It is the cause. It is the cause (was Re: Positive Feedback etc)

[Martin Taylor 2011.07.06.20.15]

···

On 2011/07/6 10:41 AM, Richard Marken wrote:

[From Rick Marken (2011.07.06.0740)]

        Martin Taylor

(2011.07.05.13.19)

            BP: I recall that the issue was

whether the button presses had to be perceived in
relation to a light turning on. Martin said it was
sufficient to imagine the button press, and that was
what made the upper system open-loop.

        MT: I don't know why my name got mixed up in this, but I

never said what either of you attribute to me. All the upper
loops in my descriptions were full control loops.

  Your name came up because you have been the inspiration for most

of the research I’ve done over the last 5 years or so. The
reaction time experiment I have been discussing with Bill is a
direct result of the discussions we had a couple years ago about
the Schouten experiment. So thank you for the inspiration and for
being a consistent defender of the status quo in psychological
research, a status quo that I am dedicated to undoing;-)

I'm glad you put in the smiley, because otherwise I might have

believed you meant what you wrote in your last sentence.

Martin

[From Bill Powers (2011.07.06.1840 MDT)]

Martin Taylor 2011.07.06.20.05 –

BP earlier: Note that the
perturbation due to the disturbance does not just “happen” at
time T, the time when the disturbing variable instantaneously changes
state. The sensory event is not a point-event occurring at a single
instant. It extends over time and takes time to reach its new state –
and while it’s reaching the new state, the perception and the action are
continually changing and altering the effect of the disturbance and
affecting the perception. When we examine the process on a time-scale at
which the dynamics of control are visible, we no longer see “events
happening.” We see variables changing.

MT:All of which I understand very well, and agree with completely. I hope
you will agree with me that it is irrelevant to what happens in the path
between the disturbance and the first effect of the disturbing influence
within the control loop.

BP: I see that I should have spelled out what I think are the
implications. The Schouten experiment purports to measure the open-loop
buildup of information in the time after some minimum delay after the
onset of the light. But if there are feedback effects developing while
the error rate is changing, what is being measured is not a pure effect
of the sensory input from the light. It’s partly an effect from the
sensory input from touching and pressing on the button. The perception
under control is not just the light; it’s a relationship between the
light and the button pressing (the “correct answer”). The loop
is open only from the turnon of the light through the place where the
intensity or sensation of light is perceived. After that, or a level or
so above (I wouldn’t press for any particular model there) the perception
is no longer a function of the light alone. I don’t know exactly what
relationship is being perceived, but it’s clearly a relationship and not
just a sensation because it is drawn from two independently variable
sources, vision and touch.

As I said, these are model-dependent deductions and the model hasn’t been
verified. But this is reason enough to be conservative about accepting
any deductions about information processing. My initial objections, two
years ago, were based on exactly this analysis, though I couldn’t
communicate it very clearly. I could see the possibility that the loop
involved a higher-order perception in which light intensity was only one
component; the loop was closed not through effects on light intensity,
but through effects on the other component of the relationship under
control at the higher level. I thought then as I do now, that the
situation isn’t as simple and clear-cut as you seem to think it is –
partly because you see light intensity as the only input at the higher
level. If you’re right about that, then your deductions are probably all
right. But you could be wrong, so the model has to be established before
taking the reasoning any farther.

I don’t know how to describe my position any better than that.

Best,

Bill P.

Best,

Bill P.

[Martin Taylor 2011.07.06.23.01]

[From Bill Powers (2011.07.06.1840 MDT)]

Martin Taylor 2011.07.06.20.05 --

BP earlier: Note that the perturbation due to the disturbance does not just "happen" at time T, the time when the disturbing variable instantaneously changes state. The sensory event is not a point-event occurring at a single instant. It extends over time and takes time to reach its new state -- and while it's reaching the new state, the perception and the action are continually changing and altering the effect of the disturbance and affecting the perception. When we examine the process on a time-scale at which the dynamics of control are visible, we no longer see "events happening." We see variables changing.

MT:All of which I understand very well, and agree with completely. I hope you will agree with me that it is irrelevant to what happens in the path between the disturbance and the first effect of the disturbing influence within the control loop.

BP: I see that I should have spelled out what I think are the implications. The Schouten experiment purports to measure the open-loop buildup of information in the time after some minimum delay after the onset of the light.

True, if one interprets your "build-up of information" charitably, and if one forgives the pejorative "purports to measure".

I would word it a little differently, to say that the mutual information between disturbance and output follows a time course calculable from the experimental data. If the output action is executed before X msec after the stimulus onset, the mutual information is zero. If the output action is executed at time T later than X msec, the mutual information is linearly proportional to T-X.

In itself, this says only one thing about the path from the disturbance to the lowest-level control loop in which the perception of the event is used. That one thing is that the channel capacity of this path is at least as much as the rate of growth of the measured mutual information. It could be greater, but it cannot be less.

But if there are feedback effects developing while the error rate is changing, what is being measured is not a pure effect of the sensory input from the light. It's partly an effect from the sensory input from touching and pressing on the button. The perception under control is not just the light; it's a relationship between the light and the button pressing (the "correct answer"). The loop is open only from the turnon of the light through the place where the intensity or sensation of light is perceived. After that, or a level or so above (I wouldn't press for any particular model there) the perception is no longer a function of the light alone. I don't know exactly what relationship is being perceived, but it's clearly a relationship and not just a sensation because it is drawn from two independently variable sources, vision and touch.

As I said, these are model-dependent deductions and the model hasn't been verified. But this is reason enough to be conservative about accepting any deductions about information processing. My initial objections, two years ago, were based on exactly this analysis, though I couldn't communicate it very clearly. I could see the possibility that the loop involved a higher-order perception in which light intensity was only one component; the loop was closed not through effects on light intensity, but through effects on the other component of the relationship under control at the higher level. I thought then as I do now, that the situation isn't as simple and clear-cut as you seem to think it is -- partly because you see light intensity as the only input at the higher level. If you're right about that, then your deductions are probably all right. But you could be wrong, so the model has to be established before taking the reasoning any farther.

I don't know how to describe my position any better than that.

And I don't know how to describe my position any better than to point out that even if all your hypotheses are correct, they are completely irrelevant to the fact that the capacity of the pre-control-system part of the path from event to action is at least as great as the rate of growth of the mutual information between event and action. The only effect that the control systems could possibly have is to introduce uncertainties through the "Stuff Happens" part of the connection between event and action output. The effect of that would be to reduce the rate of gain of mutual uncertainty between event and output. It could not increase it. However, in the Schouten experiment, the linearity of the increase suggests that if uncertainty is introduced by the "Stuff" that "happens", either it isn't much or it is simply proportional to the existing mutual information, the latter being rather difficult to describe in a model.

