Control and The Test

CSGnet Archive CSG1996
1

[From Rick Marken (960213.1330)]

Martin Taylor (960213 12:30) --

Let's not argue about which piece of the elephant we ought to look at. The
animal is all of a piece.

Huh? I'm arguing about whether control is a real phenomenon (Hans seems to
say that it's not) and whether The Test can discriminate control from cause-
effect (you seem to think that it can't).

In reply to my question:

Where do you come up this this stuff?

You said:

From trying to deal with postings _as they are written_ rather than ignoring
what is written so as to impose my _a priori_ views about what they must say.
It's not too hard. You should try it some time

Here's what was written. Hans said:

The question "what controls what" is meaningless for me.

According to my understanding of control, "what controls what" is of the
essence; there is control when a controller (the subject "what") controls a
perceptual representation of an environmental variable (the object "what").
Saying that "what controls what" is meaningless sounds (to me) an awful lot
like saying "there is no such thing as control".

In response to my comments on Hans' post you said:

Hans presented a situation rather like a ball in an angel-food cake tin
with a sharp bottom. Any Test you want to apply would look pretty
much the same as would a Test on a "real" control system

The behavior of a ball in an angel-food cake tin is obviously cause-effect;
the ball doesn't control anything. So it sounded to me like you were saying
that any Test (which, I presume, includes The Test for the Controlled
Variable) could not discriminate the behavior of a cause-effect system (like
the ball) from that of a "real" control system. I think this is a pretty
reasonable interpretation of what you wrote; based on that interpretation I
said:

this is crap; The Test can always readily distinguish a relatively high gain
control system from a no gain cause-effect system.

To which you [Martin Taylor 960213 12:30] replied:

That wasn't Hans's problem, and neither was it mine in my posting, "as you
must know", having (presumably) read the postings in question.

You go on to say:

"The Test" has several component parts...

You then carry on for several paragraphs about The Test. But I cannot seem to
find in those paragraphs a nice, simple statement of what Hans' and your
problem is. I'm sure this is an intellectual failing on my part. So perhaps
you could just tell me, as _simply_ as possible, what your (and Hans')
problem is?

Frankly, I don't see any problem at all. Hans presented an example of
behavior (Brownian motion up a gradient) that might involve control. I
suggested that he test this notion by applying a continuously varying
disturbance to the putative controlled variable. Hans has all the equations
so it's a piece of cake to calculate the expected behavior of the controlled
variable with and without the disturbance present (random changes in the
concentration gradient would seem to be a reasonable pick for a disturbance
if average rate of travel toward a goal point is the putative controlled
variable).

This Test will (I predict) quickly reveal that neither the particles nor the
concentration nor the "emergent" result of the interaction of the two is
controlling anything at all.

Have a nice day

Rick

2

[Martin Taylor 960213 17:30]

Rick Marken (960213.1330)

I'm arguing about whether control is a real phenomenon (Hans seems to
say that it's not) and whether The Test can discriminate control from cause-
effect (you seem to think that it can't).

If you would read the postings rather than quoting little out-of-context
bits as evidence that you had done so, you wouldn't say things like that.

So perhaps
you could just tell me, as _simply_ as possible, what your (and Hans')
problem is?

I don't know about "as simply as possible" but as simply as I can, quickly...

  If you have an observation of a variable that is clearly being disturbed
  by some unobservable cause, but which nevertheless persists in returning
  to or close to its original position, how can you tell whether the variable
  is or is related to the CEV of some control system?

  Which leads to the second problem: If you have determined that a variable
  is probably correlated with the CEV of some control system, how can you
  determine what the actual controlled variable is?

Frankly, I don't see any problem at all.

No, we know that. Applying disturbances to what one would like to think
would be one of your controlled perceptions usually results in no compensating
output. Outputs generally seem to be autonomous disturbances to some quite
different variable.

Hans presented an example of
behavior (Brownian motion up a gradient) that might involve control. I
suggested that he test this notion by applying a continuously varying
disturbance to the putative controlled variable.

No he didn't. He (God) knows that the behaviour is "Brownian motion up
a gradient." He (experimenter) observes a particle whose position is
subject to random disturbances, but which nevertheless always
returns quickly to a position that lies (observably) on some sphere.
If he disturbs the particle in addition to the random disturbance to
which it is ordinarily subjected, it will be by pushing it off the sphere,
and he will see that it quickly returns to the sphere. If I remember
correctly, that's exactly what Hans proposed doing. He didn't need your
suggestion, unless perhaps you want to emphasize "continuously varying."

Hans has all the equations
so it's a piece of cake to calculate the expected behavior of the controlled
variable with and without the disturbance present

He already did that, in his God persona, and posted a message about the
results. Maybe not numerically, but quite enough to show what happens.
Hans (experimenter) has no such equations, so he can't do as you suggest.

