Talkin' Perceptual Signal Blues

[Richard_Marken]

[From Rick Marken (960620.1430)]

Me:

Look, if there is information about the disturbing variable (or variables) in
the perceptual signal then it should be possible to reconstruct (as the
control system itself presumably does) the disturbing variable given ONLY the
perceptual signal -- period, amen.

Martin Taylor (960620 12:15)

This comment shows that you don't understand what it means for there to
be information about something in some observation.

Oh, really!?!?

Do you claim that you get no information about the state of an LA Freeway
when you hear "there's a 4-mile backup on the Santa Monica"?

Well, now that you mention it...

But I think this analogy is perfect. I had always thought the idea that there
is information in the perceptual signal about the disturbance is exactly
equivalent to the idea that there is information in a statement ( like
"there's a 4-mile backup on the Santa Monica") about some state of affairs
in the world (the state of the freeway). I get information about the state of
the world (the freeway) from the signal ("there's a 4-mile backup on the
Santa Monica"). The signal tells me which of the many possible states of
the world (zero to N miles of backup) actually obtains. I get information
about the 4- mile backup from the signal alone; I need no other signals or
variables.

Simlarly, if there is information about the state of the disturbance in the
perceptual signal, the control system should be able to get that information
from the perceptual signal itself; it should need no other signals or
variables. The perceptual signal should be saying things to the control
system like "the disturbance is pushing with 10 units of force from the
left", "now it's pushing with only 5 units of force from the left", "now it's
not pushing at all", "now it's pushing from the right with about 3 units of
force -- no, make that 4", etc.

Of course, the perceptual signal must communincate this information to the
control system in some other code than English words; but that's what I was
hoping you could tell us; how does the perceptual signal inform the control
system about the state of the disturbing variable -- or the effect that the
disturbing variable is having on the controlled variable. I thought you knew
that there was information about the disturbance in the perceptual signal
because you could understand the perceptual signal's "reports" about the
disturbance as well as I can understand traffic reports about the state of
the freeway.

But, instead of just explaining how the perceptual signal informs (reports
to) the control system about the state of the disturbance and/or showing how
to extract the information (report) about the disturbance (and, hence, an
approximation of the disturbance) from the perceptual signal, you started on
this program of sophistry and obfuscation that basically amounted to "if a
control system can control then there MUST be information in the perceptual
signal". I suppose that you also believe that, since people were created
there must have been a god to create them. Sorry, I'm just not smart enough
to buy this argument.

In fact, there is no information about disturbances or the net effect
disturbances in the perceptual signal. The perceptual signal doesn't talk to
the control system; it doesn't give disturbance reports to the control
system. The control system doesn't "base" its outputs on anything other than
the discrepancy between perception and reference signal -- a discrepancy that
depends, in part, on the outputs the system is generating based on that
discrepency. It's a closed loop system, remember?

Best

Rick

[_Martin_Taylor]

[Martin Taylor 960621 15:00]

Rick Marken (960620.1430)

that's what I was
hoping you could tell us; how does the perceptual signal inform the control
system about the state of the disturbing variable -- or the effect that the
disturbing variable is having on the controlled variable.

That's what I and Allan did tell you--the magical mystery function, a.k.a.
the output/feedback function.

But, instead of just explaining how the perceptual signal informs (reports
to) the control system about the state of the disturbance and/or showing how
to extract the information (report) about the disturbance (and, hence, an
approximation of the disturbance) from the perceptual signal, you started on
this program of sophistry and obfuscation that basically amounted to "if a
control system can control then there MUST be information in the perceptual
signal".

Sorry. I had thought that a system diagram description, an algebraic
description, and a simulation that showed how the perceptual system
"informs" the control system would have been considered an explanation.
Those forms did follow extensive prior verbal discussions that had been
said to be inadequate. I really don't know of any other ways to approach
an explanation than these four. If you can think of another--not verbal,
not graphical, not algebraic, and not simulation, then let me know and I'll
try it.

I had neither then nor now any intention of using sophistry or obfuscation.
Usually on CSGnet, simulations that demonstrate a point are considered to
be helpful in clarifying matters. I'm sorry that you find on this issue
a clear demonstration is tantamount to sophistry and obfuscation.

In fact, there is no information about disturbances or the net effect
disturbances in the perceptual signal. The perceptual signal doesn't talk to
the control system; it doesn't give disturbance reports to the control
system. The control system doesn't "base" its outputs on anything other than
the discrepancy between perception and reference signal

The first sentence is an expression of a strongly held opinion, not of a fact.

