[From Bill Powers (940805.1150 MDT)]
Paul George (940804.1600) --
Having finished B:CP, I think it is a good piece of work. Too bad so
much of it seems to be ignored in most of the research and the
discussions here, which seem focused on the 'worm's eye view' of PCT
(1st & 2nd order control). Perhaps because the other concepts are
deemed 'uninteresting'?
Having finished the book, you can now speak with more assurance about
what PCT is than others who haven't -- as you now realize.
Avery Andrews explained the position well. There are lots of people
guessing about how the higher levels work, but there's no unifying
principle to tie such work into the whole system. The biggest problem is
in designing research and doing simulations without understanding how
higher-level perceptions work, or even exactly what the higher levels
are. So we work where we're more sure of our ground, trying to develop
methods that will apply generally but not going too far beyond what we
can demonstrate. It will be up to smart people like you to push the
boundaries further upward in the hierarchy.
In terms of the reorganization system, why must the reference variables
be _intrinsic_ and physiological? (I equate 'intrinsic' with hardwired
or inborn).
The main reason is that I needed something that _could_ be inborn, to
provide the basis for the development of the hierarchy from infancy (or
before) onward. I see the hierarchy in the neonate as a set of possible
control systems, with the necessary sorts of computations provided for
at each level but no wiring and no organization yet. We have to learn to
perceive everything, at every level, and also how to control what we
perceive. As I said in the book, this is a worst-case model, where there
is an absolute minimum of inborn organization. The reorganizing system
has to work under these worst-case conditions, where even the _concept_
of an algorithm or systematic strategy hasn't been developed yet.
If everything from sensations to systems concepts has to be learned, we
obviously can't count on anything in those categories as an aid to
learning. The process of reorganization, in short, can't be intelligent,
because it has to work properly before any intelligence is developed. So
I asked myself what the basis for reorganization could possibly be with
those rather severe restrictions, and came up with the answer that it
has to be concerned with the status of the organism itself. That's where
the idea of intrinsic variables came from (that, and from W. Ross Ashby,
who called these "critical variables" and postulated a random switching
device which was the direct ancestor of my idea). There's no real need
to guess at what the intrinsic variables are; all we need to specify is
that they be affected by the way the organism behaves in an environment,
and that no knowledge of the external world or its laws be involved.
Ashby assumed that there were built-in upper and lower limits on
critical variables, exceeding which started the random switching
process. I substituted the idea of reference signals, but that's a minor
difference (two-way instead of one-way control). He also saw that such a
system would be "superstable" in that it would keep randomly switching
until the critical variables were once again within limits -- regardless
of _why_ they came back within limits. I simply adopted that principle;
it's still the basic principle of reorganization in PCT.
I was always somewhat unsure whether random switching could really be
efficient enough to accomplish real reorganization, until I heard of
Koshland's work on chemotaxis with E. coli, and set up some simulations
to test his idea. The results astonished me. This turns out to be an
unbelievably effective way of controlling things in a few dimensions at
once. As long as reorganization is applied to small parts of the system
at any one time, it can very quickly lead to local optimizations. The
random nature of the reorganizing process even provides for getting out
of local minima if they aren't too deep. I have shown that this method
can be used to solve 50 equations in 50 unknowns in half an hour or so.
However, I don't see anything which would preclude having learned or
otherwise generated reference levels as well. Recognizing the need to
adapt would seem to be a charateristic of higher levels of
intelligence.
But you've put your finger right on the problem, haven't you? Where do
those higher levels of intelligence come from? They have to be learned,
too. And if they are to be learned, they must be the product of some
learning process that predates them. Once each new level is organized,
many control problems that could formerly be solved only by random
reorganization can now be solved by a systematic control process, which
is far faster and more accurate, and can keep errors so small that
intrinsic errors are never seen -- at least for reasons having to do
with that level of control. It's even possible that the reorganizing
system creates (at the logical level, probably) specific strategies for
learning, although it's learning of a different kind from the random
trial-and-error of reorganization.
A general 'error level' or 'discontent' intrinsic would help to
'awaken' the reorganization function, but seems to me unsatisfying for
directing it.
