Loop decomposition, Iteration

[From Rick Marken (941001.1030)]

Martin Taylor (940929 10:30) --

good, reliable and (I think) useful data can be obtained on the
components of control loops, without studying the whole control
loop in action.

Agreed. There is good data regarding each component of the loop,
some of this data being crucial to the formulation of the model. Given
the simultaneous equations that define the canonical control loop as
follows:

(1) p = f(q)

(where q = g(o+d)) and

(2) o = h(r-p)

(where p is the perceptual signal, o is the output variable, d is the
disturbance variable, q is the controlled variable, and r is the reference
signal)

then I think we can say that we know quite a bit about g(), which
represents the physical laws that relate environmental variables to one
another (o, d and q are all in the environment). We also know
something about h() and the behavior of neural signals like r and p
from neurophysiology.

I think we also have some inklings about the nature of f(), the
perceptual function, from perceptual and psychophysical studies, like
those you describe BUT caveat emptor! There is a problem with such
studies because your conclusions about the nature of f() are based on
the relationship between o (the subject's observed responses) and
either q or d (depending on whether the subject actually has an
influence on what is considered the independent variable -- in which
case it's q -- or not -- in which case it's d).

In either case, the relationship you are looking at is in conventional
perceptual/psychophysical studies is not f(). In your "ambiguous
figure" studies, for example, the subjects outputs -- reports of changes --
have no effect on the input -- repeated presentatons of "splendid
gladiola" -- so you are looking at the relationship between d and o in an
INTACT control loop. Since the subject's output does not combine with
the disturbance to produce what the subject perceives, it is very difficult
to tell what the observed relationship between o and d tells about any of
the functions in the control loop. So, even though the data are clear (and
they are -- so I un-recant my "good data, bad theory" remark -- they are
good data in that they are very reliable) we really don't know what they
mean. The relationship you observe between d (repetitions of "splendid
gladiola") and o (indications of change) may reveal somethong about
f() -- the perceptual funciton-- but, as you can see, it is more likely
that it does not. The reliability ondicates that it reveals something
about some function in the loop but it's hard to tell which one.

All of this can be viewed as another way of supporting Bill Powers'
(940930.0655 MDT) suggestion that you replicate your perceptual
experiments as control studies. If you did this, you would have the
control model available as a basis for doing the functional analysis. It's
true that, in principle, one can study the components of a control loop
one at a time, but you are not doing that in conventional psychopysical
experiments. I think you are getting a better view of f() all by itself in
single unit studies of the response of afferent neurons to various
patterns of input on the sensory surface. But the psychophysical
studies are not looking at just one function in the loop -- like f() -- even
if they say they are. Psychophysical studies are based on observations of
responses in an intact control loop and (as we know) such responses
are not outputs (as psychophysicists assume).

Bill Powers (940930.0655 MDT) also writes a wonderful essay on
replication. In it he says:

Others who know experimental psychology and psychologists better
may not agree with me.

As the author of one of the finest (if not the most popular) textbooks
on conventional experimental psychology that was ever written;-), I
agree with everything you said in this essay, Bill.

Since you have sung the praises of _replication_ so beutifully, let me
add a brief paean to the companion to replication -- _iteration_. Most
psychological research is a one shot affair; you test tons of subjects and
see where the numbers fall; if you're lucky, you find the right averages
and get tenure. PCT research is not only based on replication (I like to
think of it as _savoring_, not just repeating), it is also based on
_iteration_; you keep testing (by distrubing hypothesized controlled
variables) until you come up with a definition of the controlled
variable that is "right". Every disturbance to a correctly identifieed
controlled variable is completely and precisely resisted. This is an
important part of the research that Chuck Tucker described; each of
Chuck's tests is just the start of what can be a difficult and frustrating
process -- one that requires creativity (the ability to think of revised
definitions of controlled variabled based on the results of previous
tests) and care (the ability to systematically test each definition of a
controlled variable until one is satisfied that it is, or is not, correct).

Iteration reflects confidence that organisms really do control perceptual
variables, that these variables can be discovered and that, when they
are, they will be protected, precisely, from disturbance. I would like to
see examples of the iterative convergence of The Test toward
controlled variables described over and over again on CSG-L. If
nothing else, it would be nice if everyone on CSG-L did "the coin
game" (p 325 of B:CP) and discussed their experiences with it. The coin
game is all about how to discover a controlled variable through
_iterative_ application of The Test.

Best

Rick