Draft for joint paper from PCTers

[From Bill Powers (930501.1010)]

The following paper is a first draft, the final product to be
submitted to an unselected journal. I am inviting comments,
revisions, deletions, additions, rewordings, new concepts, or
anything else that will make it acceptable to readers on this
net. I particularly want these things from people who will be
willing to become co-authors of the paper. I would like this
paper to appear with many authors from many disciplines in many
countries around the world: the more the better.

I will collect all suggested changes into a new version and
circulate it again, until everyone who wants to be a coauthor
feels willing to stand behind what the paper says and how it
says it. Then we can decide on the venue for publication.

···

================================================================

      Perceptual Control Theory: an idea before its time?

For the past 40 years, an idea currently termed perceptual
control theory, or PCT, has been developing out of the same roots
that gave birth to Cybernetics. During most of that time, more or
less continuous efforts have been made to introduce it to the
world of psychology through papers in the refereed literature,
seminars and colloquia, and even university courses given by a
few daring souls. Control theory applied to living systems has
appeared as the major topic in a number of books over these years
quite independently of PCT; some of its basic concepts have been
reinvented or adopted in part by many individuals working in
parallel with the line of development that is here called PCT.
Methods and concepts which are fundamental to PCT have existed in
the behavioral sciences since the 19th Century, although never
gathered into a systematic theory or methodology of the kind
found in PCT.

The concepts embodied in PCT have gradually been diffusing
throughout the life sciences. There is, however, a strong
counterpressure resisting this diffusion. The form taken by the
counterpressure is interesting, because it has nothing to do with
the merits or substance of PCT.

Most criticisms of PCT that have appeared in the literature are
based on almost complete ignorance of what it is about, and of
the experiments and mathematical analyses on which it rests. It
is as though these critics have looked into control theory just
far enough to find that it disagrees with their own common sense
and their own theories, but not far enough to grasp the actual
properties of control systems. So the criticisms are based not on
the way control systems actually operate, but on the way the
critics imagine they must operate. Yet not being acquainted with
control theory, they are not equipped to imagine correctly. They
are trying to understand a closed-loop system using only the
intellectual tools appropriate to understanding straight-line
causation. Those tools are not up to the job.

There is an intellectual threshold effect here: a little
knowledge is more dangerous than total ignorance. To grasp the
logic of control, one must invest some considerable effort in the
beginning, because normal causal concepts are simply inadequate
for an understanding a closed loop of causation. If one
approaches control theory from any conventional start, the
immediate result is repeated invalidation of one's expectations.
Cause and effect seem to work backward in control systems;
outcomes seem to determine their own causes. Until the details of
operation of control systems are fully understood, this conflict
with normal understanding seems an impassible barrier, and one
that a person sees little reason to surmount. From the standpoint
of the concepts of causation which most scientists have learned,
control systems don't seem to understand the rules of nature:
they behave perversely. If the sensivitity of a sensory receptor
is doubled, normal reasoning would predict a doubling -- or at
least a substantial increase -- in the output of that sensory
receptor. If the receptor is part of a control system, however,
the output will remain the same while the input is cut in half.
To a person versed in control theory, this is completely
understandable, but in conventional terms it makes no sense.

Those in the life sciences who have managed to grasp the basic
ideas -- and by now they probably number several hundreds --
would like very much to know what is preventing their colleagues
from getting past this intellectual barrier. It is, to say the
least, inconvenient to find oneself the only person in a
department or even a university who finds the concepts
interesting or who wants to devote time and resources to
developing them further. It is even more inconvenient to have
papers rejected for reasons having nothing to do with their
contents. And beyond inconvenience, it is alarming and
discouraging to find that one's own beloved discipline has
apparently lost the ability to handle a new idea in a normal
scientific manner.

All this is by way of saying that there is something unusual in
the response of the scientific community to PCT. We wish to
suggest an explanation in this essay: it is the predictable
response to the early stages of a scientific revolution. Let us
spend a few pages on the nature of this revolution.

                        What PCT says

A logical development of PCT would begin with its engineering
roots and the methods developed by engineers for analyzing
systems in which effects of outputs on inputs complete a closed
causal loop. For present purposes, however, a more direct
approach can be taken by focusing on one simple question: how
does behavior actually work? It is easy to show that the
assumptions about behavior on which most conventional approaches
rest are wrong, and that PCT provides the only present way of
explaining behavior when the erroneous assumptions are corrected.
Behind this explanation there is a structure of concepts that
rests rigorously on first principles, but without going that far
we can still see why PCT must be taken seriously.

