Predictive Control

CSGnet Archive CSG2016
1

[From Rick Marken (2016.02.23.0850)]

RM: Here’s a paper on object interception (dragonfly’s intercepting their prey) which purports to show evidence that inter models that predict the future position of the object (prey) are used in this control process.

https://www.dropbox.com/s/wb8w03glkazi2de/leonardo2015.pdf?dl=0

RM: Henry Yin asked me to review it for him but I’m having difficulty understanding it. In particular, I don’t understand the evidence they present for prediction being involved in this behavior. So I’m posting this paper here in the hopes that someone – particularly someone who believes that prediction is involved in the controlling done by living systems – can help me out with this. This is kind of embarrassing because object interception is one of my main areas of studies – I’ve published several papers on the subject – and yet I don’t understand what these researchers are showing. They have done some very impressive experimentation and data collection. But they don’t seem to have compared the behavior they observed to the behavior of a predictive or non-predictive model of object interception. Their conclusions seem to be based on discrepancies between what they observe and what they expect to observe if prediction were not involved in this behavior. But I don’t understand what they are observing (what their data is) or why they think it differs from what would be expected if there not prediction involved.

RM: Any help I can get on this would be greatly appreciated.

Best

Rick

···

–
Richard S. Marken

Author, with Timothy A. Carey, of Controlling People: The Paradoxical Nature of Being Human.

2

[From Adam Matic, 2016.02.23]

If I understood correctly, it seems they first established the duration of signal travel around the loop, or duration of travel from sensors to muscles. Then they say, If there is no prediction, the reaction lag should be about the same duration it takes the sensory signal to travel to muscles, or longer. If there is prediction, the reaction lag in pray tracking will be shorter than this loop time.

From the paper:

The compensatory head rotations could be driven reactively, based on
sensory feedback from prey-image drift, or predictively, based on internal
models of drift. To discriminate between these two possibilities,
we examined the delay between the summed foveation disturbances
(Fig. 4a, b) and the corrective head rotations (Fig. 4c). Insect muscle
contractions require approximately 5 ms to produce force, and any
visually driven reactive movements of the head should display an even
larger sensorimotor lag (for example, 47 ms from vision towing, Fig. 1e).
In contrast, if the dragonfly predicted foveal drift, the head could be
moved with zero delay for optimal cancellation, especially in the constant speed
prey condition in which there is no unexpected manoeuvring.
We found that head movements occurred with an average lag of 4 (+/- 4 ) ms after the disturbance they cancelled (Fig. 5c). Such an exceedingly
brief delay is strong evidence for the head being controlled predictively
rather than reactively.

AM:

Their reasoning seems ok, but this 47 ms seems rather long for a small bugger as a dragonfly, though. Even the 5 ms for muscle contraction seems long. And the zero delay seems a bit optimistic. My not very informed guess (don’t know the first thing about dragonflies) would be that there is some kind of error in these estimates and that loops are much shorter. On the other hand, drosophila visual processing lag is estimated at 30 ms, so maybe I’m completely wrong, don’t know how these estimates are made.

Another thing is that there aren’t necessarily complex models of pray behavior, but some different sort of prediction could be used, and the authors acknowledge this, but discard it as unlikely:

from the study:

“We excluded other reactive steering strategies based on prey angular velocity by evaluating the general prediction of such models that every significant prey manoeuvre is rapidly met by a corrective manoeuvre from the pursuer”

AM:

As you say, they don’t really try to make a simulation (that would also contain an internal model) of the flight pursuit. The question is of course, if there is an internal model - what kind of a model is it. That might be the main objection, since those models might prove to be biologically implausible or they might not work as well in a simulation as in the real world, or the information required for the model to work might not be available in the real brain. They only give a hypothetical diagram of the control circuit with the internal models. Maybe that is their next step in this research.

Best,

Adam

···

On Tue, Feb 23, 2016 at 5:52 PM, Richard Marken rsmarken@gmail.com wrote:

[From Rick Marken (2016.02.23.0850)]

RM: Here’s a paper on object interception (dragonfly’s intercepting their prey) which purports to show evidence that inter models that predict the future position of the object (prey) are used in this control process.

https://www.dropbox.com/s/wb8w03glkazi2de/leonardo2015.pdf?dl=0

RM: Henry Yin asked me to review it for him but I’m having difficulty understanding it. In particular, I don’t understand the evidence they present for prediction being involved in this behavior. So I’m posting this paper here in the hopes that someone – particularly someone who believes that prediction is involved in the controlling done by living systems – can help me out with this. This is kind of embarrassing because object interception is one of my main areas of studies – I’ve published several papers on the subject – and yet I don’t understand what these researchers are showing. They have done some very impressive experimentation and data collection. But they don’t seem to have compared the behavior they observed to the behavior of a predictive or non-predictive model of object interception. Their conclusions seem to be based on discrepancies between what they observe and what they expect to observe if prediction were not involved in this behavior. But I don’t understand what they are observing (what their data is) or why they think it differs from what would be expected if there not prediction involved.

RM: Any help I can get on this would be greatly appreciated.

Best

Rick

Richard S. Marken

Author, with Timothy A. Carey, of Controlling People: The Paradoxical Nature of Being Human.