To model "Stuff Happens" is an interesting exercise in itself, which is why I followed you in discussing that aspect in the earlier discussion. I'm not interested in pursuing it at the moment, and I'm not going to let your repeated introduction of it distract me from the simple point that if the history of A influences the current value of B but the history of B does not affect the current value of A, the path between A and B is truly "open-loop", no matter how many control processes constitute part of that path.

···

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

The only reason I joined this thread was that Rick said

[RM] But how do I deal with complaints like Martin's, who says that, sure, the cursor control system is closed loop but the reaction time behavior is open loop? It is possible to develop a model that is open loop for the reaction time and closed loop for the cursor that deals with the disturbance just fine.

and you responded with

[BP] I recall that the issue was whether the button presses had to be perceived in relation to a light turning on. Martin said it was sufficient to imagine the button press, and that was what made the upper system open-loop.

I wished to remind you that I said no such thing, and that in all my attempts at models of the process of generating a response, all the control loops were closed. The only open-loop element was that the response could not influence the prior event.

As I said, I have no interest now in reopening the question of modelling the generation of a response, since in the earlier discussion you got your position entrenched to the point where you actually said you wished you had never thought of the idea that reference values were generated from an associative memory, despite all the wonderful consequences that follow from working through the consequences of this idea.

All I want at this point is to set the record straight about what I did say about open-loop. I did not say that "reaction time behaviour is open-loop" and I did not say "the upper system is open-loop". At the same time, I wanted to reiterate what I did say in respect to what is NECESSARILY open-loop in a standard psychophysical experiment.

You really don't have to throw out all the babies when you throw out the bathwater. PCT is strong enough not to need the knee-jerk reaction that everything NIH is garbage.

Martin

[From Rick Marken (2011.07.06.2135)]

[Martin Taylor 2011.07.06.20.15]

  RM: So thank you for the inspiration and for

being a consistent defender of the status quo in psychological
research, a status quo that I am dedicated to undoing;-)

MT: I'm glad you put in the smiley, because otherwise I might have

believed you meant what you wrote in your last sentence.

It was my devilish winking smiley.

Best

Rick

···


Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com

[From Bill Powers (2011.07.07.0530 MDT)]

Martin Taylor 2011.07.06.23.01

BP: I see that I should have
spelled out what I think are the implications. The Schouten experiment
purports to measure the open-loop buildup of information in the time
after some minimum delay after the onset of the light.

MT: True, if one interprets your “build-up of information”
charitably, and if one forgives the pejorative “purports to
measure”.

I would word it a little differently, to say that the mutual information
between disturbance and output follows a time course calculable from the
experimental data.

I agree that something follows a time course calculable from the
experimental data. Recalling a bit more from our discussion two years
ago, I proposed that what was being measured was the rise of a perceptual
signal relative to the noise level, with the consequent increase in
signal-to-noise ratio with time and an increasing probability that the
signal is above threshold at the moment of the beep signalling the
sampling time (I just remembered that part too). I believe that,
too, would result in a linear rise of d’^2 after the minimum
delay.

This may mean, of course, that your measure of mutual information
translates directly into my measure of probable detection. If that is
true, then the question is only one of preference – whether one prefers
the information-theoretic calculation, or my more direct measure of
physical variables. I prefer the direct measure because I can imagine a
way to model it that could reproduce the phenomenon (maybe). You,
perhaps, know of a way to model mutual information that would also
reproduce the phenomenon in a non-tautological way.

The whole confused argument might be cleared up if I were to develop that
model so we could experiment with it, I to try to make it reproduce the
Schouten data, and you to show that the same information-theoretic
calculations reproduce the same measures of mutual information for the
behavior of my model. I have an idea that that might be the outcome, in
which case the argument would come to an end. I don’t know when I can get
to that modeling, but if you think it would be worth while I will put it
on the to-do list.

Best,

Bill P.

[Martin Taylor 2011.07.08.11.21]

[From Bill Powers (2011.07.07.0530 MDT)]

    Martin Taylor

2011.07.06.23.01

      BP: I see that I

should have
spelled out what I think are the implications. The Schouten
experiment
purports to measure the open-loop buildup of information in
the time
after some minimum delay after the onset of the light.

    MT: True, if one interprets your "build-up of information"

charitably, and if one forgives the pejorative “purports to
measure”.

    I would word it a little differently, to say that the mutual

information
between disturbance and output follows a time course calculable
from the
experimental data.

  I agree that *something* follows a time course calculable

from the
experimental data. Recalling a bit more from our discussion two
years
ago, I proposed that what was being measured was the rise of a
perceptual
signal relative to the noise level, with the consequent increase
in
signal-to-noise ratio with time and an increasing probability that
the
signal is above threshold at the moment of the beep signalling the
sampling time (I just remembered that part too). I believe that,
too, would result in a linear rise of d’^2 after the minimum
delay.

That was indeed the original assumption, except for the insertion of

the notion of “threshold”, which is a concept that belongs to some,
but not all, theoretical models to account for the data. The actual
data were the probability of a correct selection of the two possible
answers. Mathematically, those can be converted into a measure d’,
and since d’^2 is 8*ln(2) times mutual information, into a measure
of mutual information. That is the mathematical route I used in my
original Acta Psychologica article on the Schouten study.

The theoretical ideal for the value of d'^2 is 2*(total signal

energy)/(Noise power per unit bandwidth) where those parameters can
be determined. In cases in which they can be determined, highly
practiced human subjects generally perform some 4-6 db worse. In the
case of the Schouten display of lights, I don’t see how one would
compute noise power per unit bandwidth, but one can say that total
signal energy is proportional to the duration of the light, assuming
an instantaneous onset. This would mean that d’^2 (and mutual
information between light choice and answer selection) for the ideal
observer would increase linearly with time.

Without explicit modelling, the appearance is that it takes time for

any information at all to pass from the physical event through the
perceptual apparatus and all the “Stuff Happens” mechanism of answer
construction, but the rate of information throughput is thereafter
uniform. The data are consistent with this naive assumption, and any
acceptable model of the “Stuff” must produce the same result.

  This may mean, of course, that your measure of mutual information

translates directly into my measure of probable detection. If that
is
true, then the question is only one of preference – whether one
prefers
the information-theoretic calculation, or my more direct measure
of
physical variables.

They are actually the same.

