

They were probably doing a lot of things; it would be useful to have a
detailed record of it. I agree that Killeen's elicitation view is not very
compelling. But these pigeons were given a goodly number of sessions and in
each of them the grain only appeared once, and only in one place. Why would
they continue to expect to find food elsewhere (and elseWHEN)? In your
model, what does it mean when you say that a pigeon will continue to look
for food "until it is clear that it is no longer accessible"? That is, how
do you represent this relationship within your model of the pigeon in this
experiment?

Regards,

Bruce