The second sentence contradicts the usual circuit diagram of a control unit,
in which the perceptual signal is distinctly a part of the control circuit,
"talking" to the comparator, which is also part of the control system.

The third sentence contradicts the second.

-- a discrepancy that
depends, in part, on the outputs the system is generating based on that
discrepency. It's a closed loop system, remember?

And that's why it works, remember? The whole demonstration would fail, and
the magical mystery function would not output anything close to the
disturbing influence if it were not a closed loop syste, and one with
high gain at that. Remember?

Martin

[Bruce_Gregory9]

[From Bruce Gregory (960621.1540 EDT)]

(Martin Taylor 960621 15:00)

(Rick Marken 960620.1430)

>In fact, there is no information about disturbances or the net effect
>disturbances in the perceptual signal. The perceptual signal doesn't talk to
>the control system; it doesn't give disturbance reports to the control
>system. The control system doesn't "base" its outputs on anything other than
>the discrepancy between perception and reference signal

The first sentence is an expression of a strongly held opinion, not of a fact.

The second sentence contradicts the usual circuit diagram of a control unit,
in which the perceptual signal is distinctly a part of the control circuit,
"talking" to the comparator, which is also part of the control system.

The third sentence contradicts the second.

Martin's comments reflect my own confusion. It is very
difficult for me as an ignorant bystander to figure out what the
hell is going on here. Clearly monumental issues are at stake,
but for the life of me, I can't figure out what they are. I
doubt anyone can clear things up, since no one seems to
understand what anyone else is saying. It's a good thing we are
all in fundamental agreement, because I would be appalled to see
what would happen if real differences existed.

Bruce G.

[Richard_Marken]

[From Rick Marken (960921.1400)]

Me:

I was hoping you could tell us; how does the perceptual signal inform the
control system about the state of the disturbing variable

Martin Taylor (960621 15:00) --

That's what I and Allan did tell you--the magical mystery function, a.k.a.
the output/feedback function.

The control system itself doesn't know the output or feedback function. All
it knows is the perceptual signal. I'll send you a perceptual signal and you
send me back the disturbance. If you can't do that, then cut the crap about
information in perception about disturbances.

Sorry. I had thought that a system diagram description, an algebraic
description, and a simulation that showed how the perceptual system
"informs" the control system would have been considered an explanation.

Nope. It just shows a control model. It doesn't in any way show how the
perceptual system informs the control system. What it does show is the
lengths to which you will go to make believe that information theory has
anything to do with control theory.

Since there is information in a radio signal about freeway traffic, if I sent
you a traffic report you could send me back a description of the traffic. If
there is information about disturbances in the perceptual signal, then if I
send you a disturbance report (perceptual signal) you should be able to send
me back a description of the disturbance. It's as simple as that. So here is
a disturbance report:

13,10,10,10,10,11,12,13,12,13,13,13,13,13,14,15,15,14,12,9

If there is, indeed, information about the disturbance in this signal, then
you should be able to send me the list of numbers that represent the
disturbance that this percpetual signal is a report on. The disturbance
values, please.

I had neither then nor now any intention of using sophistry or obfuscation.

I know. The sophistry and obfuscation are unintended side effects of your
efforts to control for the idea that there is information about the
disturbance in perception. They are irrelevant side effects of control --
just like the fact that the output variable mirrors the disturbance variable
in some simple tracking tasks (one disturbing variable and linear feedback
and disturbance functions).

Usually on CSGnet, simulations that demonstrate a point are considered to
be helpful in clarifying matters.

Yes, and you can demonstrate the point that there is information about
disturbances in the perceptual signal by taking the perceptual signal above
and returning the distrubance that was present at the same time.

Me:

In fact, there is no information about disturbances or the net effect
disturbances in the perceptual signal.

Martin:

[This] sentence is an expression of a strongly held opinion, not of a fact.

That is true. I will abandon this opinion as soon as you demonstrate that
there is information about the disturbance in the sequence of numbers (the
disturbance report) above. You can demonstrate this by returning the
disturbance that these numbers were reporting to the control system.

Me:

The perceptual signal doesn't talk to the control system; it doesn't give
disturbance reports to the control system. The control system doesn't "base"
its outputs on anything other than the discrepancy between perception and
reference signal

Martin:

[This] sentence contradicts the usual circuit diagram of a control unit,
in which the perceptual signal is distinctly a part of the control circuit,
"talking" to the comparator, which is also part of the control system.

