Measuring Input-Output Characteristics of Components of a Closed Loop: Redux

[From Rick Marken (2014.03.11.1740)]

It turns out that I was wrong again. But this time I was wrong about being wrong. I thought I was wrong when I said that you can’t measure the open-loop input-output characteristics of a component of a closed control loop while that component is still part the loop. But something Bill Powers said in his 1999 muscle model paper kept nagging at the back of my mind.

What Bill said was this: “If the measured apparent spring constant is the same as that measured in muscle preparations, then the open-loop model is correct”. The “the measured apparent spring constant” is the muscle output function measured in a closed loop. Measuring this function in “muscle preparations” involves measuring the same function in an open-loop. So Bill seemed to be implying that the output function of a control loop, when measured in the loop, would differ from that function when measured open loop.

I’ll illustrate the difference using Martin’s diagram of a control loop. Measuring the input-output characteristics of the output function component of a control loop while it is in the loop would involve creating an experimental set up like this:
image3.png

The input would have to be an electrical signal that is delivered by an electrode injected into the efferent neuron that is the input to the output function (G()) that produces the observed output. If it doesn’t matter that the output function is part of a closed loop, then observed relationship between input and output will reflect the nature of the output function, G().

Doing the same thing with the loop open would look like this:

image4.png

This is a completely open-loop system; the output has no connection back to the input that is the cause of that output.

At first I thought that the measured relationship between input and output for the same function, G(), would be different in the open and closed loop situations. Then I changed my mind (based on what turned out to be an incorrect analysis of the situation) and now I have changed back to my original conclusion: the measured input-output characteristics of at least one component of a control loop - the output function component – will be different depending on whether these characteristics are measured in a closed or an open loop.

I base this conclusion on what I believe is a correct simulation of what would happen if the experiment described in the two diagrams above were actually carried out. I simulated injecting a linearly increasing electrical signal into the neuron that is the efferent input to an integrating output function. I measured the relationship between this input signal and the output response of the same output function when thist function was part of an open and a closed loop. Here are the results:

image5.png

Clearly, the measured input-output characteristic for the same output function is quite different depending on whether it is measured in an open (red line) or closed (blue line) loop. So for at least one input-output component of a control loop – the output function, G() – the apparent characteristics of this function are quite different depending on whether they are ,measured in a closed or an open loop.

The reason this happens is really no mystery; when you inject the input test signal into the closed loop it acts like a disturbance whose effect is reduced by the output it causes.

Best regards

Rick

Richard S. Marken PhD
www.mindreadings.com
The only thing that will redeem mankind is cooperation.
– Bertrand Russell

[From Adam Matic 2014.03.12.13.30 CET]

Rick Marken (2014.03.11.1740)

···

It turns out that I was wrong again. But this time I was wrong about being wrong. I thought I was wrong when I said that you can’t measure the open-loop input-output characteristics of a component of a closed control loop while that component is still part the loop.

AM:

I thought you said you were wrong about components of the loop not having the same input-output (I/O) characteristics while in a closed loop and when out of the loop, which is different from not being able to measure those characteristics. I might have misunderstood, but I think you were right then (about being wrong) and are right now about there not being a way to measure I/O properties of those components while in closed loop.

(I find the expressions ‘open-loop’ and ‘closed loop’ confusing. If it’s open, there is no loop. If there is a loop, it is closed.)

When control engineers need to tune a control system, they break the loop and measure it’s I/O characteristics. They do the same with the ‘plant’. Then they know enough about the system to make a good approximation of tuning parameters.

I don’t see any non-invasive way of breaking the loop when dealing with living control systems. You can take out a muscle from someone and measure it’s I/O properties, or a neuron, or retinal cells. That’s fine if the living thing doesn’t mind. But if you try stimulus-response measurements while the component is in the loop you get nothing relevant. At least I can’t think of a way to get anything. As you show in the plot, there is always interaction from the loop.

Adam

[Martin Taylor 2014.03.12.10.33]

In a sense you are correct, since for any set of measurements there

is an infinite number of functions that would fit the data – I
believe the number is Aleph-2, where aleph-null is the number of
rational fractions and aleph-1 the number of points on a line. But
for practical purposes your statement is too strong, because you can
choose to limit the range of functions from which to select. For
example, in PCT modelling, the selection is usually very strongly
limited, being restricted only to finding three or four parameter
values that more-or-less fit the data with a prespecified set of
function types (e.g. straight-through multipliers and leaky
integrators).
True. It’s much easier to measure the properties of a component of a
loop when you can constrain the range of its inputs. That’s not the
same as saying you can’t do it without breaking the loop.
We are talking here about reverse engineering. If it were
impossible, why would so many software licences explicitly prohibit
it? If it was as easy as building to specifications, why would it
matter?
Martin
Martin
Martin

···

On 2014/03/12 8:24 AM, Adam Matic
wrote:

[From Adam Matic 2014.03.12.13.30 CET]

Rick Marken (2014.03.11.1740)

                It turns out that I was wrong again. But this

time I was wrong about being wrong. I thought I was
wrong when I said that you can’t measure the
open-loop input-output characteristics of a
component of a closed control loop while that
component is still part the loop.

AM:

            I thought you said you were wrong about components of

the loop not having the same input-output (I/O)
characteristics while in a closed loop and when out of
the loop, which is different from not being able to
measure those characteristics. I might have
misunderstood, but I think you were right then (about
being wrong) and are right now about there not being a
way to measure I/O properties of those components while
in closed loop.

            (I find the expressions 'open-loop' and 'closed loop'

confusing. If it’s open, there is no loop. If there is a
loop, it is closed.)

            When control engineers need to tune a control system,

they break the loop and measure it’s I/O
characteristics. They do the same with the ‘plant’. Then
they know enough about the system to make a good
approximation of tuning parameters.

[Adam Matic 2014.03.12.16.00 CET]

···

(Martin Taylor 2014.03.12.10.33)

True. It's much easier to measure the properties of a component of a

loop when you can constrain the range of its inputs. That’s not the
same as saying you can’t do it without breaking the loop.

We are talking here about reverse engineering. If it were

impossible, why would so many software licences explicitly prohibit
it? If it was as easy as building to specifications, why would it
matter?

AM:

I think of modeling control systems as reverse engineering. If you can, you break the loop and measure I/O. If you can’t, you make a model fit what you measured in operation. I agree that it’s possible to reverse engineer software of hardware, I’m not sure I see how that is related to control systems.

If I understand correctly, you are saying that there is a method of reverse engineering a control system without making a model or breaking the loop. Do you have an example of such a procedure on either a live or an artificial control system?