In nearly every conventional explanation of human and higher
animal behavior there is one causal link that is uniformly
assumed: the link between the brain's activities and the global
regularities that we call the behavior of the organism. This link
is common to theories that, generically, give the environment the
ultimate controlling influence on behavior, and to those that
generically attribute behavior to cognitive, emotional,
dispositional, or hereditary processes inside the organism (and,
of course, combinations of the two generic types). If there is in
fact no regular link between the outputs of the brain and the
patterns we call behavior, then there is a fundamental flaw in
both of these generic explanations and all their permutations.

William James pointed out nearly 100 years ago that this link is
not in fact regular. He observed that the hallmark of the living
organism is that it reaches consistent ends by variable means.
This ability is not simply (as B. F. Skinner proposed) an
artifact of the way we classify outcomes of different behaviors
together according to some common way of characterizing their
consequences, nor does the regularity of outcomes hold only in a
statistical sense. In fact, any example of behavior more complex
than a single tension in a single muscle is far more repeatable
than it ought to be, considering all the direct influences that
tend to change it.

When the brain sends a command to a muscle, the response of the
muscle in terms of net force applied to a limb depends on the
state of the local energy supplies, on the immediate history of
use of the muscle, on the current angle of and angular velocity
of the limb about a joint, on the states of tension in many other
muscles, and on the dynamical state of the supporting body. To
produce a "movement" such as raising an arm from one position to
another, the net applied force must be time-integrated twice, and
the effects of orientation in the gravity field and the presence
of obstacles and loads must be taken into account. For larger
units of behavior the distance, direction, mass, stiffness, and
velocity of other objects exert an influence, as does the state
of the terrain underfoot, and the body's orientation relative to
the terrain. Many kinds of behavior involve independently mobile
objects and even other people, so that to reproduce even a simple
relationship to the environment, the commands sent to the muscles
must take the independent behavior of other objects into account.
A behavior such as "driving to work" can repeat five days per
week, 50 weeks per year, without the driver ever ending up in the
wrong parking place -- and, obviously, without the driver's ever
repeating the same pattern of motor activities during the trip.

Between the neural outputs of the brain and the patterns we can
observe from outside and name as behaviors, there are countless
points where interference from extraneous variables can alter the
effect of the neural commands. The more remote the behavioral
pattern is from the muscles, the less reliable is the connection
between the muscles and the observed behavior pattern. Yet we
observe the greatest regularities at the end of this causal
chain, and the least at its beginning.

This is exactly contrary to the assumptions of any theory in
which behavior is seen as the natural outcome of neurally-driven
muscle forces. If that were the correct picture, we would expect
the greatest regularities to occur at the level of muscle
tension, with all subsequent effects of those muscle tensions
becoming more and more randomized by unpredictable and invisible
disturbances that enter along the outward-bound causal chain. But
this is not what happens. The regularities are greatest toward
the distal end of the chain, not the proximal end.

This simple observation constitutes a time bomb that has been
ticking for several hundred years as the sciences of life
developed. In all theories of behavior so far offered, the
analysis of behavior has skipped over the details of how
behavioral patterns are produced, and have characterized behavior
in terms of the stable distal patterns rather than the variable
proximal muscle tensions. It has simply been assumed that if a
global pattern of organized behavior exists, such as running from
one end of a maze to the other, all the processes between the
brain and that final pattern must have been at least as regular
as the final pattern. Only under such an assumption could it have
been maintained that the brain produced the distal pattern by any
direct cause-effect route. It doesn't matter whether the theory
rests its ultimate assumptions of causation on cognition,
heridity, personality, or stimuli from the environment. The one
critical assumption is that there must be a regular causal chain
connecting the outputs of the nervous system to the observed
pattern. And that assumption is incorrect.

Given only this fundamental change in assumptions, PCT would
inevitably follow, in something like the following way.

Consider a simple behavior such as walking in a crosswind. We
observe that the person's path remains straight, with slight
wobbles due to gusts of wind. So we would say that this person is
"walking in a straight line."

A person is not built like a tank, made of tons of metal rolling
on cleated tracks. The person walks by balancing on jointed
sticks well-lubricated where they meet and powered by elastic
muscles. Any amount of wind will tend to deflect the direction of
walking.

The person's body follows Newton's First Law of Motion: Every
body continues in its state of rest or uniform motion in a
straight line except in so far as it may be compelled to change
that state by the action of some outside force. From observing
that the person continues to walk in a straight line, we can
deduce that the sum of forces acting sideways on the person is
essentially zero. From that, and from knowing that a crosswind
exerts a force on the person, we can deduce that there must be
another force equal and opposite to the wind, which is keeping
the net force at zero. The only place that force can be coming
from is the person's muscles, which act "outside" the inertial
properties of the body's mass.

We have here a clear example of a consistent outcome being
maintained by variable means. The walking behavior continues
essentially undisturbed. But the behavior of the leg muscles is
varying markedly, maintaining opposition to the variable effects
of the wind. We have a global pattern of behavior that is very
regular, being maintained by an irregular pattern of muscle
tensions. The greatest regularity is at the end of the causal
chain; the greatest variability is at the beginning.

