output & error

[from Tracy Harms (980219.09)]

I'd like to reiterate what Marc wrote:

i,kurtzer (980218.1830)

bruce g. replying to j.vancouver
>>Behavior is not the result of errors. Behavior is the control of
>>perception.

bruce a. replying to bruce g.

> Bruce, I take Jeff to be saying o = f(e) where o is the output of a control
>system, and e is the error signal. You seem to be denying this statement.

No, he is saying that output is not behavior. Producing output per se is not
behavior. Rather, behavior is the control of perception.

In the light of this, I was genuinely surprised to see this:

[From Bruce Gregory (980218.2211 EST)]

>Bruce, I take Jeff to be saying o = f(e) where o is the output of a control
>system, and e is the error signal. You seem to be denying this statement.

No, I am not denying that statement. You may notice that the statement
contains no reference to behavior. Behavior is not an element in the PCT
model. Strictly speaking, it belongs to another realm of discourse. This is
the point I was trying to make in my post about Newtonian physics.

Bruce

and this!

[From Rick Marken (980218.2050)]

Bruce Abbott (980218.2025 EST) --

> Jeff Vancouver is very clearly talking about the first meaning
> of "behavior," [ output which is a function of error] and you
> guys, for whatever reason, want to show him "wrong" by
> insisting on the validity only of the second definition.

While it is true that o = f(e) it is also (and simultaneously)
true that e = g(o).

It seems we all agree that output is not behavior. But how can it be
acceptable to see output as a function of error? Unless I'm grossly
misconstruing what is meant by "function" -- which I interpret in terms
of applied mathematics --- a control system can't be expected to produce
output which has any functional relationship with perceptual error. "o
= f(e)" suggests that as you change e across a range (e.g. e-sub-one
through e-sub-n) you get a particular matching value of o for each value
of e.

I do not see how this could be reasonable. "Function of" is a much
stronger claim than "is affected by".

Tracy Harms
Bend, Oregon

[from Tracy Harms (980219.10)]

[From Bill Powers (980219.0359 MST)]

...
There is no way to reconcile the closed-loop concept of behavior
with the open-loop, cause-effect, sequential concept. The two
concepts are as different as the flat earth is from the round earth.

Best,

Bill P.

This was impressed upon me a minute ago as I thought about how "o =
f(e)" might be tested with a control system implemented in computer
software. In order to vary "e" across a test range you would have to
override the system's input calculation of e, and instead provide a
"hardwired" e. But in doing this you break the system! At that point
you no longer have a control system in the test.

It must be a false claim, that the output of a control system is a
function of perceptual error.

Tracy Harms

[from Tracy Harms (980219)]

Excuse me for sending a bunch of little posts instead of something more
carefully composed. Also, I'm receiving the digest, so I don't see any
responses until the next morning.

Further reflection has let me see that you could test "o = f(e)" on a
devised control system without breaking the system. What you have to do
is put in (non-perturbing) sensors which measure o and e simultaneously
and log their values.

Significant discrepency in the value of o for multiple instances of the
same value e would refute o = f(e).

PCT leads us to anticipate exactly that sort of variation.

Tracy Harms
Bend, Oregon

[From Bruce Gregory (980219.1345 EST)]

Tracy Harms (980219)

Significant discrepency in the value of o for multiple instances of the
same value e would refute o = f(e).

PCT leads us to anticipate exactly that sort of variation.

I'm not sure why you believe this. Think of a thermostat
controlling a furnace. Do the outputs of the thermostat differ
for the same values of the error? There are three possible
outputs: on, off, do nothing.

Error signal present: is the switch on?
        Yes: do nothing.
        No: turn switch on

Error signal absent: is switch on?
        Yes: turn switch off
        No: do nothing

In no case does the action differ for the same error signal.

Bruce

i.kurtzer (980219.1500)

[from Tracy Harms (980219)]

>Further reflection has let me see that you could test "o = f(e)" on a
>devised control system without breaking the system. What you have to do
>is put in (non-perturbing) sensors which measure o and e simultaneously
>and log their values.