Adam

[Martin Taylor 2014.03.12.11.11]

I don;t see that you could reverse-engineer anything without making

a model, since to reverse-engineer something means to make a model
of it, so I’ll ignore that part of your question.
I suppose my answer depends on whether by “reverse engineering a
control system” you mean the entire system or some subsystems. In
the biological system “reverse engineering the system” is the
objective of the sciences of psychology, physiology, etc. So you
must mean some subsystem. Again if it is biological, the subsystem
must be isolated from interference from outside that subsystem. That
is difficult to arrange, but one can approximate, and that’s what
PCT modelling does. The PCT Modeller assumes a certain form of control structure,
typically a perceptual function of unknown form that produces a
scalar-valued signal, a comparator that produces a reference signal,
and an output function of prespecified form, usually a leaky
integrator, and an environmental feedback function, often a simple
connector but always well specified. The modeller then assumes that
something that the experimenter can measure (called the
“environmental variable”) corresponds exactly to the signal output
by this perceptual function (or assumes some form for the ev->p
relationship). Given these restrictive assumptions, the modeller
then determines what the output waveform would be given a known
disturbance waveform, and changes a small number of free scalar
parameters until a best match is found.
In PCT experimentation, as opposed to modelling real-life
situations, it is possible also to explicitly change the
environmental feedback function, so a tracking experiment can be
treated explicitly as a study of input-output relations between the
environmental variable and the output. This is not usually of much
value, at least if control is good, because of the highly restricted
range of variation of the environmental variable, but it can be done
and the results will be valid, if noisy.
If the assumption made by the modeller about the relation between
the environmental variable and the controlled perception is wrong,
we are into a different aspect of reverse engineering, known in this
context as “The Test for the Controlled Variable” (TCV). If the
controlled perceptual signal is not the assumed function of the
presumed environmental variable, control will be worse than if the
assumed function is correct. The reverse engineering occurs when
different functional relations are assumed between measurable
environmental variables and the perceptual signal assumed to exist
inside the organism (of course, this uses the convenient “neural
current” approximation, so is inherently inexact). The TCV, of
course, includes other manipulations, but here we are concerned only
with the reverse-engineering aspect. If you are talking about a non-biological control system in the form
of a black-box with only an input set of sensors and an output that
influences a determinate set of environmental variables that the
black box could sense, then the considerations are the same. So we
should consider an open-box system. In that case, the engineer could
measure directly the input-output characteristics for each component
within the range of input values permitted by the actions of the
control system. So that’s not a very interesting case.
I am left with one more case, the possibility of direct measurement
of components of a biological control system. Again, the
reverse-engineer has to assume some structure for the invisible
parts of the control system. HPCT provides one such assumption that
restricts the search space. Within HPCT we have the assumptions
listed above, and the engineer has to use the black-box approach
described above. The question then becomes which components can be
investigated. The obvious component is the total path from input to
output, which is assumed to consist of a perceptual function, a
comparator, and an output function. The performance of that pathway
can be determined when it is an element of a control loop that
permits considerable variation of its input and therefore of its
output.
An example of such a control loop is a question-answer session
between two people. If I call the two participants Q and R (for
"responder, since “A” is used often in normal text), we can assume R
controls a perception of Q’s knowledge of the thing questioned, with
a reference that Q should come to know it. Q asks the question,
which leads R to perceive that Q does not know the thing questioned,
so R answers the question, and with luck, R then perceives Q to now
know the thing questioned, compensating for the disturbance induced
by the perception of the question. Q could, of course, ask many different questions, and so long as Q
perceives R to be trying to answer to the best of R’s ability, a
wide range of questions will result in a wide range of answers
(unlike my Asperger’s grandchild, who spent most of one days
answering every question by saying simply “Apple”, and laughing).
Each such question uses a component of R’s loop that controls R’s
perception of Q’s kowledge, and because the range of variation of
both input and output is large, Q (and R) can determine something
about some internal property of R, though not uniquely. Because the
component contains both perceptual and output functions, Q cannot
determine whether a wrong answer is the result of a misunderstood
question or R’s lack of knowledge.
Suppose that instead of general knowledge questions, Q asks about
simpler things, such as whether R might be colour-blind. To
investigate this kind of internal property of R, Q shows R various
differently coloured patterns. If R cannot distinguish red from
green, R will provide one answer, but if he can, he will answer a
different way (“he” because most colour-blind people are male).
Again, by giving an answer, R is controlling a perception, but it is
not a perception of the way the pattern looks to him.
Q could investigate the same question by allowing R to manipulate
the colours of the pattern in a classical PCT tracking study, and
perform the TCV to see whether red-green variation was a component
of the controlled variable, but doing it in such a way that R has no
control over the input pattern allows a direct determination of R’s
perceptual ability provided Q perceives R as trying to report
correctly what he perceives. By assuming that the effect of the
output function can be eliminated, Q can evaluate the perceptual
function.
I hope this answers your question. I tried to cover as many
different conditions as I could think of, but there may be others I
forgot. Martin

···

[Adam Matic 2014.03.12.16.00 CET]

            (Martin Taylor

2014.03.12.10.33)

            True. It's much easier to measure the properties of a

component of a loop when you can constrain the range of
its inputs. That’s not the same as saying you can’t do
it without breaking the loop.

            We are talking here about reverse engineering. If it

were impossible, why would so many software licences
explicitly prohibit it? If it was as easy as building to
specifications, why would it matter?

AM:

            I think of modeling control systems as reverse

engineering. If you can, you break the loop and measure
I/O. If you can’t, you make a model fit what you
measured in operation. I agree that it’s possible to
reverse engineer software of hardware, I’m not sure I
see how that is related to control systems.

          If I understand correctly, you are saying that there is

a method of reverse engineering a control system without
making a model or breaking the loop. Do you have an
example of such a procedure on either a live or an
artificial control system?

Adam

[Martin Taylor 2014.03.12.10.49]

Your text description of what you did disagrees with your diagrams.

Your text says you used the same input to the output function in
both cases, whereas your diagrams say that the input to the output
function in the open-loop (broken-loop) case is simply the injected
signal, whereas in the closed loop case it is the injected signal
plus the error signal.
Why you would ever have imagined these two cases might give the same
result is quite beyond me.
Martin

image3.png

image4.png

image5.png

···

[From Rick Marken (2014.03.11.1740)]

      It turns out that I was wrong again. But this time I was

wrong about being wrong. I thought I was wrong when I said
that you can’t measure the open-loop input-output
characteristics of a component of a closed control loop while
that component is still part the loop. But something Bill
Powers said in his 1999 muscle model paper kept nagging at the
back of my mind.

      What Bill said was this: "If the measured apparent spring

constant is the same as that measured in muscle preparations,
then the open-loop model is correct". The “the measured
apparent spring constant” is the muscle output function
measured in a closed loop. Measuring this function in “muscle
preparations” involves measuring the same function in an
open-loop. So Bill seemed to be implying that the output
function of a control loop, when measured in the loop, would
differ from that function when measured open loop.

      I'll illustrate the difference using Martin's diagram of a

control loop. Measuring the input-output characteristics of
the output function component of a control loop while it is in
the loop would involve creating an experimental set up like
this:

          The input would have to be an electrical signal that is

delivered by an electrode injected into the efferent
neuron that is the input to the output function (G()) that
produces the observed output. If it doesn’t matter that
the output function is part of a closed loop, then
observed relationship between input and output will
reflect the nature of the output function, G().

          Doing the same thing with the loop open would look like

this:

          This is a completely open-loop system; the output has

no connection back to the input that is the cause of that
output.

          At first I thought that the measured relationship

between input and output for the same function, G(), would
be different in the open and closed loop situations. Then
I changed my mind (based on what turned out to be an
incorrect analysis of the situation) and now I have
changed back to my original conclusion: the measured
input-output characteristics of at least one component of
a control loop - the output function component – will be
different depending on whether these characteristics are
measured in a closed or an open loop.

          I base this conclusion on what I believe is a correct

simulation of what would happen if the experiment
described in the two diagrams above were actually carried
out. I simulated injecting a linearly increasing
electrical signal into the neuron that is the efferent
input to an integrating output function. I measured the
relationship between this input signal and the output
response of the same output function when thist function
was part of an open and a closed loop. Here are the
results:

          Clearly, the measured input-output characteristic for

the same output function is quite different
depending on whether it is measured in an open (red line)
or closed (blue line) loop. So for at least one
input-output component of a control loop – the output
function, G() – the apparent characteristics of this
function are quite different depending on whether they are
,measured in a closed or an open loop.

          The reason this happens is really no mystery; when you

inject the input test signal into the closed loop it acts
like a disturbance whose effect is reduced by the output
it causes.