I didn't say that the perceptual signal was not part of the control unit; I
said that it doesn't "talk" to the control system. That means that the
percetual signal is nothing more than a variable that continuously changes
value. You say that these changing values carry a report of (information
about) the disturbance acting at the time these values are occurring. I
( and Bill and Tom if he were here and anyone else who understands how a
control system works) say "bull". You can easily set us stright by
returning the disturbance reported on be the perceptual signal above.

Bruce Gregory (960621.1540 EDT) --

Martin's comments reflect my own confusion. It is very difficult for me as
an ignorant bystander to figure out what the hell is going on here. Clearly
monumental issues are at stake, but for the life of me, I can't figure out
what they are. I doubt anyone can clear things up, since no one seems to
understand what anyone else is saying. It's a good thing we are all in
fundamental agreement, because I would be appalled to see what would happen
if real differences existed.

The idea that there is information about the disturbance in perception
represents an input-output view of how a control systems works. It suggests
that perception provides the data (information) that is used by the control
system as the basis for _calculating_ the outputs it should generate to
compensate for independent influences on (disturbances to) the variables
under control. This is just the wrong way to think about how control systems
work, and it is wrong in a particularly insidious way because it suggests
that the operation of a closed loop system is no fundamentally different than
the operation of an oper-loop system.

In fact, the output of a control system is driven by error -- it is not
calculated to oppose disturbances. There is no calculation of outputs
involved in the operation of a control system (that is why Bill mentioned
that such systems would be the preferred organization even in a predictable -
- disturbance free -- world; economy of computation); disturbance resistance
is a side effect of the operation of the closed negative feedback loop:
error continuously causing outputs that reduce error. No information about
independent influences on the controlled variable are needed -- and it
wouldn't help anyway because it would be too late and it couldn't tell the
control system what kind of environment it's outputs are working through to
counter the disturbance anyway.

Information theory was developed for the analysis of open loop systems; it
was imported into psychology under the assumption that organisms were open
loop systems (communication channels). Organisms are _not_ open loop systems
so information theory, like S-R theory, cognitive theory, reinforcement
theory, psychoanalystic theory, and all other theories based on the
assumption that organisms _are_ open loop systems (or sequential state
closed loop systems) is simply irrelevant to the behavior of organisms.

Grumble. Grumble...

Best

Rick

[_Martin_Taylor]

[Martin Taylor 960621 17:30]

Rick Marken (960921.1400)

Rick's message must have come in while I was composing my response to Bruce
Gregory. I am delighted that Rick has expressed, strongly and clearly,
what I had assumed in my message to Bruce might be the case. Rick believes:

The idea that there is information about the disturbance in perception
represents an input-output view of how a control systems works.

and

Information theory was developed for the analysis of open loop systems

The first of these statements is flat-out wrong. The second is true but
misleading, in that the theory was developed with no regard to whether
a system was open-loop, and with no regard to cause and effect. It was
_used_ to analyze open-loop systems, but that does not make it any the
less applicable to closed-loop systems. To say so is rather like saying
that thermostatic control of temperature cannot work because a thermometer
simply measures a temperature so that a person can know how cold it is.

···

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

I was hoping you could tell us; how does the perceptual signal inform the
control system about the state of the disturbing variable

Martin Taylor (960621 15:00) --

That's what I and Allan did tell you--the magical mystery function, a.k.a.
the output/feedback function.

The control system itself doesn't know the output or feedback function.

I guess you didn't read my last message, then. It's much easier to continue
a discussion if we read what each other says. I don't see any benefit to the
wider CSGnet community in repeating it, so I won't.

I'll send you a perceptual signal and you
send me back the disturbance. If you can't do that, then cut the crap about
information in perception about disturbances.

There's no point in saying the same thing over and over again. I've pointed
out why this is a silly thing to ask and why the answer has nothing to say
on the question at issue. I'm not going to do it again until you come up with something new. Remember in what follows that you have said:

The control system itself doesn't know the output or feedback function. All
it knows is the perceptual signal.

Now think a bit about that, and connect it to what you also say:

Since there is information in a radio signal about freeway traffic, if I sent
you a traffic report you could send me back a description of the traffic.

