Lecture and demo

[Martin Taylor 970102 15:30]

Bill Powers (961224.1145 MST)

Bruce Abbott (961224.1310 EST)--

Happy New Year to all.

> Martin's elegant contribution
>was to suggest that in this case the CV would be "stabilized" but not
>controlled.

>Rick's response was to deny that the distinction Martin was making between
>"stabilized" and "controlled" was useful. CV, he said, either is
>controlled or is not controlled.

I would tend to agree with Rick, because of my definition of control given
above.

This may be out of date by now, but anyway...

It seems to me that if one wants to discuss (or demonstrate) whether under
some circumstances an environmental function of observables (CEV, for short)
stays stable, and wants to discuss (or demonstrate) that these circumstances
do (or do not) require the existence of perceptual control, then it is
helpful to use different words for the stabilization of the CEV and for
the existence of perceptual control.

It's hardly useful to have arguments of the form
"Under these circumstances the CEV hardly changes its value."
"Under your circumstances, it's not control, and the disturbance isn't
   really unknown."
"I know it's not control, and I know the CEV will be stable in the absence
   of control only if the disturbance is well modelled."
"Then it's not control."
"I know it's not control, but nevertheless under these conditions the CEV
   hardly departs from its nominal value."
"If the CEV is controlled, it has to be by _perceptual_ control, and I'll
   demonstrate this by showing that _perceptual_ control works to control
   the CEV." "...(!)"

It's so much easier to be able to say that (if you believe so) "The CEV is
stabilized, and it's done by an outflow model."
And to be able to make the rejoider:
"Your stabilization won't work if the disturbance has much unmodelled in it.
   If there's much unmodelled about the disturbance, you'll need _control_
   to achieve _stabilization_."

Don't you think we would all (or most of us) agree with this latter
statement, and then we could get on with finding out whether in any particular
_real-life_ circumstance some aspect of the probable disturbance might
perhaps be modellable? And if so, whether to do so might aid stabilization
both without _and_ with control?

Remember that as far as the observer is concerned, what is controlled
is ONLY the CV. The idea that this CV is represented by a perceptual signal
inside the other system is theoretical. We can observe CV, but not p. When
we apply a disturbance, we apply it to CV, not to p. The action that opposes
the effect of the disturbance acts on CV, not p. The Test does not involve p
at all. It involves only observables -- i.e., the observer's perceptions.

One might say: "Of course", but what would be the point? We are discussing
a model called PCT, aren't we? In that model, the controlled variable is
a theoretical entity called a perception. In no real world can we determine
for sure what that perception determines as a _controlled_ CEV (i.e. a CEV
whose stability is due to the fact that its defining perception is controlled).
But we can come close, using the Test to determine a CEV, determined by the
_analyst's_ perception, that is a highly correlated with the actual
controlled CEV as the experimental conditions permit.

···

-----------------

Bill,

If you program your randomly interlinked environment, may I ask two things:

1) make sure that the compute-cycle dt's are short compared with the time
it takes those inter-environmental-variable influences to take effect, and

2) Give a thought to programming it is Java.

On (2), I've been trying out Java, inspired by Rick, over the holiday, using
the free Java Development Kit from Sun (http://java.sun.com). There are kits
or links there for various Unix platforms, PCs (I'm not sure about DOS, but
various kinds of Windows), and Macs. Programs developed using the JDK should
be compatible across platforms (which seems not to be true for programs
using the Microsoft development kit). I had a notion to try to rewrite
Simcon, but my skills are as yet inadequate. I've made a neural net node
class that is intended as a generic perceptual input function, but that's
my sole success so far. All the same, I feel Java is a rather easier
language to write in than C or Pascal, once one gets the hang of it.

If the randomly linked environment is made available as a set of Java classes,
with or without source code, on the CSG Web Server, we could all use it as
part of a test suite.

What say you, Rick? And would you contemplate making your classes (and
better, your source code) available to be built on?

Martin

[From Bill Powers (970102.1700 MST)]

Martin Taylor 970102 15:30 --

It's so much easier to be able to say that (if you believe so) "The CEV is
stabilized, and it's done by an outflow model."
And to be able to make the rejoider:
"Your stabilization won't work if the disturbance has much unmodelled in
it. If there's much unmodelled about the disturbance, you'll need
_control_ to achieve _stabilization_."

