More on Schouten article

rsmarken · June 12, 2009, 7:49pm

[From Rick Marken (2009.06.12.1245)]

Martin Taylor (2009.06.11.00.16)

Rick Marken (2009.06.10.1320)

I still don't get this. Aren't the subjects in both cases having to
control something?

Most certainly they are. Why would you think I might question that?

Because you seem to think that the IV-DV relationship in the forced
reaction time experiment tells us something about the organism
transfer function (or the perceptual component thereof). PCT shows
that this is not the case if the system under study is a closed loop
control system.

What caveat of mine applies in the forced reaction time experiment
that doesn't apply in the free one?

That if the causal path between IV and DV is open loop, as it is in most
detection experiments despite the intervening and surrounding control
loops, then the results of IV-DV analysis can be valid and useful (page 2).

When I wrote this I meant to include the feedback connection from DV
to I as part of the causal path from IV to DV, if such a connection
exists, as it always does when the system under study is a control
system. I argue in the paper that the situation in _all_ psychological
experiments is as follows:

IV ---> I-->|System|-->DV
^ |
>_____________|

Where I is the proximal effect of the IV on the system. This differs
from the situation in the physical sciences, where the IV to DV path
looks like this:

IV ---> I-->|System|-->DV

The difference results from the nature of the system under study, not
from anything about the way the experiment is conducted. This was the
point of my paper (which I obviously didn't make clearly so I guess
I'll have to write another one that's even better;-) If the system
under study is closed loop then there is no way to learn about the
system transfer function (that is the nature of the behaving system
itself) or any component therof by looking only at the relationship
between IV and DV in an experiment.

After I wrote the paper Bill pointed out here that, even when the
system under study is closed loop the path from IV to DV can be called
"open loop" because the DV has no effect on the IV. This point just
shows that I could have probably been clearer in my paper about what I
meant by open vs closed loop. The point I was trying to make is that,
in an experiment, what matters is whether the system under study is
open or closed loop, not whether the connection between IV and DV in
the experiment is open or closed loop (it can always be considered
open loop). I think this point becomes clear in the paper but it
obviously didn't become clear to you so, again, I think that just
shows that I have to try again.

The bottom line of my argument in the "Revolution" paper is that if
the system under study is closed loop (a control system) then there is
simply no way to design the experiment so that the relationship
between an IV and a DV will tell you anything about the system or it's
components. What you have to do is take the feedback path into account
when analyzing the results. This means taking into account the effect
of the IV and DV on the controlled input, I. And in order to do that
you either have to determine what I is (test for the controlled
variable) or simply make a reasonable guess at what I is and then set
up the simultaneous equations that define the system-environment
interaction (DV = f(I), I = G(IV,DV)) and see how the model works.

The reason this is not the case [that IV-DV analysis can be valid and
useful] the in the free reaction experiment is that the experiment
concerns the workings of the control processes, because of the
intrinsic conflict between the control for fast response and the control
for accurate response. In the standard forced-choice detection study
there is no such conflict , because the subject has essentially infinite
time to respond (usually more than a second), and in the Schouten
study there is no conflict, or rather, the conflict is suppressed, because
the timing of the response is constrained.

I do not agree with this because it implies that an experimental
manipulation -- like, for example, "constraining" the timing of the
response by providing a cue -- can change the system from open to
closed loop. My point is that, no matter what you do, the person in
the experiment is still a closed loop system, controlling inputs to
which the experimental manipulations that affect behavior are a
disturbance, no matter what those manipulations are. Looking _only_ at
the relationship between IV and DV in an experiment on a closed-loop
control system (and ignoring the feedback effect of the system on the
input variables being controlled) will give you a misleading picture
of the nature of the system (or component functions thereof).

In both cases, the performance of the control
loops, whose eventual visible output is the button push, is unlikely to be
affected by the choice of disturbance (one light or the other, tone in first
or second interval, etc.).

Whether that is true or not makes no difference as long as the system
under study is closed loop, which it apparently is if there are
control loops involved.

Furthermore, there is no suggestion in the data other than the slight bias
toward late responses for very early bips and early responses for very late
bips that the "fast v. accurate" conflict that probably exists is affecting
the performance of the control processes.

Again, I think this is irrelevant. If the system under study is closed
loop then your analysis of the behavior under study must take into
account the feedback effects of the systems actions on controlled
input variables.

The data are much more likely to
reflect the performance of the pathway between the physical disturbance
and the input to whatever perception is being controlled

What the data "reflect" depends on the nature of the system under
study. In order to determine what the data reflect you must build a
model of the system. If you look only at the data in terms of the
relationship between IV and DV then your implicit model of the system
is the open loop model of experimental psychology (and the statistics
used to analyze the results of experiments): DV = f(IV). When this
model is applied to experiments done on closed loop systems the
results are misleading.

(I argue for the
controlled perception being a match between an abstract "position"
perception and an equally abstract remembered "answer" perception, but
whether that's correct is really irrelevant as well as being impossible to
determine from within the experiment; and just how would one do The
Test to see whether that was actually the controlled perception?).

The way to do it is to include your hypothesis about the controlled
variable in a model of the control process that is going on in the
experiment. See which version of the controlled variable in the model
provides the best fit to the data.

Anyway, thanks for the inspiration to continue my efforts to get the
message across. Actually, I think I now have a better idea of what the
message is: If the system under study is closed loop, then an analysis
of the results of the experiment must take into account the effect of
_both_ the IV and the DV on controlled input(s), I.

Best regards

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com

MartinT · June 13, 2009, 6:54pm

[Martin Taylor 2009.06.13.14.29]

[From Rick Marken (2009.06.12.1245)]

Martin Taylor (2009.06.11.00.16)

Rick Marken (2009.06.10.1320)

I still don't get this. Aren't the subjects in both cases having to
control something?

Most certainly they are. Why would you think I might question that?

Because you seem to think that the IV-DV relationship in the forced
reaction time experiment tells us something about the organism
transfer function (or the perceptual component thereof). PCT shows
that this is not the case if the system under study is a closed loop
control system.

PCT doesn’t show that. It’s a result, And it’s only a partial result,
in any case. In order to make the case in any particular instance, you
need to know how well the control system controls, because that
determines the relationship between the visible (non-side-effect) part
of the output and the visible disturbance.

Secondly, I’m glad you once again include the caveat: “if the system
under study is a closed loop control system”. So long as you keep that
in, I have no great problem, because it covers the situation at hand.

What caveat of mine applies in the forced reaction time experiment
that doesn't apply in the free one?

That if the causal path between IV and DV is open loop, as it is in most
detection experiments despite the intervening and surrounding control
loops, then the results of IV-DV analysis can be valid and useful (page 2).


When I wrote this I meant to include the feedback connection from DV
to I as part of the causal path from IV to DV, if such a connection
exists, as it always does when the system under study is a control
system. I argue in the paper that the situation in _all_ psychological
experiments is as follows:
IV ---> I-->|System|-->DV
^ |
>_____________|
Where I is the proximal effect of the IV on the system. This differs
from the situation in the physical sciences, where the IV to DV path
looks like this:
IV ---> I-->|System|-->DV
The difference results from the nature of the system under study, not
from anything about the way the experiment is conducted.

I have no problem with this. Nor, given that diagram, should you have
any problem with my analyses, which concern the pathway IV —> I in
your diagram.

If the system
under study is closed loop then there is no way to learn about the
system transfer function (that is the nature of the behaving system
itself) or any component therof by looking only at the relationship
between IV and DV in an experiment.

Correct. That was why I liked your paper.

The point I was trying to make is that,
in an experiment, what matters is whether the system under study is
open or closed loop, not whether the connection between IV and DV in
the experiment is open or closed loop (it can always be considered
open loop).

How can it always be considered open-loop? Surely if you are looking at
the functioning of pathways within the loop the close-loop property
becomes essential, does it not? The issue is hidden in those words “the
system under study”. It is always true that the DV is determined by the
output of a closed-loop process. The question is whether that
closed-loop process is the part of the IV->DV pathway that is under
study. It is not always the case that every perception is controlled,
nor is it the case that variation in the attributes of an uncontrolled
input to a controlled perception causes variation in the way the
controlled perception is controlled. In most analyses, it only affects
the value of the perceptual variable, not the properties of the loop in
which that variable is controlled.

If you really think all perceptions are controlled at all times, you
have an awful lot of elementary physics to catch up on.

The bottom line of my argument in the "Revolution" paper is that if
the system under study is closed loop (a control system) then there is
simply no way to design the experiment so that the relationship
between an IV and a DV will tell you anything about the system or it's
components.

I’ll dispute the “anything”, here. You are saying that it is impossible
to learn anything about a control loop by modelling it so that the
observable variables (the disturbance and the output) are related in
the way that the data indicate. Most of the time on CSGnet, you are
arguing exactly the opposite. I wonder why it is in this thread, and
only this thread, that you take the opposite position? …Ah, but you
contradict yourself next…

What you have to do is take the feedback path into account
when analyzing the results. This means taking into account the effect
of the IV and DV on the controlled input, I. And in order to do that
you either have to determine what I is (test for the controlled
variable) or simply make a reasonable guess at what I is and then set
up the simultaneous equations that define the system-environment
interaction (DV = f(I), I = G(IV,DV)) and see how the model works.

And none of this involves observing the way the IV relates to the DV?
Give me a break!

Think three thoughts:

(1) Almost all of our perceptions at any one moment are uncontrolled,
though they may contribute to controlled perceptions at higher levels.

(2) The pathway whereby variation in an uncontrolled perception
introduces a disturbance to a controlled perception is not part of the
loop whereby that perception is controlled.

(3) Though it is possible in a nonlinear system for the value of a
variable to affect the behaviour of a process or function, in the cases
we are discussing, nonlinearity is unlikely to be a factor, and the
control process for a perception disturbed by variation in an
uncontrolled input is unlikely to alter its performance as a function
of the value of that uncontrolled perception.

Martin

Bill_Powers1 · June 13, 2009, 8:40pm

From Bill Powers (2009.06.13.0739 MDT)]

I'm a little behind because I'm in Boulder Community Hospital getting ready to be treated for atrial fibrillation. That sounds bad, but I may have had it for years. I feel fine, if short of breath on exercise, the same symptom that comes out of, and exacerbates, having smoker's lungs. The synchronizing or "conversion" procedure will be done on Monday and we'll see what comes after that. Fortunately this is a modern hospital and I'm getting free Wi-Fi, "free" being a signal (IV) to laugh (DV).

···

=========================================================================

Rick Marken (2009.06.12.1245) --
Martin Taylor 2009.06.13.14.29 --

The discussion of closed versus open loop experiments is getting somewhere, so let's keep it going until we have agreement. There's no need to leave it at the "agree to disagree" stage, since the issue is perfectly clear.

The question is whether any experiment with a control system can be done based on external observations and manipulations in such a way that the presence of feedback makes no difference in the results compared with the same experiment done with an open-loop system.

Martin Taylor appears to be proposing that the Schouten experiment is a case where the input part of the loop is explored under conditions where the existence of a feedback loop makes no difference.

Rick Marken and I have both argued that feedback has its effects on the input variable (I) between the IV and the DV (the stimulus and response events in conventional terms), so that there is no way to tell (from knowing only the IV) what the actual input (I) to the control systems is. We can see the stimulus light (IV), and deduce or even measure the flux arriving at the pupil (I), but there is no evidence for what happens on the retina or after that in the perceptual systems. Because of the iris, there's no way to predict even the light intensity at the retina, and of course we can't measure the signal in the optic nerve.

Unless we're talking about first-order (intensity) control systems, we can't count on the controlled variable being definable strictly in terms of illumination in the plane of the pupil. In fact, we've been talking about controlling the relationship between perception of the light and perception of a button-press. To measure those perceptions we'd have to reach inside the brain to measure what I count as the fifth level of perception, or higher.