[Adam Matic 2014.03.12.18.30 cet]

MT: I am left with one more case, the possibility of direct measurement of
components of a biological control system. Again, the reverse-engineer
has to assume some structure for the invisible parts of the control
system. HPCT provides one such assumption that restricts the search
space. Within HPCT we have the assumptions listed above, and the
engineer has to use the black-box approach described above. The question
then becomes which components can be investigated. The obvious component
is the total path from input to output, which is assumed to consist of a
perceptual function, a comparator, and an output function. The
performance of that pathway can be determined when it is an element of a
control loop that permits considerable variation of its input and
therefore of its output.

AM:
OK, I suppose I agree with all of this, given that the reference is
also assumed to be constant or otherwise known.

An example of such a control loop is a question-answer session between
two people. If I call the two participants Q and R (for "responder,
since "A" is used often in normal text), we can assume R controls a
perception of Q's knowledge of the thing questioned, with a reference
that Q should come to know it. Q asks the question, which leads R to
perceive that Q does not know the thing questioned, so R answers the
question, and with luck, R then perceives Q to now know the thing
questioned, compensating for the disturbance induced by the perception
of the question.
[...]

I hope this answers your question. I tried to cover as many different
conditions as I could think of, but there may be others I forgot.

AM:
So, you're saying that we can find out which perceptual functions
exist by doing some tests, like the color blindness test? That seems
like a version of the TCV. Do you think some other psychophysics
experiments are similar or analogous to the TCV?

Adam

[Martin Taylor 2014.03.12.14.27]

[Adam Matic 2014.03.12.18.30 cet]

AM:
So, you're saying that we can find out which perceptual functions
exist by doing some tests, like the color blindness test? That seems
like a version of the TCV. Do you think some other psychophysics
experiments are similar or analogous to the TCV?

Adam

Some time ago, in one of the iterations of the psychophysics issue, I gave it as my opinion that many of the precautions taken by good psychophysicists were exactly what they would do if they were knowingly doing the TCV. I got blasted for this opinion on the grounds that "S-R" (in the same way that many pundits take any tech news to be a signal of Apple's imminent demise "because Apple" (quoting the Macalope). I still hold to the opinion that if you don't have a pretty good idea what the subject is controlling for in a psychophysics experiment, the results are likely to mislead you.

As for the actual running of the experiment, as opposed to all the set-up and instruction phase, whether there is a relation to the TCV depends on the experiment. A tracking experiment such as I was doing in the 1960s in studies of figural after-effect, simply assumes that the controlled perception is teh one specified by the experiment (keep those lines vertical, keep the rotating disk stationary, and stuff like that). The other kind is a "what do you perceive" kind of question-answer session, and the experimenter has one or two ways of estimating whether the subject is looking for what the experimenter is asking about. In those studies, in contrast to the tracking studies, a correlation between disturbance and output that varies when the experimenter makes the task more or less difficult is a good sign that the subject is doing the right thing. That's not the TCV, but it does something rather similar.

I'd prefer the label of TCV to be kept for the specific program that has these components:
   1. Imagine an environmental variable (EV) that might be functionally related to the putative controlled perception PVP.
   2. Determine whether the subject has sufficient input mechanisms for it to be possible for her to create the PVP.
   3. Determine whether the subject has the means to influence the EV.
   4. Apply a disturbance to the EV and see whether the subject acts by the means you noted in step 3 to bring it back to its presumed reference value.
   5. Do step 4 a lot (if it has to be separate discrete trials) or continuously (if possible) and note the relationship between the EV and your disturbance and/or the relationships between the EV and the subject's output and/or the relationship between the subject's output and your disturbance. A good index of control is low correlation between EV and the other two and a high correlation between disturbance and output.
   6. Return to step1, imagining a different plausible EV, and repeat the loop until some criterion is satisfied.
   7. Choose the EV with the best index of control from the various passes through step 5.

The TCV is a major task, and one that cannot actually be performed in many real-world situations. A lot of the time, one has to assume that at some particular moment a person was controlling some variable at some reference level, because at the next moment they may no longer controlling that variable, perhaps because a choice was made, or because the controlled variable was moving "the right way" so they could control something else -- we call it "multitasking", or because control at a higher level keeps varying the reference level, or ...

In the set-up for a psychophysical experiment, the experimenter tells the subject what to look for, and may test by giving an obvious example. For instance, suppose the question is how well the subject can distinguish off-white pink from off-white blue. The experimenter may test whether the subject "knows what to see and do" (i.e. to control a perception that the experimenter is satisfied) by presenting a rose and sky-blue comparison that almost everyone should distinguish easily, and checking whether the subject presses the "correct" button (says the right thing, waggles the proper finger...).

If the subject does not do what the experimenter expects, the experimenter doesn't usually use the TCV to figure out what the subject is controlling. More commonly, he acts on the subject until the subject does "do the right thing". Sometimes this takes weeks, sometimes seconds. It's not the TCV program, in which the objective is not to disturb the subject in a way that would alter what perception is being controlled. For example, in Rick's 3-car TCV demo, if the cars jumped randomly all over the screen, I suspect many subjects would cease trying to control the location of any of them.

Martin

[Adam Matic 2014.03.13.08.50]

MT: As for the actual running of the experiment, as opposed to all the set-up and instruction phase, whether there is a relation to the TCV depends on the experiment. A tracking experiment such as I was doing in the 1960s in studies of figural after-effect, simply assumes that the controlled perception is teh one specified by the experiment (keep those lines vertical, keep the rotating disk stationary, and stuff like that). The other kind is a "what do you perceive" kind of question-answer session, and the experimenter has one or two ways of estimating whether the subject is looking for what the experimenter is asking about. In those studies, in contrast to the tracking studies, a correlation between disturbance and output that varies when the experimenter makes the task more or less difficult is a good sign that the subject is doing the right thing. That's not the TCV, but it does something rather similar.

I'd prefer the label of TCV to be kept for the specific program that has these components:
1. Imagine an environmental variable (EV) that might be functionally related to the putative controlled perception PVP.
2. Determine whether the subject has sufficient input mechanisms for it to be possible for her to create the PVP.
3. Determine whether the subject has the means to influence the EV.
4. Apply a disturbance to the EV and see whether the subject acts by the means you noted in step 3 to bring it back to its presumed reference value.
5. Do step 4 a lot (if it has to be separate discrete trials) or continuously (if possible) and note the relationship between the EV and your disturbance and/or the relationships between the EV and the subject's output and/or the relationship between the subject's output and your disturbance. A good index of control is low correlation between EV and the other two and a high correlation between disturbance and output.
6. Return to step1, imagining a different plausible EV, and repeat the loop until some criterion is satisfied.
7. Choose the EV with the best index of control from the various passes through step 5.

AM:
That is a very specific program. I understand the concept of the TCV to be any procedure that tries to determine what is being perceived and controlled by the subject, such as The Coin Game or just asking "what are you trying to do?". All the time people try to find out what other people and organisms are doing, they just don't usually go at it systematically. Even when they (we) do it systematically, it's not conclusive because there always might be a slightly better approximation of the CV.
I like this R.P. Feynmans essay, it's been mentioned on CSGnet before, Cargo Cult Science: <http://neurotheory.columbia.edu/~ken/cargo_cult.html>http://neurotheory.columbia.edu/~ken/cargo_cult.html
He is very critical of experiments in psychology, and he gives an example of good science and I think it is an example of the TCV, so I'll just quote that part:

···

___________________________________________________
R.P. Feynman:
[...] there have been many experiments running rats through all kinds of mazes, and so on--with little clear result. But in 1937 a man named Young did a very interesting one. He had a long corridor with doors all along one side where the rats came in, and doors along the other side where the food was. He wanted to see if he could train the rats to go in at the third door down from wherever he started them off. No. The rats went immediately to the door where the food had been the time before.

