GA; music box; Von G.;Psychophysics; Laplace

[From Bill Powers (930112.1930)]

Greg Williams (direct) --

OK on models & worlds; take it away.

We got our driveway plowed this morning and were able to go
shopping. Ready for the next one.

Gary Cziko (930112.1620) --

From what I saw of Randall Beer's projected uses of GA,

reorganization per se is not the only factor (although it figures
in). There are also "genes" for particular chunks of behavior or
organization that get mixed between generations according to
genetics-like rules.

I'm still concerned about internal vs. external criteria. What
does Koza have to say about this? Are partially-successful
outcomes allowed to propagate? How is "success" measured? Does
the algorithm need a way of perceiving its distance from the goal
state?

ยทยทยท

--------------------------------------------------------------
Martin Taylor (930112.1212) --

RE: music boxes

Yes, one must understand something about springs, catches, and so
on to see the second explanation as an explanation. My point was
mainly to bring out the difference between dealing with the music
box behaviorally -- doing this has that effect -- and in terms of
organization -- this sort of device is connected to that sort of
device. Knowing the organization, one can deduce behavioral
effects, but not vice versa. If the music box stops playing even
when the key is wound, the person who understands the
organization can make some intelligent guesses as to what has
going wrong, because the model predicts failure modes too. The
purely behavioral approach has to rely on past experience with
music boxes: the last time that happened, my uncle pounded the
box on the table and it started working again.
-------------------------------------------------------------
Gary Cziko (930112.1800 GMT)--

Ernst von Glasersfeld is a brilliant philosopher-psychologist
interested mainly in epistemology. He's been a supporter of PCT,
and a seldom-seen friend, since the book came out. He is much
politer than I am in arguments. He is also a writer of enormous
skill; anything he has written is worth a read just for the skill
he demonstrates. His affiliation with cybernetics has been much
stronger than mine; I think he figured that in that venue his
ideas had the best chance of being heard. It was Ernst who said
"It's not the organism that's the black box; it's the
environment." He's always been much too busy to come to CSG
meetings.
--------------------------------------------------------------
Martin Taylor (930112.1400) --

RE: Psychophysics

I think Rick has made some interesting points about
psychophysics. It seems to be that with a control-system
experiment you at least make the nature of the perception you're
proposing to measure into a testable hypothesis. You don't really
know what the perception is until you've seen it defended against
disturbance. In a some such experiments I've read about, the
experimenter simply assumes that the experimenter's own
perception of the stimulus variable is the same as the subject's.
This isn't always true.

It also seems to me that psychophysics has spent a lot of time
investigating the margins of perception; difficult
discriminations, perceptions at the threshold of detection, and
so forth. This is interesting information, but it leaves the
great middle range of perception unexplored. Investigations like
Stevens' at least look at perception in the ranges where normal
perceptions occur.

The effects of output complexities are pretty much eliminated by
using simple low-level behaviors like saying yes or no, or
pressing a key, to indicate when a perception has reached some
criterion state. Unfortunately this limits the resolution with
which you can deduce what it is about the stimulus situation that
the person is monitoring. Nevertheless, I think there is probably
a lot of useful information in psychophysics. At least it's not
used to predict and control behavior.

Every control experiment is, in a sense, a psychophysical
experiment. A perception is hypothesized, and the model's
predictions depend on the correctness of the hypothesis. If the
subject isn't controlling the variable you proposed, or something
very closely and quantitatively related to it, the model isn't
going to predict very well.

I think Rick makes an important point, that people don't quit
being control systems just because you put them in an open-loop
situation. They're always controlling; that's how behavior works.
They control what is left controllable. I think it's important to
re-examine apparently open-loop experiments to look for possible
controlled variables; after all, pre-PCT experimenters didn't
spend a lot of time looking for them.

Perhaps it would be useful if you were to describe a specific
psychophysical experiment and see what ideas we can come up with.

Re the remarks on magnitude estimation. This does require
perception of a relationship between two perceptions, one of them
being the manner in which the subject indicates the experienced
magnitude. The experimenter sees this indication as an output,
but the subject is perceiving it and adjusting it until it seems
to match the magnitude of the target perception.

This gets me a little closer to remembering what that philosopher
whose name I have lost said. He was speaking of this subject, I
think. The problem is that there is a magnitude scale involved in
perceiving the behavioral indication of magnitude, too, so it
becomes impossible to give an objective or context-free magnitude
estimate. An example: suppose I indicate the experienced loudness
of a sound by holding one hand a certain distance above a table.
Clearly, I'm moving my hand until the experienced distance above
the table seems about the same as the experienced loudness of the
sound. This means that the apparent scale for loudness is being
converted to the apparent scale for distance above the table. Any
"power law" then expresses only the ratio of these two scales.

