Angles on PCT

[From Bill Powers (930825.0600 MDT)]

Avery Andrews (930825.1123) --

That was a fine presentation of PCT, from a viewpoint that is
less parochial than mine.


Rick Marken (930824.2130) --

Sometimes I think you're being too rigid and hard-headed, but
almost every time you manage to convince me that you're right.
Yes, the perceptual signal is controlled, and no, it does not
control behavior. All you have to do is to go back to the
recordings from the simplest tracking experiments. The tracing of
the controlled variable bears almost no relationship to the
tracing of the output of the system.

It's still possible, even likely, that information theory will
tell us something about the conditions required for control and
about the limits of control. But it can't tell us about the
details of control because it's not that kind of theory.
Information theory doesn't deal with the behavior of variables on
a moment-by-moment basis, but only in the aggregate, only as a
complete data set. The analyst has laid out in front of it a
record of happenings over a long period of time, as well as
knowledge of the signals and functions involved in control. None
of that is available to the control system itself, which operates
strictly in present time. From the point of view of the control
system, the perception is the dependent variable and not the
cause of anything.

Oded Maler (930824) --

I have some Deja Vu's from the recent thread going on with
Michael Fehling, and I glad to remark that certain PCTers have
progressed with the years and understood not everybody else is a
good-old-mainstream-paper-rejecting psychologists, and that the
commonalities and differences between PCT and new-age-situated-
robotics-AI are better understood and crystalized.

We learn, Oded, we learn, if only slowly. It helps a lot to see
some glimmers of support from places that never indicated any
interest before. It would also be more satisfying if some of them
were to realize that they were reinventing 40-year-old wheels.

Appearances is what you sell when you design machines. The
customer of the robotics/manufacturing systems doesn't give a
damn on the internal "mental world" of the machine (except for
prestige, e.g., this machine was implemented using neural-net
imlementation of Fuzzy PCT) but on observable output (as
perceived by the client).

Right; a point I have been trying to make for some time. But
there's another point to be made. By heading straight toward a
design that will do the job, engineers had paid far too little
attention to basic principles -- organizational principles -- of
control, and they can learn a lot from studying how organisms do

If a PCT-based design will solve a hard problem in robotics, no
one could ignore it.

Oh, yes, they could. It may be true that a PCT-based design will
solve a hard problem, but that is not the same as persuading an
engineer that it will. First the engineer has to assimilate the
PCT approach, and as we have seen on this very net, that can be a
big hurdle to get over. Engineers tend to be in love with the
approach that they have been using for many years, and see little
reason to look for something completely different. Fortunately,
this is not true of quite all engineers. Real engineers are the
ones who must demonstrate the uses of PCT; I don't have the

What is a factory controlling for? Should every perceptual
variables of the factory be encoeded using one transmission line
("one neuron for one percept") or should they be distributed in
time and space?

I prefer the parallel approach myself, as laid out in HPCT.
That's about as distributed in space as you can get. As to time-
multiplexing, I think that introduces more problems than it
solves: the demultiplexing, for example, that is needed to keep
unrelated packets of information from being confused. But the
truth is probably going to be a compromise.

In the human organism, practically everything is fed back,
starting with the most detailed processes. In control of arm
position, even the forces/accelerations are under specific
feedback control, then the velocities, then the positions.
Nothing important is assumed to happen just because a neural
signal says it should happen. The nervous system just isn't very
precise when it comes to output. Output effects are big, crude,
and sloppy. It's the perceptual system that creates precise

A factory is an output function. What makes it into a control
system is primarily the array of sensors that reports what is
happening in the factory because of those big crude output
Dag Forssell (930825 0050)

Version 4 is developing into a really good article. The opening
is now excellent, with good transitions.

A few picky details:

Missing from this intuitive understanding of control is:

a) clarity and formalization of the process.

b) a clear understanding of how control works including such
things as:
  - what is being controlled (perception, not output or
  - that control systems resist disturbances.
  - the simultaneous and continuous nature of control.
  - speed of response to changing goals.
  - sensitivity to differences from the specified goal.
    (Business: variance)
  - amplification: use of resources.

You need parallel construction here: _the_ speed of response,
_the_ sensitivity to differences. In such a list you should be
able to plug in any one of the subsidiary sentences into the
whole structure and get the same form of sentence. The last point
isn't clear: perhaps you should say "amplification: the use of
tools and other resources to increase the effects of human

Right after the diagram of the single control system, you say

In 1868 James Maxwell published equations "On Governors,"
but it was not until 1927, with the invention of the electronic
amplifier ...

The electronic amplifier was invented two decades earlier. The
reference you want is to H. S. Black, who in 1927 suddenly
understood the basic properties of a control loop using
electronic amplifiers ( on the Lackawanna Ferry, on the way to
work at Bell Labs).

Unfortunately, Wieners presentation left room for the
impression that feedback is a step by step process, internal to
the organism.

Wiener's. Also, Wiener himself didn't create these impressions;
they came from people trying to reconcile the continuous control
system with older ideas of event-based behavior models. In the
chapters that most people read, Wiener presented analog control

Engineers commonly talk about control systems as controlling
output, where output is understood to be the physical action or
behavior of a mechanism.

Engineers use "output" to refer to controlled variables, not to
the effector actions of a control system. Others are to blame for
interpreting "output" to mean "action" or "behavior". Of course
many modern engineers, as Rick Marken will attest, aren't too
clear about the difference, either. Interesting to see what you
picked out of those interchanges with Bar-Kana.

If the reading is wrong because the instrument is out of
calibration, the engineer will never know.

... until some other instrument reading begins to show that
something is radically wrong.

It's going to be a really excellent article.
Best to all,

Bill P.