Your proposal is perfectly reasonable. It still grates on me, however, for
purely subjective reasons. I think it's incumbent on anyone who proposes an
"outflow" model to put as much work in on it as we have put in on the
control models we use, and that includes proposing at least a preliminary
sketch of HOW the required information is obtained and used. Without that,
as I've been saying to Bruce Abbott, you're just presenting the easy part of
the model and assuming that the hard part will be solved some day.

For "stabilization" substitute "transmutation." We can agree that sometimes
we observe one element being transmuted into another (that's like agreeing
that we can sometimes observe variables being stabilized by actions against
disturbances). A physicist offers one model for how this is accomplished; an
alchemist a different one. The alchemist wants to talk about the
philosophical implications of changing lead to gold; he wants to get on with
that discussion, so he says, "Let's just assume we've found the right
Principles and mental attitudes that will accomplish transmutation. So now
we have lead changing into gold, a base metal changing into a noble one,
which symbolizes the perfection of base human nature through the Grace of
God ... ".

By this time the physicist is jumping up and down in frustration. He butts
in and says "Wait a minute. You haven't shown that transmutation can be
accomplished by any means but the model I've proposed. I don't want to talk
about the perfection of human nature; I want to find out how you think you
can change lead to gold." The alchemist, wounded, says "But just suppose we
have a way to create the right conditions -- is it too much to ask to grant
that for the sake of looking into the implications of having such a means
within our grasp? Aren't you interested in what would happen to the concept
of wealth? Of responsibility? Of ...". To this the physicist simply replies,
"No. I want to know how you accomplish this transmutation. Tell me that and
THEN I'll listen to the rest."

That's pretty much how I feel about outflow models.

···

------------------------------------

Bill,

If you program your randomly interlinked environment, may I ask two things:

1) make sure that the compute-cycle dt's are short compared with the time
it takes those inter-environmental-variable influences to take effect,

I always do.

and
2) Give a thought to programming it in Java.

I've downloaded the Sun Java tutorial after receiving your tip, then
winzip95 which is required in order to unpack it, and now I'm waiting for
the net to get less busy so I can download Java itself (6+ Megs).

I feel Java is a rather easier
language to write in than C or Pascal, once one gets the hang of it.

Good. You can expect calls for help.

Best,

Bill P.

[From Rick Marken (970103.0850)]

Martin Taylor (970102 15:30) --

It's so much easier to be able to say that (if you believe so) "The
CEV is stabilized, and it's done by an outflow model." And to be able
to make the rejoider: "Your stabilization won't work if the disturbance >has much unmodelled in it. If there's much unmodelled about the
disturbance, you'll need _control_ to achieve _stabilization_."

I think this is all ass-backward. If the CEV is stabilized but not
controlled then you should know this before you start proposing models
to explain the behavior of the CEV. The dialog above could only take
place between two dedicated _non-observers_ of behavior.

What say you, Rick? And would you contemplate making your [Java]
classes (and better, your [Java] source code) available to be
built on?

Yes. I will make the source for one or two of my demos available this
weekend (remind me if I forget). These can then be used as a framework
for those who want to build their own Java programs. I make no claims
regarding the quality of my programs. But I hope these programs can be
used by the really skilled programmers on the net as a starting point
for building some great new PCT demos/experiments.

Me:

Martin said that the CV in a control loop is stabilized, not
controlled.

Martin Taylor (970102 16:35) --

No, it was _you_ that put the "not controlled" in there.

Are you sure? I seem to remember both you and Bruce A. saying that the
CV in a control loop is stabilized by outputs in the same way as the CV
in Bruce's simulation-based control model is stabilized (not controlled)
by outputs. I don't have access to any of those old posts so I guess
I'll have to believe what you say. I just can't understand why I would
have reacted negatively to your claim that the CV is stabilized unless
you had implied that the CV in a normal control loop is not controlled.

Can you fairly say that a quantity z (==2x + y), which you measure,
is being controlled when in fact what is being controlled is z' (==
2.01x + 0.98y + .0003w), which you don't measure? Both are
stabilized, but only one is being controlled.

What is controlled is controlled. What is not controlled is not
controlled.

This is the fundamental distinction made by PCT. Indeed, we _must_
distinguish variables that are controlled from those that are _not_
controlled before we even try to apply PCT or we may end up applying
PCT to the wrong phenomenon. PCT is a model of the phenomenon of
control, remember. The distinction you suggest (between stabilization
that "results from" control vs stabilization that does not "result from"
control) is the distinction that led to the development of PCT in the
first place. It is a _factual_, not a theoretical, distinction. The
fact is that some variables are controlled; some are _not_. The life
sciences have failed to understand this (factual) distinction and have
proceeded as though all variables are _not_ controlled. That's why all
models of behavior in the life sciences are _cause-effect_ models;
cause-effect models are what yone uses to describe the behavior of
variables that are _not_ controlled.

