Descriptive/generative; info theory; applying PCT

[From Bill Powers (921223.0915)]

It's interesting how a defense turns into an attack, and how
letting down the defenses also reduces the attacks. Sometimes the
best defense is no defense at all.

Martin Taylor (921222.1815) --

But I'm equally impressed with the argument from evolution and
other more abstract arguments of necessity and possibility.
They all support one another.

The problem with arguments from evolution and other abstract
arguments is that while they may apply to a particular case, they
may also turn out to apply to counterfactual cases. They may
contain flaws that don't show up in the factual case (for
instance, flaws in internal logic), but those flaws can make the
generalized explanations specious, even though they "support" a
more concrete analysis.

A generative model can be wrong; it can produce right results for
wrong reasons, which show up only when circumstances change
enough to reveal a wrong prediction. This is how generative
models progress; from error to error. Generative models commit
themselves to specific proposals about underlying details of
operation. That makes them very sensitive to experimental test.

Descriptive models, however, are not very sensitive to
experimental test, some not at all. If they're cleverly put,
their predictions can remain true no matter what the outcome of
an experiment. So they are capable of "supporting" completely
contradictory generative models. This is not really support at
all. It is merely agreement. All experimental results pertaining
to behavior, whether correct or erroneous, are agreed to by the
generalization that natural selection produced the behavior we
find with our experiments. So evolutionary theory "supports"
whatever we find to be the case -- even if, later on, we discover
a mistake and find something else to be the case.

This is why I prefer generative models. They are more sensitive
to experimental test.

RE: information theory

Who is doing this presuming in the system? I asked why the
perceptual and reference signals should not both be considered
to carry an information flow appropriate to a signal varying
within a 2Hz bandwidth. You didn't answer that.

The presumption was one that you (Bill Powers) would have to
make in order to assert that the perceptual and reference
signals would convey to you (Bill Powers) information at a rate
appropriate to a 2Hz bandwidth. To anyone else, the
information rates might be different.

  Ok, so to me, the perceptual signal carries low information if I
already know the reference signal, and vice versa.

However, if we define the comparator as the receiver of the
information in both channels, doesn't this imply that to the
comparator, the same must be true? This doesn't seem satisfactory
to me. Perhaps we have to define the receiver more carefully. A
comparator receives the reference signal and the perceptual
signal at the same time, so it knows neither one before the
other. Further, what it "knows" is only amplitude, over a
relatively short span of time like 0.1 sec or less. It
immediately forgets the history of both signals, and it contains
no machinery for making extrapolations of either signal into the
future. It seems to me that this comes down to the original
Shannon application of information theory -- to defining the
information CAPACITY of a channel rather than the actual
information flow in that channel, which can be less than the
channel capacity. At times when no signal is flowing, the
information flow is zero, yet the physical channel capacity
remains the same. Channel capacity is a function of physical
design; information flow is a function of the kind of message
being carried -- not so?

···

-------------------------------------------------------------
Gary Cziko (921223.0315 GMT) --

But surely, while the modeling is very important, I would like
to think that there are lots of ways of advancing PCT without
doing the nitty-gritty modeling.

There are ways of APPLYING and TESTING the findings of PCT
without modeling and formal experimentation, but there aren't any
ways of ADVANCING it. If you're willing to take the word of
modelers and experimenters about how control systems work
(199223), then you can apply PCT all over the place. There are,
however, dangers in putting too much trust in a theory that is
underdeveloped -- in taking the pronouncements of theoreticians
and researchers on faith.

One way to guard against these dangers to some extent is to make
sure that every application of PCT that you think of is also
designed as a TEST of PCT. In psychology, theories of behavior,
once "determined" to be true by someone's study, are never tested
again. They are used to explain behavior and diagnose behavioral
problems, but nobody ever says "If this theory is still true, in
this situation, then when I do X (or X happens) I ought to
observe consequence Y." Instead, what is observed is some
consequence Y, and it is assumed, because of the theory, that X
must have occurred. If depressed white males have been found to
have difficulty with following complex instructions, then if a
person has difficulty following complex instructions, that person
must be depressed (or a depressed white male).

PCT can be used the same way, of course. But PCT is designed from
the ground up as a predictive model. It can be applied that way.
PCT says that organisms, when pushed, push back (whatever the controlled
variable being pushed upon). If you want to see if a
person's behavior fits this model, you push on what you believe
to be a controlled variable and see if the person pushes back to
counteract your push. If so, you can explain that particular
behavior using the same model.