With the technology currently in place its seems questionable whether we can
directly measure "e"; instead, tracking tasks have been devised such that if
"o" is posited as the integral of "e", and this intergal is then parameterized
for an individual, then one can predict sliding reference values knowing "o" ,
"i" , and "d", along with appropriate filtering. The self-consistancy of the
posited relations (including o=(f)e )and there predictive power is very nice.
See Powers "Measurement of Volition" in Volitional Action: Conation and
Control, Ed. Wayne Hershberger, pp315-332.

i.
p.s. i will give a free copy to anyone that attends the conference
Significant discrepency in the value of o for multiple instances of the
same value e would refute o = f(e).

PCT leads us to anticipate exactly that sort of variation.

···

[From Bill Powers (980220.0513 MST)]

Tracy Harms (980219.09)--

While it is true that o = f(e) it is also (and simultaneously)
true that e = g(o).

It seems we all agree that output is not behavior. But how can it be
acceptable to see output as a function of error? Unless I'm grossly
misconstruing what is meant by "function" -- which I interpret in terms
of applied mathematics --- a control system can't be expected to produce
output which has any functional relationship with perceptual error. "o
= f(e)" suggests that as you change e across a range (e.g. e-sub-one
through e-sub-n) you get a particular matching value of o for each value
of e.

Not exactly -- that type of function is static in time. The actual output
function would in general be described by some sort of differential
equation, as it is in our models of tracking behavior (the ones that
explain 99.5% of the variance in observed behavior). But it is indeed a
single, fixed function and the output of the model is always completely
determined by the value of the error signal, in real time. If you haven't
understood this yet, you're overdue for an enjoyable insight.

I do not see how this could be reasonable. "Function of" is a much
stronger claim than "is affected by".

Right, and "function of" is exactly what is intended.

Best,

Bill P.

[From Bill Powers (980220.0521 MST)]

Tracy Harms (980219.10)--

This was impressed upon me a minute ago as I thought about how "o =
f(e)" might be tested with a control system implemented in computer
software.

That's exactly what we do in our simulations, and in matching the behavior
of a control model to that of a real person.

In order to vary "e" across a test range you would have to

override the system's input calculation of e, and instead provide a
"hardwired" e.

... But that's not how we do it. In the model, e = r - p, where r is the
reference condition and p is the current perception of the state of the
environment. The model is given the same effect on the perceived
environment that the real person has, and its perception represents the
same variable that the real person is perceiving -- the distance between
target and cursor.
The same disturbance acts while the model is running as when the person is
doing the controlling.

But in doing this you break the system! At that point
you no longer have a control system in the test.

Nope. The model runs intact.

It must be a false claim, that the output of a control system is a
function of perceptual error.

Sorry. That's an essential part of the control model. I think you've got
some definitions in your head that need to be pulled out and looked at.

Best,

Bill P.

[From Bill Powers (980220.0528 MST)]

Tracy Harms (980219)--

Further reflection has let me see that you could test "o = f(e)" on a
devised control system without breaking the system. What you have to do
is put in (non-perturbing) sensors which measure o and e simultaneously
and log their values.

Correct, that's what we do, in effect.

Significant discrepency in the value of o for multiple instances of the
same value e would refute o = f(e).

No it wouldn't (and doesn't). One form of output function that works very
well is

do/dt = k1*e - k2*o

This function can clearly yield different values of o with the same value
of e, and the same value of o for different values of e. It's a perfectly
regular and predictable function involving time.

Best,

Bill P.

[from Tracy Harms (19980220.08)]

First, a reply to Bruce Gregory (980219.1345 EST)]

I wrote:

> Significant discrepency in the value of o for multiple instances of the
> same value e would refute o = f(e).
>
> PCT leads us to anticipate exactly that sort of variation.

Bruce responded:

I'm not sure why you believe this. Think of a thermostat
controlling a furnace. Do the outputs of the thermostat differ
for the same values of the error? There are three possible
outputs: on, off, do nothing.

Error signal present: is the switch on?
        Yes: do nothing.
        No: turn switch on

Error signal absent: is switch on?
        Yes: turn switch off
        No: do nothing

In no case does the action differ for the same error signal.