I presume you would NOT allow me to use the translation mechanisms
that convert acoustic signals to mental images (or whatever), nor to use
my understanding of what a freeway is, etc., because these items are exactly
analogous to the functions of the control system that you also won't allow
me to use. And yet..., and yet...You DO say that there is information in
a _radio_ signal (an electromagnetic fluctuation in the atmosphere) about
the all too human problem of a traffic jam.

Now, let's make a bargain. I'll send you the fluctuation waveform of a
radio signal, and you, without using knowledge of radio circuitry or
of perceptual functions in the human, will send back to me the state of
the traffic on the Santa Monica freeway. When you do that, I'll send you
an approximation of the fluctuations in your disturbance sequence without
you telling me the output/feedback function. (Remember, we only claimed
to reproduce the disturbance to the degree that the perception is controlled,
not exactly--poor control, poor reproduction; good control, good reproduction).

I can imagine you saying "that's not the same...you are asking me to do
without the perceptual function, and I'm only asking you to do without
the output function." That's true, but remember that you are asking me
to do the reproduction in a control sitation, about which you correctly
say:

disturbance resistance
is a side effect of the operation of the closed negative feedback loop:
error continuously causing outputs that reduce error.

(Side-note: But it is incorrect to say:

   >There is no calculation of outputs
   >involved in the operation of a control system

   The output function calculates the output at all times, based on the current
   and past values of the error. It is, after all, a function, at least as the
   analyst sees it.

end Side-note.)

No information about
independent influences on the controlled variable are needed

We agree on that, though. That's never been a bone of contention.

Resolution of conflict requires that some other means be found to satisfy
the higher-level references that cannot simultaneously be satisfied by the
means being used. I'm trying to find other ways of looking at the basis
of the disagreement, but I perceive you to be simply repeating what you
wrongly think to be a critical point. Look at something else for a while.

For example, perhaps you could elucidate why you think information theory
is applicable only to open-loop situations. That might help me to understand
what it is you are arguing against, because whatever that is, it sure ain't
what I understand to be information theory a la Shannon.

But as I said up top, I am delighted that you have come right out and
said that your problem is what I told Bruce Gregory I thought it might be.
It's much easier to sort out misunderstandings when their basis is brought
out into the open.

Martin

[Richard_Marken]

[From Rick Marken (960621.1600)]

Me:

The control system itself doesn't know the output or feedback function.

Martin Taylor (960621 17:30) --

I guess you didn't read my last message

I don't know what your "last message" was but if it said anything about
the control system knowing the output or feedback function then it was
more of the usual sophistry. The fact of the matter is that the perceptual
signal is the only input to a control system; period. amen. There is no
input regarding the state of the output variable, output funciton, feedback
function, or whatever.

Me:

I'll send you a perceptual signal and you send me back the disturbance.

You:

There's no point in saying the same thing over and over again. I've pointed
out why this is a silly thing to ask and why the answer has nothing to say
on the question at issue.

It's the only thing that has to do with the question at issue. If you can't
tell what the disturbance is by looking at the perceptual signal then you
have to give up this information about the disturbance in perception crap.
The perception isn't informing you or the control system about anything.

let's make a bargain. I'll send you the fluctuation waveform of a
radio signal, and you, without using knowledge of radio circuitry or
of perceptual functions in the human, will send back to me the state of
the traffic on the Santa Monica freeway. When you do that, I'll send you
an approximation of the fluctuations in your disturbance sequence without
you telling me the output/feedback function.

It's a deal. Send me the radio signal.

I'll tell you the result right now. I will be able to tell nothing about the
state of the traffic fro this signal -- even if I have a radio to play it
throughh. There is no information about the traffic in the radio signal just
as there is no information about the disturbance in the perceptual signal.
You're the one who thought that there was information in both.

I can imagine you saying "that's not the same...you are asking me to do
without the perceptual function

Nope. It's precisely the same. Perceptions don't carry information about
anything. The application of information theory to perception was just a big
mistake. It was based on the idea that perceptions "communicate" to us about
an external reality. In fact perceptions are the only reality we know; the
reality we control. Information theory is not only useless for understanding
control, it is also inconsistent with PCT epistimology. It might have value
in cryptology and the description of comm channels; but it's useless for
understanding control.

