Selection; quit

[From Bill Powers (931004.1000 MDT)]

Gary Cziko (931003.1600 GMT) --

But wasn't Avery Andrews saying, in effect, that there are TWO
different types of side effects? There are the "true side
effects" which are neither controlled by the organism nor do
they have any role in contributing to the organism's
evolutionary fitness.

Right, a good distinction. Maybe "relevant" and "irrelevant"
(with respect to stated criteria) would provide the

However, in neither case would we generally expect the side-
effect itself to be under control. The relevant side-effect, as
Rick Marken points out today, has to be fairly reliably dependent
on the variables that the organism learns to control. A pretty
reliable side-effect of eating things is to provide necessary
protein, carbohydrate, vitamins, and minerals to the organism,
but the organism doesn't learn to control for these things. It
learns to control for color, taste, smell, consistency, and so on
-- but not for what those things indirectly signify.

If we see that termite eggs left exposed to the sun
shrivel up in a few hours, can't we assume that the nest
provides important protection from the sun?

Yes, but it's doubtful that termites control for a perception
called "important protection from the sun." Maybe they inherit a
reference level for seeing eggs in a cool damp dark place, and
simply move stuff around until that's what they see. Of course
their ability to reorganize is probably limited, so they also
have to inherit control systems for doing those manipulations.

If we see an archer fish dropping insects into
the water with its built-in squirt gun and then gobbling them
up, isn't it clear that this behavior important for its

What the archer fish would inherit, perhaps, is a reference level
for bringing an image of an insect to a certain point in its
visual field (perhaps relative to an image of its shooter), and
then giving a sharp SQUEEZE that it can feel. I don't think the
archer fish would have the same concepts of trajectories,
gravity, dropping, water, etc. that we have. It probably doesn't
have the same concept of sequence that we have. If it saw a bug
drop from a leaf into the water, would it gobble it up even if it
hadn't shot at it? Or will it gobble the bug only if the bug
falls immediately after the shot? If you put up a sheet of glass
that intercepts the projectile of water, but arrange for the
insect to drop into the water anyway, will the fish THEN eat the
insect? Lots of opportunities for the Test.

... "puzzles of fit." These are often quite obvious (e.g, the
airfoil shape of a bird's wing). I don't see why The Test
would be always necessary to find these.

Without the Test, or at least the attitude required in applying
it, there is no discipline on the brain's ability to find
patterns. The airfoil shape of a bird's wing is probably not the
critical survival factor; what really matters are LIFT and DRAG.
Both of these are highly adjustable in a bird's wing. A bird's
wing is not in fact an airfoil of fixed cross-section, but a
surface that has curvature and aerodynamics that are adjustable
by use of the bird's muscles. The most important thing that is
inherited, one might say, are the control systems by which the
bird can vary the form of its wings to control lift and drag.
Wings can vary enormously in configuration, but with the right
control systems can be employed for flying. You don't have to cut
the wings off a bird to keep it from flying: just remove the
"flight feathers," the ones that are adjustable as a means of
controlling for flight. Without those, the basic configuration of
the wing is useless for flight. Even with them, flight can't
occur without the right control systems.

Your remarks on the human advantage in controlling for higher-
order variables are most interesting. The fossil record leaves no
trace of the quality or extent of the control systems that the
organism had. We say that cockroaches haven't evolved for 500
million years, but that may be entirely untrue. Today's
cockroaches might be enormously advanced over the original ones
in every way but their physical forms. A few frozen footprints
are all we have to indicate the control skills of early

I think that in many cases the survival factor involved is
quite clear.

But we can still be quite wrong about them.

I notice that when flounders come to rest on the ocean floor,
they dig into the sand a bit with only their eyes protruding
above the sand.

If the digging movements were inherited, we should observe that
the depth of digging depends on the consistency of the sand. If
the inherited reference level is for "sand felt all over the body
up to just below the eyes," then the digging behavior would VARY
depending on the kind of sand, the current, and other factors
that would change the consequences of particular digging
movements. Rick suggested some other tests.

I would bet sand dollars to sandy doughnuts that evolution is
"selecting" flounders which are well camouflaged against
predators (flounders who flounder are eaten).

That is the significant side effect, but not what flounders
control for. I wager that a flounder knows nothing of
"camouflage." That is a human conception, and a French one at
that. Flounders that are routinely eaten reproduce poorly,
leaving mainly those whose behavior protects them from being
eaten so often. If flounders understand that digging into the
sand gives them protection against predators, then flounders are
a lot smarter than I've been given them credit for. I don't think
that evolution understands that, either. In fact, nobody but a
human observer understands that, and what the human observer
understands is completely irrelevant to what happens.


