Misc subjects

[William_T_Powers2]

[From Bill Powers (930318.1015)]

Wolfgang Zocher (930318) --
(response to private post)
Sorry about that very nasty illness. I hope you will give a
higher priority to getting well than to writing programs! I'm
sure I convey the hopes of everyone else on CSGnet that you will
recover soon. Two weeks of pneumonia is no joke.

···

------------------------------------------------------------
Allan Randall (930317) --

So my version of the challenge would take Ashby's compensatory
and error-driven control systems and, assuming they were both
designed to control, make a prediction concerning which would
control better.

OK. For ideal components and linear channels, the prediction is
pretty clear. But I will be interested to see the analysis you
use for the error-driven system, in which E is partly a function
of itself (via R).

Once you have the analysis for the two cases done, it would be
useful to derive the real-world requirements on both systems for
achieving a certain degree of regulation in the presence of
noise. One channel common to both arrangements is that from R to
T. How much difference would it make to each system if the output
of R contained some specifiable amount of random variation?

Outside the scope of this challenge, there is a factor that Ashby
didn't take into account: the possibility of disturbances that
act directly on E, and are not detected by R. Under those
conditions, disturbance-driven regulation is impossible, while
error-driven regulation continues as before.

Meeting the "challenge" is less important than producing an
actual analysis that I might be able to use! Keep in mind that I
am only a humble engineer, and need to have everything spelled
out in babytalk.
------------------------------------------------------------
Hans Blom (930317) --

So sensory input is a dependent variable. I take this to mean
that sensory input has no degrees of freedom of its own; its is
dependent UPON some- thing, and that something is the reference
signal.

What we say about dependent and independent variables is always
in the context of the particular model we are proposing. In a
single elementary control system in the PCT model, the perceptual
signal is ALWAYS a one-dimensional scalar. If there is a
multidimensional external quantity being controlled, in our model
the way it currently stands more than one control system would be
needed to keep all its degrees of freedom under control. I think
we all recognize that this conception has some failings --
handling sideways interactions among control systems of the same
level would be very awkward, for example. But at our present
stage of experimental sophistication, this simple model seems to
handle everything we can understand with satisfactory precision.

So having reduced the problem to one dimension per control
system, we can ask what determines the state of a single input
variable for a single control system. By definition, the input
variable has no way of altering itself. As per Newton's laws, it
changes only when the sum of all effects on it is nonzero.

There are two determinants of an input variable: the sum of all
independent environmental physical effects acting directly on the
variable (of which the variable is a function), and the output of
the control system. As we are dealing only with one-dimensional
variables, this means that no matter how many independent
disturbances there are and by what paths they affect the input
variable, we can always express the result as a single equivalent
disturbance acting through a single equivalent path. This leaves
only two influences on the input variable: "the" disturbance, and
the control system's output. The net disturbing influence is
arbitrary and independent of the operation of the system. All the
variables in the loop, including the output quantity and the
input variable, can be solved for using the closed-loop equations
-- they are all dependent variables. As you note, the reference
signal is also an independent variable relative to the control
loop, and hence relative to the input variable.

Is this the same as saying that the reference signal controls
the perceptions?

Yes, given that the control system is capable of maintaining its
error signal very small. The action of the system will almost
completely cancel the effects of independent environmental
disturbances on the sensory input, and at the same time force the
sensory input to track the varying reference value established by
a varying reference signal.

Originally, the title "Behavior: the control of perception"
gave me the impression that all perceptions are controlled. Now
I understand that there are also uncontrolled perceptions. Is
it therefore "Behavior: the control of SOME perceptions"?

That wouldn't have made a very catchy title, but you're right.
Behavior controls only some perceptions, those that can be
systematically affected by output actions and for which the
organism has reference signals and control systems. The remainder
can be controlled in trivial ways (not looking at the moon keeps
the perception of the moon at zero), but for the most part simply
make up the world within which the things we care about happen.

Now, if some perceptions are controlled and some are not, are
there also intermediates like:

- some perceptions are sometimes controlled, but not at other
times;

- some perceptions are partially or approximately controlled;
- some perceptions are controlled in some degrees of freedom
(dimensions) but not in others?