If you want to find out whether black, brown, tan, or beige
Americans "tend to see cultural and language differences as a
type of protective barrier to maintain," you try to disturb these
barriers (or look for natural disturbances) and look for actions
that specifically counteract the disturbances. You may find that
for some differences this holds true and for others it doesn't.
"Cultural and language differences" cover a lot of territory, and
much of it may be irrelevant. It would be better to find out what
disturbances they do resist, and make the generalization
afterward. And it's important to make a prediction from control
theory (and your hypothesis about the controlled variable) FIRST,
so you're committed before the fact. If this happens, that should
happen. If _that_ doesn't happen, you have to revise your
hypotheses about the controlled variable and try again, and keep
trying until you predict correctly. Then and only then will you
know what's going on. If you NEVER can make a right prediction,
control theory is probably wrong.

I think you should always apply control theory as if you're
putting the theory itself on the line, and challenging nature to
behave the wrong way. Mind the turtle; it makes progress by
sticking its neck out.
----------------------------------------------------------------
Tom Bourbon --

You said it for me; here it is again:

" .. may Boss Reality tread lightly on the controlled variables

of you and of those you love. Happy holidays."

================================================================
Best to all,

Bill P.

[Martin Taylor 921223 14:40]
(Bill Powers 921223.0915)

But I'm equally impressed with the argument from evolution and
other more abstract arguments of necessity and possibility.
They all support one another.

The problem with arguments from evolution and other abstract
arguments is that while they may apply to a particular case, they
may also turn out to apply to counterfactual cases. They may
contain flaws that don't show up in the factual case (for
instance, flaws in internal logic), but those flaws can make the
generalized explanations specious, even though they "support" a
more concrete analysis.

What I'm talking about is sometimes called "converging operations."
Or it was when I was in graduate school. If looking at a problem from
several different viewpoints gets the same answer each time, or if an
answer turns out to be in common among many from each of the various
viewpoints, then that answer becomes much more credible than if it is
precisely determined from only one viewpoint.

I doubt that I would buy PCT simply because an evolutionary analysis
suggested that hierarchic control systems would have enhanced stability
and hence would be observed preferentially among living organisms. But
I would be much less likely to buy PCT if the evolutionary argument went
the other way. It helps the credibility of PCT a great deal that organisms
that are organized as control hierarchies are more survivable than
organisms with other proposed structures (e.g. stimulus-reactors or
simple planners).

Descriptive models, however, are not very sensitive to
experimental test, some not at all. If they're cleverly put,
their predictions can remain true no matter what the outcome of
an experiment. So they are capable of "supporting" completely
contradictory generative models. This is not really support at
all. It is merely agreement.

Yes, the degree of support depends on where the competition is. If
the descriptive model supports two contradictory generative models, it
supports neither, but it might support that class of models (the gaphroomus
could be black or white) against a different class (the gaphroomus mustn't
be spectrally coloured). The importance of agreement depends on the prior
likelihood of disagreement.

This aspect of converging operations applies to your critique of Gary. If
PCT usefully suggests ways of assisting people, and those ways work reasonably
often (not necessarily always), then that success also argues that PCT is
a valuable way of looking at the psychological world. It doesn't prove
anything, but it does reinforce the idea that PCT is worth following up
the levels hierarchy, and is not restricted to tracking tasks with
measurable errors. You have also provided examples at the level of
personality (self-image) that make the same point.

···

---------------
Occam's razor deals with the amount of specification needed in order to
desribe phenomena. If one descriptive method can be used over a wider
range of phenomena than another, or if it describes phenomena more precisely
within a given range, it is to be preferred. PCT does both. It substitutes
one descriptive method for a variety of topic-specific theories, and in
the application area of each theory it makes more precise descriptions.

To me, a model is only a concise way of making descriptions. A generative
model, as you put it, is distinct from a descriptive theory. To me, it
is not distinct. It only uses a few more parameters to make a much better
description, and if that description is accurate, Occam's razor says that
it is preferred. But if its precision allows the world to say "that's not
right," it could be that the parameters are wrong. If the parameters can
be changed and still lie within the bounds of "the theory", then the
information contained in the parameter specification has to be included
in the "size" of the description, so the case becomes less clear.

On information theory: Channel capacity is not specifiable from physical
characteristics alone. A physical channel may have many different capacities
simultaneously. And (as follows from one of your own postings a while back)
you can have continuous information flow without signal flow.

I prefer, I think, to set aside postings on information theory until I can
produce a coherent PCT-oriented tutorial. That will, I hope, provide a
reasonably good argument as to why PCT and specifically HPCT is, with
high probability, the only viable organization for thermally unstable
organisms (i.e. life) to persist. As both Tom and Rick have said directly,
and you imply, the argument may not turn out to be as strong as that, and
may provide only agreement. But at least it will be an argument from a
different viewpoint against simple planners and S-R organisms, and provide
a converging operation in support of PCT.

I have a growing suspicion that we have a different philosophy of science,
and that at some point this difference will be the theme of a posting thread.

Martin