NEC (formerly UEC)

[From Bill Powers (2002.05.20.1139 MDT)]

Bill Williams 20 May 02 11:00 CST--

>Myself, I'm not at all sure about the "universal"

element in Bill Powers' UEC.

I'n not, either. Perhaps we could refer to this as the "Nonlinear Error
Curve" or NEC, which drops any claim to universality.

>The idea has been advanced that the closer an agent come to the goal the
higher

the gain, or output of the agent will be.

That would be true outside some distance from the goal. It can't apply all
the way to zero error, or the output would be maximum when the error is zero.

It seems to me that their may be a
very good reason for this. Suppose the goal is connected to the maintaince of
the agent-- such as a source of calories. Then it becomes prudent for the
agent
to consider the relationship between the cost of attaining the goal and the
worth ( say in calories ) which the goal represents.

Nice. We can also consider that conflict itself is expensive, either in
terms of calories or in terms of lost control. The ass stuck between the
haystacks can't eat from _either_ of them.

One way out of a conflict (or "choice") is reorganization, but that
solution is slow, stochastic, and almost impossible to reverse. The
end-product of reorganization should be systems that systematically, and
much more quickly, take care of errors that can initially be dealt with
only by trial and error. An NEC type of output function would work reliably
and quickly to resolve single or multiple conflicts, and would not require
giving up the competing goals semi-permanently (I mean so you'd have to
rely on uncertain reorganization to get them back once they were abolished).

Best,

Bill P.

[From Bruce Nevin (2001.05.20 16:07 EDT)

But, transposing the problem into supposing
the ass can be represented by two

control loops does make the problem much worse because, the ass will be
stuck {somewhere between the two haystacks ) even if the ass is not
positioned

precisely at the mid-point, and even if the two control loops are
inbalanced.

Bill Powers (2002.05.19.2119 MDT)–

If we now

had a higher system that simply reduced the output gain when the error
went

above a certain level (an S-R system as I’m thinking of it right now),
the

system trying to move the ass toward the more distant pile would have
a

larger error and hence a lower gain, while the other system would have
a

lower error and hence a higher gain. This is an unstable situation (the
the

gain adjustments are sensitive enough), and one system will end up
taking

control while the other gives up.

Why two control loops? Please specify. Is there a duplication of the
hunger/eat loop? Is there a duplication of the see/smell/locate loop that
serves it, or of the walk over to hay loop that also serves it?
The bifurcation is in the environment. On what basis do you say it is in
the ass? What if there were three choices, or five, or more? Do you
postulate an instantaneous multiplication of whatever control loops are
involved?
As Martin (2002.05.20 09:38) also observed, this is about making a
decision.
Indecision does happen, but usually when to gain A I lose B
and vice versa. In Martin’s example, do I go by car, bike, or bus? Each
choice has some benefits and some detriments. These values (e.g. exercise
vs. possibly picking up a package at the post office vs. wife possibly
needing the car) are not commensurate and matched up like pairs of
haystacks. The monkey with its fist caught in a jar is another example.
Two controlled variables at a higher level (food vs. freedom) each
depending upon the employment of one effector (the monkey’s paw). When no
matter which way the ass turns it gains hay to eat, the choice lacks this
poignancy. It doesn’t matter which one it goes to, there’s hay there.
Again, why do you need two control loops? Surely there is no duplication
(or for n choices n-ary multiplication!) of e.g. the
locomotion loop so that there is conflict between locomotion loop
A which is moving the ass toward haystack A and locomotion
loop B which is moving the ass toward haystack B. The
conflict is in the fact that the ass’s resources for controlling the
perception of eating hay can be employed only in one location at a time.
It is not a problem of two control systems, each controlling one of the
possible places to deploy limited resources. It is a problem of one
control system ready for deployment wherever attention is focused – and
attention having to focus on one of several possibilities. Again,
supposing you can reduce ‘attention’ to only the direction of focus of
the eyes, there is only one control loop and two locations in the
environment.

Why turn toward one haystack rather than the other? Well, why does the
vortex over the drain start one way or the other?