Martin

[From Bill Powers (2011.07.08.1140 MDT)]

Martin Taylor 2011.07.08.11.21

MT: The theoretical ideal for the value of d'^2 is 2*(total signal energy)/(Noise power per unit bandwidth) where those parameters can be determined. In cases in which they can be determined, highly practiced human subjects generally perform some 4-6 db worse. In the case of the Schouten display of lights, I don't see how one would compute noise power per unit bandwidth, but one can say that total signal energy is proportional to the duration of the light, assuming an instantaneous onset. This would mean that d'^2 (and mutual information between light choice and answer selection) for the ideal observer would increase linearly with time.

BP: Are those db 20*log10(x)? That would make the decrement in performance only a factor of 1.5 to 2, which is easily absorbable into measurement errors in experiments like this. How much of the reaction time is simply the time it takes a limb or a finger to go into motion, touch the button, and increase the pressure enough to make the contact close? At what level (order of control) is the noise that matters? How is frequency of impulses related to signal energy? Lots of room for slop there.

I don't see how the noise level could be measured, either, since it would be neural noise, not noise in the light itself. But all that means is that you would adjust the model's noise level until the performance matches the real performance. You would then take that noise level as the model-dependent measure of the system noise (and in your case, the computation of mutual information as the model-dependent measure of the same construct in the real system). That's the same idea behind the way we compute output gain, delay, reference level, and so on for a tracking model. We're saying that IF the system is organized as the model proposes, THEN these would be the values of its parameters. That's about as far as we can go until neurological technology catches up to our questions and we find out how it's really organized in there.

I hadn't realized that you also considered a signal increasing relative to the noise level, or that "mutual information" is another way of representing the probability of detection. It's too bad that a term like "information" was picked for that measure, because the word already has a lot of other meanings that don't apply in this situation. Isn't there some more neutral term that could be used? We keep running into this problem. There are more phenomena than words, I guess.

BP earlier: This may mean, of course, that your measure of mutual information translates directly into my measure of probable detection. If that is true, then the question is only one of preference -- whether one prefers the information-theoretic calculation, or my more direct measure of physical variables.

MT: They are actually the same.

BP: "They" covers several possible referents. Do you mean that an information-theoretic calculation is the same as a direct measure of physical variables?

Best,

Bill P.

[Martin Taylor 2011.07.08.14.33]

[From Bill Powers (2011.07.08.1140 MDT)]

Martin Taylor 2011.07.08.11.21

MT: The theoretical ideal for the value of d'^2 is 2*(total signal energy)/(Noise power per unit bandwidth) where those parameters can be determined. In cases in which they can be determined, highly practiced human subjects generally perform some 4-6 db worse. In the case of the Schouten display of lights, I don't see how one would compute noise power per unit bandwidth, but one can say that total signal energy is proportional to the duration of the light, assuming an instantaneous onset. This would mean that d'^2 (and mutual information between light choice and answer selection) for the ideal observer would increase linearly with time.

BP: Are those db 20*log10(x)? That would make the decrement in performance only a factor of 1.5 to 2, which is easily absorbable into measurement errors in experiments like this.

Not so easily absorbed when you are making many hundreds of observations (I made well over a million with my best-trained subject).

How much of the reaction time is simply the time it takes a limb or a finger to go into motion, touch the button, and increase the pressure enough to make the contact close?

I have no idea. As long as I have had anything to do with psychophysics, I have never thought reaction time experiments told you much of anything, because of the factors you mention, and because there's a confounding between the subject's task to be fast and the task to be accurate, which (now I know about PCT) presumably translates into a conflict in the "Stuff Happens" complex of control loops.

You realize, I suppose, that the Schouten experiment is not a reaction time experiment? It was conceived as a way of avoiding the problems that are inherent in reaction-time experiments, and at the same time being able to deal with the different detectabilities of signals that all would yield 100% correct answers (within experimental limits) in conventional detection or discrimination experiments.

At what level (order of control) is the noise that matters? How is frequency of impulses related to signal energy? Lots of room for slop there.

We don't know. It's my guess that these questions relate to the reason that even the most highly trained and skilled observers remain so much worse than the ideal observer. They don't seem relevant to the interpretation of the mutual information rate as a lower bound on the capacity of the initial perceptual pathway to the first control loop. They probably are relevant to attempts to model the pathway to account for the data.

I don't see how the noise level could be measured, either, since it would be neural noise, not noise in the light itself.

I agree. As I said, we don't have the information necessary to define an ideal observer for the Schouten presentation. My comments about the relation between human and ideal performance reflect only those situations of which I am aware for which an ideal observer can be defined. I probably should have mentioned that the 2E/No ideal that I mentioned applies to one specific situation, when the observer knows the signal waveform exactly (though not whether the signal is present at all). Other values apply for different conditions. Mostly (maybe all) of those that we can compute are in acoustic detection. For the Schouten experiment, all we can say is that whatever the noise level, the energy in the signal increases linearly with the time the light is on before the decision must be made.

But all that means is that you would adjust the model's noise level until the performance matches the real performance. You would then take that noise level as the model-dependent measure of the system noise (and in your case, the computation of mutual information as the model-dependent measure of the same construct in the real system). That's the same idea behind the way we compute output gain, delay, reference level, and so on for a tracking model. We're saying that IF the system is organized as the model proposes, THEN these would be the values of its parameters. That's about as far as we can go until neurological technology catches up to our questions and we find out how it's really organized in there.

Yes.

I hadn't realized that you also considered a signal increasing relative to the noise level, or that "mutual information" is another way of representing the probability of detection. It's too bad that a term like "information" was picked for that measure, because the word already has a lot of other meanings that don't apply in this situation.

I don't know of any technical meaning of "information" that does not apply. In respect of everyday language, I think it is a more natural term than "probability of detection", since it means "how much we know about the event, given the response". As I use the term, "what we know about something" is very close semantically to "information about something". "Mutual information" between two things, in everyday terms, means "what we know about one thing by observing the other".

"Probability of detection" is very much model-loaded, because of the various factors that influence one's decision as to whether one thought one detected the signal on a particular occasion (technically, beta, which represents the slope on the Receiver Operating Characteristic beyond which one would say "Yes, I detected it". The parameter d' is better, but becomes a little imprecise if you don't know the shape of the ROC and are not using a forced-choice procedure. Mutual information has the virtue of being immune to those kinds of problem.

Isn't there some more neutral term that could be used? We keep running into this problem. There are more phenomena than words, I guess.

BP earlier: This may mean, of course, that your measure of mutual information translates directly into my measure of probable detection. If that is true, then the question is only one of preference -- whether one prefers the information-theoretic calculation, or my more direct measure of physical variables.

MT: They are actually the same.

BP: "They" covers several possible referents. Do you mean that an information-theoretic calculation is the same as a direct measure of physical variables?