In this model, an elementary control system does not decide
whether or not to control. It simply controls. If some variables
are controlled only part of the time, the explanation has to be
sought at a higher level in the hierarchical model. As part of a
higher-level control process, a higher-level system may change
which lower-level control systems it is using to control its own
(derived) perceptions. It must have some way, therefore, of
turning lower systems on and off. There are several ways, which
have different implications. But the main thing is that when a
control system is turned on, it controls ALL of the time. It
can't turn itself on or off: something else must do that. That's
just my basic design principle. If the external part of the loop
is lost, the control system will frantically crank up its output
trying to correct the error. It will continue to do this until a
higher-level system notices something amiss and makes the
required adjustments. Rick Marken has shown that when the sign of
the external feedback is reversed, the control system that had
been tracking runs away on an exponential curve -- for about half
a second. The curve closely matches that of the model when
feedback is reversed. Then (according to the model), a higher-
level system reverses the sign of the control system's error or
output connection and it regains control.

As to partial or approximate control, that is only a question of
how well the control system works. There is a complete spectrum
of control ranging from hardly any to very precise. If we make
the reasonable assumption that control systems evolved because it
was in the species' interest to determine for itself how certain
parts of the local environment behave, we can assume that the
less error is allowed by a control system, the greater the
advantage to the organism.

On the other hand, there are specific circumstances in which very
tight control could be a disadvantage -- a waste of energy, for
example, considering the benefit to be gained. You have mentioned
something like this. Once again, my basic design principle
applies. A control system does not decide for itself how well to
control (assuming there is any choice). If its loop gain is
lowered under certain circumstances, a higher-level system is
doing the adjustment of gain, as part of maintaining control of
higher-order perceptions.

A specific example of this appeared in my model of operant
conditioning three or four years ago. One level of control had a
reference signal set by a control system for body weight. The
reference signal specified the level of a perception of short-
term nutritional state that was immediately affected by the rate
at which food was ingested (body weight was a long-term function
of average nutritional input). This short-term state decayed
fairly rapidly with time. The action of the system was to vary
the frequency at which a bar was pressed, producing food input
through a schedule of reinforcement and thus maintaining the
perception of nutritional input level matching its given
reference signal from the weight-control system.

Another higher-level control system, acting at the same time,
compared a cost of bar-pressing proportional to the rate of
pressing with a benefit of nutritional input proportional to the
rate of ingesting food. As the cost rose above the benefit, the
output gain of the bar-pressing system was lowered to keep the
benefit at least as high as the cost. I'm sure you'll recognize
this as a primitive form of optimal control (a one-way control
system in this case).

This model did very well in fitting the bar-pressing behavior of
rats over a wide range of schedules of reinforcement and two
conditions of body-weight (forced by withholding food between
experiments in the real studies).

I was more or less forced into this model, because no matter how
I tried to make the bar-pressing control system vary its own gain
with nutritional input (still remaining an elementary control
system), I could not reproduce the double-valued function
relating the schedule of reinforcement (bar-presses per reward,
which ranged from 1 to 160) to the rate of bar-pressing. Only
when the cost-benefit control system was introduced was I able to
make the curve reverse at the right place. Then the model came
very close to all the data points from the real rats.

Is behavior FULLY in the service of the control of perceptions
or could there also be behavior that is not?

One has to wonder (a) why an organism would learn to produce
behavior that never had any feedback effects on that organism,
and (2) how any organized behavior could reliably be produced, in
a variable environment, without feedback control. My hunch is
that essentially all behaviors (that is, outputs) are learned in
order to control some perception -- that in organisms there is no
open-loop behavior of any significance.

It's possible that evolution might have created some spontaneous
emission of actions without any feedback effects on the organism
doing the acting, as a benefit to the species. But such open-loop
acts would have to be very simple and noncritical, because to
reproduce the effect of any act in a normal environment would be
almost impossible without feedback from the actual effect
created. This is not to say that a feedback control action
couldn't be inherited because of a side-effect it has on other
organisms, with evolutionary consequences. To reproduce that
side-effect in a variable environment, however, the organism
would have to control for the effect of motor acts, not the acts
themselves. There's just too much chaos and interference out
there to make any totally open-loop behavior feasible. When a
peacock spreads his tail, the actual spreading must be a control
process, and perhaps even the subsequent response of a mate is
also controlled for -- but I'm sure that the side-effect of
making more peacocks is NOT a controlled variable.
-------------------------------------------------------------
Peter Cariani (930318) --

Hi, Peter, long time no hear.

I'm afraid I don't see how anyone could call Bohr's approach
"invoking magic", when what he was calling for was the primacy
of (verifiable) results of observations and calculations over
tacit images of the nature of an underlying "Reality".