Now, however, we can see that the proximal variability is not
random. It is quantitatively related to the wind forces with
considerable accuracy, for the longer the person persists in
walking a straight line, the more nearly must the average muscle
forces quantitatively cancel the wind forces.

What remains is to explain how the muscle forces could be varying
with the kind of precision and specificity needed to make the
outcome a straight path -- or whatever sytematically-patterned
path is observed. Rather than discussing explanations that others
have offered (which all have serious problems of plausibility,
and which can always be debated at another time), and rather than
repeating explanations that are readily available in the
literature, we will, as they say, cut to the chase.

What PCT says is simple: actions are produced by organisms as a
means of keeping a perceived state of affairs congruent with an
internal specification for what is to be perceived. Behavior is
the process of controlling perceptions. In the case of walking a
straight line, we would look for some perception in the walker
that is dependent on the direction of walking, and it would not
be hard to find. Sensations of torsion in the leg joints provide
some possibilities; vestibular effects provide more; but the most
obvious perception affected by the direction of walking is
visual.

If the environment appears in a constant orientation relative to
the body's direction of movement, the body is moving in a
straight line. If the scene wheels slowly to the left or the
right, the body is moving along a curve. All we need to suppose
is that the perceived rotational motion of the visual scene
(compensated for eye and neck angle) is compared against some
internal standard for angular rate of change, with the difference
serving to adjust the sideward component of leg muscle forces. If
the internal standard specifies zero rate of change, the body
will move in a straight line. If the specified rate of change is
nonzero, the body will move in a curve. And the same system
design that accomplishes this end will also quite automatically
create whatever additional muscle forces are needed to prevent
the wind from producing any unwanted changes in direction.

This same principle applies in all cases where a variable motor
output creates a consistent effect in a variable environment. The
basic concept is that the consistent effect is directly related
to some perception in the behaving system, a perception that is
continually being compared against a reference condition, with
the difference being converted into actions. Those actions
maintain the perception near the specified reference state, and
in the process control the visible patterns on which the
perception depends.

From this concept one can develop a model showing how the nervous

system would have to be organized to produce this effect. This
model would be exactly the same control-system model that the
servo-engineers of the 1930s and 1940s created as they tried to
build machines that would reproduce human abilities to control
variables outside themselves. It is not, as some believe, a
"mechanical" model; it is a functional model that applies
wherever outputs affect inputs in real time and systematic
opposition of outputs to disturbances is found.

The most notable fact about this kind of organization is that it
does not need any information about the disturbances that
interfere with the effects of its actions. It directly monitors
those effects through perception, and acts to maintain the
perceived effects near some reference state, using only the
information in the deviations themselves. No elaborate system is
needed to sense the sources of disturbances. No supercomputer is
required to convert that information into calculated
compensations. The basic organization of a control system is very
simple, although sometimes difficult to represent with analytical
mathematics.

The basic organization is simple, but there are still plenty of
problems to solve. Most of the work going on in artificial
intelligence, neural networks, pattern recognition, and cognitive
science in general can be turned to elucidating the details of
control processes. Many phenomena studied under different names
and different methodologies are the result of legitimate control
processes. Studies of perception and interaction, insofar as they
report reproducible phenomena, are raw material for PCT as they
have been for other theoretical approaches.

But very little of the existing theory in the behavioral sciences
will be preserved if PCT is adopted. Behavior is not caused by
stimuli or perceptions, and it is not caused by plans or central
pattern generators or coordinative structures or conditioning or
situations or propensities or any of the causes proposed under
other models. It is caused by the difference between what is
perceived and what is specified to be perceived. This is the
fundamental revolutionary concept in PCT, the one that, directly
and indirectly, has been causing all the trouble for proponents
of PCT. If this concept holds up, the causal concepts that have
provided the foundations of all conventional theories of behavior
will simply collapse. All the resistance to PCT arises from the
threat of that collapse.

                        The future of PCT

A theory that has been under development for 40 years obviously
has much more to say than the brief characterization just given.
For example, HPCT (hierarchical perceptual control theory)
employs the same basic principle, but adds a hierarchical
structure in which many control systems of one level act through
adjusting goals (specifications for perceptions) for systems of
lower levels, only the lowest level using muscle forces as
outputs. There are auxiliary concepts such as a hypothetical
reorganizing system that can create and modify control systems,
and there are attempts to bring emotion and biochemistry in
general into the model in a testable way. There is a considerable
body of experimental testing of PCT, albeit in simple
circumstances, that lends strong support to the basic concepts.