No. I didn't realize you were equating measures of physical variables with a measure of detection, and took "physical variables" to mean the d'^2 measure or a transform of the probability of detection. I meant that mutual information is the same as the d'^2 measure if the experiment is forced choice, and close to being the same whatever the experimental protocol.

Martin

···

On 2011/07/8 2:11 PM, Bill Powers wrote:

[From Bill Powers (2011.07.08.1355 MDT)]

Martin Taylor 2011.07.08.14.33 --

BP: Are those db 20*log10(x)? That would make the decrement in performance only a factor of 1.5 to 2, which is easily absorbable into measurement errors in experiments like this.

MT: Not so easily absorbed when you are making many hundreds of observations (I made well over a million with my best-trained subject).

BP: I was thinking of the spread of the measure -- maximum and minimum values for one trial. The inherent limits are known only through the extremes of the measures; the averages include the effects of the best runs and the worst ones.

BP earlier: "How much of the reaction time is simply the time it takes a limb or a finger to go into motion, touch the button, and increase the pressure enough to make the contact close?

MT: I have no idea. As long as I have had anything to do with psychophysics, I have never thought reaction time experiments told you much of anything, because of the factors you mention, and because there's a confounding between the subject's task to be fast and the task to be accurate, which (now I know about PCT) presumably translates into a conflict in the "Stuff Happens" complex of control loops.

You realize, I suppose, that the Schouten experiment is not a reaction time experiment? It was conceived as a way of avoiding the problems that are inherent in reaction-time experiments, and at the same time being able to deal with the different detectabilities of signals that all would yield 100% correct answers (within experimental limits) in conventional detection or discrimination experiments.

BP: In theory, the subject samples the state of the light when the beep sounds, so the time taken subsequent to that instant for pressing the button would appear irrelevant. But how long does it take to do that sampling? What if the signal keeps increasing after the beep but before the decision? There's bound to be a spread of judgement times; it can't be an instantaneous judgment. Even if the response begins, it can undergo a course correction before the finger gets to the button, if the signal gets enough stronger during the delay to change the judgment.

But I quibble; the bottom line is the fit of the model to the data, which is pretty good even though it represents an average over a very large number of trials. At least we know how the average of a number of judgments will vary with delay. That's interesting. I prefer to do my modeling with signals and such things, but your way clearly suits your tastes better. And why not?

Best,

Bill P.

···

At what level (order of control) is the noise that matters? How is frequency of impulses related to signal energy? Lots of room for slop there.

We don't know. It's my guess that these questions relate to the reason that even the most highly trained and skilled observers remain so much worse than the ideal observer. They don't seem relevant to the interpretation of the mutual information rate as a lower bound on the capacity of the initial perceptual pathway to the first control loop. They probably are relevant to attempts to model the pathway to account for the data.

I don't see how the noise level could be measured, either, since it would be neural noise, not noise in the light itself.

I agree. As I said, we don't have the information necessary to define an ideal observer for the Schouten presentation. My comments about the relation between human and ideal performance reflect only those situations of which I am aware for which an ideal observer can be defined. I probably should have mentioned that the 2E/No ideal that I mentioned applies to one specific situation, when the observer knows the signal waveform exactly (though not whether the signal is present at all). Other values apply for different conditions. Mostly (maybe all) of those that we can compute are in acoustic detection. For the Schouten experiment, all we can say is that whatever the noise level, the energy in the signal increases linearly with the time the light is on before the decision must be made.

But all that means is that you would adjust the model's noise level until the performance matches the real performance. You would then take that noise level as the model-dependent measure of the system noise (and in your case, the computation of mutual information as the model-dependent measure of the same construct in the real system). That's the same idea behind the way we compute output gain, delay, reference level, and so on for a tracking model. We're saying that IF the system is organized as the model proposes, THEN these would be the values of its parameters. That's about as far as we can go until neurological technology catches up to our questions and we find out how it's really organized in there.

Yes.

I hadn't realized that you also considered a signal increasing relative to the noise level, or that "mutual information" is another way of representing the probability of detection. It's too bad that a term like "information" was picked for that measure, because the word already has a lot of other meanings that don't apply in this situation.

I don't know of any technical meaning of "information" that does not apply. In respect of everyday language, I think it is a more natural term than "probability of detection", since it means "how much we know about the event, given the response". As I use the term, "what we know about something" is very close semantically to "information about something". "Mutual information" between two things, in everyday terms, means "what we know about one thing by observing the other".

"Probability of detection" is very much model-loaded, because of the various factors that influence one's decision as to whether one thought one detected the signal on a particular occasion (technically, beta, which represents the slope on the Receiver Operating Characteristic beyond which one would say "Yes, I detected it". The parameter d' is better, but becomes a little imprecise if you don't know the shape of the ROC and are not using a forced-choice procedure. Mutual information has the virtue of being immune to those kinds of problem.

Isn't there some more neutral term that could be used? We keep running into this problem. There are more phenomena than words, I guess.

BP earlier: This may mean, of course, that your measure of mutual information translates directly into my measure of probable detection. If that is true, then the question is only one of preference -- whether one prefers the information-theoretic calculation, or my more direct measure of physical variables.

MT: They are actually the same.

BP: "They" covers several possible referents. Do you mean that an information-theoretic calculation is the same as a direct measure of physical variables?

No. I didn't realize you were equating measures of physical variables with a measure of detection, and took "physical variables" to mean the d'^2 measure or a transform of the probability of detection. I meant that mutual information is the same as the d'^2 measure if the experiment is forced choice, and close to being the same whatever the experimental protocol.

Martin

[From Bill Powers (2011.07.02.1324 MDT)]

Rick Marken (2011.07.02.1200) –

RM: I think that depends on what the correlation is between. I just
did a difficult pursuit tracking run and the correlation between
independent and dependent variables (target and mouse movements,
respectively) was .32. That would rate as GIGO by your criterion.
But the correlation between the output of a control model (with the
target variations as the disturbance to a controlled variable – the
distance between cursor and target) and the actual output (dependent
variable) was .996. I think what we should evaluate the model-data
correlations (if we are going to look at correlations) in terms of their
size, not the observed correlations.

I did take time delays into account. The data I reported above is not
right, though. What I actually found is this: The correlation between
model and actual mouse movements for the causal model with optimal lag
was .47; the causal model accounts for 22% of the variance in behavior.
The correlation between model and actual mouse movements for the best
fitting control model was .84; the control model accounts for 71% of the
variance in behavior.

BP: OK, that makes more sense. That 0.996 correlation has to be behavior
vs disturbance. And the correlations were not between target and mouse
movements, but between model and actual mouse movements.

