Information, alerting, levels

[From Rick Marken (950216.1200)]

I said:

They ought to have been "incontrovertible" [demonstrations that information
comes through the perceptual function], perhaps, but they were not; Bill
and I controverted them.

Martin Taylor (950215 11:10) reples:

No you didn't. You sidestepped them without dealing with the issues.

I assume that you are using the plural "you". If there was any "sidestepping"
done, it was done by Bill Powers, Tom Bourbon AND myself. Here, for example,
is a recent sidestep from Bill Powers (950215.1240 MST):

nobody knows how to analyze the information situation in a closed-loop
system (at least I've never seen it done), so it's a moot point anyway.

[Incidentally, Tom Bourbon has lost his connection to the Net so he can't
"sidestep" in person; but he told me to tell you that he was as convinced by
your "incontrovertible" proofs about information in perception as I was;
maybe even less so, if that's possible;-)]

The great debate about "information in perception" was, from my perspective,
about the fact that lineal, cause-effect thinking doesn't apply when dealing
with closed-loop systems. What started the debate( as I recall) was the
observation that the output of a high gain control system almost perfectly
counters the effects of disturbances to a controlled variable even though the
effects of these disturbances cannot be perceived by the control system
itself. I believe it was at this point that Martin introduced the claim that
there was information about such disturbance effects (or the disturbing
variables that produce them -- Martin was not clear at first) in the
perceptual signal.

Martin proceeded to go though some arcane mathmatical gymnastics to try to
show that lags between the effects of disturbance and outputs on perception
made it possible to detect the disturbance in the perceptual signal. Martin
was obviously making this stuff up as he went along in order to protect (for
himself) the notion that perception contains information regarding how to
respond in order to accomplish intended results -- the lineal cause-effect
approach to explaining control.

While Martin may have managed to prove to himself that there is information
that (as he now says) "comes through the perceptual function", he was not
able to prove it to Tom, Bill, myself or anyone else who has understood (or
cared about) what was going on in that discussion. Martin has managed (and
will undoubtedly continue to manage) to control his imagined version of
control system operation despite the fact that it is completely inconsistent
with the way control systems actually operate. Mazel tov;-)

While I no longer hope (or care) whether Martin ever abandons the idea that
perception "informs" a control system, I think some very useful points were
made in the discussion of information in control systems. Here are a few
that come to mind:

1) Since p = f(o + d) (perceptual signal is a function of the sum of output
and disturbance) AT ALL TIMES in a simple control task, there is no way a
control system can determine, at any instant, the extent of the contribution
of d and o to p; so there is no information about d (or o) in p.

2) The outputs of a control system depend on r-p, not p, which is why Bill
Powers (950215.1240 MST) just said:

I would say that it is through the perceptual function and the comparator
that the information needed for control is obtained.

Of course, what comes though the comparator is the error signal -- nothing
else. So the error signal could be called the "information needed for
control". There is no need for a control system to have any information in
its perceptual signal about the variables causing that signal because the
outputs of a control system are determined (in a closed loop) by the error
signal, not the perceptual signal.

3) There is nothing in the time variations in p that can be used as a basis
for separating out the effects of o and d on p. We knew this from the
math but it was also demonstated experimentally in three ways:

    a) Different time variations in p result in the same output variations
when disturbance variations are the same; this is the experimental result
replicated by Chuck Tucker and reported by him in a recent post to the Net.

    b) There is hardly any correlation between p and d (or between
derivatives of p and derivatives of d) at any lag between these variable.
Martin had claimed (with no data, once again; apparently just to justify
his preconceptions) that variations in p or variations in the derivative of p
could be used to infer d; they cannot.

    c) No one has been able to reconstruct variations in d from variations in
p; this is what the control system would have to do if it were using
information in p about d as the means for generating o. [Martin WAS able to
reconstruct variations in d from variations in p when he was given
information about ALL parameters and functions of a control system AND it's
environment - - information that the control system itself could not possibly
have. This is the evidence that there is information about the disturbance in
perception that Martin says I refuse to accept. Of course, Tom B. and Bill P.
refuse to accept this evidence too because it is not evidence at all. Martin
has shown that, given values of p and r and k it is possible to solve for o
in an equation like o = k(r-p); in other words, Martin has provided evidence
that algebra words. If Martin had sent back variations in d when sent nothing
other than variations in p (which is what a control system does) it would
have been STRONG evidence that there is information "coming through the
perceptual function" of a control system; but he didn't.