–

3

[From Rick Marken (2016.02.24.1220)]

Adam Matic (2016.02.23)
AM: If I understood correctly, it seems they first established the duration of signal travel around the loop, or duration of travel from sensors to muscles. Then they say, If there is no prediction, the reaction lag should be about the same duration it takes the sensory signal to travel to muscles, or longer. If there is prediction, the reaction lag in pray tracking will be shorter than this loop time.
From the paper:

The compensatory head rotations could be driven reactively, based on sensory feedback from prey-image drift, or predictively, based on internal models of drift. To discriminate between these two possibilities, we examined the delay between the summed foveation disturbances (Fig. 4a, b) and the corrective head rotations (Fig. 4c). Insect muscle contractions require approximately 5 ms to produce force, and any visually driven reactive movements of the head should display an even larger sensorimotor lag (for example, 47 ms from vision towing, Fig. 1e). In contrast, if the dragonfly predicted foveal drift, the head could be moved with zero delay for optimal cancellation, especially in the constant speed prey condition in which there is no unexpected manoeuvring. We found that head movements occurred with an average lag of 4 (+/- 4 ) ms after the disturbance they cancelled (Fig. 5c). Such an exceedingly brief delay is strong evidence for the head being controlled predictively rather than reactively.

RM: I didn't even notice this. Nice find. We get lags of nearly 0 between output and disturbance in our tracking tasks and the sensory to output transport lag is surely on the order of 100 msec. So that would imply that there is prediction involved in our tracking tasks, even though, in the compensatory task, there is no basis for prediction (you can't see the disturbance so you can't predict it) and models with no prediction account for 99.9% of the output variance in the task (with no lag). Also, my own interception models, which contain no prediction, account for over 95% of the variance in output in interception tasks. I think what these folks need, in order to show that prediction is really involved in object interception, is to show that an interception model with prediction does substantially better than one without prediction.
RM: I guess my main problem with this paper is that they don't base their conclusions on the results of modeling. Maybe that's all I need to write as a comment on this paper. My evaluation can be summarized as: Great research; needs models.
Best
Rick
>

AM:

Their reasoning seems ok, but this 47 ms seems rather long for a small bugger as a dragonfly, though. Even the 5 ms for muscle contraction seems long. And the zero delay seems a bit optimistic. My not very informed guess (don't know the first thing about dragonflies) would be that there is some kind of error in these estimates and that loops are much shorter. On the other hand, drosophila visual processing lag is estimated at 30 ms, so maybe I'm completely wrong, don't know how these estimates are made.

Another thing is that there aren't necessarily complex models of pray behavior, but some different sort of prediction could be used, and the authors acknowledge this, but discard it as unlikely:
from the study:

"We excluded other reactive steering strategies based on prey angular velocity by evaluating the general prediction of such models that every significant prey manoeuvre is rapidly met by a corrective manoeuvre from the pursuer"

AM:

As you say, they don't really try to make a simulation (that would also contain an internal model) of the flight pursuit. The question is of course, if there is an internal model - what kind of a model is it. That might be the main objection, since those models might prove to be biologically implausible or they might not work as well in a simulation as in the real world, or the information required for the model to work might not be available in the real brain. They only give a hypothetical diagram of the control circuit with the internal models. Maybe that is their next step in this research.

Best,

Adam

[From Rick Marken (2016.02.23.0850)]
RM: Here's a paper on object interception (dragonfly's intercepting their prey) which purports to show evidence that inter models that predict the future position of the object (prey) are used in this control process.
<Dropbox - Error - Simplify your life; Dropbox - Error - Simplify your life

RM: Henry Yin asked me to review it for him but I'm having difficulty understanding it. In particular, I don't understand the evidence they present for prediction being involved in this behavior. So I'm posting this paper here in the hopes that someone -- particularly someone who believes that prediction is involved in the controlling done by living systems -- can help me out with this. This is kind of embarrassing because object interception is one of my main areas of studies -- I've published several papers on the subject -- and yet I don't understand what these researchers are showing. They have done some very impressive experimentation and data collection. But they don't seem to have compared the behavior they observed to the behavior of a predictive or non-predictive model of object interception. Their conclusions seem to be based on discrepancies between what they observe and what they expect to observe if prediction were not involved in this behavior. But I don't understand what they are observing (what their data is) or why they think it differs from what would be expected if there not prediction involved.
RM: Any help I can get on this would be greatly appreciated.
Best
Rick

--
Richard S. Marken

Author, with Timothy A. Carey, of <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.amazon.com_Controlling-2DPeople-2DParadoxical-2DNature-2DBeing_dp_1922117641_ref-3Dsr-5F1-5F1-3Fs-3Dbooks-26ie-3DUTF8-26qid-3D1449541975-26sr-3D1-2D1&d=BQMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=-dJBNItYEMOLt6aj_KjGi2LMO_Q8QB-ZzxIZIF8DGyQ&m=qYbFIdOjmxVYS0cyQLMT39VHm01n-n1jCruaPc6XJJs&s=vagPR5WmFttSDpXxtdLPEfhgZuyNYnTqRG6vKPq30LA&e=&gt;Controlling People: The Paradoxical Nature of Being Human.

···

On Tue, Feb 23, 2016 at 5:52 PM, Richard Marken <<mailto:rsmarken@gmail.com>rsmarken@gmail.com> wrote:

--
Richard S. Marken
Author, with Timothy A. Carey, of <https://urldefense.proofpoint.com/v2/url?u=http-3A__www.amazon.com_Controlling-2DPeople-2DParadoxical-2DNature-2DBeing_dp_1922117641_ref-3Dsr-5F1-5F1-3Fs-3Dbooks-26ie-3DUTF8-26qid-3D1449541975-26sr-3D1-2D1&d=BQMFaQ&c=8hUWFZcy2Z-Za5rBPlktOQ&r=-dJBNItYEMOLt6aj_KjGi2LMO_Q8QB-ZzxIZIF8DGyQ&m=jBTaKLB63nMe1DYXSRIVBYHPaBMUoCyM5DMEuRRlJ58&s=UhCZwC2m5Vw7iN5vGdeS5zOhanwrnG-fwBHbbdziBlc&e=&gt;Controlling People: The Paradoxical Nature of Being Human.