The very highest correlation within-subject should be between cursor and
target in a pure pursuit tracking experiment (no disturbance of cursor
relative to mouse). Was that the 0.996? It would have to be the lagged
correlation, of course, as you say.

So the fit of the causal model
to actual behavior is quite low, but in the range of correlations
observed in conventional research, which uses the causal model as the
basis for analysis. The fit of the control model to actual behavior is
much better (but still not up to your .99 criterion).

That would be the correlation between disturbance and output variable for
the subject’s tracking performance. Models do not yield such high
correlations with real behavior. Here is a run using a two level model
for the trackanalyze program, with added computations comparing model and
human behavior at the highest level of difficulty, 5.

Emacs!

The model-human mouse correlation is only (only!) 0.96. There is a 5% RMS
error between model and human cursor (mouse) positions. Note that the
model now explains the lag between cursor and target positions (upper
plot) in terms of two levels of control, velocity and position, and no
transport lag. This model has only 3 parameters instead of 4 originally
used. The model explains 92% of the variance of the real mouse position,
which isn’t all that much higher than the 84% you’re seeing.

I’m a bit concerned about your “predictor” variable for the
causal model. I assume you’re using target position to predict mouse
position. It seems to me that there should be a very high lagged
correlation between target position and mouse position, since the mouse
is tracking the target. Perhaps the disturbance you’re using has such a
high difficulty that this correlation is not so high, and because the
model fits the behavior within the tracking error, the model-real
correlation is higher.

However, you have to be sure you’re comparing similar correlations.
Model-real correlation is not comparable to target-cursor correlation.
The dramatic refutation of SR theory is seen only when you add a second
disturbance between mouse and cursor; Now the correlation of target with
mouse will be extremely low in the causal model because the other
disturbance isn’t taken into account. Better check exactly what the
comparison is.

I agree that looking for some of these new correlation possibilities in
the market data might produce some pretty important results. Glad you’re
doing it.

Best,

Bill P.

(Attachment 17e411d.jpg is missing)

[From Bill Powers (2011.07.03.1525 MDT)]

Rick Marken (2011.07.03.1410) –

RM: I measure the amount of
behavioral noise by correlating the behavior

on separate trials which used the same disturbance. For some reason

(and this is what I have to work on unless you already know the

answer) when the task is difficult, so that the noise level is high,

the control model correlates much better with the
“predictable”

portion of the behavioral variance than does the causal model (with

optimal lag). When the task is easy, the models are basically

indistinguishable (though the control model still does slightly

better). But even in the low noise case, I think I have at least
shown

that an apparently open-loop task, like my reaction time task where

actions have no apparent effect on input, can be modeled as a closed

loop task and that the closed-loop model does just as well (or
better)

than the open-loop model.

BP: How about drawing me a diagram showing the task and indicating which
variables you’re correlating with which other variables?
Note that when a task is very difficult, the mouse movements cause
changes in the controlled variable that generate more error than
they correct. Try a run making lots of movements that are just random –
you will get a high correlation of the mouse movements with the
cursor-target distance, because the mouse is causing most of the changes
in distance. When control is good, that correlation is low; the mouse is
preventing changes in distance.

Here is the result of a run with the “ChooseControl” demo which
computes correlations. I just moved the slider irregularly up and down
without looking at the ball. As you can see, the correlations of the
mouse movements with the three disturbances are all low, while the
correlations with the controllable variables are all about the same
(bottom one in each triad) – the correlation with an actual controlled
variable would be much lower.

Emacs!

Best,

Bill P.

(Attachment 2270854.jpg is missing)

[Martin Taylor 2011.07.04.00.16]

Happy birthday, US Americans.
  [From Bill Powers (2011.07.03.2025 MDT)]
      As to acausal and non-causal, I think that bit of

hairsplitting is another excellent reason not to rely on the
word cause (or its relatives) to do any heavy lifting. Your
chances of being correctly understood are about 50-50.

    Causal, non-causal, and acausal connections:




    A -->--\


    >--->---C         Assume A and B are highly correlated.


    B -->--/




           From   A              B              C


    To


         A                     non-causal   causal


         B      non-causal                  causal


         C       acausal     acausal




    This is hardly hair-splitting. If the system is physical, it

matters greatly whether a causal link is from X to Y or from Y
to X. It doesn’t matter at all for the statistical relationship
between X and Y, but it matters greatly if there is a chain of
influence leading from one to the other. In a loop, the causal
relations go only one way round the loop, though the analysis of
the loop usually proceeds in the acausal direction. The reason
one usually does the analysis in the opposite direction to the
causal direction is precisely because the present value of, say,
output depends on past values of error, but past values of error
do not depend on the present value of output. The link
error->output is causal, while the link output->error is
acausal. It matters.

  Your explanation is impossible to interpret because where you have

A,B, and C in the diagram and table of relationships, you explain
them by referring to X and Y without saying how they relate to A,
B, and C.

They don't. I CAREFULLY used different letters, to make quite clear

that I was NOT talking about A, B, and C any more, but about the
general case.

I see that the diagram got re-edited somewhere in the circuit from

me to you and back, eliminating the leading spaces before the “C”
line. To avoid that happening again, here it is as a JPG.

<img alt="" src="cid:part1.03090806.00090404@mmtaylor.net" height="73" width="63">
  Contrary to your assertion, there is what you call a

causal link from output to error: it goes through the environment,
the input function, and the comparator.

Yes, that is true. There is a causal relation between past values of

the output and the present value of the error, going around the
loop, but there is no causal link from the output BACK to the error.
All the causal links go one way round the loop. In each leg of the
loop, the values at one end of the leg depend only on past values at
the other. Quote: " In a loop, the causal relations go only one way
round the loop, though the analysis of the loop usually proceeds in
the acausal direction." I debated whether to include the point that
past values of any loop variable are causal with respect to its own
present value, but I thought it was both self-evident and irrelevant
to the issue, the issue being the important distinctions among
“causal”, “non-causal”, and “acausal”, so I didn’t.

  I just realized that I read your diagram too hastily. Don't you

have all the arrows going backward?

No.
  If we assume A and B are actually highly correlated

as you say, looking at their joint effect on C as you indicate
doesn’t say anything about causation except that both A and B
influence C.

That's what the diagram is supposed to show, yes. It also shows that

C does not influence A or B, and that A and B do not influence each
other. Those relationships provide the table of definitions.

  But if you reverse the arrows and say that C is

affecting both A and B, then you’re explaining the apparent effect
of A on B, or B on A, as being due to a common third effect from
C. I don’t know what you’re illustrating with this diagram.

I'm illustrating which relationships are causal, which are

non-causal, and which are acausal, to make crystal clear the
difference that you find “hairsplitting” between non-causal and
acausal.

