[harnden 941118]
the dutch group sent me a memo asking for information on pct, and
where i thought it fit in. what follows is the text that i sent them.
you'll recognize a lot of the phrasing... it comes from here. i only hope
hope that i've done the ideas some justice, not injury. in retrospect,
i think there are things i could have made clearer, but they also need
to be clearer to me... i'm still learning, of course. anyway... if you
have any comments, please pass 'em on. by the way... i'm enjoying the
ecoli discussion/code swap. real stuff! i might have something to contribute
to that thread later on... but for now i'll just keep yanking and running
code.
---TEXT FOLLOWS-----------
Thank you for your interest in my initial proposal.
While I am, of course, a student of the PCT paradigm and
not yet an expert, I will attempt to address some of the
questions you bring up. I will be cribbing in places
from other sources on the topic, although I will try to
paraphrase. Please keep in mind that, as I suggested in
the original document, while this proposal represents a
significant particular interest of mine, I am willing to
engage in any task that your team would find of use. In
any case, here is a brief on PCT, following the general
outline of your inquiries:
Let me begin with a couple of references, the first
quite useful, the second critical. The first is
"Feedback Thought in Social Science and Systems Theory,"
by George P. Richardson (1991: University of Pennsylvania
Press), which, among other things, provides a much more
concise account of PCT than I could presently hope to.
The book follows two "threads" in the ongoing
thought/dialogue about feedback as a fundamental process
in the world: the "cybernetics" thread, tied more closely
to biology, natural philosophy, and information theory;
and the "servomechanisms" thread, arising from more
formal areas like economics and engineering. Richardson
identifies PCT mainly as a contribution to the
cybernetics thread, sharing that domain's features of
focusing on negative feedback as the means of control,
and of being more geared to epistemological issues than
to things like policy analysis - although Powers'
methodology is definitely that of a servomechanisms
engineer.
The second, most important reference is the original
text, "Behavior: The Control of Perception", by William
Powers. Published in 1973, it follows (althoug it does
not explicitly reference) a line of thought suggested by
Rosenblueth, Wiener, and Bigelow in a 1943 paper called
"Behavior, Purpose and Teleology," which argued that
purposiveness, being a real feature of living systems,
should not be discarded from scientific inquiry on the
grounds of the teleological fallacy. The effect may not
produce the cause, but the need for the effect may
motivate the cause. For Powers, the efforts of
behavioral science since Watson to remove the "illusion
of purposiveness" from discussions of behavior have led
to increasingly untenable models. Because they view
behavior as exactly the opposite of what it is, these
theories fail to account directly for variations in
experimental observations. Such accountability is
acheived within the framework of the dominant paradigm
only by reliance on statistical inference, which by
itself cannot validate a causal model.
According to PCT, behavior (action) is precisely not
the response of an organism to environmental stimulus.
The PCT model of behavior, at its core, is a simple
feedback model of neuromotor control, which demonstrates
that variability in behavioral output is the effect of a
goal-seeking system attempting to maintain perceived
quantities at internally-defined reference levels, even
in the presence of disturbance in the control loop (such
as environmental change). The maintenance is acheived
through negative feedback - that is, by behavior which
reduces the error between the perceived and reference
values of the controlled quantity. The simplest
schematic of this kind of system looks like this:
Internal Reference Value
>
\/
Perceived Value--->Error Comparison--->Action
/\ |
> \/
>-------------------Physical Condition<---Disturbance
With respect to stimulus-response (S-R) theories of
behavior, that which is commonly viewed as the stimulus
is seen to be the disturbance of the controlled variable.
By way of example, consider a tracking task. A subject
controls the position of a dot on a screen with a
joystick, with the task of keeping it one spot. The
position of the dot is randomly perturbed (jiggled about
a bit) by some unseen force. The subject continuously
acts to reduce the error between the dot's actual and
desired positions. It is not the disturbance but the
error signal, the perceived difference of condition and
reference, that is the cause of action. If the subject
is given a different task - say, keep the spot near the
bottom of the screen - then the reference value is
different, and although the "stimulus" (the jiggling of
the dot) remains the same, the behavior changes to
maintain the new reference condition.
Higher-order behaviors, from complex motor control to
cognitive processes, are the result of a hierarchy based
on the fundamental model, where the reference value of
one system is the condition (or rather, perception of
that condition) controlled by the next-level system.
This allows lower-order reference values, those that
control actual sensory/motor interaction with the
physical environment, to be modified by the organism.
According to Powers, this process of "reorganization", in
which an organism manifests learning by modifying the
parameters of its behavior, could even extend to
structural reorganization: control of the wiring itself -
a notion which, given what is now known about how neural
pathways are dynamically constructed, merits
consideration.