(Side-note: But it is incorrect to say:

   >There is no calculation of outputs
   >involved in the operation of a control system

I'll leave that one for Bill, if he wants it. I'll just say that the
"calculation" involved in transforming error into output is more like the
"calculation" involved when a billiard ball "calculates" how far it will move
when it is hit by another billiard ball. There is no "solution of equations"
involved in the "Calculation" of output -- certainly not of the kind that is
used in an inverse kinematic system that must calculate the output that
_would_ be required to counter a particular disturbance.

For example, perhaps you could elucidate why you think information theory
is applicable only to open-loop situations.

Because it's designed for the analysis of commumication systems where there
is a known set of possible inputs and outputs. Information is a measure of
how much the occurance of an output decreases your uncertainty about the
output that might have occurred. Your uncertainly is proportional to the
probability that each input will become the output. If the system you are
observing is closed loop, such that the output affects the probability
distribution of the inputs that might have become the outputs, then I don't
see how information theory can be applied coherently. It's just no longer
relevant.

I am delighted that you have come right out and said that your problem is
what I told Bruce Gregory I thought it might be.

I haven't seen that post yet. But now I can add another problem I have with
information theory: it's application to perception is epistimological
nonsense.

Best

Rick

[Hans_Blom]

[Hans Blom, 960621]

(Rick Marken (960921.1400) to Bruce Gregory (960621.1540 EDT)

The idea that there is information about the disturbance in perception
represents an input-output view of how a control systems works.

My two cents: there is no information in any single signal. Informat-
ion is information only when it is decoded, i.e. "compared" to some-
thing else. The language of whales or seagulls has no meaning for us
because we lack a proper decoder/comparator.

But back to control. I distinguish two cases: a fixed controller,
i.e. one with constant internal parameters, and a learning (adaptive)
controller, i.e. one with an internal mechanism that allows it to
tune itself (or be tuned by this mechanism, whatever expression you
like better).

Information is a more intuitively appealing notion in the latter case.
An adaptive controllers compares (uses the correlation between) its
actions and its perceptions in order to establish how effective its
actions are and how they can be improved. The criterium that it uses
for "goodness" is an internal model that provides predictions. If the
world in which the adaptive controller lives is noise-free and if it
is perfectly tuned, one can say that there are no disturbances: the
future can be perfectly predicted, there will never be any surprises.
If the world is not noise-free or if the tuning is not (yet) perfect,
there will turn out to be discrepancies between the predictions (of
what the future observations are going to be) and the observations
that are really forthcoming. These discrepancies can be called
disturbances (and subjectively, they would be just that, if the
controller could contemplate things ;-). But these disturbances have
two components; one is due to the inherent noise in the world that is
not modelable, the second is due to imperfect learning. But an
adaptive controller learns -- that is what adaptive control is all
about -- and it learns from its disturbances (when correlated with
its actions). And it will do that in such a way that it minimizes
those disturbances until further decrease is impossible.

So it can be said that it uses (the information in) the disturbance
in order to make itself control better. There is clearly information
in the disturbance, and this information is used, "consumed", in the
tuning process. When tuning is complete, there is no information in
the disturbance anymore, only white noise (and other components, if
the controller does not have enough internal degrees of freedom to be
able to model the world).

One can even ask the question: how much better has controller X
become during time T? The answer will be: N bits better (or
occasionally, worse). With the meaning that some (control) "decisions"
will be better than before. We have often talked about the "quality
of control". Information theory provides a gauge. You can use that
gauge if you want a more formal description of how good a controller
is.

The situation is less intuitively clear for a controller with fixed
internal parameters. If, however, we can answer the question how much
better a specific controller has become during a time T, it ought to
be clear that we can also answer the closely related question "how
much better is controller X than controller Y?". Use the gauge.

The problem is most obscure if we cannot compare. But we can! We can
compare every single controller with a reference "ideal" controller,
one that is fully tuned to its world and has a sufficient number of
internal degrees of freedom to do so. Its disturbance will be just
that part of the world that is not modelable (pure white noise). The
question "how good is controller X?" then becomes "how many bits is
controller X worse than the ideal controller?".

So is there information in the disturbance? Although I understand
what Martin says -- and he is correct and completely in line with how
information theory expresses things -- I would often say it different-
ly. But I hope that in at least one case (adaptive control) I have
been able to convince you that there IS information in the disturb-
ance. And if these words do not convince you, I have a couple of
demo's that might -- you already know them...