I don't come down on either Bohr's side or Heisenberg's. Bohr's
view is extreme, and teeters on the edge of solipsism.
Heisenberg's is naive, attributing uncertainty to the wrong
entity. I think control theory gives us a third alternative,
which I'm surprised has not shown up in physics (maybe it has).

We can easily say that our perceptions of reality (read:
instrument readings and interpretations) are a formal system that
we made up ourselves, based only on what we can perceive, not on
any objective "reality." But we don't have to stop there as Bohr
did. We also can act, produce outputs (read: experimental
manipulations). The effects of our actions are related to the
perceptions we get back only in the most indirect way. But such
effects DO OCCUR, even though we can't perceive how our output is
affecting our input. Furthermore, our perceptions often change
when we have performed no act: there are agencies out there.

To me this is a proof of existence: there is a reality out there
and it contains active agents. Unfortunately, we have to guess at
its details -- propose models of what MIGHT be there that would
account for the effects our actions have on our perceptions and
predict new effects of new actions. This guessing game works
extraordinarily well when the demands on models are exacting
enough: namely, that prediction errors should be no worse than
measurement errors. It works so well that one can reasonably
suppose that the resulting models are not inconsistent with what
is really going on. This doesn't mean they're isomorphic to
reality; it means only that something true is captured in them.

An epistemology that is based on observation alone can't lead to
such a conclusion. When you include action in the picture, and
close the loop, something different emerges.

So much of contemporary mathematical physics (and the current
wave of pop-physics pulps), having adopted a platonic-realist
approach, no longer seriously attempts to connect theory with
observation. One of the great intellectual tragedies of the
late 20th century has been this infusion of platonic mysticism
(following Godel, the later Carnap, and Tarski) into
philosophy, the foundations of mathematics, physics,
linguistics, and the cognitive sciences. We are still dealing

with the wreckage.

Platonic mysticism! Bravo. But the other side is anti-platonic
scholasticism, the triumph of pure reason over experiment. The
antidote to both sides is to include action in the picture as
well as perception.
---------------------------------------------------------------
Best to all,

Bill P.

[William_T_Powers2]

[From Bill Powers (930818.1115 MDT)]

Martin Taylor (930817.1710) --

"Fluctuation" is is, for now. More important than what we call
it, I think, is the distinction between disturbance(1) and
disturbance(2) that Allan suggested. If we keep that straight,
everything else should work out.

···

--------------------------------------------------------------
Dag Forssell (930817.1420) --

I like your paper, and the ideas for rearranging it. Good
exposition.
-------------------------------------------------------------
Hal Pepinsky (930817) --

It is human nature to be violent, and human nature to make
peace.

At the recent CSG meeting, Clark McPhail presented a paper on
"the dark side of purpose." McPhail distinguished two kinds of
violence: violence that is an innocent side-effect of a control
system trying to reach or maintain an important goal, and
violence that is a goal in itself (as in professional football or
bash movies). I suspect that both of these boil down to the same
thing: violence as merely a point on a scale, driven ultimately
by some goal other than simply being violent.

With PCT as a guide, I don't think of human nature in terms of
specific things people do, but in terms of the organization that
leads to doing specific things. Specifically, I see it as
revealed in the way people perceive the world, and in their goals
for how they would like to perceive it (including themselves, of
course, in "the world"). There is nothing special about violence
in itself. You have to become violent with a lug-wrench to get a
frozen bolt off a wheel stud. Violence is simply the upper range
of the efforts we always make in the process of controlling what
matters to us.

What calls forth violence between people is the fact that all
people are control systems, organized to produce as much effort
as necessary (within the limits of possibility) to make their
experiences match what they want them to be. When interacting
with the inanimate world (as in loosening the bolts on a car
wheel), human beings simply escalate the applied effort until
nature gives in. But when two people interact with each other as
if they were dealing with nonliving systems, this natural way of
behaving can become a conflict. A conflict exists when it is
impossible for one person to bring perceptions closer to a
desired state without forcing another person's perceptions away
from a desired state. When this occurs, the immediate natural
result is for both people to increase their efforts in the
attempt to correct the respective errors. Unless something
extraneous to the conflict happens to resolve it, the result can
only be two systems applying their maximum possible efforts to
each other, in opposition. That is violence. The party with the
greater physical resources wins, at least for the moment.

This is far from an optimal solution. While a conflict is going
on, neither party is able to achieve its goal in any normal way;
if one person is stronger than the other, the only goal-seeking
efforts available are those left over after subtracting the
efforts of the other person. This means loss of ability to
control, for much smaller disturbances than normal will be able
to cause uncorrectable errors even for the stronger party. The
weaker party has, of course, lost control altogether. And of
course both parties are wasting large amounts of resources simply
in cancelling the efforts of the other. Control is required for
survival; loss of control means loss of the ability to survive.