The question was, how did the rats know, because the corridor was so beautifully built and so uniform, that this was the same door as before? Obviously there was something about the door that was different from the other doors. So he painted the doors very carefully, arranging the textures on the faces of the doors exactly the same. Still the rats could tell. Then he thought maybe the rats were smelling the food, so he used chemicals to change the smell after each run. Still the rats could tell. Then he realized the rats might be able to tell by seeing the lights and the arrangement in the laboratory like any commonsense person. So he covered the corridor, and still the rats could tell.
He finally found that they could tell by the way the floor sounded when they ran over it. And he could only fix that by putting his corridor in sand. So he covered one after another of all possible clues and finally was able to fool the rats so that they had to learn to go in the third door. If he relaxed any of his conditions, the rats could tell.
Now, from a scientific standpoint, that is an A-number-one experiment. That is the experiment that makes rat-running experiments sensible, because it uncovers that clues that the rat is really using-- not what you think it's using. And that is the experiment that tells exactly what conditions you have to use in order to be careful and control everything in an experiment with rat-running.
I looked up the subsequent history of this research. The next experiment, and the one after that, never referred to Mr. Young. They never used any of his criteria of putting the corridor on sand, or being very careful. They just went right on running the rats in the same old way, and paid no attention to the great discoveries of Mr. Young, and his papers are not referred to, because he didn't discover anything about the rats. In fact, he discovered all the things you have to do to discover something about rats. But not paying attention to experiments like that is a characteristic example of cargo cult science.
___________________________________________

AM:
If an experiment, be it in pyschophysics, animal behavior, cognitive psychology, economics or anywhere tries to find a controlled variable, then it is finding something about the organism that is scientifically useful.

Are we still on the same topic? I don't know how I got from measuring I/O properties of a control loop to the TCV.
Adam

[Martin Taylor 2014.03.13.09.23]

Words again. I'd prefer to call :trying to determine what is being

perceived and controlled by the subject" a search for the controlled
variable rather than a test. After having narrowed down the search,
I would then do a test,if I wanted to be better assured that the
search had come close to the right answer. When you ask a subject
what he is trying to do, he may not know, or may not want to tell
you correctly, or may answer at a level diffeerent from the one that
interests you.
Yep. And in many real-life situations, they couldn’t if they wanted
to.
I imagine that it would be as rare for any search or test to find
the exact CV as it would be for a control unit to have zero error.
It might be so, but only by a fluke, and only in passing.
Great story, and I agree that it is a search for a controlled
variable. One might at that point guess that rat was controlling for
getting to the food, using some aspect of the sound of the floor as
an input to the peerceptual function at some lower level. This is
where I think the next step might be to use the TCV.
Perhaps, but since both the nature and the reference values for
controlled variables change quite rapidly over time, I’m not sure
how useful it is in determinig how the organism operates. In the
case of the rat, it is the linkage between controlling for seeing
(or whatever) the food and sensing the floor sound that is
interesting, not that the rat was controlling for getting to the
food.
You asked for an example of reverse engineering, and I proposed the
TCV as such an example, since it teases out the linkages among what
is sensed, what is controlled, and the actions used in control.
Martin

···

[Adam Matic 2014.03.13.08.50]

          MT:

As for the actual running of the experiment, as opposed to
all the set-up and instruction phase, whether there is a
relation to the TCV depends on the experiment. A tracking
experiment such as I was doing in the 1960s in studies of
figural after-effect, simply assumes that the controlled
perception is teh one specified by the experiment (keep
those lines vertical, keep the rotating disk stationary,
and stuff like that). The other kind is a “what do you
perceive” kind of question-answer session, and the
experimenter has one or two ways of estimating whether the
subject is looking for what the experimenter is asking
about. In those studies, in contrast to the tracking
studies, a correlation between disturbance and output that
varies when the experimenter makes the task more or less
difficult is a good sign that the subject is doing the
right thing. That’s not the TCV, but it does something
rather similar.

          I'd prefer the label of TCV to be kept for the specific

program that has these components:

            1. Imagine an environmental variable (EV) that might be

functionally related to the putative controlled perception
PVP.

            2. Determine whether the subject has sufficient input

mechanisms for it to be possible for her to create the
PVP.

            3. Determine whether the subject has the means to

influence the EV.

            4. Apply a disturbance to the EV and see whether the

subject acts by the means you noted in step 3 to bring it
back to its presumed reference value.

            5. Do step 4 a lot (if it has to be separate discrete

trials) or continuously (if possible) and note the
relationship between the EV and your disturbance and/or
the relationships between the EV and the subject’s output
and/or the relationship between the subject’s output and
your disturbance. A good index of control is low
correlation between EV and the other two and a high
correlation between disturbance and output.

            6. Return to step1, imagining a different plausible EV,

and repeat the loop until some criterion is satisfied.

            7. Choose the EV with the best index of control from the

various passes through step 5.

AM:

          That is a very specific program. I understand the

concept of the TCV to be any procedure that tries to
determine what is being perceived and controlled by the
subject, such as The Coin Game or just asking “what are
you trying to do?”.

          All the time people try to find out what other people

and organisms are doing, they just don’t usually go at it
systematically.

          Even when they (we) do it systematically, it's not

conclusive because there always might be a slightly better
approximation of the CV.

          I like this R.P. Feynmans essay, it's been mentioned on

CSGnet before, Cargo Cult Science: http://neurotheory.columbia.edu/~ken/cargo_cult.html

          He is very critical of experiments in psychology, and

he gives an example of good science and I think it is an
example of the TCV, so I’ll just quote that part:


R.P. Feynman:

          [...] there have been many experiments running rats

through all kinds of mazes, and so on–with little clear
result. But in 1937 a man named Young did a very
interesting one. He had a long corridor with doors all
along one side where the rats came in, and doors along the
other side where the food was. He wanted to see if he
could train the rats to go in at the third door down from
wherever he started them off. No. The rats went
immediately to the door where the food had been the time
before.

            The question was, how did the rats know, because the

corridor was so beautifully built and so uniform, that
this was the same door as before? Obviously there was
something about the door that was different from the
other doors. So he painted the doors very carefully,
arranging the textures on the faces of the doors exactly
the same. Still the rats could tell. Then he thought
maybe the rats were smelling the food, so he used
chemicals to change the smell after each run. Still the
rats could tell. Then he realized the rats might be able
to tell by seeing the lights and the arrangement in the
laboratory like any commonsense person. So he covered
the corridor, and still the rats could tell.

            He finally found that they could tell by the way the

floor sounded when they ran over it. And he could only
fix that by putting his corridor in sand. So he covered
one after another of all possible clues and finally was
able to fool the rats so that they had to learn to go in
the third door. If he relaxed any of his conditions, the
rats could tell.

            Now, from a scientific standpoint, that is an

A-number-one experiment. That is the experiment that
makes rat-running experiments sensible, because it
uncovers that clues that the rat is really using-- not
what you think it’s using. And that is the experiment
that tells exactly what conditions you have to use in
order to be careful and control everything in an
experiment with rat-running.

            I looked up the subsequent history of this research.

The next experiment, and the one after that, never
referred to Mr. Young. They never used any of his
criteria of putting the corridor on sand, or being very
careful. They just went right on running the rats in the
same old way, and paid no attention to the great
discoveries of Mr. Young, and his papers are not
referred to, because he didn’t discover anything about
the rats. In fact, he discovered all the things you have
to do to discover something about rats. But not paying
attention to experiments like that is a characteristic
example of cargo cult science.


AM:

          If an experiment, be it in pyschophysics, animal

behavior, cognitive psychology, economics or anywhere
tries to find a controlled variable, then it is finding
something about the organism that is scientifically
useful.