One problem that is related and that arises time and again: what does
the controller "know", i.e. which signals does it have access to.
Rick maintains that a controller has only access to its perceptions.
For an adaptive controller, that is not enough: it also needs access
to its actions, because actions and perceptions need to be correlated
in order for adaptation to take place. In another article -- somewhen
in the (far?) future -- I will demonstrate that a non-adaptive
controller still has (or can be thought of to have) its internal
model, but one that cannot be changed. Even such a controller has
"expectations". But, since it does not need to learn, it does not
need to know what it does. [In another sense, even a PCT controller
knows what it does. In every code there will be a line somewhat like
"output := output + ..."].

I'll wait and see for the moment whether I have been able to make
myself clear. Strange enough, I think that I am being very clear in
this post. But, strange enough, I also have the strong conviction
that this expectation will not turn out to be correct. Talk about
conflicts...

Greetings,

Hans

[Bruce_Gregory1]

[From Bruce Gregory 960622.1105 EDT)]
[Rick Marken (960921.1400)

The control system itself doesn't know the output or feedback function.

All it >knows is the perceptual signal. I'll send you a perceptual signal
and you send >me back the disturbance.

Sorry for being so dense, but I _still_ cannot see what is the issue here.
Let me give you a simple-minded example. When I was learning to fly in the
summer, the plane would undergo variations in altitude as the result of
passing through rising and falling parcels of air. These altitude
excursions were reflected in the altimeter, which I identify as a
"perceptual signal". By watching the way I was manipulating the controls,
my flight instructor inferred that I was contibuting to these oscillations
by excessive efforts to control the altitude. He instructed me to release
the controls and notice the altitude excursions in the absence of my
efforts. Indeed, excursions above the chosen altutude were often almost
matched by excursions below the chosen altitude. In fact, the plane did a
better job of flying itself than I did in "helping" it. Question: when I was
not manipulating the controls, did not the altimeter readings perceptual
signal) contain information about the disturbances? Question: When I
resumed manipulating the controls, did not the altimeter readings contain
information about the sum of the disturbances and my efforts to control for
them? (Granted that a printout of altitude would not allow you, or Martin,
to separate the two factors.)

Puzzled in Pomfret

Bruce

[Richard_Marken]

[From Rick Marken (960623.1140)]

After re-reading [Hans Blom, 960621)] I was surprised to find that
I actually agree with much of what Hans said. Hans describes an adaptive
controller which is a control system that tunes the parameters of
another control system. The adaptive controller tunes (varies the
parameters of) a control system by controlling the perceived correlation
between the output and disturbance effects on the variable controlled
by the tuned control system. I think it's fair to call the correlation
between output and disturbance the "information about the disturbance
in the output". So the adaptive controller described by Hans is trying to
control a variable that could be called "information". The goal of such
an adaptive controller would be to keep this variable at the value
indicating perfect control -- so the reference for the correlation
between disturbance and output would be -1.0.

Hans' adaptive controller is an example of what I described to Bruce
Gregory yesterday -- a separate control system (the adaptive controller)
perceiving something _about_ the controlling being done by another
control system (the system being tuned). The adaptive controller is
perceiving something that the tuned system itself cannot perceive -- the
relationship between that system's own outputs and the disturbances to
the variable being controlled. The adaptive controller perceives and controls
the correlation between these variables; the system being tuned just
perceives the variable it controls. The tuned system (like any control system,
including the adaptive controller) does not perceive its own outputs, the
disturbance to the controlled variable, the functions relating disturbance
or action to the state of the controlled variable -- nothing but the
perception of the controlled variable itself.

Of course, a real (living) adaptive controller could not work the way Has'
system works because it would have no more priviledged access to the data
needed to compute the "actual" correlation between output and disturbance
than does any other control system. But it is reasonable to say that Hans'
make-believe adaptive control system does control the information in the
output of the tuned control system about the disturbance to the variable
controlled by that system.

Best

Rick

[Hans_Blom]

[Hans Blom, 960623]

(Rick Marken (960623.1140))

After re-reading [Hans Blom, 960621)] I was surprised to find that
I actually agree with much of what Hans said.

Well, well, well...

                                        Hans describes an adaptive
controller which is a control system that tunes the parameters of
another control system. The adaptive controller tunes (varies the
parameters of) a control system by controlling the perceived correlation
between the output and disturbance effects on the variable controlled
by the tuned control system. I think it's fair to call the correlation
between output and disturbance the "information about the disturbance
in the output".

Good.