The logical expression

NOT [(light 1 AND NOT button 1) OR (light 2 AND NOT button 2)]

defines the reference condition of the relationship, any other combinations of lights and buttons being errors. Both states of the lights and both states of the buttons are, we assume in PCT, represented as states of perceptual signals, and the state of their relationship is another perceptual signal. We would not assume, as in "normal" psychology, that there is no stimulus until the light turns on. The perceptual signals just change from one state to another.

With the lights perceived as off and no button being pressed, the state of the relationship matches the reference condition and no action is needed to correct it. When a light first turns on, that creates an error condition since no button is being pressed, and the resulting error condition leads to pressing either button 1 or button 2 which removes the error and stops the press. Clearly this is a closed loop arrangement involving the variable affected by the light and the button press, but since the button press is delayed there is theoretically a short delay in which only the effect of the light has changed and the feedback from the state of the buttons has not yet changed. That is an error condition that persists until the right button is pressed. In the Schouten experiment, the length of this delay can be manipulated once the subject has learned to use the audible beeps as a signal for acting ( despite various conflicts that arise, which Schouten discusses).

There would be no disagreements here if we had a way to detect the change in the perceptual or error signals directly. We would then know when the perception or error changed before there was any indication of a response, and we would clearly be measuring the delay in part of the control loop. But the only available indication of a change in the error signal is the effect of the action that follows, the closure of a contact in one of the buttons. And that indicates only a change in the state of the DV, not revealing when any variable changed at an earlier time.

If the delay between signal detection and contact closure under a button is constant, this delay would not affect the data of interest to Martin, since the critical relationship is the fraction of erroneous responses as a function of delay time relative to the zero-error delay time. As I showed by using an alternative model, this does not pin down where in the path between input and output this relationship is determined. But no matter where it happens, it happens somewhere and according to Martin can be interpreted as an "accumulation of information." I see no objection to that proposition, other than not believing that there is some substance called information passing along the nerves. We can agree that it is possible to calculate information using certain mathematical manipulations, without having to agree that this is anything more than a perception in the observer, created by combining mathematical representations of lower-order perceptual variables. This does not guarantee that there is anything in the environment corresponding to the resulting measure.

The fact that you can calculate it doesn't mean that it exists.

So, where does that leave us? Martin is not trying to predict the state of the DV from knowing the IV, so Rick's main objection doesn't apply. There is clearly a systematic relationship between the fraction of incorrect reponses and the length of time before some critical time that the subject is required to indicate which light is on. And that is interesting, because the guesses change continuously from 50% wrong to some much smaller fraction wrong as the required reporting time becomes later after the onset of light.

While we can reverse-engineer this system and think of more than one circuit that would reproduce this phenomenon, the only way to check the model would be to trace the neural circuits inside the brain, and I don't think any of us is up for that. There's no general model that comes out of this specialized experiment; the one model I've thought of so far doesn't even require adding anything (but some background level of perceptual noise) to the existing PCT model. So unless there's something we can take from this experiment to apply more broadly, I think we may be almost finished with it.

As to the meaning of the term information, I think that is a separate subject.

Best,

Bill P.

> Martin Taylor (2009.06.11.00.16)
>
>>Rick Marken (2009.06.10.1320)

>>I still don't get this. Aren't the subjects in both cases having to
>> control something?
>
> Most certainly they are. Why would you think I might question that?

Because you seem to think that the IV-DV relationship in the forced
reaction time experiment tells us something about the organism
transfer function (or the perceptual component thereof). PCT shows
that this is not the case if the system under study is a closed loop
control system.

>> What caveat of mine applies in the forced reaction time experiment
>> that doesn't apply in the free one?
>
> That if the causal path between IV and DV is open loop, as it is in most
> detection experiments despite the intervening and surrounding control
> loops, then the results of IV-DV analysis can be valid and useful (page 2).

When I wrote this I meant to include the feedback connection from DV
to I as part of the causal path from IV to DV, if such a connection
exists, as it always does when the system under study is a control
system. I argue in the paper that the situation in _all_ psychological
experiments is as follows:

IV ---> I-->|System|-->DV
^ |
>_____________|

Where I is the proximal effect of the IV on the system. This differs
from the situation in the physical sciences, where the IV to DV path
looks like this:

IV ---> I-->|System|-->DV

The difference results from the nature of the system under study, not
from anything about the way the experiment is conducted. This was the
point of my paper (which I obviously didn't make clearly so I guess
I'll have to write another one that's even better;-) If the system
under study is closed loop then there is no way to learn about the
system transfer function (that is the nature of the behaving system
itself) or any component therof by looking only at the relationship
between IV and DV in an experiment.

After I wrote the paper Bill pointed out here that, even when the
system under study is closed loop the path from IV to DV can be called
"open loop" because the DV has no effect on the IV. This point just
shows that I could have probably been clearer in my paper about what I
meant by open vs closed loop. The point I was trying to make is that,
in an experiment, what matters is whether the system under study is
open or closed loop, not whether the connection between IV and DV in
the experiment is open or closed loop (it can always be considered
open loop). I think this point becomes clear in the paper but it
obviously didn't become clear to you so, again, I think that just
shows that I have to try again.

The bottom line of my argument in the "Revolution" paper is that if
the system under study is closed loop (a control system) then there is
simply no way to design the experiment so that the relationship
between an IV and a DV will tell you anything about the system or it's
components. What you have to do is take the feedback path into account
when analyzing the results. This means taking into account the effect
of the IV and DV on the controlled input, I. And in order to do that
you either have to determine what I is (test for the controlled
variable) or simply make a reasonable guess at what I is and then set
up the simultaneous equations that define the system-environment
interaction (DV = f(I), I = G(IV,DV)) and see how the model works.

> The reason this is not the case [that IV-DV analysis can be valid and
> useful] the in the free reaction experiment is that the experiment
> concerns the workings of the control processes, because of the
> intrinsic conflict between the control for fast response and the control
> for accurate response. In the standard forced-choice detection study
> there is no such conflict , because the subject has essentially infinite
> time to respond (usually more than a second), and in the Schouten
> study there is no conflict, or rather, the conflict is suppressed, because
> the timing of the response is constrained.

I do not agree with this because it implies that an experimental
manipulation -- like, for example, "constraining" the timing of the
response by providing a cue -- can change the system from open to
closed loop. My point is that, no matter what you do, the person in
the experiment is still a closed loop system, controlling inputs to
which the experimental manipulations that affect behavior are a
disturbance, no matter what those manipulations are. Looking _only_ at
the relationship between IV and DV in an experiment on a closed-loop
control system (and ignoring the feedback effect of the system on the
input variables being controlled) will give you a misleading picture
of the nature of the system (or component functions thereof).

> In both cases, the performance of the control
> loops, whose eventual visible output is the button push, is unlikely to be
> affected by the choice of disturbance (one light or the other, tone in first
> or second interval, etc.).

Whether that is true or not makes no difference as long as the system
under study is closed loop, which it apparently is if there are
control loops involved.

> Furthermore, there is no suggestion in the data other than the slight bias
> toward late responses for very early bips and early responses for very late
> bips that the "fast v. accurate" conflict that probably exists is affecting
> the performance of the control processes.

Again, I think this is irrelevant. If the system under study is closed
loop then your analysis of the behavior under study must take into
account the feedback effects of the systems actions on controlled
input variables.

> The data are much more likely to
> reflect the performance of the pathway between the physical disturbance
> and the input to whatever perception is being controlled

What the data "reflect" depends on the nature of the system under
study. In order to determine what the data reflect you must build a
model of the system. If you look only at the data in terms of the
relationship between IV and DV then your implicit model of the system
is the open loop model of experimental psychology (and the statistics
used to analyze the results of experiments): DV = f(IV). When this
model is applied to experiments done on closed loop systems the
results are misleading.

> (I argue for the
> controlled perception being a match between an abstract "position"
> perception and an equally abstract remembered "answer" perception, but
> whether that's correct is really irrelevant as well as being impossible to
> determine from within the experiment; and just how would one do The
> Test to see whether that was actually the controlled perception?).

The way to do it is to include your hypothesis about the controlled
variable in a model of the control process that is going on in the
experiment. See which version of the controlled variable in the model
provides the best fit to the data.

Anyway, thanks for the inspiration to continue my efforts to get the
message across. Actually, I think I now have a better idea of what the
message is: If the system under study is closed loop, then an analysis
of the results of the experiment must take into account the effect of
_both_ the IV and the DV on controlled input(s), I.

Best regards

Rick
--
Richard S. Marken PhD
rsmarken@gmail.com

--
Richard S. Marken PhD
rsmarken@gmail.com

rsmarken · June 14, 2009, 12:00am

[From Rick Marken (2009.06.13.1700)]

Martin Taylor (2009.06.13.14.29)--

Rick Marken (2009.06.12.1245)--

Because you seem to think that the IV-DV relationship in the forced
reaction time experiment tells us something about the organism
transfer function (or the perceptual component thereof). PCT shows
that this is not the case if the system under study is a closed loop
control system.

PCT doesn't show that.

I was thinking of Bill's 1978 Psych Review paper where he shows that
the relationship between disturbance and output (equivalent to IV and
DV) when studying a closed loop system reflects the inverse of the
feedback connection between output and controlled input (DV and I),
not the "organism function" relating I to DV. This is the "behavioral
illusion", of course. I say that PCT shows this because Bill's
"discovery" of this illusion depended on using the PCT mapping of
variables in a control loop to those involved in actual behavior.
Other applications of control theory to behavior did not make this
"discovery" because they did not use the PCT mapping; for example,
they mapped the reference variable in the control system to the target
input to the behaving system and they mapped the controlled variable
in the control system to the output variable in the behaving system.

The point I was trying to make is that,
in an experiment, what matters is whether the system under study is
open or closed loop, not whether the connection between IV and DV in
the experiment is open or closed loop (it can always be considered
open loop).

How can it always be considered open-loop?

For the reason I said: because the DV doesn't "loop back" and affect the IV.

Surely if you are looking at the
functioning of pathways within the loop the close-loop property becomes
essential, does it not?

Yes, but that doesn't affect the fact that one can say that the
relationship between IV and DV is open loop. It's open loop regardless
of whether the pathways between IV and DV are closed loop (as they are
in living systems) or open loop (as they are in non-living systems).

The issue is hidden in those words "the system under
study". It is always true that the DV is determined by the output of a
closed-loop process.

Only if the system is closed loop. The DV in physical experiments is
not determined by the output of a closed-loop process.

The question is whether that closed-loop process is the
part of the IV->DV pathway that is under study.

When the system under study is a closed loop system, the "IV-DV
pathway" has a closed loop in it (that closed loop is the organism,
which is controlling some variable, I, that is disturbed by the IV and
protected from disturbance by the DV). There is just no getting away
from it.

If you really think all perceptions are controlled at all times, you have an
awful lot of elementary physics to catch up on.

My point has nothing to do with assuming the "all perceptions are
controlled at all times". My point is that, when the system under
study is closed loop, there will be no relationship at all between IV
and DV unless the organism is controlling a perception (of an input
variable, I) that is disturbed by the IV and protected from that
disturbance by the DV. In the RT experiment, if the subject is not
controlling a perception like "if light on, press, otherwise don't"
then the subject will not consistently press only when the light comes
on and not press otherwise.

I'll dispute the "anything", here. You are saying that it is impossible to
learn anything about a control loop by modelling it so that the observable
variables (the disturbance and the output) are related in the way that the
data indicate. Most of the time on CSGnet, you are arguing exactly the
opposite. I wonder why it is in this thread, and only this thread, that you
take the opposite position? ...Ah, but you contradict yourself next...