Best

Rick

[Martin Taylor 970106 13:30]

Rick Marken (970103.0850)

Martin Taylor (970102 16:35) --

>No, it was _you_ that put the "not controlled" in there.

Are you sure? I seem to remember both you and Bruce A. saying that the
CV in a control loop is stabilized by outputs in the same way as the CV
in Bruce's simulation-based control model is stabilized (not controlled)
by outputs.

That's hardly something I would be likely to say, is it? Except in your
fertile imagination. Though I can well imagine that you could have construed
something in my writing to agree with your presuppositions.

. I just can't understand why I would
have reacted negatively to your claim that the CV is stabilized unless
you had implied that the CV in a normal control loop is not controlled.

No, nor could I. Neither did I understand why you put forward your
(subsequently retracted) claims that the environmental variable, rather
than the perceptual variable, was what was controlled. Of course, you
hadn't _intended_ to say that, but your writing implied it.

Don't you think that after all these years, when we each know that we both
understand well the basics of PCT, we could take a bit more charitable
interpretations of writing that might be a bit ambiguous?

>Can you fairly say that a quantity z (==2x + y), which you measure,
>is being controlled when in fact what is being controlled is z' (==
>2.01x + 0.98y + .0003w), which you don't measure? Both are
>stabilized, but only one is being controlled.

What is controlled is controlled. What is not controlled is not
controlled.

This is the fundamental distinction made by PCT. Indeed, we _must_
distinguish variables that are controlled from those that are _not_
controlled before we even try to apply PCT or we may end up applying
PCT to the wrong phenomenon.

Under what version of the Test could you distinguish the above two
variables? All you could ever know is that _both_ are extremely well
stabilized by the actions of the person who is controlling the perception
represented by _only_ one of them.

PCT is a model of the phenomenon of
control, remember. The distinction you suggest (between stabilization
that "results from" control vs stabilization that does not "result from"
control) is the distinction that led to the development of PCT in the
first place. It is a _factual_, not a theoretical, distinction.

My whole point in introducing the word "stabilization" into the discussion
was to enable the discussion to procede on a factual basis, rather than
on the theoretical faith that stabilization without control is impossible.

If you use your _words_ to deny the possibility of alternate interpretations
of observations, you make science very difficult.

Martin

[From Rick Marken (970106.1430)]

Me:

I seem to remember both you and Bruce A. saying that the CV in
a control loop is stabilized by outputs in the same way as the
CV in Bruce's simulation-based control model is stabilized
(not controlled) by outputs.

Martin Taylor (970106 13:30) --

That's hardly something I would be likely to say, is it?

I don't know. Ask yourself;-)

Neither did I understand why you put forward your (subsequently
retracted) claims that the environmental variable, rather
than the perceptual variable, was what was controlled.

I never put forth the claim that the environmental variable _rather
than_ the perceptual variable is controlled, so I never retracted it.
I claimed that both the perceptual variable _and_ the environmental
correlate of that variable (the CV) are controlled by a control system.
I still claim it;-)

Ye:

Can you fairly say that a quantity z (==2x + y), which you measure,
is being controlled when in fact what is being controlled is z'
(== 2.01x + 0.98y + .0003w), which you don't measure? Both are
stabilized, but only one is being controlled.

Me:

What is controlled is controlled. What is not controlled is not
controlled.

Ye:

Under what version of the Test could you distinguish the above two
variables?

The one where you apply a disturbance that will have a predicted effect
if the variable is _not_ controlled and far less than the predicted
effect if it _is_ controlled. It seems to me that you could distinguish
control of z from control of z' by varying w. This disturbance will be
completely effective if z is controlled but not
if z' is controlled. This seems pretty obvious to me; am I missing
something?

My whole point in introducing the word "stabilization" into the
discussion was to enable the discussion to procede on a factual
basis, rather than on the theoretical faith that stabilization
without control is impossible.

I don't understand this. If stabilization without control is impossible
then why make a distinction between stabilization and control? They
would be the same phenomenon. I thought that the distinction between
stabilization and control was being made in order to characterize the
difference between the behavior of a possible controlled variable (like
the position of a cursor) in a simulation-based and in a real control
system. This variable is stabilized by a simulation-based control system
and controlled by a control system. You can tell whether this variable
is stabilized or controlled by applying arbitrary (unpredictable and
undetectable) disturbances to the variable and looking for _lack_ of
effect.