(random changes in the
concentration gradient would seem to be a reasonable pick for a disturbance
if average rate of travel toward a goal point is the putative controlled
variable).

But why on earth would Hans-experimenter even contemplate chemical
concentration as being related to the situation? He can't see the chemistry.
He's only looking at a little dancing particle that seems to know where
it wants to stay--despite all the natural torment that it seems to be in.

If average rate of travel toward a goal point is the putative controlled
variable, surely _that_ is what he would disturb? He'd probably blow or
poke the particle, or subject it to a magnetic field, or something. It
would resist all those methods of moving it away from its preferred position.

If I want to test whether you are controlling for being at your campsite
(assuming you to be wilderness trekking), do I first try starving you so
that you can't walk? That seems to be the equivalent of changing the
chemical concentration gradient as a disturbance. It's not a disturbance
to the putative control system Hans described. It's a change to the
characteristics of the feedback loop.

This Test will (I predict) quickly reveal that neither the particles nor the
concentration nor the "emergent" result of the interaction of the two is
controlling anything at all.

Let's walk it through: (1) apply disturbance by pushing particle away from
where it seems to "want" to be. Result: pass. (2) Look for mechanism of
movement--disturbance and output. Result: can't find it (remember, our
experimenter thinks Brownian motion is a mystery, Einstein not yet having
been born). Conclusion, ambiguous--neither pass nor fail. (3) Look for
mechanism of perception of location. Result: can't find one, also ambiguous.

Where do we go from there? The Test has quickly revealed that the system
is either an equilibrium system or a control system, but it has told us
no more than that. If our experimenter intuits that the particle might be
sensing chemical concentration (as our eyes do when the rods and cones
are hit by photons when we look for the locations of things), then part
(3) of the Test is passed, and furthermore, there is a part of a passing
answer to (2), in that there is a distinct _mechanical_ response to a change
of concentration, quite like the bulking of a muscle. What is still lacking is
that the experimenter can see no way that this "pseudo-muscle" influences
the position of the particle, which is the environmental variable that
he is worrying about.

Do you still hold to your prediction that the Test will "quickly reveal"
that nothing is "controlling anything at all?" That there is no control
may be true, but is it so readily tested if you are not the God who set
up the situation?

I think Hans's example leads to a lot of possibilities for thought, which
is good. It is unfortunate that you "don't see any problem at all", because
it deprives you of those possibilities.

Have a nice day

Too cold here. Much easier for people in LaLa Land. But apart from that,
it's been pretty nice, thank you.

Now I have to go curling. Until tomorrow, then.

Martin

3

[From Rick Marken (960213.1812)]

Martin Taylor (960213 17:30) --

If you would read the postings rather than quoting little out-of-context
bits as evidence that you had done so, you wouldn't say things like that.

Believe it or not, I really do read the postings. You and Hans are just
way over my head.

But thanks for giving me a clear and concise statement of your problem:

If you have an observation of a variable that is clearly being disturbed by
some unobservable cause, but which nevertheless persists in returning to or
close to its original position, how can you tell whether the variable
is or is related to the CEV of some control system?

The solution to this problem is actually quite easy: you do The Test to
see if the variable is under control. There is no other way.

Which leads to the second problem: If you have determined that a variable
is probably correlated with the CEV of some control system, how can you
determine what the actual controlled variable is?

The solution to this problem is the same as the solution to the first: you
keep doing The Test to see if you can come up with a better definition of
the controlled variable.

Really, Martin, these "problems" are problems only for people who don't
want to get off their duffs and Test to determine whether a variable is
under control or not.

Do you still hold to your prediction that the Test will "quickly reveal"
that nothing is "controlling anything at all?"

Sure. Though it's possible that the computer simulated particles in
Hans' Brownian motion study are organized (in code) as control systems,
in which case we will discover that some variable is under control.
The important point is : IF something is controlled in the Brownian
motion simulation, The Test will reveal this fact unambigously. If there
is no control in the simulation, this will also be revealed unambigously
by The Test. There is no special problem either if the Brownian motion
simulation is an equilibrium system; equilibrium systems are just cause-
effect systems and they are readily exposed as such by The Test. No
need to worry yourself about it.

I think you are imagining problems for The Test that simply don't exist.
The only way an equilibrium system (like a pendulum) and a control system
(like a person) could possibly be confused is if you only use large
transient disturbances to the putative controlled variable and watch
only for a return of the variable to the original state after disturbance.