Even when subjects are asked to give numerical estimates, such as
"twice as loud" or "twice as long", the scale on which
"twiceness" is measured is unknown. We are still expressing one
scale in units of another scale, and lack any anchor in absolute
numbers. You could just as easily say that the loudness estimate
is being used to show what the person considers "twice as much"
to be.

This was the unknown philosopher's point: magnitude estimates are
always relative to other magnitude estimates, and no absolute
scale can be devised.

The nice thing about control-system experiments is that you don't
need a second estimator; the subject acts on the same perception
that is to be evaluated. But the question of the form of the
input function is still not answered. Maybe there's no way to
answer it without putting electrodes on the sensory nerves.

Isn't there a chance of a circular argument here? By
assumption, no S-R organization can give high correlation
between S and R, but all control organizations give high
correlations between disturbance and action. Therefore an
observed high correlation between S and R proves that they
aren't S and R after all, but disturbance and action in a
control system. This proves that there is a control system
involved always when correlations are high.

Whoa! Correlations are not the only criterion for identifying a
control system. The position of one end of a lever correlates
highly with the position of the other end, but there is no
control system there. You have the cart before the horse.

The low correlations found in S-R experiments are simply a fact.
Responses do not correlate very highly with stimuli, unless in
fact the stimulus is really a disturbance and the response is an
action that cancels the effect of the disturbance on a controlled
variable. In the latter case, one must of course demonstrate the
controlled variable; the correlation alone isn't enough. In fact,
neither is the existence of an apparent controlled variable. You
could say that the part of the lever over the fulcrum is kept
where it is by the fact that a movement of one end of the lever
is compensated by an opposite proportional movement of the other
end. To show that this is really the result of a control system's
action, you have to find the control system and show that it
senses the controlled variable and acts on the appropriate part
of the lever when the other end is disturbed. You have to show
that interrupting the path to the sensors (without affecting
anything else) brings the apparent control to an end.

In other words, a competent control theorist greets any proposal
that something is under control by immediately looking for
evidence that would prove it is not. That is why we call this
general procedure the TEST for the controlled variable, instead
of the PROOF of a controlled variable. There are many ways for a
proposed control system to fail the Test. The lever would fail it
at several points.

An S-R situation is simply one in which the S is affected by
independent variables but not by R.

I should point out that there is a way of artificially raising
correlations in S-R experiments: it is simply to divide both S
and R into bins and use a coarser number scale, the coarsest
being binary. By making the prediction much coarser than the
errors of measurement, you can eliminate all errors of prediction
the size of the bins or smaller. I have seen experimental results
shown as a line connecting just two or three points. This will
naturally give you much higher correlations than you will get by
filling in all the points between. But of course it makes the
findings irrelevant for any intemediate points.

Rick said

Psychophysical experiments are based on the assumption that
responses are caused by stimuli -- rather than the assumption
that people control sensory input.

And you replied

They may have been, but I don't see (still) where a shift in
that assumption would affect either the manner or the
interpretation of the experiments.

The main effect would be to show that the measurement is
tautological in terms of a better model. Suppose you were using
target movements as a stimulus and measuring the resulting eye
movements in a person instructed to watch the target. Ignoring
saccades, you would find that the angular position of the eye was
almost perfectly controlled by the stimulus position.

However, if you asked what the controlled variable was, you would
eventually arrive at the position of the target image on the
retina. Now you would see that your moving the target had
essentially no effect on the image position on the retina. The
actual behavior consists of using eye muscles to keep the image
centered. Naturally, because of geometric optics, the eye's
angular position follows the position of the target. But the
changes in target position are no longer stimuli; they are only a
disturbance of the real stimulus, which is the controlled
variable. And they fail to affect that stimulus because of the
very behavior they seem to elicit.

This sort of reinterpretation is always a possibility unless you
have experimentally ruled out the presence of an overlooked
controlled variable (after a reasonable search). There is little
motivation for an experimenter to go back over old experiments,
particularly those that seemed to work well, to find out whether
a disturbance has been mistaken for a stimulus. But the
possibility always remains until it has been shown to be
unreasonable.
---------------------------------------------------------------
Gary Cziko (930112.2230 GMT) --

Anybody out there want to try to explain in 100 words or less
what a Laplace transform is and why someone interested in PCT
should know about it?--Gary

A Laplace transform is a way of converting linear differential
equations into algebraic equations that are easier to solve. Once
a solution has been reached, the inverse transform will provide
the solution in terms of time variables again. As very few real
systems can be represented by solvable equations, PCTers are best
advised to use simulations rather than reduce real systems to the
nearest form that can be solved analytically, with or without
Laplace transforms.

75 words.
-------------------------------------------------------------
Best to all,

Bill P.