[From Bill Powers (990804.0806 MDT)]
Bruce Gregory (990803.1258 EDT)--
A great post about quantum mechanics (QM) and general relativity (GR),
although the implied comparison of me with Einstein is not to be taken
seriously. If you were in here where I am, you would know that.
I really think that PCT has more of an experimental base than you imply.
Remember that _neither_ QM nor GR contains any explanation of the basic
observations: both say only that if you observe nature in a particular way,
after arranging for or waiting for particular circumstances to exist, you
will observe certain phenomena. Neither one can explain _why_ this is true,
in the sense of "why the light turns on when you flip the switch." Even in
QM and GR, any models that are offered are corroborated through behavioral
experiments, not through going to a deeper level of analysis.
The basic phenomena described in PCT can be stated simply or in more
complex ways. Simply, we can say that organisms control inputs, not
outputs, and that they push back when their inputs are pushed upon. These
basic relationships can't be predicted from any other theory that I know
about but control theory.
In a more complex approach, we can describe PCT by proposing models that
fit behavioral observations. These models say the same thing as above, but
they say it in terms of quantitative relationships that allow comparisons
of the details of model behavior to the details of real behavior. Many
different detailed models can be proposed which fit the simple qualitative
statements above, but when the model becomes quantitative we can begin
talking about _how well_ their behavior fits that of the real system.
Should we use a model with a perceptual transport lag in it? An integral
perceptual lag? No perceptual lag at all? Should the output function be a
pure integral or a leaky integral, and if leaky, how much loss per second
should there be in it? Or should the output function be essentially
proportional, or digital (on-off)?
Each detailed model probably allows several alternative but equivalent
structures, so when we find the best-fit model we can say only that the
correct one will be one of the equivalent structures, and not one of the
non-equivalent structures. For example, the correct model may be one with a
lag in its perceptual, comparison, or output function, or any two or all
three, but not one with no lag at all in it.
When there are ambiguities, we have to turn to the next level of analysis,
through dissection of the underlying physical system. In QM and GR, we
don't know what that underlying physical system is, and for a long time
after 1945 it seemed that physicists had stopped looking for it. In PCT,
however, the nature of the underlying system is clear. It is the
neuromuscular system, plus the physical environment (modeled as of the 19th
Century).
Primarily through animal experiments, we know some of the details of some
of the control systems that exist at low levels of organization: the spinal
and brainstem "reflexes." So far (as far as I know) PCT and HPCT models are
the only ones in modern times that have been tested at these levels by
constructing a quantitative model to represent the qualitative hypotheses
about organization, and then letting the model interact with a model of the
physical arm for quantitative comparison with real behavior. Other models
omit the step of simulating the physical arm and letting the combined
models behave. In many experimental studies there is no explicit model at
all, so when explanations are offered it is simply assumed that the
explanations would work if actually converted into a generative model and
tested. We can, I think, dismiss such explanations as untested.
An example: for a long time it was explained that the Golgi tendon reflex
came into play only for dangerously high levels of muscle tension, and had
the function of preventing muscle damage through overuse. A decade or so
ago, it was found by Houk et. al. that the Golgi sensory signal varied
smoothly down to levels of 0.1 gram of tendon tension. The first
explanation had been offered long before this experiment took place, but
had never been subject to experimental test by means capable of accurate
measurement. Had a quantitative model been constructed with the formerly
assumed properties, the discovery of its incorrectness would soon have
followed, and perhaps the re-meassurement would have been carried out much
sooner. This is a perfect example of your quote, Bruce: don't believe any
experimental results until they
are confirmed by theory!
As it is, some physiologists still seem to believe that the Golgi tendon
signal has no effect except for very large tendon tensions, so they have
changed none of their conclusions, including the conclusion that Golgi
feedback has no effect on motor outputs except under extreme conditions. If
the custom were to construct and test models, this conclusion would have
been abandoned long ago.
Unfortunately, once a scientist has gone so far as to publish a conclusion,
retracting that conclusion becomes a very big deal (the implication of all
the drama being, I suppose, that it is incredibly rare for a scientist to
make a mistake).
At any rate, PCT is not like GR or QM because there is still another level
of experiment to which we can resort when the behavioral experiments are
not clear (or if they are clear, to subject them to still a further test).
In QM, this would be like having the ability to dissect reality and see
whether forces really are transmitted through "particles" that move from
one place to another. In GR, it would be like having a straightedge that
measures the curvature of space.
Best,
Bill P.