[From Bill Powers (2002.05.20.1156 MDT)]

you’re stuck somewhere among the goals unless
something happens to make

the situation unstable.

No, that’s backward. Something has to happen (and keep on happening) to
preserve the stability of the situation. The ass lives in the same
disturbance-filled universe as we do. Unless something is controlling to
resist disturbances, something will happen to disturb whatever
transient stability you may have set up.

As soon as you drop from the ethereal realm of logic to an actual ass in
an actual disturbance-filled environment, including factors incorporated
into the ass’s memory and imagination from its history of experience, any
perceptual differentiation can become the basis for preference. Which way
is the breeze blowing? Which way is the sun shining? is there a fly
buzzing in that ear? Is the ass right-hoofed or left-hoofed? Does it
smell water too? I don’t know about you, but when I’m faced with an
embarrassment of riches I look for differentiators. Those tomatoes in the
grocery store may look all alike, but are they really?

Having turned toward one haystack, eyes foveate it, and indeed the other
haystack is no longer in sight at all, assuming the ass was between the
two stacks and turned away from one and toward the other. There is no
gradual shift in gain as the ass moves farther from one and closer to the
other. The abandoned haystack ceases to be present in its perceptual
inputs (though presumably it persists in memory).

To simplify things, just place the ass before a pile of hay. Which
mouthful to eat? How can it possibly decide? Does the silly ass stand
there and starve to death in indecision? Of course not. Does it not just
move its mouth toward the hay by the path of least effort and eat
whatever hay its mouth meets?

So the ass between two haystacks is not stuck in indecision which way to
go because something, it matters not what, tips the balance. And no
matter how carefully the investigator has balanced the choice, it cannot
stay evenly balanced in a world filled with disturbances. And there’s hay
either way. In just the same way, the ass facing a stack of more than one
mouthful is not stuck in indecision which mouthful to eat, because it’s
all hay. All it has to do is pay attention and move toward the hay it’s
looking at.

Rick Marken (2002.05.20.0850–

···

At 07:09 PM 5/19/2002 +0300, William Williams wrote:
At 09:20 PM 5/19/2002 -0600, Bill Powers wrote:

At 08:54 AM 5/20/2002 -0500, Richard Marken wrote:

My comments about the UEC have never been influenced by anything Tom has
said about it and until very recently I had no inkling of his views about
it and no particular concern to find out. There is no campaign against
the UEC that I am aware of. If there is, it is irrelevant to my
questioning of it.

    /Bruce

[From Rick Marken (2002.05.20.1700)]

Bruce Nevin (2001.05.20 16:07 EDT)--

Why two control loops? Please specify.

I think it's because, by assumption, there are two controlled variables:
1) distance to hay stack 1 and 2) distance to hay stack 2. The same
lower level variable (ass position) is the means of controlling both
these perceptions.

What if there were three choices, or five, or more?

There is really no "choice" at the hay stack level. The systems
controlling proximity to each stack are assumed to have chosen (selected
a reference for) the distance it controls to be 0 (next to the stack).
The choice that is being made is the one regarding the ass's position.
One haystack distance control system wants ass position to be "next to
stack 1". The other haystack distance control system wants ass position
to be "next to stack 2". This conflict, which is at the level of ass
position, not haystack selection, is the reason for the difficulty in
making the "choice" regarding (set the reference for) ass position --
the ass cannot be physically in two places at once. The conflicted ass
will actually appear to have made the "virtual" choice of standing
between the two hay stacks. It will resist pushes toward either stack.

The conflict is in the fact that the ass's resources for controlling
the perception of eating hay can be employed only in one location at a
time.

I think the conflict is in the fact that the ass's resources for
controlling it's distance from each hay stack-- it's resources being its
physical location relative to the stacks -- can be in only one location
at a time.

It is not a problem of two control systems, each controlling one of
the possible places to deploy limited resources.

