Rocky Mountain High

CSGnet Archive CSG1998
1

[From Bruce Gregory (981130.1242 EDT)]

If I am bicycling westward and controlling for a perception of the
Pacific Ocean. The Rocky Mountains constitute a major disturbance to my
activities. How can I counter this disturbance? I am reminded by Rick
that "The only outputs that will work are those that precisely oppose
the net effect of disturbances to the controlled variable (controlled
outcome). Remember, o = -1/g(d)." Exactly how do I precisely oppose the
net effect of the Rock Mountains? It seems to me that I can take account
of the existence of the Rockies and plan my trip to minimize the
pedaling I must do, but this seems perilously close to model-based
control.

Bruce Gregory

2

[From Bill Powers (981130.1117 MST)]

Bruce Gregory (981130.1242 EDT)--

If I am bicycling westward and controlling for a perception of the
Pacific Ocean. The Rocky Mountains constitute a major disturbance to my
activities. How can I counter this disturbance?

Before your question can be answered, you must specify exactly how the
presence of the Rocky Mountains is affecting the present state of the
controlled variable, and by what physical means. You have defined the
controlled variable as "a perception of the Pacific ocean." Before you have
crossed the Rocky Mountains, the state of this perception is zero, and it
will remains zero until you're within a few miles of the ocean, at which
point you will be long past the Rocky Mountains, and any potential
influence they might have had on this perception will be gone. So under
almost any interpretation of your definition of the controlled variable, I
would guess that the disturbing effect of the Rocky Mountains on your
perception of the Pacific Ocean is zero.

A disturbance is not just "a major disturbance to your activities." It is,
specifically, some variable that contributes, right now, to the current
state of the controlled variable, independently of the present influence of
your actions on the same variable. I don't believe you have actually
defined a disturbance to the present-time perception of the Pacific Ocean.

Best,

Bill P.

I am reminded by Rick

···

that "The only outputs that will work are those that precisely oppose
the net effect of disturbances to the controlled variable (controlled
outcome). Remember, o = -1/g(d)." Exactly how do I precisely oppose the
net effect of the Rock Mountains? It seems to me that I can take account
of the existence of the Rockies and plan my trip to minimize the
pedaling I must do, but this seems perilously close to model-based
control.

Bruce Gregory

3

[From Bill Powers (981130.1128 MST)]

Bruce Gregory (981130.1242 EDT)--

It seems to me that I can take account
of the existence of the Rockies and plan my trip to minimize the
pedaling I must do, but this seems perilously close to model-based
control.

Are we avoiding talking about model-based control? If you can show that
some behavior is best explained as model-based control, what's to keep us
from explaining it that way?

Best,

Bill P.

4

[From Bruce Gregory (981130.1452 EDT)]

Bill Powers (981130.1117 MST)]

A disturbance is not just "a major disturbance to your
activities." It is,
specifically, some variable that contributes, right now, to
the current
state of the controlled variable, independently of the
present influence of
your actions on the same variable. I don't believe you have actually
defined a disturbance to the present-time perception of the
Pacific Ocean.

Good. What I should have said is that the Rocky Mountains are a major
disturbance to the perception that I am making progress toward my goal
of seeing the Pacific Ocean. I can counter that disturbance by picking a
route through the Rockies that will allow me to continue to the Pacific.
O.K.?

Bruce Gregory

5

[From Bruce Gregory (981130.1455 EDT)]

Bill Powers (981130.1128 MST)

Bruce Gregory (981130.1242 EDT)--

>It seems to me that I can take account
>of the existence of the Rockies and plan my trip to minimize the
>pedaling I must do, but this seems perilously close to model-based
>control.

Are we avoiding talking about model-based control? If you can
show that
some behavior is best explained as model-based control,
what's to keep us
from explaining it that way?

Fair enough. My original point was simply that it doesn't matter _how_
control is exercised as long as the system is able to compare a
representation of the world with a representation of how it wants the
world to be and react accordingly. Rick seems persuaded that control can
only be exercised in one way. This seems highly unlikely to me, but is
secondary to the point I make.

Bruce Gregory

6

[From Rick Marken (981130.1400)]

Bruce Gregory (981130.1455 EDT) --

My original point was simply that it doesn't matter _how_ control
is exercised as long as the system is able to compare a
representation of the world with a representation of how it wants
the world to be and react accordingly.