The reason I said we assume A and B are highly correlated was to

illustrate that high correlation between variables is possible even
when the relation between them is non-causal. Of course, if there is
a correlation between A and B, we know that somewhere there is a Z
that influences both of them. Z is, however, of no concern for the
demonstration at hand. The whole point would be lost if the arrows
went the other way. And the definitions table would be all wrong.

  Trying to fit traditional concepts of causation into a feedback

loop is just about impossible.

I think not. I think it is very straightforward. The output of each

link in the loop is causally related to the input of that link. The
current value of every variable in the loop is causally related to
the past values of every other variable and to its own past values.
Nothing in the loop causes itself, so there’s no physical or
philosophical difficulty at all.

The only difficulty arises when you substitute a word such as

“output” for the time function o(t), and in doing so, give yourself
the illusion that the output is causally related to the output,
without remembering that it is the past of the output that is
causally related to the present output.

As you said in a previous message, if you have the equations, it's

usually less confusing to use them and forget about the verbiage.

Martin

(Attachment AB_C.jpg is missing)

[Martin Taylor 2011.07.05.22.50]

  [From Bill Powers (2011.07.05.1845 MDT)]


  Martin Taylor 2011.07.05.19.57 --
    MT: How else do you define "open-loop",

other than the fact that although A influences B there is no
feedback connection that allows B to influence A?

  BP: The example that comes immediately to mind is the relationship

between a disturbance and an output quantity in a control system.
The disturbance influences the output, and the output has no
effect on the disturbance, yet what lies between the disturbance
and the output is not an open-loop system but a control system.
Treating what lies between as a simple one-way connection would be
a mistake, wouldn’t it?

Isn't that a little disingenuous? Aren't you the one who so

frequently points out that the disturbance itself should be
distinguished from the influence of the disturbance on the
perception? Here’s a little diagram that shows how I understand what
you so often point out when you are in your “precision” mood.

<img alt="" src="cid:part1.05010203.08010803@mmtaylor.net" height="215" width="413">

Yes, the connection from the disturbance to the output is one-way.

No, it isn’t simple. The “Stuff” that happens in the “Stuff Happens”
part of the one-way connection is not simple. The complexity or
simplicity of “Stuff Happens” is irrelevant to whether the
connection is one-way.

If you measure what happens at the disturbance and at the output,

and the link between disturbance and “Stuff Happens” is weak or
noisy, you won’t find very much correlation between them, no matter
how good the control system that is the “Stuff that happens”. If the
link from disturbance to “Stuff Happens” is noise-free and strong,
then how good the correlation is will depend on the control system
that is the “Stuff” that happens.

If the correlation is good, you can say that the link from

disturbance to “Stuff Happens” is at least as good as to permit that
correlation. If the correlation is poor, the link might be noisy or
the control might be poor. You wouldn’t know, but at least you would
know that the link could not be so noisy as to limit the correlation
to a value lower than is observed.

In standard psychophysics, the disturbance is the presented signal.

The quality of the link into the control complex is the object of
study. All we can know is that the quality of that link must be
sufficiently good as to allow for the observed correlation between
disturbance and output.

  You can always do a computation anyway, and compute how many bits

of information are being carried from one place to another. But
there is no reason to think that this calculation has any physical
significance.

The physical significance it has is that the capacity of the channel

between the disturbance and “Stuff Happens” is at least as great as
the measured channel capacity between disturbance and output. It
might be much greater, as there could be (and probably will be)
losses in the “Stuff Happens” part of the pathway, but it can’t be
less.

  It could be done, of course, but whatever analysis you

do will not apply to a simple direct connection from input to
output, because that is not present.

No. It only imposes a limiting condition on the capacity of whatever

pathway precedes the connection into the first control loop. It does
not measure the actual capacity, but does provide a floor on that
capacity.

  The output will be affecting the input before and

after the disturbance changes. There may be several places in the
loop where there is a delay. If there is a continuing flow of
information, the information from a previous input may still be
circulating around the loop while new information is entering the
loop;

All true so far.
  many assumptions have to be made in order to say

anything specific about the results.

This is the place where I differ. I dispute "anything specific".

Certainly many assumptions have to be made before you can put more
precise bounds on the capacity of the perceptual connection into the
complex of control systems. In my pre-PCT days, I published on that
point, and my understanding of PCT has given me no reason to say
that we can be any more precise than I claimed in those days.

  Can information circulate indefinitely, going around

and around the closed loop forever? How does it get used up? What
happens when information from one “message” is mingled with
information about a previous message or from a different
concurrent message? When a changing perceptual signal is
subtracted from a changing reference signal, what information is
left in the error signal? Is that information destroyed inside the
output function, or does the feedback function recieve transformed
information from the output function and inject it into the input
function along with new information from the disturbing variable?

  There are more questions than I can answer here. Perhaps you can

answer them.

All of that is for a different thread. I think I might answer some

of them, but I’m not sure I can answer them all at this point. What
I can say is that the answers to these questions would be
interesting and possibly useful, but they would be irrelevant to the
issue at hand.

Martin

-----An aside in the form of a Footnote-----

In case you are curious about the pre-PCT issue, here's the gist of

it.

In the simplest possible standard experiment in auditory

psychophysics, a subject is asked to say in which of two intervals,
both of which contained white noise, a signal (usually a tone burst)
was presented. Many, many, presentations are made at different
signal-to-noise ratios, and the probability of a correct response is
computed at each SNR.

It is possible to define a mathematically ideal observer for this

task, and to calculate the probability that the ideal observer would
give a correct response as a function of the signal-noise ratio. The
probability (or the equivalent d’ measure) can be plotted as a
function of SNR. When you ask a human to do the task, you can draw a
similar plot of p(correct) against SNR.

When some standard presentation method such as "the method of

constant stimuli" is used, the human plot is always steeper than the
mathematical ideal. At relatively high SNR, where p(correct) is
around 0.8, well-trained humans typically do about as well as the
mathematical ideal does at an SNR some 4-6 db lower. But at lower
p(correct), the human performance drops off as a function of SNR
more quickly than does the ideal. Auditory theorists called this the
“oversteep psychophysical function”, and tried to come up with
physiological explanations for it.

What Doug Creelman and I did was to consider the sequential

probabilities p(correct on trial N) if (correct/incorrect on trial
N-1). We observed that p(correct on trial N if correct on trial N-1)
was appreciably greater than p(correct on trial N if incorrect on
trial N-1). Since there was no way the correctness in trial N-1
could influence what was presented on trial N (and of course the
difference does not appear for the ideal observer), we reasoned that
the difference must be in the subject.