Bill Powers (950215.1240 MST) answered Martin's points about "alerting" and
the "goals of reorganization". I want to echo Bill's suggestion to Martin
that:

Maybe you should offer an example that you think PCT can't handle without
the "alerting" hypothesis.

But, of course, that would involve presenting (yuchhh) data and I know that
Martin prefers to reason about the horse's mouth (based on "first principles")
rather than look in it.

I said:

The next step in going up a level would be to figure out WHY you have a
control system operating with the goal of keeping the cursor on the target.

Bill Powers (950215.1240 MST) --

This can too easily be taken to mean "think up a logical rationalization
or hypothesis." When you're really doing the Method, you don't have to
look for the higher-level process: it's already there and working.

Yes. What I said above is not only misleading; it is wrong. It was based on
my own misunderstanding of the method of levels. Your comments have helped me
understand the method better.

In fact, when I do the method of levels I never really figure out "why" I
want something (like "getting Martin to see things my way" -- possibly a bad
example;-)) That would be rationalization -- controlling perceptions in order
to "make sense" of other controlling I am doing.

I would describe what I am doing when I "go up a level" as something like
"dissociation"; I become conscious of something I am doing as though "someone
else" were doing it.

It's a lot easier for me to "dissociate" myself in this way from a behavior
like the typing I am doing with my hands than from a behavior like the
"convincing people that PCT is right" that I am also doing with my hands (at
the same time). Both these behaviors are examples of controlling and I can
become conscious of both as though they were being done by "someone else"; I
can dissociate myself from them. This is what I think of as "going up a
level". Does it seem that what I'm doing when I "go up a level" is the same
as what you're doing?

Best

Rick

[Martin Taylor 950216 15:05]

Rick Marken (950216.1200)

Rick, in his normal fashion, completely misstates the facts about the
use-of-information-in-control-systems discussion. Let's start with...

Martin introduced the claim that
there was information about such disturbance effects (or the disturbing
variables that produce them -- Martin was not clear at first) in the
perceptual signal.

Throughout the discussion, I tried to make clear that there is information
IN the perceptual signal ONLY in respect of that part of the disturbance
that is NOT controlled. To the extent that there is control, as I have
throughout gone to great pains to point out, the information is not IN the
perceptual signal. I used at one early point the metaphor of the information
"leaking out of" the perceptual signal so that it was no longer detectable
in the perceptual signal. The information that has been used in control
is not IN the perceptual signal. How could it be? From Day 1 of the
discussion, Rick has persisted in repeating and repeating that I claim
the opposite, and then delights in showing that what he says I claim is
untrue. I agree that the opposite of my claim is indeed untrue.

The claim is that it is through the perceptual signal that the information
for control is passed. Control is not magical, but physical--that's another
way of making the same claim.

Martin proceeded to go though some arcane mathmatical gymnastics to try to
show that lags between the effects of disturbance and outputs on perception
made it possible to detect the disturbance in the perceptual signal.

Utter nonsense. What lags do is quite different. They ensure that control
cannot be perfect. However, to the extent that control is not perfect, there
remains some remanent effect of the disturbance IN the perceptual signal,
and that's a fact demonstrated by any amount of data contributed by Marken,
Powers, and Bourbon. They show consistent correlations, and then say that
the correlations are so small that they might as well be zero, which then
is used as proof that they ARE zero. The better the control is, the closer
to zero those correlations should be--the less information about the
disturbance remains IN the perceptual signal at any moment. (The "arcane
mathematical gymnastics," by the way, were simply the normal application
of Laplace transforms to an arbitrary linear control system. I ccan't help
it if Rick is amathetic.)

