From Mary Powers

[from Mary Powers 940602]

The recent exchanges between Martin, Bill C. and Bill Powers seem
very familiar to me - not because of the content, but because
this type of discussion has come up a number of times, and for
the same reason.

Martin said:

I thought [Bill C.] was assuming [Bill P.] had it more or less
right, for the time being, and was looking at implications of
that assumption. Some people seek out the classes and
structures, other people look at the implications of what those
pioneers find. The pioneers should really not object when their
discoveries are found useful by others who are not exploring the
same virgin territory.

But Bill has been sending you both a message you either cannot or
don't want to hear: you are trying to apply some PCT ideas to
what you already know and believe, to models you already have,
instead of _applying it directly_ to the understanding of living
systems. The pioneer has every right to object to this - just as
chemists who came to understand oxygen and combustion had every
right to object to those who wanted to look at its implications
for phlogiston - or heliocentric modellers to object to exploring
the implications for epicycles.

There are implications all right, but not ones either of you have
really taken on yet: you want to integrate PCT with the way you
already understand, the explanations you already have. But some
of the concepts you are defending, such as information and
uncertaintry, are, from a PCT perspective, irrelevant.

Mary P.

[Martin Taylor 940602 14:00]

Mary Powers 940602

But Bill has been sending you both a message you either cannot or
don't want to hear: you are trying to apply some PCT ideas to
what you already know and believe, to models you already have,
instead of _applying it directly_ to the understanding of living
systems.

Funny perception! I, at least, don't see it that way at all. I do not
see myself as applying PCT to any models I have. I do two things, as I
see it. I apply PCT, as is and undiluted, to living systems, and in
particular to groups of living systems and living systems with computers.
I am beginning even to have some success locally in getting people to
listen to the message, and we are on the verge (at the end of the contract,
unfortunately) of doing some serious simulations of learning hierarchies
leading toward the concept of "understanding" computer interfaces.

And at the same time I try to apply to PCT mathematical and physical
principles that apply to everything else in the world (living or otherwise).
The objective of the second element is not to change PCT except to provide
other ways of looking at it, and thereby to show up features that might
be overlooked when we are restricted to a single viewpoint.

My basic thesis is that there are core statements taken as true, and
these are necessary for PCT to be PCT. Added to these are statements
that seem highly reasonable, which I take to define "classical" HPCT.
And then there are statements taken as conventional, which Bill P often
points out to be probably requiring refinement or alteration, but which
we treat as usually forming part of the HPCT convention for conducting
discussions. As an incomplete list, here are some statements I take to
be in each group:

Core:
  All behaviour is the control of perception, meaning that every action
of every living system is either with the purpose of stabilizing some
perceptual variable to a reference value, or is a side effect of such a
purpose.
  All actions are the outputs of control systems that have at least the
following consituents arranged in a loop: a perceptual signal that derives
from a perceptual input function; some sensory apparatus that, together
with other signals provides input to the PIF; some real outer world that
affects the sensory apparatus; some output apparatus that affects the
real outer world; some function that generates the output; and an error
signal that feeds the output function and represents the difference between
the perceptual signal and an externally generated reference signal.
  The outer world on which the output apparatus acts and that is sensed by
the sensory apparatus is affected by factors other than the output of the
control system.
  The set of control systems that constitute a living being are coordinated
in a manner that has been set by evolution, and that can be altered by
learning in individuals of sufficient complexity.
  There exist memories, which may be implicit in the coordination structure
of the interacting control systems, or may be stored in some way such that
they can be used by the control systems.

"Classical":
  Control systems are arranged in a hierarchy, such that the direct output
of any one will contribute to the reference signals of others, and such
that there are no cycles in the connections of outputs to references; the
control systems can therefore be ordered in levels, from higher to lower.
Only the lowest send their outputs directly to the musculature and affect
the outer world directly.
  The hierarchy of output/reference connections is mirrored in the connections
of perceptual signals to the sensory inputs of higher-level control systems.
It is not necessarily true that a control system provides sensory input
to the same higher ones from which it receives its reference signal.
  At different levels, the nature of the perceptual input function is different.
  Learning occurs by altering the connections among the control systems at
different levels through a process called "reorganization." Reorganization
may affect input, output, or both.

"Conventional":
  There are eleven different levels of control system, represented by eleven
different kinds of perceptual input function. Within any one of these levels,
there may be micro-structure (sub-levels), but we normally do not include
these sub-levels explicitly in discussions.
  Control systems at higher levels work more slowly than those at lower
levels [I'm not sure whether to include this statement here, but I do so
because it seems to be taken as an unsupported statement, rather than as one
derivable from Information Theory, which it is. If that derivation were
to be accepted, I would put this statement in a list of "necessary
consequence" statements, rather than among the "conventional" statements].
  There exists a segregated system called the "reorganizing" system that
affects the structure of the control hierarchy blindly, in response to
control failures. The reorganizing system may affect the hierarchy as
a whole, or may affect in a directed manner regions of the hierarchy in
which control is poor. Whether it acts globally or locally, its effect
is to make random changes in the affected region. Its probability of acting
depends on errors in the values of a partially defined set of "intrinsic
variables" that constitute its controlled perceptions.

There are more in each list, but I'd take these as a representative sample
of what I take to be necessarily true, usefully taken as true, and taken
as true for discussion purposes but changeable without destroying the
essence of PCT, if evidence shows them to be wrong.

There are necessary consequences of these, if they are taken together with
observed facts such as the number of independent muscular outputs and sensory
inputs. Prime among these consequences is the existence of something like
the alerting system. Prior to learning about PCT, my notion of the alerting
system was rather ad-hoc. Living creatures are provided with it, so why
don't we consider what it is for, and take advantage of it in using our
instruments as well as our natural senses. PCT showed why living systems
are built that way. Alerting systems aren't an add-on to classical PCT.

There's lots to learn about how hierarchic control systems work, and lots
of ways to learn it. Simulations provide proof of principle, and surprises
in simulations lead to better understanding of principles. Consider Rick's
discussion some weeks ago of the experimentation on the Marken effect. He
said, if I recall, that it originally didn't work in simulation, until they
added a transport lag. He then said that the Marken effect was explained
by transport lag. But this isn't an explanation; it is a description of
the circumstances under which the effect is found. The simulation provided
a surprise and a boundary condition. It now leads to a place in which
there is still something to be learned about how interacting control systems
really work. What analogues of the Marken effect are there in the internal
workings of the hierarchy? Are there places where two ECSs work in concert,
one with low gain helping to stabilize the work of the "real" (high-gain
high output) controller of a perceptual signal?

Anyway, the messages I hear Bill P sending are more along the lines of
"don't try applying to PCT principles that work elsewhere; PCT is like
nothing else in the world, so your principles won't work here." Well,
I don't believe in magic. I believe a control system is a control system
and will behave as it does without magic, using perfectly normal physics
and mathematics, whether or not my skill in those disciplines is enough
for me to discover just how they apply. Over and again, I get from Bill
a statement that I can paraphrase as "I don't understand that in detail,
and control systems are simple, so it doesn't apply and I'm not interested."
"I'm not interested" is perfectly legitimate. "So it doesn't apply" is not.

I thought [Bill C.] was assuming [Bill P.] had it more or less
right, for the time being, and was looking at implications of
that assumption. Some people seek out the classes and
structures, other people look at the implications of what those
pioneers find. The pioneers should really not object when their
discoveries are found useful by others who are not exploring the
same virgin territory.

I stick with that. PCT should be useful as well as interesting.

Martin