We hypothesised that humans might sometimes "lose track of" what

they were supposed to be listening for. We observed this effect in
ourselves as subjects, so the hypothesis was not entirely out of the
blue. In part, this possibility was the reason I developed the PEST
procedure in the first place back in 1960. PEST is a funny kind of
control procedure, in that it makes the task easier if the subject
is getting too many wrong answers and harder if the subject is
getting too many right answers. The point was (in part) to try to
keep the subject aware of what the hard-to-hear signals actually
sounded like. It turned out that when PEST was used, the
oversteepness of the psychophysical function was much reduced.
People behaved more like the mathematically ideal observer would in
a SNR some 4-6 db lower, throughout the range of signal difficulty.

With that as background, Creelman and I made a simple added

assumption, that subjects are either “on” or “off” for strings of
several consecutive presentations. When they are “on”, they perform
at their best ability, and when they are “off” they make random
guesses. Using that simplifying assumption, one can calculate
p(correct when “on”) given the two conditional probabilities
mentioned above. Any other assumption about the variation of
subjects’ “losing track” of the signal would lead to a higher
estimate of the subject’s “on” performance.

Using this assumption, we computed p(correct when "on") in place of

the raw p(correct) for a couple of subjects for whom we had many
responses recorded. Plotting those values in place of the raw values
removed the oversteepness of the psychophysical function.

We presented these results at a meeting of the Acoustical Society of

America. and afterwards one of the major figures in psychoacoustics,
whom we both greatly respected, said that if our results were
correct, he was going to give up psychoacoustics. We didn’t want
that to happen, and didn’t publish until many years later.

End of story.

Martin

(Attachment StuffHappens.jpg is missing)

···

On 2011/07/5 9:11 PM, Bill Powers wrote:

[From Bill Powers (2011.07.06.0510 MDT)]

Martin Taylor 2011.07.05.22.50 –

BP earlier: The example that
comes immediately to mind is the relationship between a disturbance and
an output quantity in a control system. The disturbance influences the
output, and the output has no effect on the disturbance, yet what lies
between the disturbance and the output is not an open-loop system but a
control system. Treating what lies between as a simple one-way connection
would be a mistake, wouldn’t it?

Isn’t that a little disingenuous? Aren’t you the one who so frequently
points out that the disturbance itself should be distinguished from the
influence of the disturbance on the perception? Here’s a little diagram
that shows how I understand what you so often point out when you are in
your “precision” mood.

[]

Yes, the connection from the disturbance to the output is one-way. No, it
isn’t simple. The “Stuff” that happens in the “Stuff
Happens” part of the one-way connection is not simple. The
complexity or simplicity of “Stuff Happens” is irrelevant to
whether the connection is one-way.

I apologize for my disingenuousness. My moods are so fickle that I
can’t predict when I will be overcome by one. I will try to remain
calm.
Yes, I do distinguish between the disturbance as the independent variable
and its effect as a perturbation of the controlled quantity. There can be
a large disturbance, and next to no perturbation, even though without the
control system the perturbation would be large.
The reason for this brings us to the heart of this discussion, I think.
This is a delicate matter because clearly you are proud of the work you
did. It still means a great deal to you and you’re not going to
re-evaluate it casually. I am not going to do that for you, either. I
will speak only of control systems.
The critical notion is so simple that it’s easy to overlook. You speak of
a disturbance that occurs at time T. Your entire argument rests on the
apparently incontrovertible fact that no effect of the disturbance can
alter what happened at time T, because that event happens before anything
else happens. At first glance, this might even imply that there are no
closed-loop systems at all. But let us see.
If the disturbance happens at time T, its effect also happens at time T
unless we’re concerned with microseconds, which we’re not. But is that
true?
In fact, the effect doesn’t happen at time T, it begins to happen.
What makes continuous control possible is that after the disturbance
appears, the effect of the disturbance causes a change in the controlled
variable to start, which causes a perceptual signal inside the system to
begin changing from one value to another, which causes a feedback effect
to start rising in opposition to the effect of the disturbance.

As the perceptual signal increases, the error signal starts to increase,
too, and that causes the output function to start acting through the
feedback connection on the controlled variable. The limiting rate of
increase of the effect of the disturbance may be in the controlled
variable (the position of a mass, for example), or it may be in the
perceptual input function (perception of the acceleration of (or force
applied to) the mass, for example). Or it may be further along in the
chain, even in the output function or the feedback function. But let’s
assume we have a case where it’s in the controlled variable itself, just
as a place to start.

What we observe is not that the disturbance causes the controlled
variable to jump instantly to the final value it would reach without
control, and not that it then is brought back down to a small value by
the action of the system. Instead, we see that the perturbation starts to
increase and then reaches a limit very much less than the final value it
would reach with no control. It never does get larger than the small
value we see it reach. Before the disturbance comes anywhere near having
its potential full effect, the feedback effects have started to oppose
it, lowering the final value toward which the controlled variable appears
to be headed. That apparent final value rapidly decreases, meeting the
actual value before the controlled variable has departed significantly
from its initial state. This is how the system achieves continuous
control with no overshoots or unnecessary undershoots. The sudden onset
of a disturbance does not cause a sudden action, but a smooth change of
the action to a new level over some short time. The disturbance does
not cause a large error that is then quickly reduced. The error
never does get large.

Suppose we’re looking at a case in which a side-effect of the action is
to close a contact under a spring-loaded button. When the disturbance
occurs, the subject is supposed to press that button (I think I’m
forgetting some details about how that is supposed to occur, but you can
remind me). So the subject is trying to control the relationship between
the disturbance and the button-press.

Now we will assume that the slow buildup of the effect of the disturbance
occurs farther downstream, say on the perceived relationship in a
higher-order perceptual input function. The disturbance immediately
changes the externally visible variables from which the controlled
variable is derived, and the perception of relationship begins to change.
As soon as it begins to change, the hand begins accelerating toward the
button, meets it, and the pressure sensed in the finger begins to rise.
The finger presses harder and harder until its force matches the force
from the spring under the button, and the button begins to descend.
Eventually the contact closes and the button is felt to reach a hard
stop, which is the state necessary for “button pressed” to be
sensed. At that point, the relationship perception changes, the reference
signal is matched, and some higher system can begin the process of
releasing the button.

Of course this scenario is model-dependent, and before we could give it
much credence we’d have to try to check it out to see if it’s right. But
the model sounds plausible and would probably work, whether it fits the
real system or not.