The output "o" is not the output of the comparator, it is the output of
the *system*, i.e. everything which does not count as "environment".

To use and extend your example I direct your attention to the rather
famous book by Grady Booch, _Object Oriented Software and Design, with
Applications_. One of the application chapters is a heating system.
I've spent some time studying this chapter, in fact, because I want to
see how somebody who does not know PCT engineers such a thing, and
because I want to figure out how PCT can allow the task to be done
better. (Such efforts have convinced me that designing systems with the
help of PCT is still hard work.)

In that system there is a two-minute delay required between the turning
off of the heating elements and the turning off of the circulating fan.
Presumably this helps prevent damage to the furnace which overheating
could produce. But the incentives behind that specification are
irrelevant; the main point is that his simplified, idealized system
provides an example where output does not correlate with error signal.
The error signal can be absent while the system at large continues to
blow hot air out of its vents.

Next, Bruce Nevin (980219.1333)

Tracy Harms (980219.10)--

>It must be a false claim, that the output of a control system is a
>function of perceptual error.

You are right, it's false if taken in isolation. That is, it's incomplete.

The incompleteness is the more important point, I agree, but for now
I'll hold my ground and claim that adding the missing stuff does not
make this part true. It is just false; the "o = f(e)" equation is not a
viable part of the whole.

But Bill thinks I've misunderstood the nature of the equation, and
perhaps I have.

Bill Powers (980220.0521 MST):

>It must be a false claim, that the output of a control system is a
>function of perceptual error.

Sorry. That's an essential part of the control model. I think you've got
some definitions in your head that need to be pulled out and looked at.

Probably my notion of "function"...

Bill Powers (980220.0513 MST):

The actual output
function would in general be described by some sort of differential
equation, as it is in our models of tracking behavior (the ones that
explain 99.5% of the variance in observed behavior). But it is indeed a
single, fixed function and the output of the model is always completely
determined by the value of the error signal, in real time. If you haven't
understood this yet, you're overdue for an enjoyable insight.

Goody. Enjoyable insights are my main addiction (and they are much
harder to obtain than cigarettes...)

Along the same line I'll mention that PCT has given me my first sense of
incentive to *actually* learn calculus. (As opposed to how I dealt with
it in college.) Calculus comes up here quite regularly, and I'll admit
that I tend to get lost from the argumentation as it moves into the
math. (I'd pay for a calculus tutor who could use PCT-compliant
examples to take me through the scope of the subject. Or a
textbook/workbook.)

Bill Powers (980220.0528 MST):

Tracy Harms (980219)--
>Significant discrepency in the value of o for multiple instances of the
>same value e would refute o = f(e).

No it wouldn't (and doesn't). One form of output function that works very
well is

do/dt = k1*e - k2*o

This function can clearly yield different values of o with the same value
of e, and the same value of o for different values of e. It's a perfectly
regular and predictable function involving time.

But that equation can't be solved for o, can it? If it can't, you're
not able to talk about any o = f(e), are you?

Right about now I probably should put on my dunce cap and go sit in the
corner, but I have to ask these things if I'm going to get to the
enjoyable insights. The answer I anticipate receiving back is that when
people write "o = f(e)" they mean something like "do/dt = k1*e - k2*o",
and the fact that the former *looks* algebraic comes from that most
annoying habit of simplifying notation so that the calculus aspects are
implicit. (Such wholesale disposal of explicit detail seemed pervasive
to calculus and was a major factor in the difficulties I had with it.
Not as big a factor as neglecting to study, however.)

Finally, Marc Kurtzer (980219.1500)

With the technology currently in place its seems questionable whether we can
directly measure "e"

But I referred to a *devised* control system. When we build something
we have an opportunity to include measurement instruments at the
critical point inside the system.