Hal Pepinski (931003c) --

Good thing you came up with that "woops."

I have just read the phrase that "behavior is the control of
perception." Is it fair for me to be reminded of the problem I
have had with the sociological debate over "conflict vs.
consensus." To say that one causes the other is like saying
that a foot causes a footprint in the snow any more than the
snow does.

It's not hard to demonstrate that behavior (action) controls
perception but not vice versa. If perception controlled behavior,
then varying the perceptual input would cause behavior to change
in some predictable way. This model has been extensively
investigated, and it has been found that there is no way to
predict actions simply from knowing perceptual inputs. At best,
weak and fuzzy statistical relationships are found.

On the other hand, if behavior controlled perception, then we
would find that disturbances tending to alter perceptions would
be met by actions strongly and systematically tending to cancel
the effects of the disturbances. This relationship is in fact
observed, and is highly reliable and reproducible (orders of
magnitude more reliable than the statistical effects associated
with supposed control of actions by perceptions).

The reason for this assymetry is not hard to find, although it is
somewhat technical. The organism-environment relationship would
be symmetrical if the amplification in the closed loop were
equally distributed around the loop. In fact, the amplification
involved is concentrated almost exclusively inside the organism.
The environment part of the loop almost always introduces
_attentuation_. This places the controlling agency inside the

So the choice as to which controls which is not arbitrary; it is
forced by the properties of organism and environment.

... whether control of perception as behavior is worth studying
in its own right doesn't catch my attention nearly as much as
the underlying linearity, materialism, definition of conflict
as failure of control, and failure to link my control to yours.

"Linearity" is commonly used in two senses. The popular meaning
is that of straight-line causality: A causes B causes C and so on
(we call such systems "lineal"). The mathematical meaning is that
B depends on A in a way that can be described by a linear
equation: an equation in which B enters only as the first power
(no squares, cubes, transcendental functions, and so forth). PCT
is not linear in either sense. In what sense are you using the

"Materialism", I suppose, means something like "mechanism," the
idea that subjective experience is connected intimately to
processes in a physical brain and body. My aim in developing PCT
was to explain as much of behavior as could honestly be explained
in terms of physical brain/body processes, partly just to come up
with a decent model of behavior and partly to see what would
remain unaccounted for by an exhaustive physical model. I see the
PCT model as describing the interface between nonphysical aspects
of human experience such as awareness and volition and aspects
adequately handled by a physical model such as thought,
perception, and action.

Conflict is defined not as a "failure of control" but as the
consequence of successful control systems trying to control under
conditions that make success by one system generate failure by
the other. A simple example is two people trying to control for
enough to eat when there is only enough food for one to survive.
Most examples aren't that discouraging; in most conflict
situations one or both systems can adjust perceptions, goals, and
actions in a way that restores successful control to both. This
requires reorganization on the part of one or both systems.

I believe that your approach to conflict focusses on the
reorganizational aspects of resolving conflicts rather than on
explaining how it is that conflicts arise in the first place. If
reorganization were going on in a group of people having
conflicts with each other, something would happen much like what
you describe. All the parties' goals, perceptions, and actions
would begin to shift, the whole assemblage moving toward a state
of less conflict and more harmonious interactions, and each
individual experiencing a change of world-views and intentions
under influences from the others. The ideal outcome would be
complete resolution of all conflicts so that the group could
function with each person feeling freedom to act to achieve
personal goals without interference from, or interfering with,
any other person. Each person would also be supported maximally
by the actions of others, for control in social situations often
invites and depends upon the help of others.

PCT as a theory of control, without considering reorganization,
explains how behavior works between reorganizations. Adding the
reorganizational aspects of the model is an attempt to represent
what goes on while the characteristics of control systems are

From later post:

Any model of behavior can be made to account for all behavior.

I don't think so. That's only true if you think of a model as a
generalized description at a level of abstraction where
differences between behaviors disappear. Get Gary Cziko to show
you Demo 1. It contains many examples of behavior that, as far as
I know, can be explained ONLY by control theory.

I'm just suggesting your models don't account for the
interaction, which for me, as one little for instance,
distinguishes what comes out as a chat over someone's request
for money on the street from a mugging.

We have been able to model some very simple interactions between
real and hypothetical people. This is done not by modeling the
interactions per se, but by modeling the people and then letting
the models interact to see what the outcome will be. Tom Bourbon
has done this with real people. In the simple situation he uses,
the models interact with each other as the people do (not just
qualitatively but quantitatively with good precision from moment
to moment). He can also predict what happens when one participant
is a person and the other is a model.

If an interaction can be accurately represented by models that
define only the individuals involved and the links between them,
this goes a long way toward proving that the outcomes of
interactions are completely determined by the properties of the
individuals who are interacting, and the nature of their effects
on each other.