Martin
was obviously making this stuff up as he went along in order to protect (for
himself) the notion that perception contains information regarding how to
respond in order to accomplish intended results -- the lineal cause-effect
approach to explaining control.

What's the comparative form of "Utter"? I might have made "this stuff" up
as I went along, in third-year undergraduate control theory, but not now.
I'd almost forgotten it over 40 years of disuse.

1) Since p = f(o + d) (perceptual signal is a function of the sum of output
and disturbance) AT ALL TIMES in a simple control task, there is no way a
control system can determine, at any instant, the extent of the contribution
of d and o to p;

Correct. Well understood.

so there is no information about d (or o) in p.

Totally wrong. It's that "so" that has been at the heart of what I have
forever failed to get across to Rick--that "information" is the reduction
of uncertainty, not its elimination. Besides which, there's that "in p"
again.

There is no need for a control system to have any information in
its perceptual signal about the variables causing that signal because the
outputs of a control system are determined (in a closed loop) by the error
signal, not the perceptual signal.

Nobody on this side ever talked about the perceptual signal having information
about the VARIABLES causing the signal (the disturbing variable and the
output). However, Rick (and at times Bill P. and Tom B.) all imputed that
claim and then refuted it, in the usual way of sidestepping the issue by
choosing an issue on which the answer was to his (their) liking.

3) There is nothing in the time variations in p that can be used as a basis
for separating out the effects of o and d on p. We knew this from the
math but it was also demonstated experimentally in three ways:

Agreed. Well known. Irrelevant.

   b) There is hardly any correlation between p and d (or between
derivatives of p and derivatives of d) at any lag between these variable.

Agreed.

Martin had claimed (with no data, once again; apparently just to justify
his preconceptions) that variations in p or variations in the derivative of p
could be used to infer d; they cannot.

Hang on here. The "incontrovertible demonstration" was precisely the use of
data to show that they can, PROVIDED that the output and feedback functions
are known. The information needed to specify those functions is fixed, AND
IS INHERENT IN THE ACTUAL STRUCTURE OF THE CONTROL LOOP, whereas the
information from the disturbance approaches infinity over infinite time. The
information required to specify a FIXED output and environmental function
becomes relatively speaking infinitesimal, though the information required
to specify a VARYING environmental function limits the precision of control,
as well as limiting the information recoverable about the disturbance from the
perceptual signal. They are the same thing.

Much of the argument in the earlier discussion was on whether it is "cheating"
to use information about the output function and the environmental function
in reconstituting the disturbance waveform. The point of the demonstration
was that this information is PART OF the control loop itself, and therefore
MUST be used in the act of control. To ignore it would be wrong in
demonstrating that the perceptual signal carries the information from the
disturbance to the output.

[Martin WAS able to
reconstruct variations in d from variations in p when he was given
information about ALL parameters and functions of a control system AND it's
environment - - information that the control system itself could not possibly
have.

There's the nub. I never talked about a system that is observing the control
system being controlled. Rick insists that I must do so. Only an observer
system needs to "know" the parameters and functions. They are IN the acting
control system and used by it inherently. They are information that the
control system itself obviously has. If the output function produced only a
set of random values, unrelated from moment to moment to the input values
and to each other, control would be pretty lousy, wouldn't it? Same if the
environment made the effect of the output on the CEV random from moment
to moment. No control possible. That would be the situation of a "control
system" that did not have the information that Rick says it "could not
possibly have." Absurd.

If Martin had sent back variations in d when sent nothing
other than variations in p (which is what a control system does)

No it doesn't.

it would
have been STRONG evidence that there is information "coming through the
perceptual function" of a control system; but he didn't.