Enough for now,

Tracy Harms
Bend, Oregon

[From Bill Powers (980220.1103 MST)]

Tracy Harms (19980220.08)--

To use and extend your example I direct your attention to the rather
famous book by Grady Booch, _Object Oriented Software and Design, with
Applications_. One of the application chapters is a heating system.
I've spent some time studying this chapter, in fact, because I want to
see how somebody who does not know PCT engineers such a thing, and
because I want to figure out how PCT can allow the task to be done
better. (Such efforts have convinced me that designing systems with the
help of PCT is still hard work.)

In that system there is a two-minute delay required between the turning
off of the heating elements and the turning off of the circulating fan.
Presumably this helps prevent damage to the furnace which overheating
could produce. But the incentives behind that specification are
irrelevant; the main point is that his simplified, idealized system
provides an example where output does not correlate with error signal.
The error signal can be absent while the system at large continues to
blow hot air out of its vents.

The output does not "correlate" with the error signal; it is _driven by_
the error signal, which is a physical signal like a voltage. When the error
signal appears (it's either on or off in this kind of system), the output
function produces a signal that turns the blower and the furnace on. When
the error signal disappears, the output function turns the furnace off, and
two minutes later turns the blower off. I presume that if the temperature
falls below the reference setting before two minutes is up, so the error
signal reappears, the output function turns on both furnace and blower.

This is a digital type of control system, and differential equations
wouldn't enter into the system design.

One form of output function that works very
well is

do/dt = k1*e - k2*o

This function can clearly yield different values of o with the same value
of e, and the same value of o for different values of e. It's a perfectly
regular and predictable function involving time.

But that equation can't be solved for o, can it? If it can't, you're
not able to talk about any o = f(e), are you?

Yes, you really do need to study the calculus again. The solution of the
above equation is

o = integral( k1*e - k2*o)dt,
where the integration is from a time when the initial conditions are known
to any time after that. The variables o and e are functions of time.

Right about now I probably should put on my dunce cap and go sit in the
corner, but I have to ask these things if I'm going to get to the
enjoyable insights. The answer I anticipate receiving back is that when
people write "o = f(e)" they mean something like "do/dt = k1*e - k2*o",
and the fact that the former *looks* algebraic comes from that most
annoying habit of simplifying notation so that the calculus aspects are
implicit. (Such wholesale disposal of explicit detail seemed pervasive
to calculus and was a major factor in the difficulties I had with it.
Not as big a factor as neglecting to study, however.)

The notation o = f(e) is perfectly general. A function is simply a recipe
for computing a value given the argument or arguments. If it's not already
understood that time is involved, you can write o(t) = f(e(t)). The
notation o(t) is read " o as a function of time", and so on.

Your best bet for learning what you need to know here is to start with
elementary physics. Here you will learn how to represent physical systems
with equations, which seems to be what you're having trouble with. I
recommend going as far back as necessary to find the level of physics at
which you experience it as simple and obvious. A high-school text is
probably a good place to start (no insult intended -- you have to start
where your education stopped).

Best,

Bill P.

[From Bruce Gregory (980220.1456 EST)]

Tracy Harms (19980220.08)

The output "o" is not the output of the comparator, it is the output of
the *system*, i.e. everything which does not count as "environment".

Yes. In my example the thermostat is the *system*.

In that system there is a two-minute delay required between the turning
off of the heating elements and the turning off of the circulating fan.
Presumably this helps prevent damage to the furnace which overheating
could produce. But the incentives behind that specification are
irrelevant; the main point is that his simplified, idealized system
provides an example where output does not correlate with error signal.
The error signal can be absent while the system at large continues to
blow hot air out of its vents.

Not so. The output correlates perfectly with the error signal.
Either the error signal is present or it stopped being present
during the past two minutes. If neither of these ponditions is
true, the circulating fan is _always_ off.

Bruce

[from Tracy Harms (980221.16 PST)]

Thanks to all who provided clarification. Both Bruce Gregory and Bill
Powers have shown me that I misunderstood the role and presence of time
in these equations.

One item still seems open to discussion, re. Bill Powers (980220.1103
MST)

>But that equation can't be solved for o, can it? If it can't, you're
>not able to talk about any o = f(e), are you?