The day when we can simulate complex human interactions is a long
way off. We may never need to do so; real people are their own
best simulations.

My model accounts for the transition that occurs when Rabin and
Arafat switch from failing to recognize one another to
accommodation, as another for instance. My model as another
for instance leads me to predict that in a culture in which
adults make a point of engaging children in conversation at
adult party dinner tables, incarceration rates will be lower
than where adults instead ignore or talk about the children.

What you are calling a model is a set of generalizations that
when certain classes of relationships exist (representable
geometrically or otherwise), people's interactions will turn out
to be of various types. For all I know, these generalizations are
completely valid and useful. However, they ARE generalizations,
and they make room for many counterexamples.

For example, you say that incarceration rates are lower where
adults do not ignore or talk about children, but converse with
them. You did not say that if an adult engages in conversation
with a child, that child will never be incarcerated. In fact,
you're talking about a statistical generalization. Under this
generalization, it is almost a certainty that there are many
cases in which an adult converses with a child and avoids
ignoring or talking about the child, and the child very quickly
becomes incarcerated. All you know is that on the average, adults
who behave in a certain way seem to have an effect of reducing
the average incarceration rate of children, a statistical effect.
If you were presented with an individual case in which the
generalization failed to hold true, this would not be critical;
all you need is for more people to fit the hypothesis than fail
to fit it.

This is not the sort of model that PCT is. PCT does not say that
on the average, most of the time, people tend to behave so as to
control their perceptions. It says that all people, all the time
and under all circumstances, behave to control perceptions and
for no other reason, ever. PCT applies to every individual case.
If any case is found in which any behavior is NOT organized
around the control of some specific perception, then PCT must be
revised. So far in every case where we have been able to subject
PCT to a formal experimental test, the model has correctly
predicted the behavior of every individual tested, with excellent
(if not perfect) accuracy. The first failure might occur
tomorrow. But so far there have been no failures.

That is one reason why PCT, in its current state of development,
can't make many interesting statements about complex behavior.
Before we can make such statements we have to learn how to apply
the model quantitatively, and that means we must build up the
model level by level until we know enough about such
applications. If we continue to maintain the same standard,
requiring accurate predictions for every subject on every test,
we will eventually arrive at the higher levels very well equipped
to do the job -- considerable better equipped than we would be if
we simply wanted statistical generalizations that hold true in
some cases but fail in others.
Chuck Tucker (931001 et seq) --

The best course I took in college was from a physics professor
trying to find some rational way to handle grades. This teacher
based the grade in the course on four tests during the quarter.
But tests were given far more often, sometimes every day. Most of
these tests were returned with the errors noted, but were not
given a grade and did not count toward the course grade. The
professor told us not to consider these extra tests a burden, but
a service. Their only purpose was to inform us how we were doing.
If we got the right answers to most of the questions, we could
take it that we had understood the course material to that point.
Where we missed the answers, we could use that information as a
way of knowing what we had to study more thoroughly. That was the
only purpose of the extra tests.

I hate to think of how much time and labor this professor put
into making up those dozens of tests and marking them with
conscientious comments. I also wonder whether he had forecast one
of the main side-effects of this policy, which was to remove
completely any sense of threat from the tests that counted. I
think all of us came to look forward to these little quizzes,
because they really made it possible to know how well we were
grasping the course material. By the time the real tests came
along, we knew what we knew, and we could predict how well we
would do in the critical tests. The biggest problem for the
professor was in how to grade the tests that counted, because the
class average was in the nineties. He ended up giving all A's and
B's because he had the documentation to show that they were
Avery Andrews (931095.0830) -- RE: dormitive principle.

In many cases you can't have a real science of something (i.e.
good explanations), until the `natural history' is done right.

I agree. I'm not objecting to observing phenomena and developing
taxonomies to put them into order. Neither was Bateson; he keep
criticizing cyberneticists for not paying enough attention to

My objection is to offering explanations that are not
explanations. Even B. F. Skinner had this same objection to what
psychologists were calling explanations. An intervening variable
that has the sole purpose of explaining one phenomenon is not an
explanation. It's explaining X by envoking nothing more than
"whatever cause is required to explain X." This cause can be
given a name like guilt or thirst or propensity to purchase that
makes it sound separate from that which is to be explained, but
in the final analysis such names simply assert that there must be
a cause, without giving the cause any properties but that of
producing the phenomenon in question.

If we don't develop a sensitivity to dormitive principles, we
will think we have explained a phenomenon when all we've done is
to think up another name for it. And thinking that we have
explained it, we'll stop looking for a real explanation.
Best to all,

Bill P.