4

[From MK (2016.02.26.1800 CET)]

Rick Marken (2016.02.23.0850)--

RM: Henry Yin asked me to review it for him but I'm having difficulty
understanding it. In particular, I don't understand the evidence they
present for prediction being involved in this behavior.

Rick, read the paper while occasionally substituting "predictive" with
"purposive" and "cancellation" with "control" and you should be able
to make more sense of the paper. The "reactive vs predictive" in this
paper roughly corresponds to a PCT influenced researcher's "S-R vs
purposive."

Like Sawtell's group at Janelia, the people at Antonio's lab seem to
be thinking of what a PCTer would refer to as "perceptual control" as
"(sensory) cancellation." Their "predictions" correspond to PCTers
"purposes" or "references". Both groups have produced papers that
contain graphs that look like run-of-the-mill tracking runs. This
paper contains a labeled CV. I'd be surprised if the "negative copy"
in this paper doesn't correspond to the "negative images" of the
papers from NS's group. MM. et al/ ("Antonio, 2015") have identified
the goal-tied nature of the control activity, but not related it
clearly enough to their "predictions", which of course _are_ the
goals. Both groups have provided at least some indirect support
against the mostly strawmannish "feedback is too slow" objection,
although I think that point was made a bit more explicitly in one of
the papers from NS's group. I know that NS's former grad student, who
did some of the modeling/control work, had not read Powers and it's
clear that Michiani hasn't done so, either. It would not surprise me
if what people speak about as a "inverse model" is a normally
functioning control system. I'm not familiar enough with their work to
tell how much of the verbiage is there for the appeasement of
reviewers, but the "internal models" in this paper seems to be spoken
about as "something that isn't a fictionally stereotypical I-O
system."

M

5

[From MK (2016.02.26.1945 CET)]

MK (2016.02.26.1800 CET) --

I know that NS's former grad student, who
did some of the modeling/control work, had not read Powers and it's
clear that Michiani hasn't done so, either.

Sorry, it looks like I managed to send a draft version of the post.
That should read Mischiati.

M

6

[From MK (2016.02.26.2120 CET)]

That should read Mischiati.

...and "Antonio" should read "Anthony" and should have actually been
written as "Leonardo," and Sawtell isn't actually at Janelia, although
I, for a reason I can't figure out, seem to have him mentally placed
there.

I think I'm in need of some sleep. :slight_smile:

M

7

[From Rick Marken (2016.02.27.1600)]

···

MK (2016.02.26.1800 CET)

Rick Marken (2016.02.23.0850)–

RM: Henry Yin asked me to review it for him but I’m having difficulty

understanding it. In particular, I don’t understand the evidence they

present for prediction being involved in this behavior.

MK: Rick, read the paper while occasionally substituting “predictive” with

“purposive” and “cancellation” with “control” and you should be able

to make more sense of the paper. The “reactive vs predictive” in this

paper roughly corresponds to a PCT influenced researcher’s "S-R vs

purposive."

RM: It seems like a bit of a stretch. They don’t describe any kind of model of the dragonfly behavior. I see nothing that looks like proposed controlled variables. It’s just observational data, some of which they say is inconsistent with a “reactive” approach to predation (cancelling changes in the range vector with body/head movement) and some of which they say is consistent with an “internal” predictive model (which predicts limb and target dynamics). They never describe that model or show how it could work for intercepting prey. It sure doesn’t sound like they understand that the dragonfly is controlling perceptual variables.

MK: Like Sawtell’s group at Janelia, the people at Antonio’s lab seem to

be thinking of what a PCTer would refer to as “perceptual control” as

“(sensory) cancellation.” Their “predictions” correspond to PCTers

“purposes” or “references”.

RM: That may be true. But I’d sure like to see how that works. In PCT references are set by higher level systems as the means of controlling their perception. They are not set by a predictive model.

RM: But we could test to see whether they are controlling for something like PCT. Just explain that an internal predictive model is not needed for successful interception of an object that is dodging the the pursuer and see how they take it. My guess is that they would come up with all kinds of reasons why that couldn’t be the case, even if we showed them that such a “non-predictive” model can intercept a dodging target.

RM: Since you seem to know these folks, Matti, how about you conducting the test.

Best

Rick

Both groups have produced papers that

contain graphs that look like run-of-the-mill tracking runs. This

paper contains a labeled CV. I’d be surprised if the “negative copy”

in this paper doesn’t correspond to the “negative images” of the

papers from NS’s group. MM. et al/ (“Antonio, 2015”) have identified

the goal-tied nature of the control activity, but not related it

clearly enough to their “predictions”, which of course are the

goals. Both groups have provided at least some indirect support

against the mostly strawmannish “feedback is too slow” objection,

although I think that point was made a bit more explicitly in one of

the papers from NS’s group. I know that NS’s former grad student, who

did some of the modeling/control work, had not read Powers and it’s

clear that Michiani hasn’t done so, either. It would not surprise me

if what people speak about as a “inverse model” is a normally

functioning control system. I’m not familiar enough with their work to

tell how much of the verbiage is there for the appeasement of

reviewers, but the “internal models” in this paper seems to be spoken

about as "something that isn’t a fictionally stereotypical I-O

system."

M

–

Richard S. Marken

Author, with Timothy A. Carey, of Controlling People: The Paradoxical Nature of Being Human.

8

[From MK (2016.02.28.2315 CET)]

It might be fruitful to treat the parts of the paper separately and to
review the segment from pg. 335 and onwards independently from what
preceded it.