          Are we still on the same topic? I don't know how I got

from measuring I/O properties of a control loop to the
TCV.

[Adam Matic 2014.03.13.16.10]

···
MT: Words again. I'd prefer to call :trying to determine what is being

perceived and controlled by the subject" a search for the controlled
variable rather than a test. After having narrowed down the search,
I would then do a test,if I wanted to be better assured that the
search had come close to the right answer. When you ask a subject
what he is trying to do, he may not know, or may not want to tell
you correctly, or may answer at a level different from the one that
interests you.

AM:

Always the words :). Sure, I guess the ‘search for the control variable’ is an appropriate term, covering a wider range of procedures than the TCV. I imagine there can be many valid varieties of the TCV, since it’s not always possible to get the human subject or an animal to do a tracking task of some kind.

MT: I imagine that it would be as rare for any search or test to find

the exact CV as it would be for a control unit to have zero error.
It might be so, but only by a fluke, and only in passing.

AM: That sounds like a good analogy. We could say that doing the TCV or the Search for the CV involves controlling the difference between a hypothetical CV and the ‘real’ CV.

MT: You asked for an example of reverse engineering, and I proposed the

TCV as such an example, since it teases out the linkages among what
is sensed, what is controlled, and the actions used in control.

AM:

I might have misunderstood something from your previous messages because I expected measurements of I/O properties of loop components.

Is there an example of a procedure that is not similar to the TCV (or the SCV) and is a solid scientific procedure that helps with reverse engineering control systems?

Or was the point that sometimes SR experiments give clues about the CV?

Adam

          AM:Even when they (we) do it systematically, it's not

conclusive because there always might be a slightly better
approximation of the CV.

[From Rick Marken (2014.03.13.1315 PDT)]

···

Adam Matic (2014.03.12.13.30 CET)–

RM: Sorry for the late reply. This post of yours was wonderful and makes it clear that studying intact control systems of any kind (living or artifactual) can’t be done by measuring the I/O characteristics of its components. As you point out, this is something that is understood by control engineers and should be something that is understood by psychologists who know that living systems are control systems. As you point out in later posts, psychologists are in the position of trying to reverse engineer control systems that have already been “built”, unlike control engineers, who are trying to build control systems from scratch, which, I guess, could be called forward engineering.

RM: A paper I wrote recently for what will become LCS IV is all about this distinction between forward and reverse engineering of control systems. I conclude that the most important difference between these approaches is in terms ofcontrolled variables. The forward engineer knows what variable(s) the system is to control and he has to build the system using I/O components that will result in a system that controls that variable well. The reverse engineer is dealing with a system that has already been built with I/O components that allow it to control well (well enough to survive,anyway) and he has to figure out what variables it is controlling. So the essential component of reverse engineering control systems is the test for controlled variables, a procedure that is not even a part of the tool box of the engineers who build control systems because they don’t need it; they already know exactly what variables the system is designed to control because they built it to control those variables.

RM: I believe that the psychologists who first applied control theory to understanding human behavior made the mistake of applying the forward engineering methods of control engineers to the task of reverse engineering living control systems. It was a drastic mistake that was eventually pointed out by Bill Powers (in his 1978 Psych Review paper, reprinted in LCS I), but the damage had already been done. So we still have psychologists trying to reverse engineer living control system by studying the I/O characteristics of these systems, something that, as you point out Adam, can’t even be done in an intact control system. I think the only way to fix this is to show psychologists the proper way to reverse engineer living control systems. And I think the best way this can be done is by starting to carry out Bill’s vision for a research program based on PCT, the one described in LCS I (p. 216-218).

Best regards

Rick


Richard S. Marken PhD
www.mindreadings.com
The only thing that will redeem mankind is cooperation.

                                               -- Bertrand Russell

AM: When control engineers need to tune a control system, they break the loop and measure it’s I/O characteristics. They do the same with the ‘plant’. Then they know enough about the system to make a good approximation of tuning parameters.

AM: I don’t see any non-invasive way of breaking the loop when dealing with living control systems. You can take out a muscle from someone and measure it’s I/O properties, or a neuron, or retinal cells. That’s fine if the living thing doesn’t mind. But if you try stimulus-response measurements while the component is in the loop you get nothing relevant. At least I can’t think of a way to get anything. As you show in the plot, there is always interaction from the loop.

[Martin Taylor 2014.03.13.15.28]

That is indeed a point I mention every few months. Even if you

could, why would you expect what they were controlling at that time
to be what they were controlling at the point you became interested?
Two questions, two different themes. I’ll take the second one first.
No, the point was that sometimes S-R experiments give information
about how the organism operates. The critical point is that there
must be no way that anything the subject does can feed back into
what the subject will perceive on the pathway being analyzed.
Now the first question. Psychophysical experiments provide limits on
the properties that can properly be used in control models. Suppose
there was an experiment (I’m making this up) that uses two different
configurations at different locations not too far apart, shown one
after the other with a short time-gap between them. Suppose that if
the time gap is negative or near zero, or if it is fairly large, you
see two distinct configurations, but with an intermediate time gap
you see one configuration smoothly moving and flowing into the
other. That would tell you something about the workings of the
perceptual system, perhaps about configuration and event
perceptions. (At least for the different perceptions with moderate
and large time gaps, this is actually the change of perception that
happens, as I can attest from personal experience, and has been
reported by Paul Kolers’s subjects several decades ago – it’s
amazing how two very different configurations can flow smoothly into
one another if the time gap is right.).
Now suppose in the experiment you determine the likelihood of seeing
one smooth change of configuration as opposed to two distinct
configurations as a function of the gap, and that the 50-50 time is
consistently different between two people. Now we use those same
subjects in a tracking study such as the one in Chapter 2 of LCSIII,
and we find that people who have a shorter 50-50 time gap also have
higher gain when their tracking is modelled for control of shape,
but not with the other two aspects. That would be a correlational
measure, often denigrated on CSGnet, but would it not suggest that
something about perception of configuration might relate the two
phenomena? Might it not suggest, perhaps, that the modelling might
be extended to partition the loop gain between the output and the
perceptual function rather than include it all in the output?
Remember that this experiment is simply made up. I have no idea
whether any of it would come out as I suggest. It’s just a prototype
of the kind of way an S-R experiment might be useful. Since it is
part of the Church of PCT catechism to deny that any S-R experiment
can be useful, I doubt that any such experiments have been done.
However, that doesn’t change the fact that parameters properly
measured can and should be used in PCT modelling, and that they
might well limit the kinds of structures in which control systems
are actually connected.
In any one S-R study, the component being measured includes the
entire pathway from sensory input to action output. You need a
variety of measures to determine which contributes most to your
measure. In psychophysical studies, it is recognized that even if “I
see it” and “No I didn’t see it” are very easily answered, as by
pushing a “Yes” or a “no” button or by saying “Yes” nad “No”,
nevertheless with most subjects you will never get more than around
99% correct no matter how easy the task. Since this doesn’t change
when the apparent difficulty of an extremely easy task changes, the
psychophysicist usually ascribes this limit to the output function.
On the other hand, when the easy task is made more difficult, the
99% correct goes down as the apparent difficulty increases. Since
the output mechanism hasn’t changed, that limit is ascribed to the
perceptual function. In this way, the gross subcomponents of the
large I-O component can often be teased apart.
Martin

···

On 2014/03/13 11:10 AM, Adam Matic
wrote:

[Adam Matic 2014.03.13.16.10]

            MT: Words again. I'd prefer to call :trying to determine

what is being perceived and controlled by the subject" a
search for the controlled variable rather than a test.
After having narrowed down the search, I would then do a
test,if I wanted to be better assured that the search
had come close to the right answer. When you ask a
subject what he is trying to do, he may not know, or may
not want to tell you correctly, or may answer at a level
different from the one that interests you.