> What you have to do is take the feedback path into account
>when analyzing the results. This means taking into account the effect

of the IV and DV on the controlled input, I. And in order to do that

>you either have to determine what I is (test for the controlled
>variable) or simply make a reasonable guess at what I is and then set

up the simultaneous equations that define the system-environment
interaction (DV = f(I), I = G(IV,DV)) and see how the model works.

And none of this involves observing the way the IV relates to the DV? Give
me a break!

It involves a bit more than observing the way the IV relates to the
DV. As I said in the last sentence before your last point, you have to
include in the model something that is never included in conventional
models of the relationship between IV and DV: a controlled variable
(CV). Conventional models of experimental results are (as I said)
based on the general linear model of statistics:

DV = aIV+e

A control model of the experiment would consist of two simultaneous equations:

CV = aIV+bDV
DV = r - kCV

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com

Bill_Powers1 · June 14, 2009, 3:52am

[From Bill Powers (2009.06.13.2005 MDT)]

Rick Marken (2009.06.13.1700)–

Martin Taylor
(2009.06.13.14.29)–

RM: you either have to determine what I is (test for the
controlled

variable) or simply make a reasonable guess at what I is and
then set

up the simultaneous equations that define the
system-environment

interaction (DV = f(I), I = G(IV,DV)) and see how the model
works.

MT: And none of this involves observing the way the IV relates to
the DV? Give me a break!

Take my word for this: the whole problem in this argument come from not
defining terms and using them consistently. If we would all stop and say
“Wait a minute, I think I disagree with X, but exactly what did you
mean by X?” most of our go-arounds would stop.
Martin asks, “And none of this involves observing the way the IV
relates to the DV?” We could ask, "Involves in any way at all
whether related to PCT or not? If that’s your question, of course we
observe the “way” the IV relates to the DV. But if he is asking
whether the apparent dependence of the DV on the IV (which is the
“way” we obtain by the observation he enquires about) has the
same form as the actual dependence determined by the forward path inside
the organism, the answer is, NO, the apparent relationship is determined
primarily by the form of the feedback path from DV back to I, the input
variable that is also affected by the DV. That is generally different
from the form of the forward path.

One hint as to the natura of the problem was revealed in a recent post
when Martin spoke of “I”, the input variable, as if it were
something inside the organism. In effect, he was assuming that the
controlled variable had been tested for and identified, while Rick and I
were assuming that we could see a proximal stimulus I only in the
environment outside the subject, and therefore could not know what the
actual controlled variable affected by the feedback was. Not without
doing The Test.

If my model of the Schouten experiment is correct, which I suggest but do
not claim, the button-press is apparently related to the light intensity
by a straight-through causal path with a delay in it. Light on, button
press. Light off, no button press. I propose that the actual forward path
through the organism is

light on and no button press —>

relationship false —> error signal —> start button press
—>

relationship true —> no error signal —> release button
press.

In other words, the apparent causal function is approximately the
derivative of the forward causal function, in fact its temporal inverse.
If the light being on causes the button to be pressed, the button turns
the error off, which says the apparent forward function is like an
impulse, not a continuing condition, even though the light stays on. If
the button turns the light off as in the experiments Dick Robertson
reported, the appearance of an impulse relationship is even stronger. I
have found nothing in the Schouten data to tell us when the button was
released on each trial. For all I know, “button press” is
defined as down-up.

RM:

CV = aIV+bDV

DV = r - kCV

A control model of the experiment would consist of two simultaneous
equations:

CV = aIV+bDV

DV = r - kCV

This is fine, but what did Martin say, exactly, that makes you think he
doesn’t know this? I think it’s in the ambiguity of “the way the IV
relates to the DV.” It actually relates in two ways, not one, as
shown by the simultaneous equations above. You can rewrite them
as

CV = aIV+bDV

CV = (r - DV)/k

Which shows that we can eliminate CV and leave only a relationship
between IV and DV. It is

aIV + bDV = r/k - DV/k

or

aIV + (b + 1/k)DV = r/k

If k, the output gain, is a large number, this becomes close to

aIV + bDV = 0, or

DV = -(a/b)IV

The path from disturbance to CV is the factor a, and from action to the
CV through the environmental feedback path is the factor b.

Notice that the forward gain, k, disappears from the solution. So if k is
large enough, the relationship between IV and DV is determined almost
entirely by the two environmental connections: IV to CV, and DV to
CV.

This is in fact, exactly how operational amplifiers are used in analog
computing. Op amps have gains ranging from 20,000 to a billion, so the
input-output function is determined by the component in the feedback path
and the component in series with the input path, both connected to the
inverting input terminal. The computation accuracy when the gain is a
billion is better than one part per million over a modest band of
frequencies. I saw this one in a self-balancing amplifier used for
millidegree temperature control.

When gains are not quite so high, the forward properties begin to have
some small effect, which gets larger when the gain is under 10 or so or
the paths are complicated by time delays, thresholds, limits, or
hysteresis. The Schouten situation involves some rather extreme
complications including having to overcome strong tendencies to avoid
obeying the instructions. The usual performance in reaction-time
experiments is to respond to the stimulus as fast as possible, so the
controlled variable remains in an error state for the least possible time
and is then restored to the no-error state. So the forward signal,
generated by combining the stimulus signal with the feedback signal, is
more or less a one-shot square wave of short duration, leading some
experimenters to conjecture that it is the onset of the stimulus that
causes the response, not the steady-state value. At one time, I recall,
many psychologists seemed to think that organisms responded only to
changes in stimulation and not to the steady state at all. That’s because
as soon as the response occurred it stopped again. But that was because
the steady-state error was corrected, not because the organism was
sensing only the first derivative of the stimulus. And the experimenters
were looking only at the first derivatives of the responses! A
“button press” was down-up, not just down.

Threshing these things out is time-consuming and exasperated, but it’s
worth the trouble to get everyone synchronized. Loose talk sinks ships,
and also theories.

Best,

Bill P.

rsmarken · June 14, 2009, 6:04pm

[From Rick Marken (2009.06.14.1100)]

Bill Powers (2009.06.13.2005 MDT)–

Take my word for this: the whole problem in this argument come from not
defining terms and using them consistently. If we would all stop and say

“Wait a minute, I think I disagree with X, but exactly what did you mean by
X?” most of our go-arounds would stop.

Isn’t this the case in all discussion; trying to get at what another person means? I think we do the “Wait a minute…” implicitly each time we answer, hoping to see how our answer (which reflects our understanding of the meaning of X) is handled. Maybe putting the “Wait a minute…” into our discussions explicitly would help. But I kind of doubt it.

Martin asks, “And none of this involves observing the way the IV relates to
the DV?” We could ask, "Involves in any way at all whether related to PCT or
not? If that’s your question, of course we observe the “way” the IV relates

to the DV. But if he is asking whether the apparent dependence of the DV on
the IV (which is the “way” we obtain by the observation he enquires about)
has the same form as the actual dependence determined by the forward path

inside the organism, the answer is, NO, the apparent relationship is
determined primarily by the form of the feedback path from DV back to I, the
input variable that is also affected by the DV. That is generally different

from the form of the forward path.

This makes sense to me. I think I clearly understand what you mean. And I agree. So what did you mean Martin, when you asked “And none of this involves observing the way the IV relates to the DV?”

RM:
CV = aIV+bDV
DV = r - kCV

A control model of the experiment would consist of two simultaneous
equations:

CV = aIV+bDV
DV = r - kCV

This is fine, but what did Martin say, exactly, that makes you think he
doesn’t know this?

It’s not that I didn’t think Martin didn’t know this. I was trying to answer him in a way that would help him understand my point of view.

I think it’s in the ambiguity of “the way the IV relates
to the DV.” …

If k, the output gain, is a large number, this becomes close to

aIV + bDV = 0, or

DV = -(a/b)IV

Right, the behavioral illusion. This, of course, has been at the heart of my argument about the shortcoming of looking only at the relationship between IV and DV in experiments on closed loop systems. Maybe I’m not understanding Martin but he seems to be agreeing that, in experiments on closed loop systems, this behavioral illusion is a problem in some experimental (those like the free RT task) but not in others (like the forced RT task). I think the distinction is captured (somehow) in this paragraph from Martin:

Surely if you are looking at the functioning of pathways within the loop the close-loop property becomes essential, does it not? The issue is hidden in those words “the system under study”. It is always true that the DV is determined by the output of a closed-loop process. The question is whether that closed-loop process is the part of the IV->DV pathway that is under study. It is not always the case that every perception is controlled, nor is it the case that variation in the attributes of an uncontrolled input to a controlled perception causes variation in the way the controlled perception is controlled. In most analyses, it only affects the value of the perceptual variable, not the properties of the loop in which that variable is controlled.

I guess I should have said “Wait a minute…” after I read this. Actually, I can’t really make out what this means and it seems to be central to Martin’s point about the usefulness of the results of some experiments and the uselessness (as argued in my “Revolution” paper) of the results of others. Maybe if Martin could draw a little diagram of what he means it would help. I’ll give it a try and then he can set me straight. In these diagrams, IV is a disturbance. Here goes:

IV1 → I1----- >DV
^ |
>_____ |
^
>

IV2 → I2

So here we have two IVs (disturbances) each affecting a different sensory input, I1 and I2. I1 is a controlled variable and I2 is not controlled. DV is the output of the control process that controls I1. But the DV is also caused, in an open-loop manner, by the disturbance to I1 via another variable, I2, that is in turn caused by IV2. Is this a reasonable picture of what you, Martin, think is going on in those “useful” kinds of conventional experiements, like the forced RT experiment?

Threshing these things out is time-consuming and exasperated, but it’s worth
the trouble to get everyone synchronized. Loose talk sinks ships, and also
theories.

I’m not exasperated at all. This is lots of fun and it has already led to some cool ideas, for me anyway.

Best

Rick

···

–
Richard S. Marken PhD
rsmarken@gmail.com

Bill_Powers1 · June 14, 2009, 10:00pm

[From Bill Powers (2009.06.14.1346 MDS)]

Rick Marken (2009.06.14.1100)–

Maybe I’m not understanding
Martin but he seems to be agreeing that, in experiments on closed loop
systems, this behavioral illusion is a problem in some experimental
(those like the free RT task) but not in others (like the forced RT
task). I think the distinction is captured (somehow) in this paragraph
from Martin:

Surely if you are looking at the functioning of pathways within the
loop the close-loop property becomes essential, does it not? The issue is
hidden in those words “the system under study”. It is always
true that the DV is determined by the output of a closed-loop process.
The question is whether that closed-loop process is the part of the
IV->DV pathway that is under study. It is not always the case that
every perception is controlled, nor is it the case that variation in the
attributes of an uncontrolled input to a controlled perception causes
variation in the way the controlled perception is
controlled.

BP: Notice that there are one unspoken assumption here and one ambiguity,
in the last sentence: the assumption that there can be a behavior (DV)
that is not part of a control loop, and the ambiguous use of
“way” in stating that a second input to a controlled perception
can change without causing any change in the way the controlled
perception is controlled.

The assumption I reject without qualification. There is no behavior, and
no apparent response to an apparent stimulus, that is not part of a
negative feedback control loop (the First Principle of PCT). If any
verifiable exception to that principle is found, I want to know the
details immediately. MArtrin seems to say the same thing, yet tries to
leave the door open for an IV-DV effect that does not operate through
disturbing the controlled perception. It is, of course, possible that he
has not said this so as to convey what he really means. The first step in
achieving clarity of communication is to realize that one has not been
clear.

The ambiguity is less of a problem. We can say that the way the
controlled perception is controlled is by pressing a button, or we can
say that the way the controlled perception is controlled is by perceiving
it, comparing the perception to a reference signal, and routing the
resulting error signal to the correct lower-order system’s reference
input. The latter “way” then includes the system parameters
including the definition of the input function.