M

9

[From Rick Marken (2003.09.06.1100)]

Either our Commander in Thief is continuing his lying ways or he really
believes in predictive control, the very thing we were discussing last
month when we talked about the role of planning in control. In a recent
speech, Bush argued that his tax cuts for the rich have been successful
because they avoided a far worse recession than we would have had
without them. So the job loss and huge deficits we are seeing would
have been even worse if there had been no tax cut. This is like saying
that you would have hit even more pedestrians if you hadn't accelerated
through the intersection where many pedestrians were _bound_ to be
crossing soon.

Why control your perceptions when its so much easier to control your
imaginations? And the economic data suggests that Republican
administrations have been controlling their imaginations since Reagan.
If you look at the federal deficit/surplus data since 1945, what you
see is a variable that is being held rather nicely under control until
about 1980. The deficit (as a proportion of GDP) is virtually constant
(and about 0) until 1980. Then the deficit increases constantly until
1992. Clinton cleans up the mess and actually overshoots a bit,
creating a small surplus. Then in comes Bush and the deficit instantly
explodes. The data show that this results from a decline in revenue and
an _increase_ in spending.

  In less than three years, the current administration, using it's
predictive approach to control, has taken us from surpluses to deficits
even larger than those that existed at the end of Bush I. It looks
pretty clear from the data that all presidents since 1945 -- except
Reagan and Bush I and II -- have been in control of their present time
perceptions of the the deficit and brought it back to zero. Reagan and
Bushes I and II were _not_ controlling their present time perception of
the deficit at all. It looks like they were simply controlling
perceptions that had to do with their ideologies: perceptions of lower
taxes and higher defense spending. And now Bush II has justified this
irresponsibility in terms of predictive control. Since America, god
bless it, is still a democracy, it is apparently getting exactly what
it wants.

Best regards

Rick

10

[From Steve Dennis (2003.09.06.1330)]

Predicting the consequences of one’s own actions has value if the perception of those consequences is relevant to a reference standard that has priority. I think you are correct that ideological standards are more important to this bunch than deficit reduction. I suspect their predictions of the outcome of their policy were just as accurate as everyone else’s. That vision of the future simply had no relevance to their control agenda. For them, the important prediction is the movement of voters in a year or so…

SD

···

----- Original Message -----

From:
Rick Marken

To: CSGNET@listserv.uiuc.edu

Sent: Saturday, September 06, 2003 12:37 PM

Subject: Predictive Control

[From Rick Marken (2003.09.06.1100)]

Either our Commander in Thief is continuing his lying ways or he really
believes in predictive control, the very thing we were discussing last
month when we talked about the role of planning in control. In a recent
speech, Bush argued that his tax cuts for the rich have been successful
because they avoided a far worse recession than we would have had
without them. So the job loss and huge deficits we are seeing would
have been even worse if there had been no tax cut. This is like saying
that you would have hit even more pedestrians if you hadn’t accelerated
through the intersection where many pedestrians were bound to be
crossing soon.

Why control your perceptions when its so much easier to control your
imaginations? And the economic data suggests that Republican
administrations have been controlling their imaginations since Reagan.
If you look at the federal deficit/surplus data since 1945, what you
see is a variable that is being held rather nicely under control until
about 1980. The deficit (as a proportion of GDP) is virtually constant
(and about 0) until 1980. Then the deficit increases constantly until
1992. Clinton cleans up the mess and actually overshoots a bit,
creating a small surplus. Then in comes Bush and the deficit instantly
explodes. The data show that this results from a decline in revenue and
an increase in spending.

In  less than three years, the current administration, using it's

predictive approach to control, has taken us from surpluses to deficits
even larger than those that existed at the end of Bush I. It looks
pretty clear from the data that all presidents since 1945 – except
Reagan and Bush I and II – have been in control of their present time
perceptions of the the deficit and brought it back to zero. Reagan and
Bushes I and II were not controlling their present time perception of
the deficit at all. It looks like they were simply controlling
perceptions that had to do with their ideologies: perceptions of lower
taxes and higher defense spending. And now Bush II has justified this
irresponsibility in terms of predictive control. Since America, god
bless it, is still a democracy, it is apparently getting exactly what
it wants.

Best regards

Rick

11

[From Rick Marken (2003.09.06.2330)]

Steve Dennis (2003.09.06.1330)--

Predicting the consequences of one's own actions has value if the perception of those consequences is relevant to a reference standard that has priority.

I don't know if I understand this. If a perception is under control then it is relevant to a reference standard because the reference standard defines the state to which the perception is brought and maintained. So I think what you are saying is that predicting the (perceived) consequences of one's own actions has value if the consequences are controlled. I think we can show, using modeling, that this is not true. Prediction of the consequences of action on a controlled variable does not improve control of that variable. It seems like it should but it really doesn't.

I think you are correct that ideological standards are more important to this bunch than deficit reduction. I suspect their predictions of the outcome of their policy were just as accurate as everyone else's. That vision of the future simply had no relevance to their control agenda. For them, the important prediction is the movement of voters in a year or so...

I think I agree with this, though I can't believe that a Republican administration is really not concerned about the deficit. I take Bush at his word. I think he (or his handlers) is actually doing predictive control of the deficit. He is predicting what should happen based on his actions (tax reductions) and when what should have happened doesn't happen he continues to believe in the predicted result (increased revenue) and to assume that that result did happen, but that it's not perceived because another prediction, made post hoc (that the economy was getting bad faster than imagined) results in what we actually perceive. The result, as is clear from the data, is a huge debt to be paid by off by our children.