Control is not a necessary result of comparing a representation
of the world (p) with a representation of how the system wants
the world to be (r). This comparison must be turned into output
variations (o) that have the appropriate effect on the controlled
variable: the appropriate effect is the one that satisfies
o = -1/g(d).

Think about it in terms of a simple tracking task. If you (the
tracker) are set up so that a comparison of r (zero distance from
target to cursor, say) and p (actual distance from target to cursor)
leads to mouse movements in the wrong direction then there is no
control (despite the comparison). Similarly, if you are set up so
that a comparison of r and p leads to eye movements rather than
mouse movements then there will be no control.

Rick seems persuaded that control can only be exercised in one
way.

It depends on what you mean by "way". In a tracking task, for
example, there is only one "way" to control the cursor and that
is by having effects (o) on cursor position that are equal and
opposite to the net effects (d) of disturbances to this variable:
that is, the only way to control the cursor is by generating outputs
such that o = -1/g(d). But there may be several different "ways" to
exert these outputs. For example, there are many different hand
positions that can be used to hold the mouse. But there is still
only one way any of these hand positions can be used to _control_
the cursor; any hand position must end up generating effects on
the controlled variable that counter the effects of the disturbance;
that is, they must generate outputs such that o = -1/g(d). These
outputs are the _only_ right answer to the question "how do I
exercise control over cursor position".

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates e-mail: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

7

[From Bruce Gregory (981130.1825 EDT)]

Rick Marken (981130.1400)

Bruce Gregory (981130.1455 EDT) --

> My original point was simply that it doesn't matter _how_ control
> is exercised as long as the system is able to compare a
> representation of the world with a representation of how it wants
> the world to be and react accordingly.

Control is not a necessary result of comparing a representation
of the world (p) with a representation of how the system wants
the world to be (r). This comparison must be turned into output
variations (o) that have the appropriate effect on the controlled
variable: the appropriate effect is the one that satisfies
o = -1/g(d).

This is what I meant by "react accordingly."

Think about it in terms of a simple tracking task. If you (the
tracker) are set up so that a comparison of r (zero distance from
target to cursor, say) and p (actual distance from target to cursor)
leads to mouse movements in the wrong direction then there is no
control (despite the comparison). Similarly, if you are set up so
that a comparison of r and p leads to eye movements rather than
mouse movements then there will be no control.

True, you will be unable to "react accordingly."

> Rick seems persuaded that control can only be exercised in one
> way.

It depends on what you mean by "way". In a tracking task, for
example, there is only one "way" to control the cursor and that
is by having effects (o) on cursor position that are equal and
opposite to the net effects (d) of disturbances to this variable:
that is, the only way to control the cursor is by generating outputs
such that o = -1/g(d). But there may be several different "ways" to
exert these outputs. For example, there are many different hand
positions that can be used to hold the mouse. But there is still
only one way any of these hand positions can be used to _control_
the cursor; any hand position must end up generating effects on
the controlled variable that counter the effects of the disturbance;
that is, they must generate outputs such that o = -1/g(d). These
outputs are the _only_ right answer to the question "how do I
exercise control over cursor position".

Let's look at the Rocky Mountain case. There are many routes through the
mountains. Each will counter the effects of the mountains in disturbing
my perception that I am on course to see the Pacific. No path is the
"right" path. Or am I wrong?

Bruce Gregory

8

[From Bill Powers (981201.0331 MST)]

Bruce Gregory (981130.1452 EDT)--

I don't believe you have actually
defined a disturbance to the present-time perception of the
Pacific Ocean.

Good. What I should have said is that the Rocky Mountains are a major
disturbance to the perception that I am making progress toward my goal
of seeing the Pacific Ocean. I can counter that disturbance by picking a
route through the Rockies that will allow me to continue to the Pacific.
O.K.?

Still too indefinite. _Something about_ the Rocky Mountains is a
disturbance to the perception of progress toward the Pacific Ocean -- that
is, a disturbance of your rate of movement, which is the immediate goal.
Their mere existence is not a disturbance. Their beauty is not a
disturbance (to progress). And if the controlled variable is qualitative
(progress vs no progress) there is no disturbance until the effect of the
terrain is to prevent progress altogether.