PCT is not currently a widely accepted view... but
then, neither is much of cybernetic thinking in the
social sciences, it seems to me. While it has a
publication history, mainstream psychological journals
tend to reject papers on the direct topic either because
of behaviorist hostility to the feedback paradigm or,
where feedback is recognized by some branches of
psychology, because its implications in placing the locus
of control inside the organism are misunderstood. The
closest I have seen things come to it are in the
"subsumption architecture" idea in robotics and in the
Swarm/Sugarscape-type models. My familiarity with those
experiments is not sufficient to provide a decent
overview of PCT's overlap/disjunct with them. I am
aware, however, that within the PCT community the
question that persists about these approaches is whether
"control" is actually what is implemented. In other
words, do the entities in these models act purposively to
maintain the perception of an external condition in
accordance with an internal reference? This community,
by the way, is the Control Systems Group Network, a
discussion group mediated by electronic mail, distributed
by a server in Indiana (information below). The group
includes Powers himself, as well as about 130 people in
17 countries.
Crowd models, and the study of collective behavior in
general, happen to be one of the applications of PCT. If
one has a causal model of individual behavior, then one
has a foundation for a theory of collective behavior
that, like good theories in classical physics, relies on
first principles. PCT is also used in human factors
engineering, management science, education, and practical
psychology. One implementation appears in the last
category under the name of "reality therapy", described
by William Glasser in "Stations of the Mind" (1981:
Harper and Row). Concrete implementations of the theory
exist as designs for experiments in behavioral
psychology, and as computer programs which seek to
emulate some aspect of behavior (simple, complex, or
collective) under the assumptions of the paradigm.
As you have probably gathered by now, PCT does not
have much to do directly with genetic algorithms, a-life,
and the like. Current experiments in a-life, as I
understand them, based on the patterned behavior of
interacting automata, doesn't put much inside those
automata. It seems to be more the interaction of an
environment with a set of rules than an organism with an
environment. It may yet turn out that, as a-life systems
evolve, these rules will turn out to map to the
behavioral effects of identifiable controlled variables
in living systems. Similarly, while genetic algorithms
do not immediately lend themselves to an implementation
of PCT (since they do not seem to account for the
organism's own role in the co-evolution of ecologies),
they might, for example, be used to model on a higher
order the process by which "memes" (Richard Dawkins' term
for the atomic unit of culture, which must ultimately
have a behavioral definition) are transmitted among PCT-
based individuals. I do in fact have a copy of your
paper on "The Battle of the Perspectives", and it was
largely due to the acknowledgement of the importance of
these kinds of questions that I found there that I was
inspired to put this proposal forward.
My idea for the inclusion of a PCT model in the
TARGETS framework is to position it so that it either
subsumes or drives the steering models. In my roughest
initial conception, the goals which the steering models
seek to achieve are the reference values of the PCT
system. The outputs of the PCT system either influence
or become the steering system's investment, allocation,
and other action-oriented instruments. The global
system's response models provide input to the PCT system,
acting as the perceptual signals that it is controlling.
As I said in the initial proposal, I see the initial
implementation simply as a proof of concept. The PCT
sector would have a relationship with the global system
on the order of a "world-mind", an aggregate high-order
model of human behavior attempting to control the world.
A more theoretically useful tool would evolve in the next
step: the reconstruction of the PCT system as a sector,
consisting of multiple individuated PCT systems whose
collective behaviors would influence the steering
variables, perhaps by weighted "vote". Certainly, as
TARGETS itself disaggregates in future versions, then
each individual system would varying degrees of control
over its own and other sectors of the globe. I expect
that the initial reference values for these individual
systems could be parameterized according to some
sociological taxonomy, such as Thompson's Culture Theory.
A single sectoral unit could then be seen as representing
general national characters, each unit emitting behaviors
in accordance with the values that it holds.
Along with the complex dynamics that I expect to see
the entire coupled system demonstrate, there is probably
going to be something even more interesting to look for,
pointed out to me by Powers after I posted my proposal to
his discussion group. A control system acts to maintain
a perceived state of affairs. Disturbances of its
environment or direct attempts to thwart its goals are
met with behaviors that either seek to bring that
environment back under control or, under extreme
conditions, reorganize the goals themselves. What should
be interesting is to observe how the behavioral sector
adjusts the means at its disposal to maintain the
outcomes it wants. In the same concrete terms that
Powers put it to me, remember that a population faced
with starvation does not want to starve - it will do what
it has to in order to avoid it. Sometimes the behaviors
can be quite unexpected, or violent. If my PCT sector
works, the question I might be looking at could well
shift from "what effect will global change, brought about
by the population, have on that population?" to "what
will a population do, faced with global change that it
has brought about?"
Reference Note:
The Control Systems Group Network : CSG-
L@VMD.CSO.UIUC.EDU - for information contact Gary Cziko,
the moderator, at CZIKO@VMD.CSO.UIUC.EDU. The group's
discussion also appears on the Usenet, in
bit.listserv.csg-l. If you like, I can provide you with
an electronic copy of their introductory document, which
has the full publication reference and more information
on means of access.