          So the adaptive controller described by Hans is trying to
control a variable that could be called "information". The goal of such
an adaptive controller would be to keep this variable at the value
indicating perfect control -- so the reference for the correlation
between disturbance and output would be -1.0.

Yes.

Hans' adaptive controller is an example of what I described to Bruce
Gregory yesterday -- a separate control system (the adaptive controller)
perceiving something _about_ the controlling being done by another
control system (the system being tuned). The adaptive controller is
perceiving something that the tuned system itself cannot perceive -- the
relationship between that system's own outputs and the disturbances to
the variable being controlled.

Yes, it would be fair to characterize this as a hierarchical
controller, the lowest level being control (matching perceptions with
reference levels), and the next level control OF control (generating
the "best" match, dynamically). This leaves out some "minor" details,
but let that be...

Of course, a real (living) adaptive controller could not work the way Has'
system works because it would have no more priviledged access to the data
needed to compute the "actual" correlation between output and disturbance
than does any other control system.

The basic requirement for a "living" adaptive controller that you
point at is whether the human controller, say, "has access to" what
it does, to its actions. Physiology shows that that is so. Numerous
signals, generated at the final points of the action chain, are
available to higher levels; signals that make "knowledge" available
about muscle length, muscle tension, muscle load, etc. I maintain
that it must be these "internal perceptual signals" that, correlated
with the perceptions of/about the outside world (vision, audition,
proprioception), allow a "second (?) level controller" to tune the
first level controller. And this all the way up, throughout the
hierarchy.

                            But it is reasonable to say that Hans'
make-believe adaptive control system does control the information in the
output of the tuned control system about the disturbance to the variable
controlled by that system.

To me it does! And the implementation isn't that difficult, as I have
demonstrated before.

Now let me point at a consequence of my previous post that might make
clearer what Martin has been saying all along. If we can assign a
"number of goodness" (in bits) to a controller by comparing it to the
best controller possible, we can also compare a particular controller
to a "controller" that does not control at all. This is The Test, as
you know. This comparison again yields a number of bits as a measure,
by using the tools that information theory provides. It is actually
this number of bits that the controller "uses" in its control, and
again the result is that this information is "consumed" in that the
controller's output is more negatively correlated (i.e. closer to -
1.0) with the disturbance than it would be if no control occurred.

Does this make sense as well?

Greetings,

Hans

[_Martin_Taylor]

[Martin Taylor 960624 16:00]

Rick Marken (960621.1600)

I am delighted that you have come right out and said that your problem is
what I told Bruce Gregory I thought it might be.

I haven't seen that post yet.

I'm not surprised. In an inversion of my usual fumble-fingered mode, I
replied privately to Bruce, thinking I was sending it to CSGnet. If it
is still of any interest, here it is:

···

----------------------------
[Martin Taylor 960621 16:45]

Bruce Gregory (960621.1540 EDT)

On "information":

Clearly monumental issues are at stake,
but for the life of me, I can't figure out what they are.

Neither can I, but I have a hypothesis.

One of the precepts of PCT is that of circular causation. PCT is directly
contradictory to an S-R view of the organism, in which causation is linear.
Information theory has historically often been wrongly cast in
the form: "Signals emanating from point A go through a noisy channel C
to cause values to be detected at point B, and the signals convey information
from A to B." That's a straight linear cause-effect description, and
antithetical to PCT.

If someone who is attuned to PCT has learned about information theory in
this linear cause-effect framework, they are naturally going to resist
strongly any attempt to suggest that IT can be used as a tool to look
at control systems. They'd say: "It's cause-effect, isn't it? That's
dead wrong, isn't it?" I never learned about information theory in that
context. I learned it from reading Shannon and Weaver, originally, not
from derivative works.

When I innocently tried to use information theory in analyzing control
those years ago, I ran into this immediate (I might almost say "reflex"
if I didn't know better opposition to the idea, an antagonism I
simply didn't understand, since I have never treated information as
relating to cause-effect systems. Neither did Claude Shannon in the theory,
though he did apply it only to cause-effect systems. The theory is based
around uncertainty at one point about what is/was happening at another
point. It's, in a way, a theory of measurement. Information is the
reduction of uncertainty.