We do not control _things_. We control _variables_. Disturbances do not
disturb just by existing. They disturb by having quantitative effects on a
variable that is being controlled. Logical variables
(disturbance/no-disturbance) are either undisturbed or totally disturbed,
since a logical controlled variable can be only true or false.

We could say that the slope of the terrain on which we're bicycling is a
disturbance of the speed with which we can travel. If the slope is so great
that we can't maintain the desired speed toward the Pacific Ocean, then
some higher-level variable might be disturbed -- for example, a perception
of the number of days it will take to get to the Pacific coast (a
perception of (miles remaining) / (speed in miles per day)). At a higher
level, we might be able to compute the distance remaining if we go by
different routes (Berthoud Pass versus a course that takes us over the peak
of Mt. Evans), and the speed we can maintain on each route, depending on
its average slope, and from this derive perceptions of the estimated time
to the coast by each route. That obviously involves previous experience and
the use of maps and paper-and-pencil calculations. We would then choose the
route that gets us to the coast the soonest. Or the most scenic route that
gets us to the coast no later than a week from next Thursday, or whatever
other higher-level goals we try to satisfy at the same time.

Beware oversimplification.

Best,

Bill P.

9

[From Rick Marken (981201.0745)]

Me:

Control is not a necessary result of comparing a representation
of the world (p) with a representation of how the system wants
the world to be (r). This comparison must be turned into output
variations (o) that have the appropriate effect on the controlled
variable: the appropriate effect is the one that satisfies
o = -1/g(d).

Bruce Gregory (981130.1825 EDT) --

This is what I meant by "react accordingly."

OK, so "react accordingly" means "have the appropriate effect". My
point is that a comparison between r and p does not necessarily
lead to a person "reacting accordingly" (appropriately). A person
has to _learn_ how to turn the comparison of r and p into effects
in the world that keep p under control. The real world places
constraints on what the results of this learning must be if one
is to successfully "react accordingly" (appropriately).

If the comparison between r and p leads a person to react
"unaccordingly" (inappropriately) then p will not be controlled;
the control system has not _learned_ how to react accordingly
(it has not learned how produce outputs that satisfy o = -1/g(d);
it hasn't learned how to produce the _right_ outputs). The system
is comparing r and p but reality does not allow the person's
outputs to have the appropriate ("accordingly") effects; effects
that result in p staying approximately equal to r).

My point is that learning is constrained by the the nature of
the reality "behind" our perceptions (we have a pretty good idea
of the nature of this reality from the natural sciences). Learning
to compare r and p is just the first step in learning. The tough
part is learning how to convert that comparison into outputs (which
are often references for lower level perceptions) that keep p
under control.

For example, when I was younger I quickly learned one perception
I wanted to control when I road a bicycle: angle of the plane of
the frame with respect to the ground while moving forward. I also
quickly learned my preferred reference state for this variable: 90
degrees. What I didn't know was how to generate outputs (actually,
lower level perceptions) that would keep this bicycle angle perception
under control.

It took some time to learn how to control bicycle angle (p); and I
have no idea _what_ I learned -- how the outputs I generated had
effects that kept bicycle angle (p) at 90 degrees (r). But I know
that, during the learning process, even though I was comparing r
to p, I was unable to keep p equal to r until I made that mysterious
transformation (via learning) into a bicyclist; I learned how to
transform r-p into effects on the bike's angle with respect to
the ground that kept my perception of this variable at the reference.

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates e-mail: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

10

[From Bruce Gregory (981201.1100 EDT)]

Rick Marken (981201.0745)

OK, so "react accordingly" means "have the appropriate effect". My
point is that a comparison between r and p does not necessarily
lead to a person "reacting accordingly" (appropriately). A person
has to _learn_ how to turn the comparison of r and p into effects
in the world that keep p under control. The real world places
constraints on what the results of this learning must be if one
is to successfully "react accordingly" (appropriately).

I agree.

If the comparison between r and p leads a person to react
"unaccordingly" (inappropriately) then p will not be controlled;
the control system has not _learned_ how to react accordingly
(it has not learned how produce outputs that satisfy o = -1/g(d);
it hasn't learned how to produce the _right_ outputs). The system
is comparing r and p but reality does not allow the person's
outputs to have the appropriate ("accordingly") effects; effects
that result in p staying approximately equal to r).

Yes, this is my understanding, too.