Best regards

Rick

12

[From Steve Dennis (2003.09.06.2345)]

I meant simply that in many situations we seem to have the power to decide which perceptions we will control and which we won’t. The projected consequences of those choices help us decide. If we control perceptions around reducing tax burdens and increasing budgets for defense projects, then we seemingly cannot also successfully control perceptions around maintaining or reducing the deficit. The mechanics of the external economic world render the respective reference specifications contradictory. So we can focus on one set of perceptions or the other, but not both. The way we choose has something to do with how we evaluate the matrix of consequences of controlling or not controlling each of the relevant variables. If I choose tax cuts and big defense budgets, I anticipate getting certain political benefits with my constituents, but also certain fiscal penalties. If I can live with that vision of the future, however flawed it might turn out to be, I will act to bring tax rates down and defense budgets up, leaving the deficit to drift out of control. At some future point, the actual consequences, as opposed to my predictions, will emerge and I may have to deal with some serious disturbances, or not, depending on how good my predictions were. As in evolution, survival is the only relevant issue in politics, not efficiency.

In complex human arenas like politics and economics, it seems like our supposed control systems are constantly undergoing major reorganization. As we have discussed previously, prediction of possible consequences may have its place more in guiding reorganization than in modulating actual control.

SD

···

----- Original Message -----

From:
Rick Marken

To: CSGNET@listserv.uiuc.edu

Sent: Saturday, September 06, 2003 11:27 PM

Subject: Re: Predictive Control

[From Rick Marken (2003.09.06.2330)]

Steve Dennis (2003.09.06.1330)--

Predicting the consequences of one's own actions has value if the perception of those consequences is relevant to a reference standard that has priority.

/smaller>

I don’t know if I understand this. If a perception is under control then it is relevant to a reference standard because the reference standard defines the state to which the perception is brought and maintained. So I think what you are saying is that predicting the (perceived) consequences of one’s own actions has value if the consequences are controlled. I think we can show, using modeling, that this is not true. Prediction of the consequences of action on a controlled variable does not improve control of that variable. It seems like it should but it really doesn’t.

I think you are correct that ideological standards are more important to this bunch than deficit reduction.  I suspect their predictions of the outcome of their policy were just as accurate as everyone else's.  That vision of the future simply had no relevance to their control agenda.  For them, the important prediction is the movement of voters in a year or so.../smaller>

I think I agree with this, though I can’t believe that a Republican administration is really not concerned about the deficit. I take Bush at his word. I think he (or his handlers) is actually doing predictive control of the deficit. He is predicting what should happen based on his actions (tax reductions) and when what should have happened doesn’t happen he continues to believe in the predicted result (increased revenue) and to assume that that result did happen, but that it’s not perceived because another prediction, made post hoc (that the economy was getting bad faster than imagined) results in what we actually perceive. The result, as is clear from the data, is a huge debt to be paid by off by our children.

Best regards

Rick

13

[From Rick Marken (2003.09.07.1200)]

Steve Dennis (2003.09.06.2345)--

I meant simply that in many situations we seem to have the power to decide which perceptions we will control and which we won't. The projected consequences of those choices help us decide.

Yes. I agree.

If we control perceptions around reducing tax burdens and increasing budgets for defense projects, then we seemingly cannot also successfully control perceptions around maintaining or reducing the deficit.

But I think Bush really believes that lowering tax burdens will _increase_ revenue. And there is some evidence that lowering tax burdens can increase revenue. So I think Bush (well, his handlers) predicted that the effect of lowering the tax burden would be to _increase_ revenue and, thus, maintain the surplus. The problem with predictive control comes when you don't revise actions based on the actual state of the controlled perceptual variable, as seems to be the case at the moment. It's what would happen in a tracking task if you predicted that the effect of a rightward movement of the handle (the action) would be to move the cursor rightward -- closer to the target. Usually rightward movement of the handle _does_ move the cursor toward the target. If, however, rightward action is not having the usual (predicted) effect you might be inclined to persist in moving the handle toward the right since this action _should_ have the predicted effect. Eventually, environmental circumstances (disturbances) may change so that rightward movement _does_ again have this effect, which would _confirm_ the prediction. But this ignores the fact that, during much of the tracking task the cursor was _not_ on the target. A person who is simply controlling the perceived position of the cursor and not trying to predict what effect their own actions _should_ be having would do _much_ better at controlling the cursor than the person using the predictive approach.

The mechanics of the external economic world render the respective reference specifications contradictory.

The mechanics of the external world have apparently made it impossible to _always_ control the deficit by decreasing the tax burden. But when you are controlling perceptions this effect of external constraints on controlled variables is handled automatically; the system simply changes its actions as necessary to reduce the discrepancy between perception (deficit/surplus) and reference (zero). Predictive control is always trumped by control of perception. If the predicted result of lowered taxes and increased spending is a lowered deficit but the actual (perceived) result is an increased deficit then the system controlling for a lowered deficit will simply change actions in a way that reduces the error that causes those actions. Prediction is unnecessary, at best, and an interference, at worst.

So we can focus on one set of perceptions or the other, but not both. The way we choose has something to do with how we evaluate the matrix of consequences of controlling or not controlling each of the relevant variables.

These choices are made automatically as the outputs of higher order control systems change the references for lower order control systems. See my "Spreadsheet analysis..." paper in _Mind Readings_ for an explanation of how this works in the PCT model.

If I choose tax cuts and big defense budgets, I anticipate getting certain political benefits with my constituents, but also certain fiscal penalties. If I can live with that vision of the future, however flawed it might turn out to be, I will act to bring tax rates down and defense budgets up, leaving the deficit to drift out of control.

This is a hypothesis about the perceptions Bush controls and why they are controlled (the higher order goals achieved by controlling them). The hypothesis is that Bush is controlling for political benefit (higher order goal) by controlling for lower taxes and higher defense spending. This hierarchical explanation of _why_ people control certain perceptions at certain levels does not involve anticipation and planning.

At some future point, the actual consequences, as opposed to my predictions, will emerge and I may have to deal with some serious disturbances, or not, depending on how good my predictions were. As in evolution, survival is the only relevant issue in politics, not efficiency.