In other words the “way” can be a reference to the forward path
inside the organism or to the feedback path outside of it. Of course it
could also refer to the whole loop but that’s not one of the meanings I
see here.

If there is no behavior that is not part of a control loop, then there is
no second variable that contributes to the state of a controlled variable
that causes behavior to change but does not change the “way”
the perception is controlled.

If the second input does not change the controlled variable, there will
be no change in the perception, the error, the output, or the observable
feedback effect on the controlled variable. In other words, no
disturbance and no action. If the second input does change the controlled
perception, then there will be an error and an action and a feedback
effect. Of course, as Martin says, there are perceptual variables that
are not under control. But if they don’t contribute to higher-order
controlled perceptions eventually, they cause no behavior (by the First
Principle of PCT).

There is one small loophole. By using large unnaturally rapid or
instantaneous step disturbances or impulse disturbances it is possible to
make delays in the system visible, and by that means to observe effects
of phenomena (but not the phenomena themselves) after the perception has
been disturbed but before the corrective action has been completed.
During this brief period of time the loop appears to be open, because the
action is on the verge of occurring but has not yet had any feedback
effect. Many psychological experiments are designed this way – I don’t
know whether conciously or not – because it prevents the problem of
having the behavior start to modify the effects of the stimulus before
the stimulus has gone more than a small part of the way through its
pattern of change. Psychologists are always trying to surprise their
experimental subjects, catch them off guard or give them a swift jolt
they can’t prevent – for good reason. S-R experiments don’t work right
under any other conditions. Imagine trying to measure the withdrawal
reflex by advancing the point of the pin toward a person’s skin at one
tenth of an inch per second, if the person can see the pin and can move.
Imagine increasing the light intensity over a period of ten seconds in
the Schouten experiment, and letting the subject control the perceived
light level with a continuous potentiometer rather than just pressing a
button for someone else to record.

This has also led to artificially restricting the nature of the
corrective action so instead of being smoothly variable and having graded
effects on the controlled variable, it too is turned into an
instantaneous jump from one state to another – the button or key press
is a ubiquitous example. So a continuous analog world is turned into a
discrete digital world to make the assumptions of S-R theory seem more
plausible. One legacy of years of this sort of strategy is the
proliferation of computer gaming devices which can be activated only by
rapid button-pressing – yet are used to attempt to control analog
variables on the screen. Like it or not, they are forced to allow analog
inputs in the form of smoothly-variable rates of pressing.

Oh yes, one more thing. At one point early in this discussion, Martin
seemed to allow an open-loop closed-loop system by having the reference
signal copied back into the perceptual path as an imagination signal at
the same time it was acting to produce the real button press, which was
not directly sensed. This is a version of the Von Holst-Mittlestaedt
“reafference copy” theory of feedback control (don’t know how I
managed to dredge that one up), and is also much like the idea that
“modern control theory” is trying to sell us along with dead
parrots. It would be interesting some day to try to collect all the
alterative attempts to explain control processes by people who knew
essentially nothing about it, starting with “compensatory
responses” or even Watson’s explanation of how the bird removed from
its nest gets back to the nest (the “sight of the nest” made it
struggle in that direction).

The principle is apparently that psychologists are certainly smart enough
to figure it out for themselves and don’t want those pesky engineers to
get their noses into the tents. A quick read of the Executive Summary and
we’re off.

Wasn’t Gary Czico working on something like that for a book – “Not
quite PCT”? Wonder how it’s progressing.

And the one explanation that really works is so simple and
elegant!

Best,

Bill P.

CARTER_CLOYD · June 15, 2009, 1:37pm

Get well soon,

Carter

···

— On Sat, 6/13/09, Bill Powers powers_w@FRONTIER.NET wrote:

From: Bill Powers powers_w@FRONTIER.NET
Subject: Re: More on Schouten article
To: CSGNET@LISTSERV.ILLINOIS.EDU
Date: Saturday, June 13, 2009, 11:52 PM

[From Bill Powers (2009.06.13.2005 MDT)]

Rick Marken (2009.06.13.1700)–

Martin Taylor (2009.06.13.14.29)–

RM: you either have to determine what I is (test for the controlled
variable) or simply make a reasonable guess at what I is and then set
up the simultaneous equations that define the system-environment
interaction (DV = f(I), I = G(IV,DV)) and see how the model works.

MT: And none of this involves observing the way the IV relates to the DV? Give me a break!

Take my word for this: the whole problem in this argument come from not defining terms and using them consistently. If we would all stop and say “Wait a minute, I think I disagree with X, but exactly what did you mean by X?” most of our go-arounds would stop.
Martin asks, “And none of this involves observing the way the IV relates to the DV?” We could ask, "Involves in any way at all whether related to PCT or not? If that’s your
question, of course we observe the “way” the IV relates to the DV. But if he is asking whether the apparent dependence of the DV on the IV (which is the “way” we obtain by the observation he enquires about) has the same form as the actual dependence determined by the forward path inside the organism, the answer is, NO, the apparent relationship is determined primarily by the form of the feedback path from DV back to I, the input variable that is also affected by the DV. That is generally different from the form of the forward path.

One hint as to the natura of the problem was revealed in a recent post when Martin spoke of “I”, the input variable, as if it were something inside the organism. In effect, he was assuming that the controlled variable had been tested for and identified, while Rick and I were assuming that we could see a proximal stimulus I only in the environment outside the subject, and therefore could not know what the actual
controlled variable affected by the feedback was. Not without doing The Test.

If my model of the Schouten experiment is correct, which I suggest but do not claim, the button-press is apparently related to the light intensity by a straight-through causal path with a delay in it. Light on, button press. Light off, no button press. I propose that the actual forward path through the organism is

light on and no button press —>
relationship false —> error signal —> start button press —>
relationship true —> no error signal —> release button press.

In other words, the apparent causal function is approximately the derivative of the forward causal function, in fact its temporal inverse. If the light being on causes the button to be pressed, the button turns the error off, which says the apparent forward function is like an impulse, not a continuing condition, even though the light stays on. If the
button turns the light off as in the experiments Dick Robertson reported, the appearance of an impulse relationship is even stronger. I have found nothing in the Schouten data to tell us when the button was released on each trial. For all I know, “button press” is defined as down-up.

RM:
CV = aIV+bDV
DV = r - kCV

A control model of the experiment would consist of two simultaneous equations:

CV = aIV+bDV
DV = r - kCV

This is fine, but what did Martin say, exactly, that makes you think he doesn’t know this? I think it’s in the ambiguity of “the way the IV relates to the DV.” It actually relates in two ways, not one, as shown by the simultaneous equations above. You can rewrite them as

CV = aIV+bDV
CV = (r - DV)/k

Which shows that we can eliminate CV and leave only a relationship between IV and DV. It is

aIV + bDV = r/k - DV/k

or

aIV + (b + 1/k)DV = r/k

If k, the output gain, is a large number, this becomes close to

aIV + bDV = 0, or

DV = -(a/b)IV

The path from disturbance to CV is the factor a, and from action to the CV through the environmental feedback path is the factor
b.

Notice that the forward gain, k, disappears from the solution. So if k is large enough, the relationship between IV and DV is determined almost entirely by the two environmental connections: IV to CV, and DV to CV.

This is in fact, exactly how operational amplifiers are used in analog computing. Op amps have gains ranging from 20,000 to a billion, so the input-output function is determined by the component in the feedback path and the component in series with the input path, both connected to the inverting input terminal. The computation accuracy when the gain is a billion is better than one part per million over a modest band of frequencies. I saw this one in a self-balancing amplifier used for millidegree temperature control.

When gains are not quite so high, the forward properties begin to have some small effect, which gets larger when the gain is under 10 or so or the paths are complicated by time delays, thresholds, limits,
or hysteresis. The Schouten situation involves some rather extreme complications including having to overcome strong tendencies to avoid obeying the instructions. The usual performance in reaction-time experiments is to respond to the stimulus as fast as possible, so the controlled variable remains in an error state for the least possible time and is then restored to the no-error state. So the forward signal, generated by combining the stimulus signal with the feedback signal, is more or less a one-shot square wave of short duration, leading some experimenters to conjecture that it is the onset of the stimulus that causes the response, not the steady-state value. At one time, I recall, many psychologists seemed to think that organisms responded only to changes in stimulation and not to the steady state at all. That’s because as soon as the response occurred it stopped again. But that was because the steady-state error was corrected, not because the
organism was sensing only the first derivative of the stimulus. And the experimenters were looking only at the first derivatives of the responses! A “button press” was down-up, not just down.

Threshing these things out is time-consuming and exasperated, but it’s worth the trouble to get everyone synchronized. Loose talk sinks ships, and also theories.

Best,

Bill P.

Bill_Powers1 · June 6, 2009, 9:18pm

[From Bill Powers (2009.06.06.1316 MDT)]
It’s D-day, the sixth of June. In 1944, I had just finished High School,
and was headed soon toward boot camp and then Electronic Tech training in
the Navy, after which I would have to go off to war. Fourteen months
later I was waiting in San Franciso Bay for a troop ship to take me and
my fellow techs to the invasion of Japan. My job was going to be to plant
radio beacons on the beach where we invaders would land, not quite what I
had been trained for. It didn’t seem likely that I would survive that. I
was therefore both horrified and relieved, sitting in a barracks on
Treasure Island, when the Hiroshima bomb went off, then a week later the
Nagasaki bomb. My life was saved at a very high price which I wasn’t sure
it was worth. My love of physics was considerably reduced. The troop ship
came and we all went to Japan as planned, but as seasick occupiers
instead of a suicide squad. I ended up on Landing Craft Support 68,
code-named Sweetgum Eight, anchored in Tokyo Bay at the naval base town
of Yokosuka (pronounced Yo-Kuss-Ka). On one weekend liberty I took a
train up to Tokyo, 50 miles to the north, and ended up on The Ginza where
I played boogie-woogie piano, sitting in with a Japanese jazz band in a
bar. It was very strange to be a 19-year-old conquerer.
Yes, my ship was named “LCS-68”, painted on both sides of the
bow. So why am I only up to LCS-III?

···

============================================================================
The three beeps in the Schouten experiment are timed so the third beep
comes 150 milliseconds after the first one. This means that the person
can wait for the first beep and react to it, and will then press the
button about 150 milliseconds later, as instructed. That’s about a normal
well-practiced reaction time to the sudden onset of a stimulus intensity.
The right answer therefore has to be selected by the time of the first
beep, which at the earliest (150 msec for the 3rd beep)) is essentially
the time of the light’s turning on, where the mistakes occur about 50% of
the time – pure chance.
The actual forced reaction times recorded (figs 2 and 3) show a
distribution around the intended mean (the time of the third beep)
similar to that for the free reaction time, showing that there is a
source of randomness in the timing no matter what the reference delay
time is. We could treat this as a random disturbance of the actual delay
reference signal (say, in the output function of the system controlling
the intended reaction time).
I neglected to mention in the previous post that the reference signal for
delay in the free-reaction cases was apparently set to specify about 300
msec, after some practice. Different subjects set it
differently.
The parameter adjusted by the delay-controlling system could simply be
the gain of an integrating control system that reacts to the light by
pressing the bar when the output exceeds some fixed level, triggering a
pressing event. With low gain, the apparent reaction time would be
longer.
For the mistake-controlling system, it would be necessary for a higher
system to perceive whether the button-press was correct. Since the trials
were spaced three seconds apart and each light was left on for one
second, there was plenty of time to realize that a mistake had been made.
The spacing of the trials supports my guess in the previous post that the
channel capacity was not measured, but was only estimated from averaged
single-reaction data. That calculated channel capacity would be far
higher than possible under the conditions of the experiment. A light
could not be turned on every 300 milliseconds or so to measure
(successfully) what the real channel capacity would be. There is no way
to verify the correctness of the calculation.
There is nothing in all this either to support or deny the idea that the
subject was perceiving uncertainty or probability or likelihood or any
such thing. A working model doesn’t require any such labels to be applied
to any of its variables or parameters. Even if such uncertainty were
perceived (some people are evidently very sensitive to it), there
wouldn’t seem to be time for such a perception to be recognized and
accurately measured in the few tenths of a second available before a
response is initiated. Schouten says (p. 143) that the usual fraction of
errors in this binary reaction task varies between 0.3% and 6%, implying
that even one estimate would require 16 to 300 trials to approximate
roughly the actual fraction wrong. The current correct percentage would
have to be an average that could change only slowly.
I have to remind myself that this experiment was reported in 1967, when
PCT was also in a relatively primitive state, 6 years before B:CP.
Information theory was very popular then, and played a dominant role in
cybernetics. I don’t see it mentioned much any more in the literature I
see.
The main thing that hasn’t changed since then is the tendency to
overinterpret experimental results, attributing abstract traits and
motives to the behaving system which go far beyond anything supportable
by the experimental design. It’s not easy to kick that habit even if one
wants to. Apparently, many people just don’t want to, yet they seem to
have no trouble getting published in Science.