Looked at this way, to me the application to control was obvious. In words,
the "perception" is a continuous version of Shannon's "observation". By making
the perception/observation, one is less uncertain about the current state
of the thing observed (the CEV) than one would otherwise have been. With
this reduced uncertainty, one is better placed to push or pull on the CEV
to bring it to where one wants it to be than if one didn't observe. This
seems to me to incontrovertible, and it says that PCT has to work better
than a planning-outflow model. I also argued that an S-R model couldn't
work, on similar grounds--this time because without observing the changeable
effect of a push or pull, one would not be able to counter the disturbances
to the CEV (that's a bit more subtle, but I think correct). Information
theory seemed to me to demonstrate the necessity of PCT as opposed to the
two competitors discussed in "Models and their worlds."

I also tried to demonstrate that in a chaotic world like our own, the nature
of control--to smooth the environment of the organism--was equivalent to
reducing the information rate from the world into the organism, and that
was seen as paradoxical rather than inevitable. Again, I suspect that
behind the problem is an identification of "information" with "cause-effect."
Anyway, I remain bewildered as to the background to the opposition. In my
mind there is no contradiction between an information-theoretic view of
control and any other view. Those who choose not to use IT don't lose
if other people do use it, so I'm confused as to why they worry about it.
I assume there is some other perception for which they are controlling, one
that hasn't yet come to light, one that I haven't guessed to the extent that
I could try the Test on it.

For my own part, I'm controlling, at least in part, for a perception that
other people see me as honestly reporting what happened in the earlier
discussion, including that the transmission of information about the
disturbing influence through the perceptual signal had been demonstrated.
This is why I spent considerable valuable time to compile and send two
long messages consisting mostly of verbatim, chronological, extracts.

I don't care whether other people use an information-theoretic view, but
I am disturbed by statements that it had been demonstrated by the
experiment that the view was wrong, and by statements that I and my
collaborator had changed the conditions of the experiment after Bill and
Rick had agreed that they were well chosen and that our demonstration
would not work under those conditions. We didn't, and it did, and that
should have been an end to the whole affair.

There may be monumental issues involved, but I think they are only
tangentially relevant to the theory of PCT. I'd like to feel free to
mention information-based issues when they seem to be appropriate to an
ongoing discussion on CSGnet, but I don't, because I have learned that to
do so is to risk becoming embroiled in fruitless discussions based in
rigidly maintained misapprehensions that I don't fully understand.

It's a good thing we are
all in fundamental agreement, because I would be appalled to see
what would happen if real differences existed.

Perhaps not. Real difference are often easier to see and to eliminate.
The worst wars are the Civil wars (think of England 1640s, US 1860s,
Cambodia, Northern Ireland, Bosnia, Ruanda...). People on opposite sides of an
international war are often (not always) more civilized to their opponents
than they are to people essentially indistinguishable from themselves
except for some minor element of culture such as whether or not to kneel
to the Pope.

Martin

[Richard_Marken1]

[From Rick Marken (960623.1600)]

Hans Blom (960623) --

Now let me point at a consequence of my previous post that might >make clearer what Martin has been saying all along.

If we can assign a "number of goodness" (in bits) to a controller by ><comparing it to the best controller possible, we can also compare a >particular controller to a "controller" that does not control at all. > This is The Test, as you know. This comparison again yields a number >of bits as a measure, by using the tools that information theory >provides. It is actually this number of bits that the controller >"uses" in its control, and again the result is that this information >is "consumed" in that the controller's output is more negatively >correlated (i.e. closer to - 1.0) with the disturbance than it would >be if no control occurred.

Does this make sense as well?

It made sense up to the point where you talked about information
the controller "uses" being "consumed" by the controller. Well,
the part about "This is The Test" is wrong too, but that's
irrelevant here.

It is certainly possible to measure, in "bits", how well a control
system "could" control and how well it actually does control. Though
I can't understand what the benefit of this is over measuring control
in physical units using RMS error or the stability factor; but, if
computing -log2 of variables makes you feel more scientific then
it's sure OK with me.

Now if you find that control, measured in bits, is better when a
control system is present than when it is not, and you conclude
that the bits were "consumed" by the control system, you are just
using a metaphor -- and a bad one at that. Control systems
operate in loops so if they consume anything that are consuming
themselves (oy vey, that reminds me of cybernetics meetings with
control loops illustrated by snakes eating their tails; is that
how you think of control loops, Hans?). Anyway, we know how contol
systems work -- and they don't consume anything except the energy
needed to convert error into output. It might sound cool and trendy
to talk about bits being "consumed" -- but it ain't happenin' in
control loops. Sorry.

Best

Rick