Understanding how violence arises and how conflict can be
resolved requires considering the hierarchical aspects of human
control systems. When we set a goal, it does not exist in
isolation. We pick a particular goal as a means toward
controlling something else -- as part of a higher-level control
process. What's important here is that conflicts, which arise
from incompatible goals, are not created at the level where they
are expressed. The immediate cause of conflict is a disparity of
goals, but the real cause lies in the higher systems that are
selecting those goals.

A conflict therefore can't be resolved at the level where the
conflict is visible -- where we see the opposing forces clashing.
It can be resolved only at the higher levels which are choosing
those opposed goals for some other more general reason. The
resolution of conflict requires going up a level on the part of
at least one party to the conflict. Somebody has to ask, "Why was
it that I wanted this so much? Is there another way to get it? Is
this way really working? Is the benefit greater than the cost?"
That's looking at the conflict from a level where something can
be done to change a goal (or a perception). At least one party to
the conflict has to reorganize at a level higher than the actual
conflict.

A phenomenon of consciousness:

It seems that awareness is mobile in the human hierarchy of
perception and control. When awareness is not associated with a
given control system, the system goes right on working, but
automatically and in a fixed way. The world that is experienced
consciously is the world as represented at some level in the
hierarchy of perceptions, the level with which awareness is
associated. We project into the world of experience the
interpretations typical of the level from which we are being
aware. Both lower levels and higher levels remain out of
consciousness, working without conscious direction. The lower
levels work automatically; the higher levels work unconsciously.

It seems to be true that we are never aware OF the level FROM
which we are consciously operating. To become aware OF the
operative level is to move up a level, but to be unaware of that
new level while operating from it. One technique for helping
another person to move up a level is to call attention to the
operative level and ask the other to describe it, to characterize
it, to become aware of it in some way. This is known as the
"method of levels" among PCT psychotherapists. "Tell me more
about hating your mother. How does it feel to hate her, what do
you think when you're with her" and so forth. The more detailed
the description, the harder it is to stay focussed in the level
where the hate is expressed, and the more one shifts to the
levels where the hatred is merely a means to something else.
Signs of the higher level will show up as the shift occurs.

The last principle is simply this: reorganization seems to follow
the locus of awareness. To reorganize the systems that are
obeying conflicting goals will do no good; the conflict will
simply change forms, and probably change right back again. The
goals will still be in conflict, because they are not set by the
systems that receive them. Resolving conflict requires
reorganizing at the levels that are setting the goals, so one
must shift to become aware from the levels that are setting them.
And shifting levels will bring reorganization to bear where it
can do some good.

The steps toward conflict resolution are therefore (1) to become
aware of the conflict and the goals that are producing it, (2) to
describe the goals and perceptions involved in as much detail as
necessary to cause one (3) to move the center of awareness into
the higher systems that are setting the goals, focussing
reorganization where it can be effective. If one step doesn't do
the trick, do it again, and again, until the level is reached
where a change will be possible and effective.
                       * * * * *
Hal says:

Where we differ, I gather, is in feeling that any occasion
demands violent rather than Golden-Rule response. You
apparently feel others' violence may demand one's deviation
from the Golden Rule. I certainly acknowledge that I DO
respond violently, that I do treat others with a disrespect I
would never seek for myself. But the question I keep asking--
my research question if you will--is what do we get for
choosing one way or the other.

When I say that the Golden Rule may not always work, I do not
mean that violence will work instead. The Golden Rule is not the
only alternative to violence.

Among my understandings of the principle of respect for the will
of others is the understanding that I can't direct the
reorganizations of another person. I can call attention to things
in a way that might facilitate the other in seeing a conflict
from a higher level, and that might lead to reorganization at
that level. But I can't predict what that reorganization will
create, or that I will like it, or that the other will like it.
All I can do is to be willing for both myself and the other to
reorganize in whatever way makes the interaction work
differently, and to accept the result if I can.