AM:

          Always the words :). Sure, I guess the 'search for the

control variable’ is an appropriate term, covering a wider
range of procedures than the TCV. I imagine there can be
many valid varieties of the TCV, since it’s not always
possible to get the human subject or an animal to do a
tracking task of some kind.

            MT: You asked for an example of reverse engineering, and

I proposed the TCV as such an example, since it teases
out the linkages among what is sensed, what is
controlled, and the actions used in control.

AM:

          I might have misunderstood something from your previous

messages because I expected measurements of I/O properties
of loop components.

          Is there an example of a procedure that is not similar

to the TCV (or the SCV) and is a solid scientific
procedure that helps with reverse engineering control
systems?

          Or was the point that sometimes SR experiments give

clues about the CV?

So, Martin seems to acknowledge that S-R experiments are in no way ideal and are comprised in many ways, and that the TCV is the method of choice. I agree and I also agree with Martin that in these remnants of studies there may be gems that PCT can use. But I also feel that Rick’s purist stance, despite his stubborn refusal to acknowledge the worth of S-R research, brings the clarity of purpose that is needed to head the direction of science clearly towards what we learn through PCT. So, I really value these debates, a dose of Rick with a steadying reminder from Martin, and then a rejoinder from Rick. What a team! Almost akin to the hard and fast direction that Bill combined with his own openness to ideas even to the end.

I have read Rick’s chapter which makes exactly the right point we need to address at this point in time, and I am really excited to see Martin and Bruce’s work. I need to get on with my chapter!

All the best,

Warren

···

On 2014/03/13 11:10 AM, Adam Matic
wrote:

[Adam Matic 2014.03.13.16.10]

            MT: Words again. I'd prefer to call :trying to determine

what is being perceived and controlled by the subject" a
search for the controlled variable rather than a test.
After having narrowed down the search, I would then do a
test,if I wanted to be better assured that the search
had come close to the right answer. When you ask a
subject what he is trying to do, he may not know, or may
not want to tell you correctly, or may answer at a level
different from the one that interests you.

AM:

          Always the words :). Sure, I guess the 'search for the

control variable’ is an appropriate term, covering a wider
range of procedures than the TCV. I imagine there can be
many valid varieties of the TCV, since it’s not always
possible to get the human subject or an animal to do a
tracking task of some kind.

            MT: You asked for an example of reverse engineering, and

I proposed the TCV as such an example, since it teases
out the linkages among what is sensed, what is
controlled, and the actions used in control.

AM:

          I might have misunderstood something from your previous

messages because I expected measurements of I/O properties
of loop components.

          Is there an example of a procedure that is not similar

to the TCV (or the SCV) and is a solid scientific
procedure that helps with reverse engineering control
systems?

          Or was the point that sometimes SR experiments give

clues about the CV?

[Adam Matic 2014.03.14.16.00]

···

MT:

No, the point was that sometimes S-R experiments give information

about how the organism operates. The critical point is that there
must be no way that anything the subject does can feed back into
what the subject will perceive on the pathway being analyzed.

AM:

You are talking about breaking the loop after the output function? Do you have an example of an actual experiment in which this was done?

MT:

Now the first question. Psychophysical experiments provide limits on

the properties that can properly be used in control models. Suppose
there was an experiment (I’m making this up) that uses two different
configurations at different locations not too far apart, shown one
after the other with a short time-gap between them. Suppose that if
the time gap is negative or near zero, or if it is fairly large, you
see two distinct configurations, but with an intermediate time gap
you see one configuration smoothly moving and flowing into the
other. That would tell you something about the workings of the
perceptual system, perhaps about configuration and event
perceptions. (At least for the different perceptions with moderate
and large time gaps, this is actually the change of perception that
happens, as I can attest from personal experience, and has been
reported by Paul Kolers’s subjects several decades ago – it’s
amazing how two very different configurations can flow smoothly into
one another if the time gap is right.).

AM:

That looks to me like a search for the controlled variable (which is defined in the perceptual function). Sure, lots of things can be used when you’re trying to find out what people perceive, since there are so many kinds of variables that we can perceive, on so many levels. Disturbing intensities might be done simply, but disturbing system concepts might be more complex.

If we talk about scientific value of past experiments, I don’t see the what they could possibly say about the properties of the elements of the control system, other than whether a certain variable can or can not be perceived, and if done right, that is just TCV. So, I think each experiment would have to be carefully analysed to see if the study actually found out something about control systems.

Adam

[From Bruce Abbott (2014.03.14.1305 EDT)]

Adam Matic 2014.03.14.16.00 –

MT:

No, the point was that sometimes S-R experiments give information about how the organism operates. The critical point is that there must be no way that anything the subject does can feed back into what the subject will perceive on the pathway being analyzed.

AM:

You are talking about breaking the loop after the output function? Do you have an example of an actual experiment in which this was done?

MT:

Now the first question. Psychophysical experiments provide limits on the properties that can properly be used in control models. Suppose there was an experiment (I’m making this up) that uses two different configurations at different locations not too far apart, shown one after the other with a short time-gap between them. Suppose that if the time gap is negative or near zero, or if it is fairly large, you see two distinct configurations, but with an intermediate time gap you see one configuration smoothly moving and flowing into the other. That would tell you something about the workings of the perceptual system, perhaps about configuration and event perceptions. (At least for the different perceptions with moderate and large time gaps, this is actually the change of perception that happens, as I can attest from personal experience, and has been reported by Paul Kolers’s subjects several decades ago – it’s amazing how two very different configurations can flow smoothly into one another if the time gap is right.).

AM:

That looks to me like a search for the controlled variable (which is defined in the perceptual function). Sure, lots of things can be used when you’re trying to find out what people perceive, since there are so many kinds of variables that we can perceive, on so many levels. Disturbing intensities might be done simply, but disturbing system concepts might be more complex.

If we talk about scientific value of past experiments, I don’t see the what they could possibly say about the properties of the elements of the control system, other than whether a certain variable can or can not be perceived, and if done right, that is just TCV. So, I think each experiment would have to be carefully analysed to see if the study actually found out something about control systems.

BA: Keep in mind that most perceptions are not controlled perceptions; for example, you may see many people walking along a street. The retinas of your eyes convert the patterns of illumination produced by this scene into neural signals that are analyzed at higher levels within the brain to produce perceptions of the people, street, other objects. You are not controlling the expressions on those people’s faces, nor the style of clothing they are wearing, to mention just a couple aspects of the scene that you are not controlling. It is possible to test the perceptual system to determine what basic properties of the image go into the analysis that produces those recognizable objects, their apparent motions, colors, shading, and so on. Such tests can be done by giving participants control over those aspects, but it is not necessary to do so. Thus I disagree with your statement above that investigating perceptual functions is “like a search for the controlled variable” –unless by “like” you mean that it shares certain features of the Test. To control a perception you must be able to perceive it; investigations of perceptual functions, whether or not under control during the investigation, can help to establish exactly how perceptions are created by the relevant sensory and brain systems.

Bruce

HI Bruce, but what about when we think of those perceptual properties during development, not as they are put together in this current, unique scene, but how they were first perceived in development. The first perception of a happy face; the first perceptions of colour; perceptions of velocity - potentially the closed loop was important, maybe even essential to the formation of the perceptual functions that allows us to perceive those features now, as adults, and take it for granted, but as a child those perceptions might have been non-existent until the child started to learn how to control them, or at least influence them (e.g. to make people smile; to move an object at different speeds), etc?
Warren

···

On Fri, Mar 14, 2014 at 5:06 PM, Bruce Abbott bbabbott@frontier.com wrote:

[From Bruce Abbott (2014.03.14.1305 EDT)]

Adam Matic 2014.03.14.16.00 –

MT:

No, the point was that sometimes S-R experiments give information about how the organism operates. The critical point is that there must be no way that anything the subject does can feed back into what the subject will perceive on the pathway being analyzed.