My point is that learning is constrained by the nature of
the reality "behind" our perceptions (we have a pretty good idea
of the nature of this reality from the natural sciences). Learning
to compare r and p is just the first step in learning. The tough
part is learning how to convert that comparison into outputs (which
are often references for lower level perceptions) that keep p
under control.

Yes.

For example, when I was younger I quickly learned one perception
I wanted to control when I road a bicycle: angle of the plane of
the frame with respect to the ground while moving forward. I also
quickly learned my preferred reference state for this variable: 90
degrees. What I didn't know was how to generate outputs (actually,
lower level perceptions) that would keep this bicycle angle perception
under control.

Funny, I used exactly this example in just this way recently. Chilling
thought, isn't it? [Attempt at humor.] This is the point I was trying to
get at, knowing what "outputs to generate" (even if one could) is no
help unless one can tell that they are reducing the error.

It took some time to learn how to control bicycle angle (p); and I
have no idea _what_ I learned -- how the outputs I generated had
effects that kept bicycle angle (p) at 90 degrees (r). But I know
that, during the learning process, even though I was comparing r
to p, I was unable to keep p equal to r until I made that mysterious
transformation (via learning) into a bicyclist; I learned how to
transform r-p into effects on the bike's angle with respect to
the ground that kept my perception of this variable at the reference.

Again, I agree completely. A bicycle is a good example because we are
quite clear that at some level we don't "know what we are doing." Thanks
for sticking with this. [No humor, sarcastic or otherwise, intended.]

Bruce Gregory

11

[From Bruce Gregory (981201.1105 EDT)]

Bill Powers (981201.0331 MST)

Beware oversimplification.

Thanks, I'll do my best.

Bruce Gregory

12

[From Bruce Gregory (981201.1700 EDT)]

Rick Marken (981201.1245)

Bruce Gregory (981201.1100 EDT)--

> Again, I agree completely. A bicycle is a good example because
> we are quite clear that at some level we don't "know what we are
> doing." Thanks for sticking with this. [No humor, sarcastic or
> otherwise, intended.]

At the risk of ending this beautiful friendship before it begins,
I would like to point out that your agreement with my comments
suggest that you would now disagree with your own comments on
this issue, specifically:

I knew it would never last. [Attempt at humor]

> Some patterns of output may be more efficient than others,
> but all will _eventually_ work if you let the direction of
> error change determine the next output. This process is
> sometimes called learning.

Ordinarily, people learn how to control by random trial and
error (if they learn at all). This is how I learned to ride
a bike; I kept varying my outputs until I found an output
function (f(r-p)) that generated outputs that satisfied
o = -1/g(d), thus keeping p = r under all circumstances.

This is the point I was trying to make. You can start out with a firm
idea of how to ride a bike and persist until you are clear that it
doesn't work. At which point you need to try something else.

But I think learning can be done more efficiently than trial
and error; I think that's the hope, anyway, that underlies
the fact that we have an educational system. Schools and
universities exist because people think we can do more to teach
control than just saying "figure it out for yourself".

It's not obvious to me that we learn _anything_ except by trial and
error (isn't it obvious that this is the way I'm learning PCT?) Gary
Cziko has argued this point and I happen to believe he is right.

I think what educators should be doing is 1) trying to figure
out what people want to control so they know what to teach
them and

I agree.

2) trying to figure out how to represent the output

function that allows control as instructions about to vary
lower level perceptions in order to achieve control of the
higher level variable.

More directly, figure out what the student needs to perceive rather than
what the student needs to do.

Telling a person who is learning to ride a bike that everything
they do "will _eventually_ work if [they] let the direction of
error change determine the next output" is really not very helpful;

I can't think of what to tell a person who is learning to ride a bike
that would be any help at all. What were you told that you found
helpful? Telling new pilots how to move the controls is very
ineffective. Telling them what they should perceive works very well (the
FAA calls this attitude--the plane's not the pilot's)control.

in fact, some things they do will _never_ work, even if those things
are leading to the right change in the direction of error. An
educator's message should not be "everything you do is right";

I don't know of who would say this, but I agree that they shouldn't.

the
educator's message should be encouragement ("stick to it; you will
get it eventually") or, even better, suggestions about lower
level perceptions to control that might help out ("you might try
turning into the fall").

Again advice on what to do is far less valuable than advice on what to
perceive.

Bruce Gregory