Best,

Bill P.

rsmarken · June 7, 2009, 4:51pm

[From Rick Marken (2009.06.07.0950)]

Bill Powers (2009.06.06.1316 MDT)--

It's D-day, the sixth of June. In 1944, I had just finished High School, and
was headed soon toward boot camp and then Electronic Tech training in the
Navy, after which I would have to go off to war. Fourteen months later I was
waiting in San Franciso Bay for a troop ship to take me and my fellow techs
to the invasion of Japan. My job was going to be to plant radio beacons on
the beach where we invaders would land, not quite what I had been trained
for. It didn't seem likely that I would survive that. I was therefore both
horrified and relieved, sitting in a barracks on Treasure Island, when the
Hiroshima bomb went off, then a week later the Nagasaki bomb. My life was
saved at a very high price which I wasn't sure it was worth.

I'm personally very glad you survived to develop PCT (which may
eventually save the world) and very sad for the horror rained on
Japan. Linda has been developing a hobby of making bonsai trees and we
learned that one of her sensei (teachers) was a grammar school student
in Hiroshima when the bomb fell. He survived because he ditched school
that day and went boating; he saw the explosion from his boat. He lost
his entire family, of course. So it goes.

I neglected to mention in the previous post that the reference signal for
delay in the free-reaction cases was apparently set to specify about 300
msec, after some practice. Different subjects set it differently...

For the mistake-controlling system, it would be necessary for a higher
system to perceive whether the button-press was correct. Since the trials
were spaced three seconds apart and each light was left on for one second,
there was plenty of time to realize that a mistake had been made.

I was going to give up on modeling this RT study because of
disappointing experiences we have had modeling data collected in
experiments where there was no real test for the controlled
variable(s). I'm thinking, for example, of your attempts to model some
of the "operant" behavior observed in scheduling studies. The basic
modeling approach looked promising and accounted for the data at the
level that the data was presented; press rate varies as necessary to
keep food delivery rate constant in the face of schedule changes. But
then I believe you and Bruce A did some studies suggesting that things
didn't actually work that way; rats didn't just sit at the bar and
press at a rate that produced a regular delivery rate. It only looked
that way in the averaged data. Also, I've had problems finding data
that would appropriately test my models of fly ball catching since
there is never any systematic effort to test for the controlled
variable in any of the published studies. I've also done some modeling
of the behavior observed in the very nifty studies performed by
Metzler and his colleagues. That work is up at my (again functional)
MindReadings site:

http://www.mindreadings.com/Coordination.html

But I was never able to get the model to explain what Metzler thinks
of as their main finding -- the effect of handle turning rate on error
rate (the data graph presented on the web page) -- so I kind of gave
up.

But now I think it might be worth it to do the RT task modeling
because even if the model is not perfect at least it would show how a
control model would explain the apparently "open loop" behavior in
this kind of task. If I could get a nice working model of the simple
"free" RT task then that model could be used to show how the apparent
S-R relationship between light onset (IV) and button press (DV) is
actually the disturbance-output relationship in a control loop with
something like the light/press relationship as the CV. Indeed, the
model could be used to show what things would look like if the
relationship between IV and DV were completely open-loop, with no CV
involved.

Best regards

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com

Bill_Powers1 · June 8, 2009, 8:39am

[From Bill Powers (2009.06.08.0109 MDT)]

From Rick Marken (2009.06.07.0950) --

I was going to give up on modeling this RT study because of
disappointing experiences we have had modeling data collected in
experiments where there was no real test for the controlled
variable(s). I'm thinking, for example, of your attempts to model some
of the "operant" behavior observed in scheduling studies. The basic
modeling approach looked promising and accounted for the data at the
level that the data was presented; press rate varies as necessary to
keep food delivery rate constant in the face of schedule changes. But
then I believe you and Bruce A did some studies suggesting that things
didn't actually work that way; rats didn't just sit at the bar and
press at a rate that produced a regular delivery rate. It only looked
that way in the averaged data.

What Bruce found was that when he corrected the apparent rate of lever pressing for the food collection time, the differences in rate disappeared. The rats apparently just pressed as fast as they could except when collecting the food, when they didn't press at all.

A lot of the problems I have found with old data aren't necessarily due to failure to carry out The Test. It's simply the experimenter's failure to pay attention to all the relevant details of what is going on, or the failure of someone citing an experiment to pass on everything that the experimenter reported. There seems to be a lot of pressure to get the experiment designed and under way, and then to get the data collected and published as fast as possible. I have seen this a number of times, and Bruce Abbott can cite some of his own examples, things he found when trying to replicate someone else's experiment. There is some pretty hasty work out there, and some sloppy reporting.

In the present case (Schouten experiment) there is one big glaring fact that is evident only when you look at the original data that Richard Kennaway got for us. I refer to Fig. 1 showing the free reaction times for subjects A and B.

For subject B, we have a histogram showing the number of reponses at each of a range of delay times. About 180 responses (the peak) occurred at a reaction time of 300 millisec, about 20 at 200 milliseconds, and about 18 at 400 milliseconds. The histogram makes a nice bell-shaped curve and contains points for 1000 responses.

Now we have to ask ourselves: why does the reaction time vary so much? For subject A it varies between about 220 and nearly 700 milliseconds, and for Subject B, between about 120 and 500 milliseconds. That is a huge range. And why does it vary on both sides of the peak in about the same way, though skewed toward the long side? When I think of modeling this behavior, I have to think of how to introduce that kind of variation. We could just say that there is a noisy timing mechanism inside the subject, and that would give us a model, but it would be nice if we could think of some mechanisms that would introduce that kind of noise. What random processes are going on that could account for this spread? One, of course, is the random selection of which light goes on. But others may be introduced by the subject, for example by looking in different directions.

Some answers might come out of knowing more about the physical setup. How bright were the lights? How bright was the background against which they were seen? How far apart were they -- could they both be foveated at once, or did the subject's eyes have to move (and how much) to look at one, then the other? What I'm trying to do here is see whether we should be looking at thresholds of detection of a foveated light source, or perhaps should be asking what happens if the subject's gaze is aimed more toward one light when the other turns on.

The fact that we don't know the answers to these questions tells us that the experimenters didn't think of them, or didn't think they were important. And as in so many experiments I have read about, it tells me that I can't come up with a good model unless I just do the experiment again myself.

As you said once, Rick, proposing a model and matching its behavior to real behavior is actually an advanced form of the Test. The model says "I think this kind of system is behaving," and matching it to behavior says "This model accounts for what is observed" (well or poorly). So when we find that an experiment isn't done or described well enough for us to construct a model and fit it to behavior, that tells us we can't do the Test. In fact, all we have to do is say that we don't have enough information to construct a model. That doesn't mean the experiment is worthless; it just means it has to be done all over again, this time recording everything that is important.

I think some triage is needed here. We don't need to go through all the foofahraw of beeps and forced reactions: there's enough to occupy us in simply modeling the free reaction experiments. That much we can set up easily on a computer.

I suggest that we keep the gaze fixed in one place, and simply present a short bar vertically or horizontally (or two different colors or small shapes). That way we won't have the subject looking at one light when the other goes on and everything can be foveated. If this reduces the variability of the reaction times, we'll have found one source of irrelevant random interference.

I also suggest that we try at least two levels of brightness of the target images, on a black background in a dimly-lit room with no screen reflections.

I suggest that we use two keys operated by two hands, so the hands don't have to move from wherever they are to the right key. An electromyograph should be used, if feasible, to find the actual neural reaction time uncontaminated by physical dynamics of the arm, hand, and key. I'm dreaming, of course.

I suggest that we make the response turn the display off so the subject knows every time whether the response was correct. This could tighten up control of the reaction time, if that is going on.

I suggest that we record any additional responses after the initial one during each trial, and that we signify the start of a new trial in some way -- perhaps by turning on a dot where the subject is to look. There would have to be a variable delay before turning the target on.

It would be interesting, if it can easily be done, to check the reaction time to a beep as well as to the optical target. And if someone has a lab handy, or can find some data sheets, it would be most welcome if we knew how the intensity of the optical target changes with time when it is turned on.

Also, I've had problems finding data
that would appropriately test my models of fly ball catching since
there is never any systematic effort to test for the controlled
variable in any of the published studies. I've also done some modeling
of the behavior observed in the very nifty studies performed by
Metzler and his colleagues. That work is up at my (again functional)
MindReadings site:

Bimanual Coordination

But I was never able to get the model to explain what Metzler thinks
of as their main finding -- the effect of handle turning rate on error
rate (the data graph presented on the web page) -- so I kind of gave
up.

Same problems, right? You have to do the experiment yourself. You could test to see whether it's the interference between two different rates of perceived movements (hands and flags) that creates the variability, just by giving the subjects a knob to turn that varies the speed of one flag. If the purpose of the experiment is to see whether the hand movements or the flag movements are under control, there's no need to have both hands going at once: just rotate one flag while the subject adjusts the other's speed to match, with or without a gear train or pulley to separate the flag's speed from that of the hand.

It's even simpler if the subject just moves a mouse fore and aft to cause the flag's speed to vary. That would make it pretty obvious whether the subject is controlling the action or the result, especially if you insert a slowly varying disturbance between the mouse and the speed of the flag.

Experiments like Metzler's are too complicated for the point they're trying to investigate. The tendency to couple movements of the hands makes the task unnecessarily difficult and just introduces variability for no reason. The main point can be made in a far simpler way.

But now I think it might be worth it to do the RT task modeling
because even if the model is not perfect at least it would show how a
control model would explain the apparently "open loop" behavior in
this kind of task.

You will notice that in the model I sketched in, there is nothing open-loop about the system at any time, except of course for the light turning on and off which is an independent variable anyway and would never be considered as part of a loop.

If I could get a nice working model of the simple
"free" RT task then that model could be used to show how the apparent
S-R relationship between light onset (IV) and button press (DV) is
actually the disturbance-output relationship in a control loop with
something like the light/press relationship as the CV. Indeed, the
model could be used to show what things would look like if the
relationship between IV and DV were completely open-loop, with no CV
involved.

Yes, I agree, though we hardly need to prove to ourselves that the light is a disturbance. We ought to demonstrate it formally, of course.

My suggestions don't all have to be followed to get this off the ground.