AM:

You are talking about breaking the loop after the output function? Do you have an example of an actual experiment in which this was done?

MT:

Now the first question. Psychophysical experiments provide limits on the properties that can properly be used in control models. Suppose there was an experiment (I’m making this up) that uses two different configurations at different locations not too far apart, shown one after the other with a short time-gap between them. Suppose that if the time gap is negative or near zero, or if it is fairly large, you see two distinct configurations, but with an intermediate time gap you see one configuration smoothly moving and flowing into the other. That would tell you something about the workings of the perceptual system, perhaps about configuration and event perceptions. (At least for the different perceptions with moderate and large time gaps, this is actually the change of perception that happens, as I can attest from personal experience, and has been reported by Paul Kolers’s subjects several decades ago – it’s amazing how two very different configurations can flow smoothly into one another if the time gap is right.).

AM:

That looks to me like a search for the controlled variable (which is defined in the perceptual function). Sure, lots of things can be used when you’re trying to find out what people perceive, since there are so many kinds of variables that we can perceive, on so many levels. Disturbing intensities might be done simply, but disturbing system concepts might be more complex.

If we talk about scientific value of past experiments, I don’t see the what they could possibly say about the properties of the elements of the control system, other than whether a certain variable can or can not be perceived, and if done right, that is just TCV. So, I think each experiment would have to be carefully analysed to see if the study actually found out something about control systems.

BA: Keep in mind that most perceptions are not controlled perceptions; for example, you may see many people walking along a street. The retinas of your eyes convert the patterns of illumination produced by this scene into neural signals that are analyzed at higher levels within the brain to produce perceptions of the people, street, other objects. You are not controlling the expressions on those people’s faces, nor the style of clothing they are wearing, to mention just a couple aspects of the scene that you are not controlling. It is possible to test the perceptual system to determine what basic properties of the image go into the analysis that produces those recognizable objects, their apparent motions, colors, shading, and so on. Such tests can be done by giving participants control over those aspects, but it is not necessary to do so. Thus I disagree with your statement above that investigating perceptual functions is “like a search for the controlled variable” –unless by “like” you mean that it shares certain features of the Test. To control a perception you must be able to perceive it; investigations of perceptual functions, whether or not under control during the investigation, can help to establish exactly how perceptions are created by the relevant sensory and brain systems.

Bruce


Dr Warren Mansell
Reader in Psychology
Cognitive Behavioural Therapist & Chartered Clinical Psychologist

School of Psychological Sciences
Coupland I
University of Manchester
Oxford Road
Manchester M13 9PL
Email: warren.mansell@manchester.ac.uk

Tel: +44 (0) 161 275 8589

Website: http://www.psych-sci.manchester.ac.uk/staff/131406

See teamstrial.net for further information on our trial of CBT for Bipolar Disorders in NW England

The highly acclaimed therapy manual on A Transdiagnostic Approach to CBT using Method of Levels is available now.

Check www.pctweb.org for further information on Perceptual Control Theory

[Martin Taylor 2014.03.14.12.20]

There is a VERY large literature on such experiments, going back

over a century. The experimenter (human or machine) did something
that might or might not influence some perception in the subject. Of
course, the subject could not control THAT perception, since there
would be no opportunity to influence the input that the experimenter
had already provided. But because the subject was controlling a
DIFFERENT and unrelated perception (perhaps a perception of the
experimenter’s perception that the subject was following
instructions), the subject does tell the experimenter what he
perceives. The I-O component from experimenter’s input presentation
to the subject’s answer is a small component of the larger loop that
goes through both subject and experimenter.
It is true that very similar studies are done in which the subject’s
response could influence what the experimenter provided as input on
the next trial, and the literature on them may be as large as that
on experiments in which it doesn’t. Continuing the line of thought,
there are tracking studies that look at the same kinds of question.
The first kind of study is exemplified by a “2 Alternative Forced
Choice”. Here’s a low level example and a high-level example.
At a low level, the experimenter presents two bursts of noise
separated by a short gap, identical except that a 500 Hz tone is
embedded in one noise burst. The subject says “First” or “Second”.
When the tone is loud enough compared to the noise, the subject is
correct around 99% of the time, but in trials using much softer
tones, the subject gets less and less correct, until there comes a
point at which only 50% of the answers correspond to the interval in
which the tone was presented. Whether the tone is presented loud or
soft is quite independent of the subject’s earlier answers. For such
a low-level auditory study, a mathematical analysis is available to
describe the best possible performance achievable, and very highly
trained subjects usually perform within 4 db of this ideal under
different presentation conditions. I once did such a study with two
subjects over a whole summer, recording over a million trials from
each. They got feedback after each trial as to whether they had been
correct, and their performance improved and was continuing to
improve over the entire summer. I take this as evidence of
reorganization occurring within a very low-level perceptual process.
At a high level, the “experimenters” are two political parties, and
the experimental input presentation (I) consists of everything the
voter learns about the parties. The output (O) is a vote “A”, “B”,
or “don’t care” (don’t bother voting). The choice of vote cannot
influence the input, but the relationship between input and output
may suggest something about the internal operations of the voter.
The second kind of study is sometimes called “adaptive”. The
subject’s answer does influence the difficulty of the next input
presentation. In a trivial example, if the answer is correct, the
next presentation is made more difficult, and vice-versa. Most
actual experiments use a statistical criterion based on the
correctness of several previous answers, but however it is done, the
subject’s commandable choice of “First” or “Second”, does not
influence whether the next input would correctly be called “First”.
Only the correctness of the answer influences the difficulty of the
next presentation, and the correctness of the answer is something
the subject can perceive only when the correct answer is obvious.
Since the correctness of the answer is a perception in the
experimenter, not the subject, and the difficulty of the next
presentation is determined by the experimenter, it is the
experimenter who is controlling a perception of the difficulty of
the task, and the subject’s I-O pathway is an essential component of
this control loop. Of course, since the situation is formally
identical to a measurement of weight using a scale balance, the
“experimenter” could well be an algorithm, but the subject would
still be controlling for following instructions in a way he would
perceive to be satisfactory to the human who set up the experiment
(conceivably himself – I was a subject in many of my auditory
experiments).
The third kind of study seems on the surface to be a continuum
version of the second, but there is a subtle difference. The
presentation and the response are the same, but it is the subject
who determines whether the task is “too easy” or “too hard”, and
controls that perception by directly adjusting the difficulty of the
task. Bekesy audiometry is an example.
The three different kinds of experiment give very much the same, but
not quite the same results, especially when the task is very
difficult. If you are interested in this, you could look at
What perception is being controlled here? The subject is presented
with a sequence of two configurations and sees whatever that
sequence produces in his perceptual system. He doesn’t have to
answer anything, and there’s no way he can alter what was presented
or what will be presented if and when there might be another trial.
Who is searching for a controlled variable? The experimenter depends
on the fact that the subject cannot control whether a presentation
is seen as a smooth flow between two configurations or as two
discrete configurations. The subject sees it the way it seems to be.
Why does the subject tell the experimenter whether is is a smooth
flow or two discrete configurations? Sometimes we say we answer a
question “to be nice”, which is to control a perception of how we
are seen by others, and I guess that is a perception the subject is
likely to control. Closer to the actual experiment, the subject
probably is controlling for perceiving himself to be following
instructions, which include looking carefully at the presentation
and answering honestly what he sees.
Does the experimenter do the TCV to determine whether the subject is
controlling this “following instruction” perception? Good
experimenters do, without knowing that this is actually what they
are doing.
I don’t understand the leap you make between determining what people
are able to perceive and the TCV. All the TCV can do is home in on
one particular perception a person is actually controlling at some
moment when that choice is guaranteed not to change. What would
Figure 2-2 of LCSIII look like if the subject frequently and
randomly changed which aspect of the red football was being
controlled during the run? Why does it not look like that? Because
Bill asked the subject to follow an instruction to control the same
perception for the whole duration of the run, and the subject
controlled another perception – to perceive himself as following
instructions – while controlling a perception of the red football.
Without that instruction, could the TCV operate in that display
situation? Without that instruction, would what the subject had the
ability to perceive change? I think both questions must be answered
“no”.
I’ll return your “I don’t see”, and say “If we talk about scientific
value of the TCV, I don’t see the what it could possibly say about
the properties of the elements of the control system, other than
whether a certain variable can or can not be controlled.”
If something can’t be perceived, it can’t be controlled. Since the
first step in doing a Test for the Controlled Variable is to
determine whether the candidate environmental variable can be
perceived, it seems to me that the psychophysics is a prerequisite
for the TCV, even if it is done informally, by saying “of course he
can see it if he wants.”
Understanding the properties of capacitors, resistors, diodes, and
complexes such as op-amps may not tell you anything about the
circuits in which they are employed, but it sure tells you about how
they might be employed. The situation seems to me to be the same. Martin