When you do The Test, you apply a _continuously varying_ disturbance
to the putative controlled variable and watch for _lack of effect_
of this variation on the controlled variable. Try this Test with a
pendulum or a marble on the side of a bowl or a mass on a spring
and see if you have ANY trouble seeing INSTANTLY that the position
of the bob, the marble and the mass (or any other variables involved
in these systms) are NOT controlled. Then try it with a person who
is asked to hold her arm at her side and see if you have ANY trouble
seeing (feeling, too) that the position of the arm is under control.

Oh, and it was 78 degrees and clear as a bell here today;-)

Have a nice day

Rick

4

[Hans Blom, 960214g]

(Rick Marken (960213.1812))

Martin:

If you have an observation of a variable that is clearly being
disturbed by some unobservable cause, but which nevertheless
persists in returning to or close to its original position, how can
you tell whether the variable is or is related to the CEV of some
control system?

The solution to this problem is actually quite easy: you do The Test
to see if the variable is under control. There is no other way.

Wouldn't you have to compare with how the variable would vary if it
were NOT under control? And, not being able to create nor _perceive_
such presumably uncontrolled behavior, don't you need some theory
before you can generate a prediction of what it would look like? And
where does that theory come from? And how do you know whether it's
correct?

Really, Martin, these "problems" are problems only for people who
don't want to get off their duffs and Test to determine whether a
variable is under control or not.

DOING a test is no difficult. Thinking of a good one is...

Greetings,

Hans

5

[Martin Taylor 960215 11:00]

Rick Marken (960213.1812)

+Bill Powers (960214.0100 MST)

Rick, may I introduce you to Bill Powers? Bill, may I introduce you to
Rick Marken? You really should meet some day.

[Rick]

If you have an observation of a variable that is clearly being disturbed by
some unobservable cause, but which nevertheless persists in returning to or
close to its original position, how can you tell whether the variable
is or is related to the CEV of some control system?

The solution to this problem is actually quite easy: you do The Test to
see if the variable is under control. There is no other way.
...
Really, Martin, these "problems" are problems only for people who don't
want to get off their duffs and Test to determine whether a variable is
under control or not.

[Bill]
+Martin Taylor (960213 17:30) --

ยทยทยท

+
+In answering Rick Marken you said
+
+ If you have an observation of a variable that is clearly being
+ disturbed by some unobservable cause, but which nevertheless
+ persists in returning to or close to its original position, how can
+ you tell whether the variable is or is related to the CEV of some
+ control system?
+
+If the cause of the disturbance is unobservable, you can't do the Test.

I don't know that I agree with either of you, but you do seem to represent
opposite ends of the spectrum of possible answers.

Rick, instead of grandly stating that it's simple, and since you _do_ read
other people's postings, how about commenting on the results of the stages
of applying the Test to Hans's Brownish Particle, as I wrote them? If
you don't see any problem, that should be an easy task. Perhaps by trying
it, you might see where the problem lies.

Bill, are you asserting that it you see something being apparently
disturbed and then returning to its original value, you cannot do the Test
(stage 1) by disturbing it yourself to see whether it resists?

Wasn't Hans proposing to do just that--noting that the particle moved
away from and then back to it's equilibrium distance from the origin, he
would himself disturb it in like manner to see whether it would resist
the disturbances he introduced as well as the ones apparently influencing
the particle naturally.

Also, I once again remind everyone that the Test does not consist only
of the application of disturbances and observations of perturbations.
The lack of an effect of an applied disturbance could be due to an
incomplete understanding of the physical situation, or some other system
beside the one you have in mind might be doing the controlling. To
complete the Test, you have to verify that the appropriate input and
output connections are present, so you know which system is doing the
controlling and what the input and output pathways are.

That issue seems to me to be what lies at the heart of Hans's thought
experiment, and is why I believe that the situation provides an interesting
opportunity for investigating the nature of how one observes and deals
with control systems at the border between obvious control and obvious
equilibrium.

Martin

6

[From Rick Marken (960215.0950)]

Martin Taylor (960215 11:00) points out that, in reply to his statement:

If you have an observation of a variable that is clearly being disturbed by
some unobservable cause, but which nevertheless persists in returning to or
close to its original position, how can you tell whether the variable
is or is related to the CEV of some control system?

I said:

The solution to this problem is actually quite easy: you do The Test to
see if the variable is under control. There is no other way.

and Bill said:

If the cause of the disturbance is unobservable, you can't do the Test.

Martin seems to think that Bill and I are contradicting each other:

I don't know that I agree with either of you, but you do seem to represent
opposite ends of the spectrum of possible answers.