If you use your _words_ to deny the possibility of alternate
interpretations of observations, you make science very difficult.

I don't understand this. I thought stabilization and control refer
to two different phenomena. Don't they? Stabilization can _look
like_ control; it can vary less than expected based on an analysis
of the causal influences on it. But we can tell whether what we are
seeing is a stabilized or controlled variable by doing The Test.

If this is not what you meant by your distiction between stabilization
and control could you please explain what you _did_ mean.

Thanks

Rick

[Martin Taylor 970107 12:00]

Rick Marken (970106.1430)

> Neither did I understand why you put forward your (subsequently
> retracted) claims that the environmental variable, rather
> than the perceptual variable, was what was controlled.

I never put forth the claim that the environmental variable _rather
than_ the perceptual variable is controlled, so I never retracted it.
I claimed that both the perceptual variable _and_ the environmental
correlate of that variable (the CV) are controlled by a control system.
I still claim it;-)

OK. As I said in the part you didn't quote, some writing is easily
misunderstood. And I didn't really _believe_ that you were excluding
control of the perceptual variable. It's just that your writing _seemed_
to say so. Just as you mistook my intention. But let's both try not to
do that in future, OK?

By the way, you haven't yet explained what you mean by "CV". Is it
"controlled variable"? If not, what is it? If it is, then why make
tautological statements about the controlled variable being controlled.
And if you mean "Controlled Complex Environmental Variable", why not
use the full acronym CCEV rather than the confusing CV?

Ye:

>Can you fairly say that a quantity z (==2x + y), which you measure,
>is being controlled when in fact what is being controlled is z'
>(== 2.01x + 0.98y + .0003w), which you don't measure? Both are
>stabilized, but only one is being controlled.

Me:

> What is controlled is controlled. What is not controlled is not
> controlled.

Ye:

> Under what version of the Test could you distinguish the above two
> variables?

The one where you apply a disturbance that will have a predicted effect
if the variable is _not_ controlled and far less than the predicted
effect if it _is_ controlled. It seems to me that you could distinguish
control of z from control of z' by varying w. This disturbance will be
completely effective if z is controlled but not
if z' is controlled. This seems pretty obvious to me; am I missing
something?

Yes. My presumption, which I had thought clear from the presentation, was
that "w" was a variable not perceived by the experimenter. It was supposed
to be a variable that might make small changes to the value of z' but that
the experimenter did not consider. Such as the possibility that the
fluctuations in the earth's magnetic field in solar storms might affect
control of the perception of colour (choose your own absurdity). Changes
in "w" would make the measurements different on different occasions when
the experimenter-induced disturbance was the same. Some experimenters
might see measurement noise, some might see prolonged effects of prior
patterns of disturbance, some might seek out ways in which the experimenter's
model of the subject failed. The last kind might eventually happen on
the notion that the subject's perception incorporates "w". But the Test
would not show it, until the experimenter happened on "w", and used it
to differentiate between z and z'.

The question works quite apart from what "w" represents. Suppose the
experimenter applies the Test on the assumption that z (==2x + y) is
controlled when in fact it isn't, but z' (==2.01x + 0.99y) is (forget w
since it confused you). You say that the Test is done as follows:

... you apply a disturbance that will have a predicted effect
if the variable is _not_ controlled and far less than the predicted
effect if it _is_ controlled.

The experimenter will find that the variable z _is_ controlled, for all
disturbances that might be tried. But we know that it isn't. A correlated
variable z' is actually being controlled. The controlled variable is what is
defined by the perceptual function "2.01x + 0.99y". The Tester observes the
_environmental_ variable "2x + y", which is different. But 2x + y is
stabilized, very nearly as closely as is the true controlled variable.
It's stabilized very well _by the control system_ but it is not the
controlled variable.

> My whole point in introducing the word "stabilization" into the
> discussion was to enable the discussion to procede on a factual
> basis, rather than on the theoretical faith that stabilization
> without control is impossible.

I don't understand this. If stabilization without control is impossible
then why make a distinction between stabilization and control?

Because some people assert that it is not impossible, and if you are to
argue with them, you have to have the words to do so.

I thought that the distinction between
stabilization and control was being made in order to characterize the
difference between the behavior of a possible controlled variable (like
the position of a cursor) in a simulation-based and in a real control
system.