The actual consequences (state of the controlled perceptual variable) are really all that matters. Predictions and plans go aglay about 50% of the time in a world where disturbances vary randomly. In life, _control_ is the only relevant issue. And the best control is achieved by a system that is properly organized (in terms of gain and slowing parameters) so that it always keeps the controlled perception matching the reference. Prediction will not help this process.

In complex human arenas like politics and economics, it seems like our supposed control systems are constantly undergoing major reorganization. As we have discussed previously, prediction of possible consequences may have its place more in guiding reorganization than in modulating actual control.

I agree with that last sentence. I can see how prediction (imagined consequences of action) could make reorganization (learning) much more efficient. One can probably learn to do a tracking task more quickly if, from experience, one can _predict_ that rightward movement of the mouse is likely to move the cursor rightward and leftward mouse movement is likely to move it leftward. This prediction (which is usually correct) will help by establishing immediately the correct polarity of the connection between error and output. This possible benefit of prediction might fit into your thesis regarding the superior control exhibited by humans. It's not so much that people control better than other species but that they can learn new ways of controlling (people reorganize) more effectively than other species. While reorganization is basically a random, trial and error process in other species, people can _use their imagination_ to guide reorganization and quickly develop new ways of controlling their perceptions.

Best regards

Rick

marken@mindreadings
Home 310 474-0313
Cell 310 729-1400

14

[From Steve Dennis (2003.09.07.1625)]

I suppose there are still supply-side economists who believe in the less-is-more theory, despite several administration’s worth of contrary data. Personally, I think it is simply a fig leaf for tax reduction and selective spending increases. I am not sure what the Bushies really believe. Their veracity on such matters tends to fall short of my standards.

I also think complex matters like the national economy and political machinations are far from stable control processes. The output functions remain crude and often only marginally tested by experience. What controls the target perceptions in one economic cycle sometimes fails in the next. While politicians and economists would like us to believe they are under control, and they certainly take credit when things work, the output functions sometimes seem to have only passing correlation with changes in the controlled variables. When control fails, the responsible parties are forced either to invent new actions, or behave randomly, or simply make stuff up. In systems which have not yet reached stability as control organizations, prediction may represent the only rational choice on what to do next. If the actors (no pun intended) can assemble enough data, prediction may work better than random action, but it is certainly no substitute for stable output functions in a fully organized control system.

SD

···

----- Original Message -----

From:
Rick Marken

To: CSGNET@listserv.uiuc.edu

Sent: Sunday, September 07, 2003 12:02 PM

Subject: Re: Predictive Control

[From Rick Marken (2003.09.07.1200)]

Steve Dennis (2003.09.06.2345)--/smaller>

I meant simply that in many situations we seem to have the power to decide which perceptions we will control and which we won't.  The projected consequences of those choices help us decide.

/smaller>

Yes. I agree.

If we control perceptions around reducing tax burdens and increasing budgets for defense projects, then we seemingly cannot also successfully control perceptions around maintaining or reducing the deficit.

/smaller>

But I think Bush really believes that lowering tax burdens will increase revenue. And there is some evidence that lowering tax burdens can increase revenue. So I think Bush (well, his handlers) predicted that the effect of lowering the tax burden would be to increase revenue and, thus, maintain the surplus. The problem with predictive control comes when you don’t revise actions based on the actual state of the controlled perceptual variable, as seems to be the case at the moment. It’s what would happen in a tracking task if you predicted that the effect of a rightward movement of the handle (the action) would be to move the cursor rightward – closer to the target. Usually rightward movement of the handle does move the cursor toward the target. If, however, rightward action is not having the usual (predicted) effect you might be inclined to persist in moving the handle toward the right since this action should have the predicted effect. Eventually, environmental circumstances (disturbances) may change so that rightward movement does again have this effect, which would confirm the prediction. But this ignores the fact that, during much of the tracking task the cursor was not on the target. A person who is simply controlling the perceived position of the cursor and not trying to predict what effect their own actions should be having would do much better at controlling the cursor than the person using the predictive approach.

The mechanics of the external economic world render the respective reference specifications contradictory.

/smaller>

The mechanics of the external world have apparently made it impossible to always control the deficit by decreasing the tax burden. But when you are controlling perceptions this effect of external constraints on controlled variables is handled automatically; the system simply changes its actions as necessary to reduce the discrepancy between perception (deficit/surplus) and reference (zero). Predictive control is always trumped by control of perception. If the predicted result of lowered taxes and increased spending is a lowered deficit but the actual (perceived) result is an increased deficit then the system controlling for a lowered deficit will simply change actions in a way that reduces the error that causes those actions. Prediction is unnecessary, at best, and an interference, at worst.

So we can focus on one set of perceptions or the other, but not both.  The way we choose has something to do with how we evaluate the matrix of consequences of controlling or not controlling each of the relevant variables.

/smaller>

These choices are made automatically as the outputs of higher order control systems change the references for lower order control systems. See my “Spreadsheet analysis…” paper in Mind Readings for an explanation of how this works in the PCT model.

If I choose tax cuts and big defense budgets, I anticipate getting certain political benefits with my constituents, but also certain fiscal penalties.  If I can live with that vision of the future, however flawed it might turn out to be, I will act to bring tax rates down and defense budgets up, leaving the deficit to drift out of control.

/smaller>

This is a hypothesis about the perceptions Bush controls and why they are controlled (the higher order goals achieved by controlling them). The hypothesis is that Bush is controlling for political benefit (higher order goal) by controlling for lower taxes and higher defense spending. This hierarchical explanation of why people control certain perceptions at certain levels does not involve anticipation and planning.