The two "stack proximity" control systems are actually the _cause_ of
the conflict. The outputs of these two control systems are "pushing"
the ass's position in opposite directions. The result is an ass frozen
between the two stacks -- unless there is an NEC in both systems. If
there is a NEC, then if one system manages to _push_ the error in the
other system past the non-linear inflection point, then that system
"wins" (gets the distance it is controlling to 0).

Best regards

Rick

···

--
Richard S. Marken, Ph.D.
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

[From Bill Williams 20 May 02 7: ]

[From Bruce Nevin (2001.05.20 16:07 EDT)

>But, transposing the problem into supposing the ass can be represented by two
>control loops does make the problem much worse because, the ass will be
>stuck {somewhere between the two haystacks ) even if the ass is not

positioned

>precisely at the mid point betweeen the two haystacks.

Why two control loops? Please specify. Is there a duplication of the
hunger/eat loop? Is there a duplication of the see/smell/locate loop that
serves it, or of the walk over to hay loop that also serves it?

The bifurcation is in the environment. On what basis do you say it is in
the ass? What if there were three choices, or five, or more? Do you
postulate an instantaneous multiplication of whatever control loops are
involved?

As Martin (2002.05.20 09:38) also observed, this is about making a decision.

Again, why do you need two control loops?

The answer, if what I have amounts to an answer, is that I wrote the routine
for Buridan's Ass the same way I did a number of previous "behaviors" such as
various versions of the Giffen effect, the consumer in terms of the Veblen
/Dusessenberry effect, and everything else. In one version of the Giffen effect
there are three loops, one each for the budget, calories and meat. So, when
the old paradox of the Ass popped into mind, I said to myself I'll translate
the scholastic paradox into control theory terms. So I assigned the Ass two
control loops with a reference level attached to each haystack. The only
response I can provide ( right now ) is that's the way I've always done it.

What I understand you to be saying is that this isn't an adaquate way in which
to represent the Ass, or any of the other situations which have been modeled in
the past. Maybe it isn't. There's a long standing dispute in economics
concerning the justifications or lack there of for economic modeling. (Most of
what the economists call modeling wouldn't pass inspection here as modeling--
but this brings in another question.) The people doing the modeling in
economics tell critics, If you don't like the way I do it, show me a better
way. The critics insist that, whether or not there is any alternative, the
existing models are simply too unrealistic. Upto this point when thinking
about the control theory models I've managed to construct, I've been rather
smug. Compared to the orthodox economic models, the control theory based models
are I think better because they a based upon a much better conception of
behavior. And, they seem to me to generate some useful insights into how a
consumer actually behaves. Insights that outcompete the orthodox conception of
a consumer. So the justification, if there is one, in my view derives from the
models capacity to predict what will happen when the conditions in which an
agent is placed are changed-- as in an experiment. But, I can't think, right
now, of a way to do an experiment with the model of an Ass. So, maybe for the
time being it is a "just so" story written in Pascal. This may not be an
adaquate defense, but then I don't feel all that defensive.

regardsz,

Bill williams

···

At 07:09 PM 5/19/2002 +0300, William Williams wrote:

______________________________________________________________________
Do you want a free e-mail for life ? Get it at http://www.email.ro/

[From Bruce Nevin (2001.05.21 01:31 EDT)]

Rick Marken (2002.05.20.1700)--

···

At 04:57 PM 5/20/2002 -0500, Richard Marken wrote:

by assumption, there are two controlled variables:
1) distance to hay stack 1 and 2) distance to hay stack 2.

The ass is controlling a relationship perception where one signal comes from memory at any given moment as it turns its head and eyes to regard the other haystack. There is no conflict between two separate control systems.

         /Bruce

[From Rick Marken (2002.05.21.0820)]

Bruce Nevin (2001.05.21 01:31 EDT) --

Rick Marken (2002.05.20.1700)--

>by assumption, there are two controlled variables:
>1) distance to hay stack 1 and 2) distance to hay stack 2.

The ass is controlling a relationship perception where one signal comes
from memory at any given moment as it turns its head and eyes to regard the
other haystack. There is no conflict between two separate control systems.