Yes, you really do need to study the calculus again. The solution of
the above equation is

o = integral( k1*e - k2*o)dt,
where the integration is from a time when the initial conditions
are known to any time after that. The variables o and e are
functions of time.

Yes, thank you, this solves for o. But what's that I see on the right
hand side of the "="? It is an o! I understand how we can drop t from
the explicit equation, but I *don't* see how you can drop o. If you
don't drop it, then the generic version has to be "o = f(e,o)" {or
however it is considered proper to write that}, not "o = f(e)". If you
can just drop the o from the right-hand side, you can equally well drop
the e; what does that leave?

All this is, I'd guess, right in line with the claims that o = f(e) is
incomplete. If I'm on target here (despite my general mathematical
ineptness) then the incompleteness in question means that the claim
communicated by o = f(e) just isn't so. The incompleteness isn't to be
remedied by adding another equation, it must be remedied by including o
in the input to the function.

···

--
Tracy Harms
Bend, Oregon

             "The day the flesh shapes and the flesh the day shapes."

[From Bill Powers (980221.0315 MST)]

Tracy Harms (980221.16 PST)--

The solution of
the above equation is

o = integral( k1*e - k2*o)dt,

Yes, thank you, this solves for o. But what's that I see on the right
hand side of the "="? It is an o! I understand how we can drop t from
the explicit equation, but I *don't* see how you can drop o.

Consider the (arbitrary) algebraic equation

o = a + b*o

Subtract b*o from both sides:

o - b*o = a

o*(1 - b) = a

o = a/(1-b)

In a differential equation you can't get rid of the o on the right so
easily, but through the methods of solving such equations, you can arrive
at an expression with o appearing only once, on the left. If the error
signal is a "step function" (zero before some time, constant and nonzero
after that time), the solution will be that o follows a rising curve that
levels out after a while. The solutions of differential equations are not
fixed values, but time-functions that depend on how the independent
variables change with time.

If you
don't drop it, then the generic version has to be "o = f(e,o)" {or
however it is considered proper to write that}, not "o = f(e)". If you
can just drop the o from the right-hand side, you can equally well drop
the e; what does that leave?

You're quite right about all this, which shows that your mathematical
intuition is working. However, it is also possible to solve the equation
for the behavior of o given the behavior of e -- the easiest way, and
sometimes the only way, is by simulating the system. Your first equation, o
= f(e,o), refers to instantaneous values of o and e. The second, o = f(e),
should more properly be written o = g(e) to show that the function g is not
the same as f. The second equation would be the solution of the first, and
the second function would relate the behavior of o through time to the
behavior of e through time.

All this is, I'd guess, right in line with the claims that o = f(e) is
incomplete. If I'm on target here (despite my general mathematical
ineptness) then the incompleteness in question means that the claim
communicated by o = f(e) just isn't so. The incompleteness isn't to be
remedied by adding another equation, it must be remedied by including o
in the input to the function.

You've decided that "o = f(e) just isn't so," and you're looking for
evidence to support this conclusion. Give up: it is so. An output function
is a physical device. An error signal and an output signal are physical
signals. The output signal is generated by the output function and the
output function is caused to act by the error signal. This means that the
output is a function of the error signal. No way around it.

Ouch, reorganization hurts. Grin and bear it.

Best,

Bill P.

[Fred Nickols (980222.1650 EST)]

   Bill Powers (980221.0315 MST)

     responding to Tracy Harms (980221.16 PST)--

Me:
I'm not trying to butt in here but I have a question about the comments
below:

Bill:

An output function
is a physical device. An error signal and an output signal are physical
signals. The output signal is generated by the output function and the
output function is caused to act by the error signal. This means that the
output is a function of the error signal. No way around it.

Me again:
I'm checking my understanding here, not challenging. As I understand this
stuff, an error signal is the output of a comparator. This output is the
difference, if any, between two inputs to the comparator, one of which is
the reference condition, the other of which is the perceived condition (I
don't think I have that last term quite right). An error signal is not
only an output of the comparator but also an input to the output function.
The output function, in turn, generates an output signal.