At some future point, the actual consequences, as opposed to my predictions, will emerge and I may have to deal with some serious disturbances, or not, depending on how good my predictions were.  As in evolution, survival is the only relevant issue in politics, not efficiency./smaller>

The actual consequences (state of the controlled perceptual variable) are really all that matters. Predictions and plans go aglay about 50% of the time in a world where disturbances vary randomly. In life, control is the only relevant issue. And the best control is achieved by a system that is properly organized (in terms of gain and slowing parameters) so that it always keeps the controlled perception matching the reference. Prediction will not help this process.

In complex human arenas like politics and economics, it seems like our supposed control systems are constantly undergoing major reorganization.  As we have discussed previously, prediction of possible consequences may have its place more in guiding reorganization than in modulating actual control./smaller>

I agree with that last sentence. I can see how prediction (imagined consequences of action) could make reorganization (learning) much more efficient. One can probably learn to do a tracking task more quickly if, from experience, one can predict that rightward movement of the mouse is likely to move the cursor rightward and leftward mouse movement is likely to move it leftward. This prediction (which is usually correct) will help by establishing immediately the correct polarity of the connection between error and output. This possible benefit of prediction might fit into your thesis regarding the superior control exhibited by humans. It’s not so much that people control better than other species but that they can learn new ways of controlling (people reorganize) more effectively than other species. While reorganization is basically a random, trial and error process in other species, people can use their imagination to guide reorganization and quickly develop new ways of controlling their perceptions.

Best regards

Rick

marken@mindreadings
Home 310 474-0313
Cell 310 729-1400

15

[From Rick Marken (2003.09.08.0850)]

Steve Dennis (2003.09.07.1625)I suppose
there are still supply-side economists who believe in the less-is-more
theory, despite several administration’s worth of contrary data.
Personally, I think it is simply a fig leaf for tax reduction and selective
spending increases. I am not sure what the Bushies really believe.
Their veracity on such matters tends to fall short of my standards.

Agree.

I also think complex matters like the national economy
and

political machinations are far from stable control
processes.

The historical data on the deficit suggests otherwise.
from 1947- 1980 the deficit/surplus was kept very near zero; the variations
were like those you see in a typical tracking task. So it looks to me that
people (the president and congress) were able to control the deficit extremely
well in that period. Then in 1981 the deficit (as % of GDP) increases substantially,
and continues to increase until 1993 at which point it starts decreasing
each year until there is an overshoot (surplus) starting in about 1998.
So in 6 years Clinton turned the largest deficit ever into the largest
surplus ever. In 2 years Bush II turned the largest surplus ever into the
largest deficit ever. It looks like the deficit stopped being a controlled
variable in 1981 and started being a controlled variable again in 1992
and then stopped being a controlled variable again in 2001.

The output functions remain crude and often only marginally

tested by experience.

It’s the connection of output to controlled variable
(the feedback function) that is not well understood for many controlled
economic variables. But in the case of the deficit I think the effect
of the main output functions are very well understood. Deficit occurs
when spending exceeds revenue collection. Even a Republican can probably
understand that you will have a deficit is you spend more than you take
in.

In systems which have not yet reached stability as control
organizations,

prediction may represent the only rational choice on
what to do next. If

the actors (no pun intended) can assemble enough data,
prediction may

work better than random action, but it is certainly
no substitute for

stable output functions in a fully organized control
system. SD

When the relationship between actions and output is really
unknown or random then prediction is (if possible) even less useful than
it would be if that relationship were well understood. That’s because
the only way to know whether an action is helping ort hindering is to
monitor the actual state of the controlled variable. If the variable is
moving toward the goal (reference) state then the action is having the
desired effect and it is continued; when the variable is moving away from
the goal then the action is no longer having the desired effect effect
and it should be change. But to what? Since the relationship between action
and consequence is not well understood, selection of new actions has to
be random (or, perhaps, biased by imagined predicted consequences). Once
the new action is selected, the state of the controlled consequence is
monitored; if that consequence does not start moving toward the goal, then
the action must be abandoned (regardless of what its predicted effect was)
and a new action selected. This approach to control is called the E. coli
strategy because it seem to be the way the E. coli bacteria moves toward
a source of nutrient. It is a biased random walk strategy that is
presumed to be the way reorganization works. It turns out to be a
very effective way to control when the effect of one’s outputs
on the controlled variable is unknown (as in the case of control of economic
variables). It is successful only when the system is willing to try
a new action when it is clear that the controlled variable is no longer
moving toward the goal.

Best regards

Rick

···

–

Richard S. Marken, Ph.D.

Senior Behavioral Scientist

The RAND Corporation

PO Box 2138

1700 Main Street

Santa Monica, CA 90407-2138

Tel: 310-393-0411 x7971

Fax: 310-451-7018

E-mail: rmarken@rand.org

16

[From Steve Dennis (2003.09.08.0940)]

I think you framed it well: “Since the relationship between action and consequence is not well understood, selection of new actions has to be random (or, perhaps, biased by imagined predicted consequences). Once the new action is selected, the state of the controlled consequence is monitored; if that consequence does not start moving toward the goal, then the action must be abandoned (regardless of what its predicted effect was) and a new action selected.”

The outcome (perception) should always dictate the continuation or abandonment of action (although true ideologues might disagree even here). Continuation is usually easy: things are working, let’s not change. It’s when control fails that life gets tough. Sometimes the control linkages are poorly defined if at all. Random action (E.coli) will eventually work, but it is probably not politically astute as a strategy for managing the national economy. Such politicians would have a shorter life-span than the bacteria.