We always have to start from where we are. Respect for the will
of others includes the right to assume that respect from others.
When we are in situations of conflict that we don't know how to
resolve, the conflict will simply continue until we do know how
to resolve it. There is no a priori reason for one party rather
than the other to submit to loss of control. Often one person who
understands about levels can dissolve a conflict one-sidedly, by
deciding that it's not important to win, in terms of higher-level
goals. But sometimes one is forced into conflict by a genuine and
unpostponable threat to survival. In that case one simply does
what is necessary to survive, or decides not to survive. The
decision not to survive has ramifications that go far beyond the
immediate situation; it is seldom chosen by any sane person, and
is not a workable general principle. So we do what is necessary.
We protect ourselves, so that we can try again for the right
solution another day.

In the case of clear and present danger, what you choose to do is
almost irrelevant. What you *will* do comes out of all your
goals, conscious and unconscious, at levels both higher and lower
than the levels where you customarily reside. The best you can do
against that time is to explore all of your hierarchy of goals,
including the forgotten but still quite active ones, so that when
you do find yourself acting, it will be as one person and not a
conflicted rabble.

                       * * * * *

Thpoughts a little off your track, but I hope interesting.

---------------------------------------------------------------
Bob Clark (930917.1700) --

A nice summary of the portable demos. I think that the latter
ones need more work, however. As you recall from the hasty
attempt to demonstrate them at the end of the conference, the
participant doesn't always do what is expected. There are some
loopholes that need to be fixed. I think it's important for these
demos to work exactly as predicted EVERY TIME and for EVERY
PARTICIPANT. The lower-level ones do. Getting the higher-level
demos to work that well may require altering details to eliminate
the unwanted alternative solutions that people think up. I think
we ought not put on any demo in public until we are
overwhelmingly confident about what will happen. Otherwise we'll
be sliding back into the world of ordinary psychology where
theories have to work only once or twice a day.
--------------------------------------------------------------
Avery.Andrews (930818) --

Yeah, nice note about the beavers and the loudspeakers. That's
behavior that works directly on a perception! Whoever thought of
using the loudspeakers was really doing the test for the
controlled variable. There must be tons of that sort of stuff
lying around in the literature undigested -- like Tom Bourbon's
wonderful report on Kavanaugh's rats (mice?).
-------------------------------------------------------------
Back to work.

Best,

Bill P.

[William_T_Powers2]

[From Bill Powers (921002.0600)]
Postcard:
Back from a trip with Mary that stitched back and forth and up and
down New Mexico for four days. Saw the Very Large Array again, and one
radio telescope of the Very Long Baseline Array in Pietown, NM. Saw
them both move this time. Visited the observatory I designed in Las
Cruces -- locked up with nobody around, as it's run by amateur
astronomers now, but it's still a beautiful setting. No signs of
Georgia O'Keefe at Abiquiu, but you can see her colors everywhere in
the Chama River valley. Camping at the end of September gets a little
chilly at night. There's a lot of geology in New Mexico and you can
usually see about 40 miles in any direction, on the ground, because
the valleys all rise at the edges. Interesting place where we humans
live.

···

------------------------------------------------------------------
Dag Forssell (920927-1) --

Added suggestions for System Concept level: entities, organizations,
persons, realities, disciplines.
------------------------------------------------------------------
Cliff Joslyn (920922.1900) --
RE: Turing Test and intelligence.

The TT helps us understand when we're PERCEIVING intelligence. But
like redness, that perception is intuitive, and automatic; and
frequently faulty.

It's also parochial -- that is, academics tend to rate verbal skills
high, while others consider a preoccupation with words (as opposed to,
say, financial manipulations) rather stupid. I don't think the TT
helps us to understand anything but the problem, which is that a word
like "intelligence" can't possibly wrap up all the dimensions of human
Being.
-------------------------------------------------------------------
Greg Williams (920928) --

I agree with

1. A disturbs particular perceptions being controlled by B so that
B compensates for the disturbances with actions which A wants to
perceive.

2. A arranges B's environment so that when B controls for
particular perceptions, A perceives what he/she wants to perceive.

but I have a problem with

3. A arranges B's environment so as to trigger learning
/reorganization in B's control system resulting in actions which A
wants to perceive.

4. A applies physical constraints or threatens to apply physical
constraints to B so that B's actions are as A wants to perceive.

According to my model of what triggers reorganization, these would
both mean arranging the environment so that B suffers critical error
(you notice my return to Ashby's term) such as hunger, thirst, pain,
illness, suffocation, "stimulus deprivation," or whatever you want to
put on the list. By definition, reorganization is unsystematic. This
means that you can't predict what behavior will be used to correct the
error unless you have removed all means of correcting it but one,
which is within B's capacity to learn. That's easy to do with a lower
animal or a child, but hard to do with an adult human being. You do
note that these methods involve conflict, but you don't mention that
the outcome is largely unpredictable because reorganization is
involved.