That's one context, sure. It's been the focus of recent discussion, in which
the counter-argument has not been that the simulation-based whatever-it-is
fails to stabilize the observed environmental variable. The counter-argument
has been "It's not control, because control is control of perception." Bad
argument in a good cause. The argument is better that control stabilizes
under a far greater range of conditions than does simulation-based outflow
action, and the conditions under which simulation-based outflow action
stabilizes are unlikely to occur in real life (note: I'not taking sides
in this argument; merely showing how wording matters in stating the argument).

> If you use your _words_ to deny the possibility of alternate
> interpretations of observations, you make science very difficult.

I don't understand this. I thought stabilization and control refer
to two different phenomena. Don't they? Stabilization can _look
like_ control; it can vary less than expected based on an analysis
of the causal influences on it. But we can tell whether what we are
seeing is a stabilized or controlled variable by doing The Test.

You are making a statement that makes sense if and only if you use
different words for the stabilization of an observed variable and the
control of that variable. So now, having used the words appropriately
to make an argument, check back to see whether you understand why the
argument would have been impossible if you insisted on using
"control" for both situations, eliminating the possibility of using
"stabilize" or any of its synonyms. I think you will find you _do_
understand.

Martin

[From Rick Marken (970107.1300)]

Martin Taylor (970107 12:00) --

As I said in the part you didn't quote, some writing is easily
misunderstood.

I'd say _all_ writing is easily misunderstood. We do the best we can, which
is usually surprisingly good, I think.

By the way, you haven't yet explained what you mean by "CV".

I think of CV as the controlled variable (q) in the disgram of a
control system. Verbally, I would call it "the environmental correlate
of the perceptual signal" but I understand the problems with saying it that
way. Maybe a better technical definition of CV is "an observer's
representation of the perception controlled by the control system".

This whole discussion came up in the context of Bruce's model of compensatory
tracking with a sinusoidal disturbance. In that situation the CV seems rather
unambiguous; it is the position of the cursor on
the display. The model (like the real controller) controls a perception of
this CV. But it is rather easy to specify the CV as the position of the
cursor -- the environmental correlate of the controlled percpetion.

My presumption, which I had thought clear from the presentation, was
that "w" was a variable not perceived by the experimenter.

Then it would, indeed, be difficult for the experimenter to determine that
the controlled variable is a function of w. But the experimenter would be
able to tell that all his guesses about the controlled variable (since they
don't include w) are wrong, becuase, as you say:

Changes in "w" would make the measurements [of z'] different on
different occasions...

You go on:

Some experimenters might see measurement noise, some might see
prolonged effects of prior patterns of disturbance, some might
seek out ways in which the experimenter's model of the subject
failed.

Only the last experimenter is a PCTer. This experimenter knows that
s/he has not yet determined what the subject is controlling. This
experimenter is the only one with any hope of discovering that the controlled
variable depends on w -- which you corrrectly note:

The last kind might eventually happen on the notion that the subject's
perception incorporates "w". But the Test would not show it, until
the experimenter happened on "w", and used it to differentiate between
z and z'.

Righto! Testing for controlled variables is hard work; you have to discover
what a person is _really_ controlling. Conventional psychological research is
a piece of cake by comnparison, All you have to do in conventinoal research
is find statistically significant relationships between variables.

The Tester observes the _environmental_ variable "2x + y", which is
different [from the actual controlled variable: 2.01x + 0.99y].
But 2x + y is stabilized, very nearly as closely as is the true
controlled variable.

This is a question of level of accuracy. It's true that 2.01x + 0.99y isn't
_really_ what is controlled but it's awfully closed to the controlled
variable. If 2.01x + 0.99y were really just stabilized (like the CV in
Bruce's model) then the level of control would go to zero (as measured in
terms of the "Stability measure" I use in the Nature of Control Demo).

Me:

I thought stabilization and control refer to two different phenomena.

Martin:

You are making a statement that makes sense if and only if you use
different words for the stabilization of an observed variable and the
control of that variable.So now, having used the words appropriately
to make an argument, check back to see whether you understand why the
argument would have been impossible if you insisted on using
"control" for both situations, eliminating the possibility of using
"stabilize" or any of its synonyms. I think you will find you _do_
understand.

So these words (stabilize and control) _do_ refer to two different
phenomena? The CV in a simulation based control system may be
stabilized (if the system is lucky enoough to have computed the right
outputs) but it is not controlled. The CV in a control system is controlled.
Do I understand correctly now?

Best

Rick