I'm sorry. I thought we were talking about Bill Williams' simulation (ass.exe)
of an ass caught in a conflict between two haystacks. In Bill's program the
ass is characterized by two control systems: one controlling the distance
between the ass and the haystack (circle) on the left (A) and the other
controlling the distance between the ass and the haystack on the right (B).
The program let's you adjust the relative gains of the two control systems.
When the gains of the two systems are the same the ass (small red circle)
remains between the two haystacks and equidistant from each for each iteration
of the control cycle (press of the space bar). When the gain of one system is
higher than the other the ass still remains between the two haystacks but
closer to the one for which the control loop has the higher gain. The are
apparently no disturbances to the ass's position so once the ass reaches the
d�tente point between the two haystacks it doesn't move at all.

I agree that there is not necessarily any conflict when a single relationship
(such as the ass's relationship between itself and one haystack) is
controlled.

Best regards

Rick

···

--
Richard S. Marken, Ph.D.
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

[From Bill Powers (2002.05.21. 0957 MDT)]

Bruce Nevin (2001.05.20 16:07 EDT) --

Why two control loops? Please specify. Is there a duplication of the
hunger/eat loop? Is there a duplication of the see/smell/locate loop that
serves it, or of the walk over to hay loop that also serves it?

You bring up a relevant question, to which I have only a conjectural answer.

I see one level at which the food locations are perceived and controlled,
and a lower level in charge of moving to and being in different spatial
positions. If two independent systems are controlling for food at two
locations, they both would operate by sending reference signals to a
location-controlling system (if we assume that that is where the conflict
would be expressed). The reference signals would be added, or perhaps be
averaged, at the reference input of the lower system, where the virtual
reference level would be established. At that level we have normal control,
but the final goal position would not be at the location of either food source.

The basic question is, why two control systems at the higher level? This is
a result of a basic assumption I made long ago. Rather than thinking of one
perceptual signal that can convey many different kinds of perceptual
information (the "encoding" concept), I opted for a model in which each
perceptual channel conveys only one dimension of experience, with
variations in signal magnitude only, so a given signal never means any
different _kind_ of perception. This may seem wasteful, until you look at
the other problems it does away with, such as decoding that can recognize
one signal pattern out of many different patterns that might be present in
the same complex signal. Instead of having, for example, a
phoneme-recognizer that can put out one signal pattern for "a" and another
pattern, in the same physical channel, for "e", we would have separate
_simple_ signal channels for different phonemes, with one channel always
representing the same phoneme. If we go for the "encoding" idea, we then
have to provide, at the input of every system that receives complex phoneme
signals, a device that can pick one particular pattern out of a stream of
many possible (or even overlapping) patterns -- so we have simply moved the
pattern-recognition problem into the receiving system without solving it.

If each perceptual channel always carries a signal representing a specific
dimension or attribute of perception, then the receiving system needs no
decoder. However, the control of any given perceptual signal requires its
own control system, which might seem like an undesirable tradeoff. We get
it all back, with interest, when we look at how complex the control systems
must be. The comparator consists of a simple magnitude-subtractor: it
receives two signals and outputs a signal having a magnitude proportional
to the difference between the input magnitudes (magnitude == frequency).
The output function may or may not be complex, but it doesn't need to
recognize different _kinds_ of error signals, so it's far less complex than
the output system needed with the "encoding" concept. In fact, the minimum
control system,. at any level, consists of just two neurons: the input
neuron that creates a signal that is a function of multiple inputs, and the
comparator, which receives a reference signal and the perceptual signal and
outputs the difference signal. These functions could even be combined into
a single neuron. The complexity of the whole system is then brought about
by having a large number of cheap and simple control systems rather than a
small number of very complicated and neurally expensive ones.

This is basically how I can visualize two control systems, one for
controlling the relative position of each food source (by moving either the
body or the food source). However, since this implies a huge conflict, we
need yet another level that activates the reference signals for just one of
the food-source controllers at a time. If that level exists, the donkey
will simply walk up to one haystack (the biggest? The closest? The
greenest? The first one noticed?) and eat it first, and then if it's still
hungry, will move to the other and eat it (sequence control).