In the second section, I don't understand

3. Onset of learning/reorganization at a particular time is a
function of reference signals, input/output functions, and
environmental disturbances at that time.

4. The path of learning/reorganization is a function of (possibly
randomly generated) successive sets of changes in reference signals
and/or input/output functions ...

It seems that you're allowing for systematic reorganization here. Why
isn't that just the operation of a higher level control system, which
itself has to be learned?

5. Whether or not the criteria for ceasing learning/reorganization
are met by the reference signals and/or input/output functions at
any point on the path of learning/reorganization is a function of
reference signals, input/output functions, and environmental
disturbances at that point.

Are you proposing some particular mechanism here, or is this just in
general? What is it that judges whether the criteria are met? How does
that judgment affect the continuation or cessation of reorganization?
I can't visualize the arrangement you're talking about. Can you boil
it down to a specific model?

6. At any time, the criteria for ceasing learning/reorganization
are functions of reference signals and input/output functions at
that time.

The CRITERIA are functions of reference signal and input/output
functions? Now I'm thoroughly confused. How do these things affect the
criteria? Do you have a mechanism in mind?

The main motive for my simple model of reorganization was a need to
explain how animals learn such things as walking in a figure eight to
get food -- situations where what is learned has no necessary or a
priori connection with the reasons that it has to be learned. I then
realized that NOTHING has a neccessary or a priori connection with the
need for learning a specific behavior. Even learning to eat certain
items having particular appearances or smells or tastes has nothing
guaranteed to do with assuaging hunger or correcting the underlying
nutritional state. So that's where random reorganization based on
critical errors came from.

It seems to me that you're proposing something different here. What is
it? What phenomena of learning does it explain that my version of
reorganization doesn't explain? And how does your explanation work?
-------------------------------------------------------------------
Martin Taylor (920927.1830) --

I said:

The "impedance" concept is sort of ingenious, but I can't see how
to model it so that a particular output from a control system
would be spread out among all the different possibilities -- what
would keep all of them from trying to happen at once?

You said:

Well, I must have misinterpreted you some months or more ago,
because I thought it was your idea. It's "always" been a part of
my concept of PCT. Anyway, if it wasn't your idea, here's how I
see it. They all DO try to happen at once, but they can't. They
inhibit one another.

I did speak of impedance-matching in a different context, having to do
with power gain. But the idea of "competing behaviors" inhibiting each
other doesn't seem very plausible to me. How does one behaving
subsystem know which other behaviors to inhibit and which to leave
alone? I don't think you could make a runnable model out of this, but
you're welcome to prove me wrong. Your idea isn't even self-
consistent, because you follow the above by saying

It is the world that stops them all happening at once. If they
could, they would.

But that means that they are NOT mutually inhibiting each other inside
the system. They're all in a state of perpetual conflict. This seems
like a lousy design for a control-system hierarchy where you would
like the loop gains all to be quite high.

Outputs are going in all directions, but the world prohibits some
percepts from actually being controlled. So the "taking bicycle"
percept/reference cannot be satisfied if you are sitting in the
car.

So when you're driving along in the car, you're wishing you could be
riding your bicycle, and vice versa? And how could your outputs be
going in all directions while you're using them to drive a car?

I guess in the background there is another point I meant to discuss
at some time--"giving up." When there is persistent error, one
possible and often used response of an ECS is simply to reduce its
gain to zero, to give up on a hopeless situation.

I have already brought up this idea on the net, but not by supposing
that every ECS can perceive and judge "situations" as being
"hopeless." My solution was a comparator that has a curve that I drew
like this:

     ^ * |
     > * * |
Error>signal * * |
            * * |
* - *| +
-------------------------------*-----------------------------------
         <-- actual error--> |* *
                               > * *
                               > * *
                               > * *
                               > *

In the central region, feedback is negative. When a disturbance gets
so large that the peak of the curve is reached (in either direction),
the error signal begins to fall for further increases in disturbance.
This results in a drop of output, still more error, still less error
signal, and so on to the "giving up" regions at the ends. I had
previously proposed this as a possible explanation of operant
conditioning under conditions of high deprivation. The giving up
process is reversible; if the disturbance falls, the error will drop
and the error signal will rise, producing more output, until the
system passes over the peak of the curve and snaps into the region of
negative feedback control again (the whole system may be unstable in a
region near the peaks, the size of the region depending on the loop
gain).