Best,

Bill P.

rsmarken · June 9, 2009, 3:21pm

[From Rick Marken (2009.06.09.0820)]

Bill Powers (2009.06.08.0109 MDT)--

As you said once, Rick, proposing a model and matching its behavior to real
behavior is actually an advanced form of the Test. The model says "I think
this kind of system is behaving," and matching it to behavior says "This
model accounts for what is observed" (well or poorly). So when we find that
an experiment isn't done or described well enough for us to construct a
model and fit it to behavior, that tells us we can't do the Test. In fact,
all we have to do is say that we don't have enough information to construct
a model. That doesn't mean the experiment is worthless; it just means it has
to be done all over again, this time recording everything that is important.

I guess the "worth" of the original experiment is to show that it has
to be done all over again. If that's what you consider "worth" then,
sure, experiments done using conventional methods (like Schouten's)
are very worthwhile. I actually think they have more cost than worth
since they were designed to test things that are simply not of any
particular interest to a control theorist.

I think some triage is needed here. We don't need to go through all the
foofahraw of beeps and forced reactions: there's enough to occupy us in
simply modeling the free reaction experiments. That much we can set up
easily on a computer.

That is precisely what I plan to do, not so much in order to
understand the "processing" going on in RT experiment but because I
think it would be useful as a way of showing what is going on in such
experiments from a PCT perspective.

I suggest that we use two keys operated by two hands, so the hands don't
have to move from wherever they are to the right key.

I was thinking of just modeling a simple RT experiment. I'm sure we
can find a nice simple one in the literature to use as a prototype.

I suggest that we record any additional responses after the initial one
during each trial, and that we signify the start of a new trial in some way
-- perhaps by turning on a dot where the subject is to look. There would
have to be a variable delay before turning the target on.

I'm not really interested in doing this. I would just l like to have a
nice, simple model of behavior in an RT experiment. One model would be
open loop and the other closed loop.

If I could get a nice working model of the simple
"free" RT task then that model could be used to show how the apparent
S-R relationship between light onset (IV) and button press (DV) is
actually the disturbance-output relationship in a control loop with
something like the light/press relationship as the CV. Indeed, the
model could be used to show what things would look like if the
relationship between IV and DV were completely open-loop, with no CV
involved.

Yes, I agree, though we hardly need to prove to ourselves that the light is
a disturbance. We ought to demonstrate it formally, of course.

Good, that's what I'll be working on, if I ever get the time.

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com

MartinT · June 10, 2009, 2:15pm

[Martin Taylor 2009.06.09.11.58]

[From Bill Powers (2009.06.08.0109 MDT)]

A lot of the problems I have found with old data aren't necessarily due to failure to carry out The Test. It's simply the experimenter's failure to pay attention to all the relevant details of what is going on, or the failure of someone citing an experiment to pass on everything that the experimenter reported. ...
In the present case (Schouten experiment) there is one big glaring fact that is evident only when you look at the original data that Richard Kennaway got for us. I refer to Fig. 1 showing the free reaction times for subjects A and B.

I don't think it is very fair to criticize a study for asking a question different from the one that interests you. It is indeed fair to criticize a study for not properly addressing a question that the study claims to be asking. That's the criterion I would use in assessing whether Rick's "Revolution" paper applies. The issue, for me, is similar to the question of whether it matters in addressing a particular question of mechanics whether their temperature is determined by the bodies' phlogiston content or by the motion of their molecules. When you talk about "relevant details" you have to be clear about the question to which the details might be relevant.

As soon as you talk about free reaction times, you are talking about the control loops involved in making the output button push, and you need to model what those loops might be controlling and how they might conflict. As I have said so many times over the months, that's an interesting question, but irrelevant to the question Schouten asked, and for which I initially introduced the Schouten experiment to CSGnet. Even if you know PCT, Schouten's question does not require any real consideration of those control loops, other than to question whether their performance differs as a function of the designated response moment. I believe Schouten used the free reaction times only to determine the range of times to be used in the main experiment, for which the analysis need only assume that those control loops exist and act the same way whatever the delay between light onset and third bip. It matters not a jot that Schouten had never heard of PCT, any more than it matters that Newton had never heard of Boltzmann.

The data I have been arguing to be useful and valid are those from timed responses, not from free reaction times. There is certainly a distribution of actual response times for each designated response time, but these distributions are very narrow compared to the free response time, and the only way in which the "button-pressing" control system enters into the analysis is in the assumption that its functioning is the same no matter what the designated response time. Clearly this assumption isn't true in exact detail, since there is a bias toward the middle in the actual response times for very short and very long designated response times. But in the middle range where the linear trend is observed, there is no obvious reason for asserting (and necesssarily modelling) an influence of the designated response time on the functioning of the control loop that connects the perception of which light is on with which button is pressed. The consistency and linearity of the information measures argues that this influence, though probably present, is negligible in practice.

Going back to the free reaction question, I find it quite encouraging that you are considering a PCT-based study of the speed-accuracy trade-off, which has a long history in conventional research, going back to at least Paul Fitts in the 50's. These are the kinds of study to which Rick's "Revolution" paper is likely to apply, where the action of the control loop (probably two in conflict here) are central to the results, in contrast to the Schouten study, where the caveat Rick subsequently said he had mis-stated does apply.

For subject B, we have a histogram showing the number of reponses at each of a range of delay times. About 180 responses (the peak) occurred at a reaction time of 300 millisec, about 20 at 200 milliseconds, and about 18 at 400 milliseconds. The histogram makes a nice bell-shaped curve and contains points for 1000 responses.

Now we have to ask ourselves: why does the reaction time vary so much? For subject A it varies between about 220 and nearly 700 milliseconds, and for Subject B, between about 120 and 500 milliseconds. That is a huge range. And why does it vary on both sides of the peak in about the same way, though skewed toward the long side?

Does a subject in a free reaction experiment respond when the perception of sureness about the correctness of a choice builds to a critical level? How would you test whether this is even a meaningful question?

When I think of modeling this behavior, I have to think of how to introduce that kind of variation. We could just say that there is a noisy timing mechanism inside the subject, and that would give us a model, but it would be nice if we could think of some mechanisms that would introduce that kind of noise. What random processes are going on that could account for this spread? One, of course, is the random selection of which light goes on. But others may be introduced by the subject, for example by looking in different directions.

Some answers might come out of knowing more about the physical setup. How bright were the lights? How bright was the background against which they were seen? How far apart were they -- could they both be foveated at once, or did the subject's eyes have to move (and how much) to look at one, then the other? What I'm trying to do here is see whether we should be looking at thresholds of detection of a foveated light source, or perhaps should be asking what happens if the subject's gaze is aimed more toward one light when the other turns on.

The fact that we don't know the answers to these questions tells us that the experimenters didn't think of them, or didn't think they were important.

Or that they were asking a different question than the one you ask, or that they had no way of measuring them, as Schouten would not for the gaze direction. If Schouten had been asking how the slope of the detection vs response time was affected by the brightness of the light, he would have thought of it and reported it. If I were doing the experiment, I would be likely to record the time of day for each trial, since I might wonder whether a subject's performance would be affected by that subject's diurnal cycle. So far as I know, Schouten didn't ask that, just as he didn't ask them whether they were worried about their current bank statement or a myriad of other things that might affect performance. Those influences or potential influences all are potential sources of variation in the data, but when the data are as tight as they are for the linear information gain curves, what we know is that they don't matter very much for the question being asked, even though they might be critical for some other question that might be asked.

I think some triage is needed here. We don't need to go through all the foofahraw of beeps and forced reactions: there's enough to occupy us in simply modeling the free reaction experiments. That much we can set up easily on a computer.

Good. But by eliminating the key element of Schouten's study you will be answering a question completely unrelated to the question Schouten asked, which concerns the way in which increased opportunity to observe the light affects the accuracy of perceptual judgments. You also are setting up a simpler experimental situation but (I think) a more complex psychological one.

I suggest that we keep the gaze fixed in one place, and simply present a short bar vertically or horizontally (or two different colors or small shapes). That way we won't have the subject looking at one light when the other goes on and everything can be foveated. If this reduces the variability of the reaction times, we'll have found one source of irrelevant random interference.

Good idea.

I also suggest that we try at least two levels of brightness of the target images, on a black background in a dimly-lit room with no screen reflections.

Good idea.

I suggest that we use two keys operated by two hands, so the hands don't have to move from wherever they are to the right key. An electromyograph should be used, if feasible, to find the actual neural reaction time uncontaminated by physical dynamics of the arm, hand, and key. I'm dreaming, of course.

Or you could dream further and get even closer to the events within the control loop, and monitor single neurons, as Kiani and Schadlen did in a closely related study that you did not like. Monitoring the final output muscular events really isn't a lot closer to the real action than monitoring the buttons, is it?

I suggest that we make the response turn the display off so the subject knows every time whether the response was correct. This could tighten up control of the reaction time, if that is going on.

I don't see what you mean, here. How would turning off the display after the response tell the subject whether the response was correct?

I suggest that we record any additional responses after the initial one during each trial, and that we signify the start of a new trial in some way -- perhaps by turning on a dot where the subject is to look. There would have to be a variable delay before turning the target on.

Yes. That is good practice. But what is an "additional response"? Are you suggesting that the subject will press both buttons, because if that is the case, wouldn't a sensible subject just press both buttons as fast as possible every time?

It would be interesting, if it can easily be done, to check the reaction time to a beep as well as to the optical target. And if someone has a lab handy, or can find some data sheets, it would be most welcome if we knew how the intensity of the optical target changes with time when it is turned on.

I rather think that Schouten's subjects would have known every time whether their response was correct. After all, the lights were clearly visible, and they knew what button they had pushed! Only if they had been confused as to which button went with which light would they not have known whether they were correct, and in such a case, I think any normal experimenter would have given them a little more training to eliminate the confusion. And what would be "additional responses" in this kind of experiment?

As I have said from the beginning of the discussion on Schouten's experiment so many months ago, to study the control loops involved in making the responses is a very worthwhile project, and I'm happy to see you getting involved in it. But it is almost completely irrelevant to the information rate question addressed by Schouten (and by Kiani and Shadlen).

Martin

Bill_Powers1 · June 10, 2009, 6:40pm

[From Bill Powers (2009.06.10.0922 MDT)]

Martin Taylor 2009.06.09.11.58 –

[From Bill Powers
(2009.06.08.0109 MDT)]

BP, earlier: A lot of the
problems I have found with old data aren’t necessarily due to failure to
carry out The Test. It’s simply the experimenter’s failure to pay
attention to all the relevant details of what is going on, or the failure
of someone citing an experiment to pass on everything that the
experimenter reported. …

In the present case (Schouten experiment) there is one big glaring fact
that is evident only when you look at the original data that Richard
Kennaway got for us. I refer to Fig. 1 showing the free reaction times
for subjects A and B.

MT: I don’t think it is very fair to criticize a study for asking a
question different from the one that interests you.

BP: You’ve said that before, and I didn’t agree with it then, either,
though I said nothing. There is a valid criticism of any study to the
effect that the right questions were not asked. My interest is in the
phenomenon, not in one person’s incomplete ideas about it. Conclusions
can very easily be altered by new observations that the investigator did
not realize were important.

In Schouten and Bekker’s Introduction, we find this:

“Experiments on binary stimulus-response reactions … seen to
indicate that the fraction of erroneous responses rises toward lower
reactions times (Fig. 1). Theoretically this is highly intriguing (see
e.g. Rapaport 1959). If we suppose that a subject reacts if and when he
has obtained a certain degree of certainty, say 98%, and if the time
needed for reaching that uncertainty varies, then the fraction of errors
of 2% should be independent of the actual reaction time.”

It’s not obvious at that point, but the Schouten paper shows that the
error rate DOES depend on reaction time, and thus that the reaction time
is NOT determined by reaching some fixed level of uncertainty. My
proposed model is an attempt to show what the reaction time does depend
on. Yours would seem to be based on the idea that a certain amount of
information must accumulate to produce a reaction, which would seem
contrary to what Schouten says he found.