If, somehow, the donkey failed to use a higher level of control to resolve
the conflict, then both control systems would turn on at the same time and
would try to move the body to two different locations at once. The rest
follows.

Am I satisfied with this? Well, no. There must be some compromise idea that
retains the simplicity of the one-perception-one-channel approach, but
without the awkwardness of supposing that there is a control system for
every possible perceptual signal. On trhe other hand, we must recognize
that there are many situations in which we actually notice multiple
attractive things, and for some time at least dither about which one to
pick first (assuming, of course, that we can't have all of them at once).
Why do we dither instead of just picking one and being done with it? That's
really the situation we're trying to analyze, even if it's short-lived.

I quite agree with you about actual donkeys, and would agree if you said
that if any real donkeys behaved like the apocryphal donkey between the
haystacks, there would be no real donkeys. We have probably wasted enough
time on this invented and imaginary problem which psychologists, but no
real donkeys, have had.

>As Martin (2002.05.20 09:38) also observed, this is about making a decision.

Well yes, that names the problem but doesn't illuminate much. Exactly what
is a decision? How is a decision made? I think that what we usually call
decision-making situations are simply conflicts, where, for example two
desirable goals exist but we can't achieve either without precluding the
other, or where picking the desirable one means we get the undesirable
consequence, too, and so forth.

To say that a decision is needed doesn't tell us how we make it. We've
already considered the main possibilities: reorganization, or the
application of some learned algorithm that automatically makes the
selection of one goal for us. Either way, flipping a coin or turning a
logical crank, we have avoided the need to "make a decision." I don't
really think there is any special human capacity for "deciding." There are
just various miscellaneous tricks for resolving conflicts that we have
learned, or that are built in to us. When there are no conflicts, no
decisions are needed.

Indecision does happen, but usually when to gain A I lose B and vice
versa. In Martin's example, do I go by car, bike, or bus? Each choice has
some benefits and some detriments. These values (e.g. exercise vs.
possibly picking up a package at the post office vs. wife possibly needing
the car) are not commensurate and matched up like pairs of haystacks. The
monkey with its fist caught in a jar is another example. Two controlled
variables at a higher level (food vs. freedom) each depending upon the
employment of one effector (the monkey's paw). When no matter which way
the ass turns it gains hay to eat, the choice lacks this poignancy. It
doesn't matter which one it goes to, there's hay there.

That's a good discussion of conflict, which is the same thing I'm talking
about. But no particular "poignancy" is needed to create a conflict. When
no matter which theater you go to, you will see a good movie, you don't
generally just turn at random or go into the nearest one. You stop and
discuss it for a moment, looking for some "reason" for picking one over the
other -- for example, you want to see both, and could see either one, but
one will be playing tomorrow while this is the last day for the other one.
Ah, no decision required. When you're in conflict between equally desirable
alternatives, almost any reason will do, however, superficial (better popcorn).

>Again, why do you need two control loops?

One reason is to explain why we feel a pull in two different directions at
once, even though we know that any such "pull" is really our own doing
(unless we prefer the S-R explanation). Somehow it seems that we are really
trying to go two ways at once, if not with our muscles, then at a higher
level, in picking where we want to go. Such situations don't last long,
normally, but when they do we can be in a real pickle. Stay with the wife
and kids, or fly off to Tahiti with the gorgeous bimbo? I have that problem
all the time.

>Why turn toward one haystack rather than the other? Well, why does the
vortex over >the drain start one way or the other?

If you mean that any slight disturbance could tip the balance one way or
the other, you're at least acknowledging that there are simultaneous
opposing tendendies or balanced forces. But the crucial question is what
you assume about the consequences of a perturbation. If you assume that the
equilibrium is unstable, then anything that creates a tiny unbalance
determines the outcome. But if the equilibrium is stable, like a marble in
a bowl, then perturbations will be resisted and the equilibrium will be
restored. So your argument tends to get circular. If, as you say, the
imaginary donkey will go increasingly in whichever way it is perturbed,
then it follows that the donkey's system is of the type that is in unstable
equilibrium, which is the kind it has to be if it is to behave that way.
But if it's the other kind, it will return to the equilibriuum point after
the perturbation. Whatever you assume determines the outcome. It's not
valid to make the assumption, deduce how the donkey would behave, and then
use that behavior to prove that your assumption is correct.