-------------------------------------------------------------------

When two ECSs are in conflict, it is impossible for both to bring
their errors to zero simultaneously. That's almost a definition of
conflict.

That's the definition I've always recommended and used.

If one ECS, for example, has a reference to perceive the body as
bicycling to work, and the other has a reference to perceive it as
driving, there is no compromise intermediate position (possibly a
motorcycle, but let's suppose none is at hand). The mechanism is
that one ECS gives up attampts to control. Either the person
cycles, or the person drives, but not part of each.

Unless you use something equivalent to my comparator function above,
you will require that an ECS do something other than control its own
perception. It seems simpler to me to invoke a higher-level system
that selects one of several means of locomotion (the kind of
explanation you used for explaining why the wastebasket wasn't moved
on the third day). This has the advantage of not pitting the outputs
of high-gain (competent) control systems against each other, which
always leads to a lessening of the capacity to control if not its
complete destruction. I don't think you're taking full advantage of
the concept of hierarchical control, at least not consistently.

My claim is that The Test is always ambiguous. P can tell that Q is
controlling for a percept that incorporates some CEV that P has
disturbed, but P can never tell that the percept Q is controlling
for corresponds exactly to the CEV that P's percept corresponds to.

One application of one hypothesis leaves the Test ambiguous. There is
nothing to prevent you, however, from trying out more hypotheses aimed
at reducing the ambiguity (as you pointed out to your colleague). I
don't think there is any more ambiguity in applying the Test than
there is in guessing at the causes of any natural phenomenon. A single
measurement is always ambiguous. One employs strategies aimed at
eliminating alternative hypotheses until no more alternatives can be
found. Then you go with what remains. Your statement that you then
resort to statistics is gratuitous; you may or you may not, depending
on how unclear the hypothesis is when you're done and whether you
happen to like statistics. Lots of people will think up one
explanation on the basis of the flimsiest evidence and assume it's
absolutely correct. They wouldn't do any better using the Test.
-----------------------------------------------------------------

RE: applying disturbances (stuck phonograph record division):

Yes, ideally any observer should avoid disturbing the thing
observed. But any disturbance to a controlled variable causes
error in the controller, even if momentarily. The tester is
inevitably controlling for perceiving a change in a variable, or a
resistance to change (if the circumstances are appropriate).

The disturber should NOT control for a visible change in the variable
being disturbed. That simply creates conflict. What the disturber
should do (where possible) is alter some OTHER variable that is
loosely coupled to the putative controlled variable. This will elicit
an opposing change in the controller's actions even if the controlled
variable doesn't visibly change. In low-gain situations (like the coin
game) this doesn't matter so much. But when the control system
involved is a very good one, insistence on seeing the controlled
variable actually change will result in applying very large forces to
the controlled variable, with a probable change in what variables the
controller is controlling. The Test is most accurate when the
controlled variable doesn't change at all (that you can see).
----------------------------------------------------------------
Dennis Delprato (920929) --

Request for lab experiments noted. In the queue.

----------------------------------------------------------------



[_Martin_Taylor]

[Martin Taylor 921002 22:30]
(Bill Powers 921002.0600)

But the idea of "competing behaviors" inhibiting each
other doesn't seem very plausible to me. How does one behaving
subsystem know which other behaviors to inhibit and which to leave
alone? I don't think you could make a runnable model out of this, but
you're welcome to prove me wrong. Your idea isn't even self-
consistent, because you follow the above by saying

It is the world that stops them all happening at once. If they
could, they would.

But that means that they are NOT mutually inhibiting each other inside
the system. They're all in a state of perpetual conflict. This seems
like a lousy design for a control-system hierarchy where you would
like the loop gains all to be quite high.

Quite right. I never intended to imply that the competing behaviours had
any connection internal to the hierarchy other than that they all had
references to which a higher-level ECS contributed.
They certainly should not inhibit one another internal to the hierarchy,
unless we are getting into the realm of pre-planning, which was the kind
of choice behaviour I was trying to show was not always required (if ever).

I see you still don't buy the notion that not all controllable percepts are
at any one moment controlled, and that most ECS gains either are very low
or the ECS is in some way disconnected from the physical world. It still
seems to me inevitable that there exists a very small actively controlling
subset of the ECSs at any one level at any moment, the rest not, at that
moment being in active control of anything except their imaginations. The
idea of "giving up" is inherent in the normal situation.

I guess in the background there is another point I meant to discuss
at some time--"giving up." When there is persistent error, one
possible and often used response of an ECS is simply to reduce its
gain to zero, to give up on a hopeless situation.