Back to the good old keep-the-auto-in-the-lane example again. The desired
state of affairs (the reference condition) is the auto roughly in the center
of the lane. If the perceived position is too far to the left or right of
center, there is an error signal. There's where I lose it. Is the output
signal my turning of the wheel or is that the output function? Or, am I
the output function and the output signal is whatever results in my body
turning the wheel?

help me, help me

Regards,

Fred Nickols
The Distance Consulting Company
nickols@worldnet.att.net
http://home.att.net/~nickols/distance.htm

[Bruce Gregory (98022.1827 EST)]

Fred Nickols (980222.1650 EST)

Back to the good old keep-the-auto-in-the-lane example again. The desired
state of affairs (the reference condition) is the auto roughly in the center
of the lane. If the perceived position is too far to the left or right of
center, there is an error signal. There's where I lose it. Is the output
signal my turning of the wheel or is that the output function? Or, am I
the output function and the output signal is whatever results in my body
turning the wheel?

Turning the wheel is part of the feedback function that eliminates the error
arising from a difference between the perceived position and the reference
condition. As far as this error is concerned, your actions are part of the
environmental feedback function. On a more detailed examination, the motions
you make are part of other feedback loops as higher order system set reference
levels for muscle movements, etc.

Bruce

[From Bill Powers (980222.0428 MST)]

Fred Nickols (980222.1650 EST)--

I'm checking my understanding here, not challenging. As I understand this
stuff, an error signal is the output of a comparator. This output is the
difference, if any, between two inputs to the comparator, one of which is
the reference condition, the other of which is the perceived condition (I
don't think I have that last term quite right). An error signal is not
only an output of the comparator but also an input to the output function.
The output function, in turn, generates an output signal.

Right all the way, so far.

Back to the good old keep-the-auto-in-the-lane example again. The desired
state of affairs (the reference condition) is the auto roughly in the center
of the lane. If the perceived position is too far to the left or right of
center, there is an error signal. There's where I lose it. Is the output
signal my turning of the wheel or is that the output function? Or, am I
the output function and the output signal is whatever results in my body
turning the wheel?

The error signal becomes the input to a "device" (in your brain) that
converts it into lower-level reference signals for the angular position of
the steering wheel. That device is your output function. The output of that
function is either an angle or a torque; not having worked out a model of
steering a car, I don't know which would fit real behavior the better.

The steering wheel is connected to the front wheels which, when they cock
left or right, exert a sideways force on the car tending to alter its path.
At the same time, other influences like wind and bumps also produce forces
tending to alter the cars's path. The sum of all these forces determines
the curvature of the path.

As the car turns, the driver's picture of its relation to the road changes.
One of these pictures has been stored and is the current desired position
relative to the road, as seen by the driver. This is the picture currently
serving as the reference signal. The actual picture is the current
perceptual signal. When these signals are compared, we get the error
signal, which is where we started two paragraphs above.

In analyzing any control loop, you can define the internal functions in any
convenient way. However, I usually try to define the system so the control
system part is neural, and the muscles form the output interface to the
outside world. Everything between this motor output and the perceptual
(sensory) input is part of the environment, and constitutes the
environmental feedback function. Disturbances can act anywhere in the
environmental feedback function, including directly on the sensory input or
parallel to the motor output.

Is this what you wanted?

Best,

Bill P.

···

help me, help me

Regards,

Fred Nickols
The Distance Consulting Company
nickols@worldnet.att.net
http://home.att.net/~nickols/distance.htm

[From Bruce Gregory (980223.1015 EST)]

Bill Powers (980222.0428 MST)

As the car turns, the driver's picture of its relation to the road changes.
One of these pictures has been stored and is the current desired position
relative to the road, as seen by the driver. This is the picture currently
serving as the reference signal. The actual picture is the current
perceptual signal. When these signals are compared, we get the error
signal, which is where we started two paragraphs above.