MT: It is indeed fair to
criticize a study for not properly addressing a question that the study
claims to be asking. That’s the criterion I would use in assessing
whether Rick’s “Revolution” paper applies. The issue, for me,
is similar to the question of whether it matters in addressing a
particular question of mechanics whether their temperature is determined
by the bodies’ phlogiston content or by the motion of their molecules.
When you talk about “relevant details” you have to be clear
about the question to which the

details might be relevant.

BP: Schouten makes clear what the highest-order consideration is:
“Therefore a corroboration of the suspected dependence of the
fraction of errors upon reaction time would would rule out any theory
based on the strategy of the fixed certainty norm.” The main thrust
of this paper is to determine whether, in fact, the fraction of errors
depends on reaction time, as it appears to do in the lower panel of Fig.

If I understand the rather ambiguous writing, the authors believe
their results refute the idea that the reaction occurs when uncertainty
reduction reaches some fixed norm like 98%.
The above-cited statements by the authors would seem to make it clear
that a detail of importance is the fact that even in the free-reaction
case, reaction times shorter than some amount result in an increase of
errnoeous responses. The authors make it quite clear why they did the
other experiments in this paper:
"Even for histograms of 1000 reactions, as shown in Fig. 1, the
fraction of errors f(t) is rather accurate in the middle regions but not
in the lower tail end where one would like to check the suspected
rise.
“In order to investigate this lower tail end, we used the
method of forced reaction time.”

By forcing reaction times other than the free-reaction time, the authors
expected to demonstrate whether the error fraction did depend on reaction
time. In that they succeeded, but I find none of their explanations of
this phenomenon believable. I think a control-system model provides a
much better explanation.

MT: As soon as you talk about
free reaction times, you are talking about the control loops involved in
making the output button push, and you need to model what those loops
might be controlling and how they might conflict. As I have said so many
times over the months, that’s an interesting question, but irrelevant to
the question Schouten asked, and for which I initially introduced the
Schouten experiment to CSGnet.

BP: When you invite inspection of a report on an experiment, you can’t
dictate which aspects of it the reader is suppose to ignore. I say that
the control loops and particularly their interaction is highly relevant,
and in fact offer an explanation for some and possibly all of the
observed phenomena in this experiment. Schouten himself raised the point
that the requirements of speed and accuracy are in conflict, and that
further conflicts arise when subjects are asked to delay their reactions.
He didn’t call these requirements reference conditions, but that is what
they are. These are not true conflicts because they don’t prevent
successful control, but they are definitely interactions, and those
interactions account rather nicely for the fact, noted by the authors,
that the reaction times at given error rates are shifted toward longer
intervals in the forced-reaction case.

MT: Even if you know PCT,
Schouten’s question does not require any real consideration of those
control loops, other than to question whether their performance differs
as a function of the designated response moment. I believe Schouten used
the free reaction times only to determine the range of times to be used
in the main experiment, for which the analysis need only assume that
those control loops exist and act the same way whatever the delay between
light onset and third bip. It matters not a jot that Schouten had never
heard of PCT, any more than it matters that Newton had never heard of
Boltzmann.

BP: Of course nothing in Schouten’s question “requires” any
consideration of these control loops, just as nothing in it
“requires” calculating information flow. But if we want to
understand the phenomenon, we need to propose a testable model of it and
make sure it fits the data. You misintepret Schouten’s reason for the
free reaction times – in fact he used a much wider range of forced
reaction times than found in the free-reaction case (compare Fig. 4 with
Fig. 1 lower panel). Unfortunately Fig. 4 is group data and does not tell
us anything about any individual’s responses.

Schouten simply wanted to investigate whether reaction time affects the
error rate at the low end of the range. His conclusion was not that the
systems acted the same way regardless of the delay – it was to show just
the opposite, that the systems made more errors when the reaction times
were shorter than some amount. By forcing reaction times other than the
free-reaction time, the authors expected to demonstrate whether the error
fraction did depend on reaction time. In that they succeeded, but I find
none of their explanations of this phenomenon believable (or even very
seriously proposed).

MT: The data I have been arguing
to be useful and valid are those from timed responses, not from free
reaction times. There is certainly a distribution of actual response
times for each designated response time, but these distributions are very
narrow compared to the free response time,

BP: They are narrowest near the range where the forced reaction time is
the same as the free reaction time. Here the loop gain would be highest
because the tendency toward conflict is the least. The width increases
both below and above this reaction time (and would increase more if the
curves were normalized to the same peaks). You may not be interested in
that, but I am.

MT: …and the only way in which
the “button-pressing” control system enters into the analysis
is in the assumption that its functioning is the same no matter what the
designated response time. Clearly this assumption isn’t true in exact
detail, since there is a bias toward the middle in the actual response
times for very short and very long designated response times.

BP: That’s because of the conflict, I would guess. With the forced
reaction time, the intended reaction time is not zero, but some
longer time. I have concluded that subjects probably don’t attend much to
the third beep: they just react to the first one, and it take about 150
milliseconds to press the button after that, which happens to be the time
between the first and third beeps.

MT: But in the middle range
where the linear trend is observed, there is no obvious reason for
asserting (and necesssarily modelling) an influence of the designated
response time on the functioning of the control loop that connects the
perception of which light is on with which button is pressed. The
consistency and linearity of the information measures argues that this
influence, though probably present, is negligible in
practice.

BP: The word “functioning” covers a lot of territory. If you
count choosing a target for a button press as part of the functioning,
there’s a definite effect on choosing wrongly.

MT: Going back to the free
reaction question, I find it quite encouraging that you are considering a
PCT-based study of the speed-accuracy trade-off, which has a long history
in conventional research, going back to at least Paul Fitts in the 50’s.
These are the kinds of study to which Rick’s “Revolution” paper
is likely to apply, where the action of the control loop (probably two in
conflict here) are central to the results, in contrast to the Schouten
study, where the caveat Rick subsequently said he had mis-stated does
apply.

BP: The tricky part will be to decide how to get those mistakes to appear
at shorter reaction times – and how to get a distribution of reaction
times instead of just one constant time.

BP earlier: Now we have to ask
ourselves: why does the reaction time vary so much? For subject A it
varies between about 220 and nearly 700 milliseconds, and for Subject B,
between about 120 and 500 milliseconds. That is a huge range. And why
does it vary on both sides of the peak in about the same way, though
skewed toward the long side?

MT: Does a subject in a free reaction experiment respond when the
perception of sureness about the correctness of a choice builds to a
critical level? How would you test whether this is even a meaningful
question?

BP: I don’t know, but Schouten raised it. The trouble is that you can
imagine that there’s an uncertainty inside the subject, but there’s no
way to check to see if there really is. And why uncertainty and not just
a noisy signal with a threshold of detection? It’s pretty sure that there
is some noise level associated with pulse-frequency-coded signals, but
whether that is subjectively experienced as uncertainty is hard to
say.

Are you saying that sureness about correctness builds to a critical level
and then a reaction occurs? If so, that is the hypothesis that Schouten
thinks he has refuted.

BP earlier: I think some triage
is needed here. We don’t need to go through all the foofahraw of beeps
and forced reactions: there’s enough to occupy us in simply modeling the
free reaction experiments. That much we can set up easily on a
computer.

MT: Good. But by eliminating the key element of Schouten’s study you will
be answering a question completely unrelated to the question Schouten
asked, which concerns the way in which increased opportunity to observe
the light affects the accuracy of perceptual judgments. You also are
setting up a simpler experimental situation but (I think) a more complex
psychological one.

BP: The brightness question has to do with the illumination level at
which photon detection becomes an issue. I doubt that this will be the
answer – the light levels just wouldn’t be that low. Eliminating the
forced reaction experiments just means trying to reproduce the original
observations that led Schouten to do this experiment. I suspect that once
we get a model of free responding (with errors) to work, it will be
fairly simple to introduce the forced reaction times.

…

MT: Or you could dream further
and get even closer to the events within the control loop, and monitor
single neurons, as Kiani and Schadlen did in a closely related study that
you did not like. Monitoring the final output muscular events really
isn’t a lot closer to the real action than monitoring the buttons, is
it?

BP: The emg signals aren’t the muscle events (forces applied at tendons)
and the muscle events start considerably before the contact closure. By
using the earliest emg signals we find out when the first signals from
the nervous system arrive at the muscle (within a millisecond or so). As
I mentioned a few days ago, I found long ago that when the contact
closure takes 150 milliseconds after a visual stimulus, the first emg
signal shows up in about 50 milliseconds. And that was with a setup that
minimized the moment of inertia of the arm (forearm rotation about its
long axis).

BP earlier: I suggest that we
make the response turn the display off so the subject knows every time
whether the response was correct. This could tighten up control of the
reaction time, if that is going on.

MT: I don’t see what you mean, here. How would turning off the display
after the response tell the subject whether the response was
correct?

BP: You don’t turn off “the display”, you turn off the light
that was turned on – if the right button is pressed. If the light
doesn’t go out when you press the button, you pressed the wrong button.
As the experiment was done, the light stays on for one second whether the
button press was right or wrong.

BP earlier: I suggest that we
record any additional responses after the initial one during each trial,
and that we signify the start of a new trial in some way – perhaps by
turning on a dot where the subject is to look. There would have to be a
variable delay before turning the target on.

MT: Yes. That is good practice. But what is an “additional
response”? Are you suggesting that the subject will press both
buttons, because if that is the case, wouldn’t a sensible subject just
press both buttons as fast as possible every time?

BP: No, I’m suggesting that if the subject presses the wrong button and
knows immediately that it was wrong, the subject may well then press the
right button as quickly as possible. This would tell you how long it
takes to detect that the wrong button was pressed, and gives some idea of
the reaction time.

BP earlier: It would be
interesting, if it can easily be done, to check the reaction time
to a beep as well as to the optical target. And if someone has a lab
handy, or can find some data sheets, it would be most welcome if we knew
how the intensity of the optical stimulus changes with time when it is
turned on.

MT: I rather think that
Schouten’s subjects would have known every time whether their response
was correct. After all, the lights were clearly visible, and they knew
what button they had pushed!

BP: But they wouldn’t have known it until after sufficient delay to let
them be sure of which light was on. For the shortest delays that would be
one or two hundred milliseconds after the response. The lights were left
on for just one second, and then turned off whether the response was
right or wrong.

MT: Only if they had been
confused as to which button went with which light would they not have
known whether they were correct, and in such a case, I think any normal
experimenter would have given them a little more training to eliminate
the confusion. And what would be “additional responses” in this
kind of experiment?

BP: Again, the “additional responses” that I was thinking of
would be attempts to substitute a right button press for a wrong
one.

MT: As I have said from the
beginning of the discussion on Schouten’s experiment so many months ago,
to study the control loops involved in making the responses is a very
worthwhile project, and I’m happy to see you getting involved in it. But
it is almost completely irrelevant to the information rate question
addressed by Schouten (and by Kiani and Shadlen).

BP: It wouldn’t be irrelevant if all the phenomena turned out to be
explainable by a model in which information rate is not
considered.

But this is not the point here. If you are interested in information
rates, who am I to say you shouldn’t be? Something very interesting might
turn up if you investigate information rates, and it won’t turn up if you
don’t investigate them. If there’s any problem here, it’s that you think
I should be interested in information rates, and so far I’m not. But I am
interested in the phenomena in this experiment, and think that trying to
model them might be of interest both inside and outside of PCT. It’s
perfectly possible that once a good PCT model is put together, you will
see applications of information theory to it, and why not? You might even
get results sufficient to make me get interested even against my
will.