Not really. I was taking Bill's point for granted; if you can't observe the
cause of disturbance (what we have been calling the disturbance itself,
which is a _potential_ cause of perturbation), you can't do the Test, which
would involve looking for lack of effect of disturbance (lack of a resulting
perturbation) on the putative controlled variable. Since we can't do the Test
with what, according to you, has been observed (simply the behavior of a
putative controlled variable) I suggested that you DO the Test, which would
involve introducing a known disturbance and monitoring the relationship
between that disturbance and the behavior of the putative controlled
variable.

So there is no contradiction. Bill pointed out that the case you describe
does not allow one to Test for a controlled variable. I said that one can
solve this problem by doing the Test for the controlled variable. Do you
still see these statements as being at "pposite ends of the spectrum"?

Rick, instead of grandly stating that it's simple, and since you _do_ read
other people's postings, how about commenting on the results of the stages
of applying the Test to Hans's Brownish Particle, as I wrote them?

I already did, in my (exceptionally clear, if I do say so myself) post
called "Hans' Amazing Test" (960214.0845). If you read that post carefully
you will see that your problems with the Test stem from concentrating only on
what happens to a putative controlled variable _after_ a disturbance (a
_potential_ cause of perturbation) has been _removed_. When you realize that
the Test is about how variables behave _while_ they are being disturbed, you
will see why it is so easy to Test to see if any variable is being controlled
in Hans' Brownian motion experiment.

Best

Rick

7

[Martin Taylor 960215 16:45]

Rick Marken (960215.0950)

So there is no contradiction. Bill pointed out that the case you describe
does not allow one to Test for a controlled variable. I said that one can
solve this problem by doing the Test for the controlled variable. Do you
still see these statements as being at "pposite ends of the spectrum"?

Yes (despite reading the rest of your posting).

But you don't, so we obviously have a different understanding of one of the
two messages. I admit to being confused as to the agreement between Bill's
pointing out that one cannot do the Test for the controlled variable in
this situation and your saying that one can solve the problem by doing the
Test for the controlled variable, but I'll take your word for the fact that
these statements are not contradictory. It's a religious fact that what
Marken and Powers say cannot be in contradiction, by definition. But no
matter. I don't care much whether you agree with each other. I'd just like
to understand what is going on.

Let's get to the _really_ simple stuff.

Me:

... how about commenting on the results of the stages
of applying the Test to Hans's Brownish Particle, as I wrote them?

Rick:

I already did, in my (exceptionally clear, if I do say so myself) post
called "Hans' Amazing Test" (960214.0845).

Just to humour my obviously deficient brain, I'll repeat the walk-through,
with spaces left for you to insert what you wrote in that posting, or
elsewhere, that deals with each phase of the Test about which I asked
(a paraphrase, or even a totally new statement, will do). I do not remember
reading the answer to any of them, not a single one:

Let's walk it through: (1) apply disturbance by pushing particle away from
where it seems to "want" to be. Result: pass.

Rick's answer 1: ...?
        (By the way, in your answer, notice that the word is "pushing",
not "displacing" or "picking up and putting down" or any other discrete
variant. Pushing happens while it happens, and varies at the experimenter's
will.)

(2) Look for mechanism of
movement--disturbance and output. Result: can't find it (remember, our
experimenter thinks Brownian motion is a mystery, Einstein not yet having
been born). Conclusion, ambiguous--neither pass nor fail.

Rick's answer 2: ...?

(3) Look for

mechanism of perception of location. Result: can't find one, also ambiguous.

Rick's answer 3: ...?

Where do we go from there? The Test has quickly revealed that the system
is either an equilibrium system or a control system, but it has told us
no more than that.

Rick's answer 4: ...?

If our experimenter intuits that the particle might be
sensing chemical concentration (as our eyes do when the rods and cones
are hit by photons when we look for the locations of things), then part
(3) of the Test is passed, and furthermore, there is a part of a passing
answer to (2), in that there is a distinct _mechanical_ response to a change
of concentration, quite like the bulking of a muscle.

Rick's answer 5: ...?

If you read that post carefully
you will see that your problems with the Test stem from concentrating only on
what happens to a putative controlled variable _after_ a disturbance (a
_potential_ cause of perturbation) has been _removed_. When you realize that
the Test is about how variables behave _while_ they are being disturbed, you
will see why it is so easy to Test to see if any variable is being controlled
in Hans' Brownian motion experiment.

If you read my questions repeated above, you will see that at no place was
there any suggestion of what happens after a disturbance has been removed.
When you realize that we _know_ that the Test is about how variables behave
while _or after_ they are being disturbed, you will see why it should be easy
to respond to messages _as they are written_ rather than responding to
fantasies about what you would like them to be so you can knock down the
straw men you would like to see therein described.

Now, how about responding to the phases of the Test. Would they be as I
suggest, or if not, how would they go? There are 5 distinct questions, to
which my stupid brain would like 5 distinct answers.

Is that too much to ask?

Martin