No question you can control the deficit fairly precisely. The control levers are pretty obvious: revenue and expense. Where it gets tough is the effect on other perceptions. During the period you mention (late 40’s to early 80’s), I suspect people in the upper income brackets would have considered tax rates to be “out of control,” meaning that deficit control applied by tax-and-spend Democrats fell disproportionately on the wealthy, at least in their view. As the American population ascended the income strata during this period, they created a powerful new political focus on tax control as the target perception. Logically, you could have gotten both tax control and deficit control by using the expense lever properly: reducing government spending. I think the Reaganites tried to do this, at least in the early going, but that too had a political price. Recession, the Evil Empire, and a pork-inclined congress eventually overwhelmed spending discipline.

So there are really two kinds of control failures. First, when you simply do not have well-defined output functions in control systems, and the only alternative is random behavior or some sort of predictive strategy. Second, when you have multiple target perceptions, each with perhaps very well defined output functions, but which are incompatible within the collective “ownership” of the system. Randomness and prediction do little good here, unless they lead to an alternative output that resolves the conflict. As you point out, the prediction of such a resolution seems to be dominating the current domestic political rhetoric.

Best,

SD

···

----- Original Message -----

From:
Richard Marken

To: CSGNET@listserv.uiuc.edu

Sent: Monday, September 08, 2003 5:49 AM

Subject: Re: Predictive Control

[From Rick Marken (2003.09.08.0850)]

Steve Dennis (2003.09.07.1625)I suppose there are still supply-side economists who believe in the less-is-more theory, despite several administration's worth of contrary data.  Personally, I think it is simply a fig leaf for tax reduction and selective spending increases.  I am not sure what the Bushies really believe.  Their veracity on such matters tends to fall short of my standards.

Agree.

I also think complex matters like the national economy and
political machinations are far from stable control processes.

The historical data on the deficit suggests otherwise. From 1947- 1980 the deficit/surplus was kept very near zero; the variations were like those you see in a typical tracking task. So it looks to me that people (the president and congress) were able to control the deficit extremely well in that period. Then in 1981 the deficit (as % of GDP) increases substantially, and continues to increase until 1993 at which point it starts decreasing each year until there is an overshoot (surplus) starting in about 1998. So in 6 years Clinton turned the largest deficit ever into the largest surplus ever. In 2 years Bush II turned the largest surplus ever into the largest deficit ever. It looks like the deficit stopped being a controlled variable in 1981 and started being a controlled variable again in 1992 and then stopped being a controlled variable again in 2001.

The output functions remain crude and often only marginally
tested by experience.

It’s the connection of output to controlled variable (the feedback function) that is not well understood for many controlled economic variables. But in the case of the deficit I think the effect of the main output functions are very well understood. Deficit occurs when spending exceeds revenue collection. Even a Republican can probably understand that you will have a deficit is you spend more than you take in.

In systems which have not yet reached stability as control organizations,
prediction may represent the only rational choice on what to do next. If
the actors (no pun intended) can assemble enough data, prediction may
work better than random action, but it is certainly no substitute for
stable output functions in a fully organized control system. SD

When the relationship between actions and output is really unknown or random then prediction is (if possible) even less useful than it would be if that relationship were well understood. That’s because the only way to know whether an action is helping ort hindering is to monitor the actual state of the controlled variable. If the variable is moving toward the goal (reference) state then the action is having the desired effect and it is continued; when the variable is moving away from the goal then the action is no longer having the desired effect effect and it should be change. But to what? Since the relationship between action and consequence is not well understood, selection of new actions has to be random (or, perhaps, biased by imagined predicted consequences). Once the new action is selected, the state of the controlled consequence is monitored; if that consequence does not start moving toward the goal, then the action must be abandoned (regardless of what its predicted effect was) and a new action selected. This approach to control is called the E. coli strategy because it seem to be the way the E. coli bacteria moves toward a source of nutrient. It is a biased random walk strategy that is presumed to be the way reorganization works. It turns out to be a very effective way to control when the effect of one’s outputs on the controlled variable is unknown (as in the case of control of economic variables). It is successful only when the system is willing to try a new action when it is clear that the controlled variable is no longer moving toward the goal.

Best regards

Rick

Richard S. Marken, Ph.D.
Senior Behavioral Scientist
The RAND Corporation
PO Box 2138
1700 Main Street
Santa Monica, CA 90407-2138
Tel: 310-393-0411 x7971
Fax: 310-451-7018
E-mail: rmarken@rand.org

17

[From Rick Marken (2003.09.08.1950)]

Steve Dennis (2003.09.08.0940)--

The outcome (perception) should always dictate the continuation or abandonment of action (although true ideologues might disagree even here).

The discrepancy between reference and perception (which is itself dependent on action) must dictate the continuation or abandonment of action in a control loop or there is no control.

Continuation is usually easy: things are working, let's not change. It's when control fails that life gets tough. Sometimes the control linkages are poorly defined if at all. Random action (E.coli) will eventually work, but it is probably not politically astute as a strategy for managing the national economy. Such politicians would have a shorter life-span than the bacteria.

It's very efficient with the bacteria (see my "Selection of consequences" demo at http://www.mindreadings.com/ControlDemo/Select.html to see it in action) and I think it could works fine in politics. Once a policy (like the tax cut) doesn't work try something new (like making the income tax more progressive again or doing nothing for a while).

So there are really two kinds of control failures. First, when you simply do not have well-defined output functions in control systems, and the only alternative is random behavior or some sort of predictive strategy. Second, when you have multiple target perceptions, each with perhaps very well defined output functions, but which are incompatible within the collective "ownership" of the system.

Yes! In the first case you simply don't have a control system organized to control anything. I the second you have two perfectly functional control systems that are unable to control because they are in conflict (a demonstration of loss of control through conflict is available at http://www.mindreadings.com/ControlDemo/Conflict.html).

Best

Rick
marken@mindreadings
Home 310 474-0313
Cell 310 729-1400