Best,

Bill P.

rsmarken · June 10, 2009, 8:19pm

[From Rick Marken (2009.06.10.1320)]

Martin Taylor (2009.06.09.11.58])--

Going back to the free reaction question, I find it quite encouraging that
you are considering a PCT-based study of the speed-accuracy trade-off, which
has a long history in conventional research, going back to at least Paul
Fitts in the 50's. These are the kinds of study to which Rick's "Revolution"
paper is likely to apply, where the action of the control loop (probably two
in conflict here) are central to the results, in contrast to the Schouten
study, where the caveat Rick subsequently said he had mis-stated does apply.

I still don't get this. Aren't the subjects in both cases having to
control something? In the free reaction time experiment the subject
must control for hitting a button when the light comes on. If the
subjects didn't control for this, the light would just be another
thing that happens; lights coming on do not ordinarily lead to button
pressing. The same is true in the forced reaction time task, but now
there is more to control: the subject has to control for pressing the
button when the light comes on but now the subject also has to control
for pressing when the third beep comes on. What caveat of mine
applies in the forced reaction time experiment that doesn't apply in
the free one?

Best

Rick

···

--
Richard S. Marken PhD
rsmarken@gmail.com

Dick_Robertson · June 10, 2009, 8:34pm

[From Dick Robertson,2009.06.10.1525CDT]

BP, earlier: A lot of the problems I have found with old data aren’t necessarily due to failure to carry out The Test. It’s simply the experimenter’s failure to pay attention to all the relevant details of what is going on, or the failure of someone citing an experiment to pass on everything that the experimenter reported. …

"Experiments on binary stimulus-response reactions … seen to indicate that the fraction of erroneous responses rises toward lower reactions times (Fig. 1).

(Shouten) “Theoretically this is highly intriguing (see e.g. Rapaport 1959).”

I reacted quickly, with incredulity, when I first read this in Shouten’s paper. Since it has come up again, I can’t resist a comment. Ask any tennis player, fencer, hunter, typist (I could name dozens more examples): What happens when you have to react faster? If he doesn’t say, “You make more mistakes,”
I’d faint dead away, or something.

If my understanding of the point at issue is oversimplified, I’ be grateful to be enlightened.

Best,
Dick R

MartinT · June 11, 2009, 4:50am

[Martin Taylor 2009.06.11.00.16]

[From Rick Marken (2009.06.10.1320)]

Martin Taylor (2009.06.09.11.58])--

Going back to the free reaction question, I find it quite encouraging that
you are considering a PCT-based study of the speed-accuracy trade-off, which
has a long history in conventional research, going back to at least Paul
Fitts in the 50's. These are the kinds of study to which Rick's "Revolution"
paper is likely to apply, where the action of the control loop (probably two
in conflict here) are central to the results, in contrast to the Schouten
study, where the caveat Rick subsequently said he had mis-stated does apply.

I still don't get this. Aren't the subjects in both cases having to
control something?

Most certainly they are. Why would you think I might question that?

In the free reaction time experiment the subject
must control for hitting a button when the light comes on. If the
subjects didn't control for this, the light would just be another
thing that happens; lights coming on do not ordinarily lead to button
pressing.

Yes. The reverse is more usual in everyday life

The same is true in the forced reaction time task, but now
there is more to control: the subject has to control for pressing the
button when the light comes on but now the subject also has to control
for pressing when the third beep comes on.

Yes.

What caveat of mine
applies in the forced reaction time experiment that doesn't apply in
the free one?

That if the causal path between IV and DV is open loop, as it is in
most detection experiments despite the intervening and surrounding
control loops, then the results of IV-DV analysis can be valid and
useful (page 2).
The reason this is not the case in the free reaction experiment is that
the experiment concerns the workings of the control processes, because
of the intrinsic conflict between the control for fast response and the
control for accurate response. In the standard forced-choice detection
study there is no such conflict , because the subject has essentially
infinite time to respond (usually more than a second), and in the
Schouten study there is no conflict, or rather, the conflict is
suppressed, because the timing of the response is constrained. In both
cases, the performance of the control loops, whose eventual visible
output is the button push, is unlikely to be affected by the choice of
disturbance (one light or the other, tone in first or second interval,
etc.). That statement is, of course, unprovable, but a model of the
process that incorporated an effect of the choice of disturbance on the
performance of the various control systems would be quite different
from any model yet suggested for the Schouten experiment. Furthermore,
there is no suggestion in the data other than the slight bias toward
late responses for very early bips and early responses for very late
bips that the “fast v. accurate” conflict that probably exists is
affecting the performance of the control processes. The data are much
more likely to reflect the performance of the pathway between the
physical disturbance and the input to whatever perception is being
controlled (I argue for the controlled perception being a match between
an abstract “position” perception and an equally abstract remembered
“answer” perception, but whether that’s correct is really irrelevant as
well as being impossible to determine from within the experiment; and
just how would one do The Test to see whether that was actually
the controlled perception?).

The free-reaction condition is quite different. In this study, the
actual ability to distinguish one light from the other is not at issue.
If the subject waits long enough before responding, the answers will
almost always be correct. What is at issue is how long it takes for the
subject to resolve the conflict between a control system whose error
increases as time goes by (controlled perception “fast response”) and
one whose error decreases as time goes (controlled perception “accurate
response”). That means that the experiment is not concerned with the
pathway between the physical manifestation of the disturbance and its
contribution as input to the Perceptual Input Function of a controlled
perception, except insofar as that pathway introduces a time delay
between disturbance and output. The experiment is concerned with the
functioning of the operating control systems. Even if one uses varied
disturbances such as lights of different intensities, the resolution of
the conflict is still going to be important in modelling the results.

Anyway, I hope the free-reaction experiments do go forward, and that
modelling the conflict (or any other model) provides interesting
information. It would be even more interesting if under the same visual
presentation conditions, a timed response experiment was used with the
same subjects to provide a measured pathway to plug into at least part
of the model.

Martin

rsmarken · June 16, 2009, 4:02am

[From Rick Marken (2009.06.15.2100)]

Bill Powers (2009.06.14.1346 MDS)–

I’m really glad to hear that the operation went well.

Rick Marken (2009.06.14.1100)–

Maybe I’m not understanding
Martin but he seems to be agreeing that, in experiments on closed loop
systems, this behavioral illusion is a problem in some experimental
(those like the free RT task) but not in others (like the forced RT
task). I think the distinction is captured (somehow) in this paragraph
from Martin:

Surely if you are looking at the functioning of pathways within the
loop the close-loop property becomes essential, does it not? The issue is
hidden in those words “the system under study”. It is always
true that the DV is determined by the output of a closed-loop process.
The question is whether that closed-loop process is the part of the
IV->DV pathway that is under study. It is not always the case that
every perception is controlled, nor is it the case that variation in the
attributes of an uncontrolled input to a controlled perception causes
variation in the way the controlled perception is
controlled.

BP: Notice that there are one unspoken assumption here and one ambiguity,
in the last sentence: the assumption that there can be a behavior (DV)
that is not part of a control loop, and the ambiguous use of
“way” in stating that a second input to a controlled perception
can change without causing any change in the way the controlled
perception is controlled.

Let’s see what Martin has to say about this.

Take care of yourself and rest so that you’re sharp and ready to go.

Best regards

Rick

···

–
Richard S. Marken PhD
rsmarken@gmail.com

Bill_Powers1 · June 8, 2009, 4:09pm

[From Bill Powers (2009.06.08.0839 MDT)]

Rick Marken (2009.06.07.0950) –

The question of open vs. closed loop still nags at me. How can we explain
what is meant by “closed loop?” We need a clear and unambiguous
way to talk about this.

What we observe in the relevant kind of experiment is this:

independent --------> Organism -----> dependent

variable
variable

When the organism is a control system, the actual case is this:

independent ----> [qi ----> Organism] ----->
dependent

variable
^
variable

----------<----------------

The input quantity qi is the controlled quantity that is affected by both
the independent variable and the dependent variable. Before qi has been
identified (it is some unknown function of environmental variables
affected by the independent variable), there is no way to see that the
independent and dependent variables are affecting the controlled quantity
equally and oppositely. The square brackets are meant to indicate that qi
can’t be seen independently of the organism.

From outside the organism, we do not see qi, but only a set of physical
variables that the organism can sense. The controlled input quantity is
some function of those proximal variables. We can write

qi = Fi(v1, v2, … vn].

where v = a proximal variable

This was shown in the 1972 Science article (reprinted in LCS-1):

Emacs!

If we happen to perceive those proximal variables using a “Sensor
function” (input function Fi) like the one in the organism, we will
see a quantity being controlled; otherwise we will just see changes in
various proximal variables that are not opposed by effects of the
dependent variable. A large infinity of possible functions of those
variables exists; the chance of finding the right one by random guessing
is zero. There’s a progressively better chance if the observer and the
behaving system are of the same species, have similar histories, and have
communicated with each other about things they control.

If this is a diagram of a higher-order control system, the sensor
function will include the input functions of all lower levels, where the
v’s will now represent such lower-order variables, some controlled and
some not. The controlled ones will vary less than the uncontrolled ones
and so will be less affected by the disturbance.

The effect of the disturbance is “open-loop” in the two ASCII
diagrams above from the disturbance to the point where its effect enters
qi, or enters the circle around the v’s in the Science diagram. That is,
there is no return path from the v’s directly to the disturbance.
However, the v’s are in the feedback loop, and that loop is always
closed. It remains closed in the Schouten experiment, too. This
means that there is no way for the disturbance to affect the input
quantity without its effects on the closed loop being modified by
feedback from the dependent variable. The dependent variable is not a
simple function of the independent variable, but the kind of function
that can be described only by the control-system equations – a
simultaneous solution of the equations describing the forward and
feedback relationships.

Even when the basic variables are discrete events, perceptions can be
formed which depend on repeated occurrances of the events, cumulative
effects such as running averages or sums, and functions of such
time-spanning effects. Thus Schouten can speak of subjects adjusting
their reaction times to accomodate the conflicting requirements of the
task. This adjustment is not done during a single trial (that would be
impossible), but over many trials, and the variable being controlled is
derived from the results over many trials and affected by the successive
adjustments, a little at a time, over many trials. Therefore the fact
that one response can’t change the effect of a just-previous stimulus
event is irrelevant to the question of whether the control loop is
closed. The loop remains closed with respect to time-spanning functions
of the variables, and those are what is being controlled.

Some of our problems in discussing these matters have arisen because of
omitting time-spanning functions from consideration. In bowling, for
example, there is no way to correct the aim once the ball has been
released; it travels “ballistically.” The same applies to
firing artillery shells or hitting tennis balls. But this does not mean
those behaviors occur open-loop. The patterns of reference signals for
component control systems are slowly varied over many trials so as to
bring the average result as close as possible to the reference level. The
function defining the controlled quantity computes a time-spanning
function of individual events. There is no attempt to control each
individual outcome for the simple reason that the physical situation
makes that impossible. But the average effect can be very tightly
controlled. That’s how Roger Federer won the French Open yesterday – at
last.

In the Schouten experiment, I am proposing (after Schouten, who
identified some controlled variables without knowing he was doing that)
that the average reaction time is varied so as to satisfy two control
systems as nearly as possible: one controlling for zero reaction time and
the other for zero mistakes. When the beeps are introduced, the reference
level for reaction time is forced to increase from zero to a setting
between 150 and 700 milliseconds, inclusive. The article didn’t mention
whether the timing of the beeps was randomized (an exasperated remark was
considered at this point and omitted), so we don’t know how to model the
small adjustments of reaction time required when the beep timing
changes.

I think we can conclude that there is never a time in any condition of
the Schouten experiment when any of the control loops is broken. As in
bowling, tennis, and the aiming of artillery fire, the loops are always
closed and the controlled variables are defined as functions with values
that vary slowly over many repetitions of the discrete events. The
discrete events, individually, are uncontrolled.

Best,

Bill P.

(Attachment 6c7aa1.jpg is missing)