I have already brought up this idea on the net, but not by supposing
that every ECS can perceive and judge "situations" as being
"hopeless." My solution was a comparator that has a curve that I drew
like this:

    ^ * |
    > * * |
Error>signal * * |
           * * |
* - *| +
-------------------------------*-----------------------------------
        <-- actual error--> |* *
                              > * *
                              > * *
                              > * *
                              > *

In the central region, feedback is negative. When a disturbance gets
so large that the peak of the curve is reached (in either direction),
the error signal begins to fall for further increases in disturbance.
This results in a drop of output, still more error, still less error
signal, and so on to the "giving up" regions at the ends.

This is a different situation, in itself quite legitimate. In the situation
I am considering, the error need not be desperate, but it is consistently
in the same direction. If the gain depends on the integrated error, I guess
your curve would deal with the same situation, but the way I see it, the
evolved solution should be expected to contain a switch. It would, as you
say, be a lousy way to build a control hierarchy if every unachievable
reference led to a continuous "kicking against the pricks." And in your
system, it looks as if what would happen in the car-bicycle-walk scenario
is that when the destination was near (and the higher reference-percept
difference small) the gains of the unselected transport-medium-percept
controllers would increase, giving rise to renewed conflict.

It seems simpler to me to invoke a higher-level system
that selects one of several means of locomotion (the kind of
explanation you used for explaining why the wastebasket wasn't moved
on the third day). This has the advantage of not pitting the outputs
of high-gain (competent) control systems against each other, which
always leads to a lessening of the capacity to control if not its
complete destruction. I don't think you're taking full advantage of
the concept of hierarchical control, at least not consistently.

I thought I was pointing out another instance of the power of hierarchic
control. I am perfectly willing to believe that I'm not taking full advantage
of its power, since I keep finding out new aspects of that power, and I
doubt that discovery process has come to an end. But putting the onus of
selection onto a combination of the relative insistences of the "competing"
ECSs and the impedances of the relevant (mutually inhibitory) CEVs seems
to me to be consistent with a lot of feelings one has when actually making
that decision, besides not requiring any special mechanism of choice in
the higher-level ECS. If the bike is handy and the distance not too far,
one might use it, whereas if the bike were hidden behind piles of junk that
had to be moved, one might not. On the other side of the person-world
interface, if there were a reference level for keeping fit and using the
bike would reduce that error as well as the error of not being at the
destination place, then one might take the bike even if it were hidden
by a pile of junk.

Perhaps it is worth noting that the concept of "impedance" applies not only
to the CEVs of the world, but also to the reference-percept relation of a
single ECS as seen from a higher level ECS that contributes to its reference
signal and receives its perceptual signal. That impedance will be very low
if the lower ECS has a high gain and the CEV it is controlling has a low
impedance. And it will be near zero if the lower ECS is imagining. Conflict
at the lower level may well lead to the higher ECS seeing the impedance of
the lower as being rather high. And for good measure, one should mention the
possibility of negative impedance (which implies positive feedback, and
trouble).

The world may well be much more closely coupled than the perceptual control
hierarchy is. I think that may be what people mean by "leaning on the world."

Martin

[_Martin_Taylor]

(Martin Taylor 921009 1800]
(Bill Powers 921002.0600)

On The Test

RE: applying disturbances (stuck phonograph record division):

....

The disturber should NOT control for a visible change in the variable
being disturbed. That simply creates conflict. What the disturber
should do (where possible) is alter some OTHER variable that is
loosely coupled to the putative controlled variable. This will elicit
an opposing change in the controller's actions even if the controlled
variable doesn't visibly change. In low-gain situations (like the coin
game) this doesn't matter so much. But when the control system
involved is a very good one, insistence on seeing the controlled
variable actually change will result in applying very large forces to
the controlled variable, with a probable change in what variables the
controller is controlling. The Test is most accurate when the
controlled variable doesn't change at all (that you can see).

I wouldn't have thought it made any difference whether the observer affected
the supposed controlled CEV directly or through a "loosely coupled variable."
And I wouldn't have thought that actually seeing a visible effect of the
disturbance on the test CEV was important. If you do, the subject probably
isn't controlling it with any substantial gain. What the observer should
observe is a failure of the test CEV to change as much as it should have
been expected to do, given the magnitude of the applied disturbance. If
the observer wants to discover the limits of the subject's control, then
what you talk about will presumably come to pass--large forces and a change
in the subject's behaviour (if not organization).

Martin