But all these embellishments on the model are only a feeble start
toward developing the full potential of PCT. Others who have
worked and are working on control processes in behavior will make
their contributions as they see fit. Experimenters in every
discipline of the life sciences will put PCT to tests of kinds
impossible now for lack of hands to share the work and support to
supply the needed resources. No doubt there will be debates and
controversy over the proper way to implement the model and
investigate its implications. In the open and free arena of
scientific striving, PCT will be modified and tested and modified
again, taking on forms that its present custodians are unable to
anticipate.

For this to happen, however, PCT must first be recognized as a
legitimate line of enquiry on an equal footing with all the
traditional approaches that it challenges. It is not now accepted
in that way by the scientific community.

This is not to say that scientists with other ideas should
suddenly start believing in PCT like converts to some new
religion. It is only to say that the grounds for judging PCT in
the open market of science should be scientific. The methods,
experiments, and conclusions of PCT should be evaluated in terms
of the underlying system of ideas and the validity of
experimental tests, not on the basis that these methods,
experiments, and ideas are different from what is generally
accepted now. The conclusions drawn from PCT analyses should be
judged on whether they follow correctly from the premises and the
evidence, not on whether they differ from conclusions that follow
from some other premises.

As most of those scientists who are now engaged in developing PCT
will attest, a scientist in any mature discipline who seriously
investigates this new field will be faced with very difficult
intellectual conflicts of interest. The brief introduction to PCT
given above may hint as to the reasons. In many respects, to
accept the validity of a PCT analysis of behavior is to deny the
validity of some well-established way of interpreting behavior --
and not only the interpretation, but the implications, the
methodology, the experimental procedures, the deductions, the
reputations, and the careers that depend on the interpretation.
It may well be that some of the people who now understand PCT
wish that they could return to the easier state of affairs
existing before they had heard of it. This is the price of a
scientific revolution.

But that price need not be as high as it is. There is little
standing in the way but the barriers put up by those who see PCT
not as a new tool, but as a monkey-wrench in the machinery; who
see it as wrong simply because it is new; who play the role of
gatekeepers and forget their obligations to the good of science
as a principled human endeavor.

New ideas in science that fail the test of rigorous experiment
soon disappear, and ought to disappear. Those that survive enrich
all of science now and in the future. The proponents of PCT ask
only that this new approach be given normal scientific treatment
and its developers be granted a legitimate place in the forum. If
these new concepts prove out, everyone will benefit in the long
run, whatever the apparent short-term costs.

From Tom Bourbon (930403.1616)

[From Bill Powers (930501.1010)]

The following paper is a first draft, the final product to be
submitted to an unselected journal. I am inviting comments,
revisions, deletions, additions, rewordings, new concepts, or
anything else that will make it acceptable to readers on this
net. I particularly want these things from people who will be
willing to become co-authors of the paper. I would like this
paper to appear with many authors from many disciplines in many
countries around the world: the more the better.

David Goldstein's observations about the proposed paper seem apt: why will
anyone pay attention, or if they do, why will they take it as anything but
sour grapes? But having said that, I wonder if this might not be the time
for another attempt at BBS. To assuage my lingering bad feelings over
having encouraged some ealrier attempts there, I am willing to go along as a
coauthor. One of Harnad's major put downs of PCT, on those earlier tries,
was to claim that it was, "just another way of looking at things" and ask
"Where are the data?" (Somewhere in the files, I still have my copies of the
rejection letters he sent to Bill Powers -- I could find his exact words if
necessary.) I doubt that he would ract differently this time around, but it
might be worth a try, just for the sake of learning if anything has changed.
If that is the targeted journal, I think the paper should include a range of
demonstrations -- quantitative, of course, to assuage the editor's thirst for
data. There is an impressive enough array to draw from -- E. coli, crowds,
pointing, economic theory, mind reading, and -- for old time's sake --
tracking, of various degrees of complexity and sociality. (Maybe even blind
men and worlds could be tossed in.)

Maybe it is just the latent logical positivism that reviewers have claimed
to see in my manuscripts, but I still believe a few good demonstrations
make a nice tonic in these days of deconstructed behavioral and cognitive
science. I don't think we will get very far by simply pleading our case
and asking for mercy and fairness. The "proof" is in the behavior of the
model. It should be featured in as wide a range of applications as
possible. That way, the commentators in BBS would have numerous targets to
shoot at -- but they would need to shoot at all of them at the same time,
with words thrown up against the versatlity of the PCT model in its
simulations.

Of course, such an article would never get in -- too diffuse and unfocused,
not directed to topics that are of interest or that are the subject of
discussion and controversy in behavioral and cognitive science, directed to
issues that are already resolved and understood. There could be dozens of
reasons. All the more reason to try.

Until later,
  Tom Bourbon
Until later,
  Tom Bourbon