For reasons that should have been obvious.

I want to echo Bill's suggestion to Martin that:

Maybe you should offer an example that you think PCT can't handle without
the "alerting" hypothesis.

But, of course, that would involve presenting (yuchhh) data and I know that
Martin prefers to reason about the horse's mouth (based on "first principles")
rather than look in it.

Further _ad hominem_ misrepresentation, but I don't mind. I know Rick is
defending reference levels for perceptions that don't correspond to the
outer world--lots of imagination in them--and that some day the evidence
will shine through. And I've answered Bill, who seldom misrepresents me,
though he often misunderstands what I try to say.

Rick, if you want another of my famous "truths" introduced into PCT, why
don't you take on The Bomb? You seem to be exemplifying the early parts
of its action:-)

I think that this particular reprise on the "information" thread has been
stretched far enough, and I propose to continue it only if I can see some
technical advance as being encouraged by doing so. I anticipate that
there will be at least one response to this message, and if it contains
matters of technical interest that have not previously been addressed, I
will respond. Otherwise I hope I will have enough restraint to let it lie.
"Yes you did"-"No I didn't" discussions are not very fruitful. But Rick
deserves an opportunity to respond.

Martin

<[Bill Leach 950217.18:19 EST(EDT)]

[Rick Marken (950216.1200)]

Oh how I know I'm gonna almost hate myself for stepping in on this one
but then I might end up with high levels of conflict if I don't too!

Rick;

While I absolutely agree with you, Bill, Tom, etc. concerning the
usefulness of the concept that there is information in the perceptual
signal about the disturbance, Martin IS NOT absolutely wrong in his
assertions.

Only an UNREAL control system -- the theoritically perfect control system
with infinate gain and zero response speed will the presence of
disturbances to the controlled variable be absolutely undetectable.

In any REAL control system, gain is less than one and therefore any
disturbance MUST result in a NON-ZERO net error output change (even if
response speed is still instantaneous). Since response speed is NOT
instantaneous then there will also be transient effects from any
disturbance. Naturally, the point that Martin did not emphasize (though
admitted to) is that also contained within this "information stream" are
the effects of "system drift", "noise" and reference changes.

Martin also (at least eventually) admitted that what he was talking about
was related to "measuring" system performance with external equipment
connected at the relevent points WITHIN the system being studied. Ie:
Perceptual signal, reference signal, error signal and output signal all
being measured. In the discussion it was of course admitted that while
possibly we currently have the equipment with the noise figure and gain
necessary to actually test the idea, we do not have the ability to even
find the connection points much less make the connections.

Are any of these absolute truths concerning ALL closed loop negative
feedback systems relevent to PCT? No, definately not! Will they ever
be? Maybe, but it is probably several generations too early to even
guess.

So, while from a practical standpoint AND from a PCT theoritical
standpoint: "INFORMATION ABOUT THE DISTURBANCE IN THE PERCEPTION IS A
NON-SEQUITR", BUT even I can't stand seeing Martin being "hit" about
something in which he is factually correct.

Additionally, Bill P. DID admit that "information about the disturbance
would be present in the perceptual function", saying as well that such
information is "in the noise" for the control system itself and is
otherwise of no use (at least currently) for PCT.

As to Martin "abandoning the idea that the perception 'informs' a control
system" or not, he did at one time admit that the control system itself
could not make use of such 'information.'

Naturally enough, this WAR has been raging for longer than I have been
around. When I first joined the battle, I too read Martin as claiming
that looking only at the perceptual function, one could determine
something about the disturbance. I am still not certain he was not
claiming just that -- at that time.

My perception (right or wrong) is that Martin is controlling for
disturbance (to others, especially Rick). It is again my perception,
that one of his favorite perceptions is to make a "flat statement of
fact" without presenting solid reasoning and then sit back and watch the
fur fly (stirring as necessary).

-bill