For example:

No, that's backward. Something has to happen (and keep on happening) to
preserve the stability of the situation. The ass lives in the same
disturbance-filled universe as we do. Unless something is controlling to
resist disturbances, something will happen to disturb whatever transient
stability you may have set up.

That assumes that the ass's behavior is not determined by a pair of
conflicting control systems (which keep on acting to preserve stability),
but is in an unstable equilibrium such that a push in any direction will
cause it to go further in that direction. Fine. But you could have assumed
the other thing, too.

As soon as you drop from the ethereal realm of logic to an actual ass in
an actual disturbance-filled environment, including factors incorporated
into the ass's memory and imagination from its history of experience, any
perceptual differentiation can become the basis for preference. Which way
is the breeze blowing? Which way is the sun shining? is there a fly
buzzing in that ear? Is the ass right-hoofed or left-hoofed? Does it smell
water too?

All this starts with the assumption that there are no control systems in
conflict, operating at the same time in different directions. Instead, the
proposal is that our actions are caused by the sum of stimuli acting on our
nerve-endings, which trigger our behavior to go one way or another. Is
that really what you mean to propose?

I don't know about you, but when I'm faced with an embarrassment of riches
I look for differentiators. Those tomatoes in the grocery store may look
all alike, but are they really?

That's what I've been saying here. When there's conflict, we look for some
way to resolve it -- rejecting one tomato that has a micosopic blemish that
the other doesn't have, or one more blemish than the other has. The last
thing we want is to have to make an actual arbitrary decision. But of
course no such problem would arise if we didn't have any conflict. We'd
just pick up the first tomato to hand and go on our way. We create our own
conflict by setting up multiple goals; if we can't see any basis for
choice, we make up some basis. It's as though we have to suffer at least
some conflict before we can feel that action is "justified." If we happen
to be trying to remember where the bread aisle is, we just pick up the
tomato at hand and press on, forgetting to make the decision.

So the ass between two haystacks is not stuck in indecision which way to
go because something, it matters not what, tips the balance. And no matter
how carefully the investigator has balanced the choice, it cannot stay
evenly balanced in a world filled with disturbances. And there's hay
either way. In just the same way, the ass facing a stack of more than one
mouthful is not stuck in indecision which mouthful to eat, because it's
all hay. All it has to do is pay attention and move toward the hay it's
looking at.

I think that's about enough of that. You opt for unstable equilibrium and
the impossibility of conflict, as well as the deciding influence of
stimuli. I opt differently. And we must remember that no such asses,
organized in any way mentioned here, have actually been observed to behave
in either way, whether on the head or the point of a pin.

The only way out of arguments like these is continued modeling. Every set
of assumptions implies its own set of predicted behaviors. The best we can
do, lacking data, is to set up models covering the various possibilities,
and wait for a time when we can put them all to the test.

Best,

Bill P.

[From Bill Williams 21 May 02 CST ]

[From Bill Powers (2002.05.21. 0957 MDT)]

Bruce Nevin (2001.05.20 16:07 EDT) --

>Why two control loops?

This is
a result of a basic assumption I made long ago. Rather than thinking of one
perceptual signal that can convey many different kinds of perceptual
information (the "encoding" concept), I opted for a model in which each
perceptual channel conveys only one dimension of experience.....

If each perceptual channel always carries a signal representing a specific
dimension or attribute of perception. These functions could even be combined

into

a single neuron. The complexity of the whole system is then brought about
by having a large number of cheap and simple control systems rather than a
small number of very complicated and neurally expensive ones.

So, thanks to Bruce's skepticism, I've been reminded of what I've previously
known but since forgotten the "one perception: one neural current" formulation.

Bill Williams

···

______________________________________________________________________
Do you want a free e-mail for life ? Get it at http://www.email.ro/