···

On 2014/03/14 11:20 AM, Adam Matic
wrote:

[Adam Matic 2014.03.14.16.00]

MT:

            No, the point was

that sometimes S-R experiments give information about
how the organism operates. The critical point is that
there must be no way that anything the subject does can
feed back into what the subject will perceive on the
pathway being analyzed.

AM:

          You are talking about breaking the loop after the

output function? Do you have an example of an actual
experiment in which this was done?

     <Taylor, M.M., Forbes, S.M., and Creelman, C.D.,

* PEST
reduces bias in forced choice psychophysics*
.
J. Acous. Soc. Amer., 1983, 74,
1367-1374>, but I don’t think we need to discuss it further
here.

MT:

            Now the first

question. Psychophysical experiments provide limits on
the properties that can properly be used in control
models. Suppose there was an experiment (I’m making this
up) that uses two different configurations at different
locations not too far apart, shown one after the other
with a short time-gap between them. Suppose that if the
time gap is negative or near zero, or if it is fairly
large, you see two distinct configurations, but with an
intermediate time gap you see one configuration smoothly
moving and flowing into the other. That would tell you
something about the workings of the perceptual system,
perhaps about configuration and event perceptions.

AM:

          That looks to me like a search for the controlled

variable (which is defined in the perceptual function).

          If we talk about scientific value of past experiments,

I don’t see the what they could possibly say about the
properties of the elements of the control system, other
than whether a certain variable can or can not be
perceived, and if done right, that is just TCV.

          So, I think each experiment would have to be carefully

analysed to see if the study actually found out something
about control systems.

[From Bruce Abbott (2014.03.14.1505 EDT)]

···

From: Control Systems Group Network (CSGnet) [mailto:CSGNET@LISTSERV.ILLINOIS.EDU] On Behalf Of Warren Mansell
Sent: Friday, March 14, 2014 1:31 PM
To: CSGNET@LISTSERV.ILLINOIS.EDU
Subject: Re: Measuring Input-Output Characteristics of Components of a Closed Loop: Redux

WM: HI Bruce, but what about when we think of those perceptual properties during development, not as they are put together in this current, unique scene, but how they were first perceived in development. The first perception of a happy face; the first perceptions of colour; perceptions of velocity - potentially the closed loop was important, maybe even essential to the formation of the perceptual functions that allows us to perceive those features now, as adults, and take it for granted, but as a child those perceptions might have been non-existent until the child started to learn how to control them, or at least influence them (e.g. to make people smile; to move an object at different speeds), etc?

BA: Perhaps.

BA: When Bill developed PCT, he was trying to develop a system that, from a very minimal starting point, could develop all the different levels of control to be found in the adult individual. The key was Ashby’s essential variables and the reorganizing system that could find solutions to the problems of control needed to keep those essential variables within survivable ranges, plus the notion that higher-level systems could emerge from the same process once there were lower-level systems whose references could be set by the higher-level systems as their means to control perceptions at their own level. My own view is that genetically/biochemically orchestrated development provides far more structure than seems to be allowed under PCT. Bill and I discussed this issue privately some time ago during one of our Skype sessions, and he was willing to agree that this may be the case, although of course he’d have to see the evidence for that. Current understanding is that basic perceptual mechanisms (such as those in vision that create perceptions of objects and their apparent properties such as color and motion) are present at birth but will deteriorate if not given adequate input during the first years – thus the importance of early restoration of vision to those born with cataracts or “lazy eye.” On the other hand, having experience interacting with those products of perception must be crucially important as well. A person who has learned the shape of a ball by touch usually cannot initially relate that experience to the appearance of the ball if deprived of vision from birth and having vision subsequently made available. We learn from experience what to expect of the things we perceive and how perceptions from different senses relate, especially when attempting to control them.

Years ago when my daughter was born, the received wisdom was that newborns were unable to make any sense of the visual world around them, that it was just a blur of shapes and colors. This notion was quickly put to rest for me when I watched her move her eyes and head to track the movement of a nurse only minutes after birth. Clearly, more is available to the newborn than perception of and control over intensities.

How much is “given” and how much must be acquired is still an open question. Do we learn to judge depth or distance by interacting with objects and moving through a scene, with attendant changes in perspective and differences in apparent motion of the objects at different visual depths? Or are our visual systems already preprogrammed to render depth in the visual scene, requiring experience only for “fine tuning”? I don’t think we know for sure as yet, but the answer is likely to be “it’s a bit of both” – nature and nurture (genetics and reorganization) interacting during development in complex ways to produce the end product.

Bruce

[From Rick Marken (2014.03.14.1220)]
Warren Mansell writes:

WM: But I also feel that Rick's purist stance, despite his stubborn refusal to acknowledge the worth of S-R research, brings the clarity of purpose that is needed to head the direction of science clearly towards what we learn through PCT.

RM: I'm glad that you appreciate my point of view but I've got to tell you that referring to it as "purist" makes me very uncomfortable. I think of purist stances as those taken by ideologues (or racists) and such stances encourage the purging (or worse) of the "impure". I think of myself as taking a scientific rather than an ideological stance -- the same scientific stance taken by Powers in the 1978 Psych Review article where he showed that the S-R approach to studying living control systems tells you very little about what is important about the behavior of these systems: the perceptual variables that are being controlled when we observe various behaviors. And I have no interest in purging (or worse) people who don't agree with the "stance" Powers took in that brilliant paper.
I don't think Bruce and Martin, the main defenders of the S-R approach to research, are "impure". I think they are simply wrong. I can understand why they are defending the S-R approach but I think their stance is impeding (or, at least, not helping with) the development of a PCT-based science of life. So I will continue to try to convince them that they should drop the S-R stuff , not because the S-R stuff is impure but because it's misguided (as explained by Bill in the 1978 Psych Review paper). And I will continue to encourage them (and other researchers) to start developing ideas for research based on PCT so that we can start implementing Bill's vision of a science of living control systems, as described in the "Cybernetic Model for Research" paper in LCS I.
Best
Rick

···

--
Richard S. Marken PhD
<http://www.mindreadings.com>www.mindreadings.com
The only thing that will redeem mankind is cooperation.
-- Bertrand Russell