I suspect that "one of these pictures has been stored" is too
simplified a description of what has actually happened. My
thinking about this was triggered by a recent post by Martin
[Martin Taylor 980222 02:00]. Just as we do not have a stored
"picture" of the letter A to which we compare a perception (if
this were true, pattern recognition would be a piece of cake), we
probably do not have a stored memory of the appearance of the
road. It seems more likely, that learning to drive requires
reorganization of our perceptual apparatus. This reorganization
allows many different road conditions to "show up" for as
"being in the desired position relative to the road". (I've been
using the term "distinction" as a way to refer to this
reorganization process and its outcome.) Once we have the
distinction "being in the desired position relative to the road"
it becomes possible to exercise control with respect to a
reference position that varies with changing environment and
different roads, but always signals "being in the desired
position with respect to the road." I am beginning to suspect
that most reorganization takes place in the perceptual inputs
rather than in the mechanism linking error to output, and that
the process of creating new distinctions lies at the heart of
learning (as contrasted with practising) control.

Does any of this make sense? What have I overlooked?

Bruce

[From Bill Powers (980223.1313 MST)]

Bruce Gregory (980223.1015 EST)--

As the car turns, the driver's picture of its relation to the road changes.
One of these pictures has been stored and is the current desired position
relative to the road, as seen by the driver. This is the picture currently
serving as the reference signal. The actual picture is the current
perceptual signal. When these signals are compared, we get the error
signal, which is where we started two paragraphs above.

I suspect that "one of these pictures has been stored" is too
simplified a description of what has actually happened.

Yes, it is. It is almost as oversimplified as the concept of "distinctions"
or "discriminations." But not quite.

What is required for continuous control is a perceptual signal that is a
continuous analog of a physical quantity, in this case the lateral position
of the car relative to the road. As the actual distance of the car from the
edge or center of the road varies, the perceptual signal must also vary, in
the same way a voltmeter reading varies as the voltage being measured changes.

The reference signal is just a magnitude representing a point somewhere in
the range of continuous variation of the perceptual signal. As the
reference signal changes, the control system steering the car sees to it
that the perceptual signal changes right along with it. Thus you can move
to the left to see around the truck ahead, and if the way is clear move all
the way into the left lane until you've passed the truck and then back into
the original lane, controlling the car's relationship to the road (and the
truck) all the time. This sort of control requires a continuous graded
perception, not merely a distinction or discrimination of "right position"
from "wrong position". Steering a car is not like a game of hot and cold.

Forget, please, that I ever mentioned "pictures in the head."

Reading the posts for the past few days has left me with an overwhelming
sense of futility. It is futile even to try to explain why.

Best,

Bill P.

[From Bruce Gregory (980223.1830 EST)]

Bill Powers (980223.1313 MST)]

What is required for continuous control is a perceptual signal that is a
continuous analog of a physical quantity, in this case the lateral position
of the car relative to the road. As the actual distance of the car from the
edge or center of the road varies, the perceptual signal must also vary, in
the same way a voltmeter reading varies as the voltage being measured changes.

Yes. This is just the way I would model the phenomenon. I was
talking about the phenomenology, not the physics.

The reference signal is just a magnitude representing a point somewhere in
the range of continuous variation of the perceptual signal. As the
reference signal changes, the control system steering the car sees to it
that the perceptual signal changes right along with it. Thus you can move
to the left to see around the truck ahead, and if the way is clear move all
the way into the left lane until you've passed the truck and then back into
the original lane, controlling the car's relationship to the road (and the
truck) all the time. This sort of control requires a continuous graded
perception, not merely a distinction or discrimination of "right position"
from "wrong position". Steering a car is not like a game of hot and cold.

Yes, I agree.

Forget, please, that I ever mentioned "pictures in the head."

Consider it forgotten.

Reading the posts for the past few days has left me with an overwhelming
sense of futility. It is futile even to try to explain why.

I'm sorry you feel that way. I don't share your pessimism. I
agree with everything you say in this post. I suspect my
language is offputting and I apologize for that. For me it
turns out to be a fruitful way to describe the phenomenology of
the reorganization of perceptions that is necessary if control
is to be extended to new domains. It obviously simply
turns you off. I can only assure you that nothing _I_ am saying
is inconsistent with what _you_ are saying -- at least as far as
I can tell. I will stop talking this way.

Bruce