Note that the perturbation due to the disturbance does not just
“happen” at time T, the time when the disturbing variable
instantaneously changes state. The sensory event is not a point-event
occurring at a single instant. It extends over time and takes time to
reach its new state – and while it’s reaching the new state, the
perception and the action are continually changing and altering the
effect of the disturbance and affecting the perception. When we examine
the process on a time-scale at which the dynamics of control are visible,
we no longer see “events happening.” We see variables
changing.

That may or may not affect the analysis of the Schauten
experiment.

Best,

Bill P.

(Attachment 5673b6.jpg is missing)

[Martin Taylor 2011.07.06.20.05]

[From Bill Powers (2011.07.06.0510 MDT)]

  Martin Taylor 2011.07.05.22.50 --
      BP earlier: The

example that
comes immediately to mind is the relationship between a
disturbance and
an output quantity in a control system. The disturbance
influences the
output, and the output has no effect on the disturbance, yet
what lies
between the disturbance and the output is not an open-loop
system but a
control system. Treating what lies between as a simple one-way
connection
would be a mistake, wouldn’t it?

    Isn't that a little disingenuous? Aren't you the one who so

frequently
points out that the disturbance itself should be distinguished
from the
influence of the disturbance on the perception? Here’s a little
diagram
that shows how I understand what you so often point out when you
are in
your “precision” mood.

    <img src="cid:part1.00060606.03070102@mmtaylor.net" alt="[]" height="215" width="413">



    Yes, the connection from the disturbance to the output is

one-way. No, it
isn’t simple. The “Stuff” that happens in the “Stuff
Happens” part of the one-way connection is not simple. The
complexity or simplicity of “Stuff Happens” is irrelevant to
whether the connection is one-way.

  I apologize for my disingenuousness. My moods are so fickle that I

can’t predict when I will be overcome by one. I will try to
remain
calm.

Sorry about that, but you are often rather frustrating to try to

interact with because of your changes of what you believe in…

  Yes, I do distinguish between the disturbance as the independent

variable
and its effect as a perturbation of the controlled quantity. There
can be
a large disturbance, and next to no perturbation, even though
without the
control system the perturbation would be large.

  The reason for this brings us to the heart of this discussion, I

think.
This is a delicate matter because clearly you are proud of the
work you
did. It still means a great deal to you and you’re not going to
re-evaluate it casually. I am not going to do that for you,
either. I
will speak only of control systems.

  The critical notion is so simple that it's easy to overlook. You

speak of
a disturbance that occurs at time T. Your entire argument rests on
the
apparently incontrovertible fact that no effect of the disturbance
can
alter what happened at time T, because that event happens before
anything
else happens.

Not at all. I can't see how anything I said would lead you to that

conclusion. My argument hinges only on the fact that no action
output from the “Stuff Happens” region (the complex of control
systems that produce the output) can influence anything in the
pathway between the disturbance and the place where the disturbance
influence is added to the influence of the output. What happens in
“Stuff Happens” can reduce the correlation between the disturbance
and the output, but it cannot increase it.

That's all.
  At first glance, this might even imply that there are

no
closed-loop systems at all. But let us see.

  If the disturbance happens at time T, its effect also happens at

time T
unless we’re concerned with microseconds, which we’re not. But is
that
true?

  In fact, the effect doesn't happen at time T, it *        begins to

happen* .
What makes continuous control possible is that after the
disturbance
appears, the effect of the disturbance causes a change in the
controlled
variable to start, which causes a perceptual signal inside the
system to
begin changing from one value to another, which causes a feedback
effect
to start rising in opposition to the effect of the disturbance.

  As the perceptual signal increases, the error signal starts to

increase,
too, and that causes the output function to start acting through
the
feedback connection on the controlled variable. The limiting rate
of
increase of the effect of the disturbance may be in the controlled
variable (the position of a mass, for example), or it may be in
the
perceptual input function (perception of the acceleration of (or
force
applied to) the mass, for example). Or it may be further along in
the
chain, even in the output function or the feedback function. But
let’s
assume we have a case where it’s in the controlled variable
itself, just
as a place to start.

  What we observe is not that the disturbance causes the controlled

variable to jump instantly to the final value it would reach
without
control, and not that it then is brought back down to a small
value by
the action of the system. Instead, we see that the perturbation
starts to
increase and then reaches a limit very much less than the final
value it
would reach with no control. It never does get larger than the
small
value we see it reach. Before the disturbance comes anywhere near
having
its potential full effect, the feedback effects have started to
oppose
it, lowering the final value toward which the controlled variable
appears
to be headed. That apparent final value rapidly decreases, meeting
the
actual value before the controlled variable has departed
significantly
from its initial state. This is how the system achieves continuous
control with no overshoots or unnecessary undershoots. The sudden
onset
of a disturbance does not cause a sudden action, but a smooth
change of
the action to a new level over some short time. The disturbance
does
not cause a large error that is then quickly reduced. The error
never does get large.

  Suppose we're looking at a case in which a side-effect of the

action is
to close a contact under a spring-loaded button. When the
disturbance
occurs, the subject is supposed to press that button (I think I’m
forgetting some details about how that is supposed to occur, but
you can
remind me). So the subject is trying to control the relationship
between
the disturbance and the button-press.

  Now we will assume that the slow buildup of the effect of the

disturbance
occurs farther downstream, say on the perceived relationship in a
higher-order perceptual input function. The disturbance
immediately
changes the externally visible variables from which the controlled
variable is derived, and the perception of relationship begins to
change.
As soon as it begins to change, the hand begins accelerating
toward the
button, meets it, and the pressure sensed in the finger begins to
rise.
The finger presses harder and harder until its force matches the
force
from the spring under the button, and the button begins to
descend.
Eventually the contact closes and the button is felt to reach a
hard
stop, which is the state necessary for “button pressed” to be
sensed. At that point, the relationship perception changes, the
reference
signal is matched, and some higher system can begin the process of
releasing the button.

  Of course this scenario is model-dependent, and before we could

give it
much credence we’d have to try to check it out to see if it’s
right. But
the model sounds plausible and would probably work, whether it
fits the
real system or not.

  Note that the perturbation due to the disturbance does not just

“happen” at time T, the time when the disturbing variable
instantaneously changes state. The sensory event is not a
point-event
occurring at a single instant. It extends over time and takes time
to
reach its new state – and while it’s reaching the new state, the
perception and the action are continually changing and altering
the
effect of the disturbance and affecting the perception. When we
examine
the process on a time-scale at which the dynamics of control are
visible,
we no longer see “events happening.” We see variables
changing.

All of which I understand very well, and agree with completely. I

hope you will agree with me that it is irrelevant to what happens in
the path between the disturbance and the first effect of the
disturbing influence within the control loop.

Martin

(Attachment Re It is the cause. It is the .jpg is missing)