Yes, general error level in the hierarchy (along with other types of
variables inside the organism) is considered an intrinsic variable, with
a built-in reference level of zero. When you say that reorganization is
unsatisfying for "directing" the results, however, you're not
recognizing the surprising property that random reorganization CAN be
directed toward achieving a specific end-state. Whatever the variable
under control, reorganization is driven by the difference between that
variable and a corresponding reference state. This is the "intrinsic
error" signal. Specifically, the interval between reorganizations
depends on the rate of change of intrinsic error times the magnitude of
that error -- which amounts to the first derivative of the square of the
error. That is the model that seems to work the best.
What's hardest to grasp here is that the reorganizing system doesn't
care what organization it produces as it acts on the system being
reorganized. The ONLY thing it is concerned with is the intrinsic error.
It will stop reorganizing only when the intrinsic error gets small
enough. There many be better organizations, or the intrinsic error might
drop to zero because of the arbitrary action of an external agency. None
of that matters to the reorganizing system. It just wants its own
intrinsic error to be zero. If it is zero, or small enough,
reorganization stops, leaving _whatever_ organization exists at that
point to go on operating.
It is this extreme pragmatism of the reorganizing system that makes it
so effective. It does not need to know anything about the environment,
or the organism, or the brain, or anything else besides the states of
its own input variables in relation to their respective reference
signals. It will try literally anything, and at random, until intrinsic
error disappears, and then it will stop acting. How, during this
process, it has changed the relationship between the behaving system and
its environment is completely irrelevant to it. It can't either know or
care about that. That is why it can produce a workable organization in
any kind of environment. Its very stupidity is what makes it so
powerful.
RE: experimental work on reorganization.
Very little has been done with real people, although Dick Robertson has
done some experiments in which reorganization appears to occur, and
Frans Plooij has recorded data on infant learning in chimpazees and
humans that seem to show definite periods of reorganization. We have
done a number of simulations, one of them being a control system model
that matches itself to behavior of a real person by treating the
difference between the model's and the person's behavior as an intrinsic
variable and randomly reorganizing the integration factor of the model's
output function to reach a minimum in the difference. As mentioned, I've
tested some models involving multiple control systems controlling a
shared world of multiple variables, and reorganizing until the system of
simultaneous equations is solved for independent control by each system
of a different aspect of the shared environment. That's not a lot, but
such things take time to work out and at least it's something.
RE: presenting the theory
I personally think you do your 'cause' a bit of harm by splitting off
the reorganization hierarchy from the central control hierarchy, and
then ignoring it in most of your posts and articles.
Well, sheesh, how much can we talk about at once? We talk about the
things we have done the most work with. Actually, if I had responded to
your initial inquiries with a complete summary of everything that is in
the PCT model, you would have thought I was some kind of nut. Haven't
you ever seen real nut mail? If somebody you had never heard of sent you
a two-inch-thick packet of drawings and text with the title, "A complete
theory of everything that all organisms do under all conditions,"
wouldn't you alert the nearest looney bin? Most people who first hear
about PCT have the impression that I thought it all up last week, so
they're surprised that I haven't cited someone who said something
somewhat similar last month. The idea that there's been this whole long
development going on for 40 years, and that they've never heard of it,
never occurs to them (and why should it?). And when I can't explain it
all in one breath, they start fidgiting.
Please accept a suggestion from a newbee who recently went through
trying to get up to speed with PCT, even given a fairly strong
understanding of control. Given the paucity of 'public domain' sources,
I would recommend writing a little whitepaper and incorporating it in
the monthey post (as well as archiving it (you could really use a FAQ).
I'd love to have an FAQ. How much do they cost, and can I get a used one
cheaper? By the way, what is an FAQ?
I'm still leery of trying to cram the whole theory into the monthly
intro, for the reasons cited above. Would you like to try it, being
fresh from the learning experience? You seem to have a pretty clear view
of what might make the whole thing more quickly understandable. Why not
post a first try and let some of the other newcomers chime in. It might
turn out to be a really effective document. "How I finally got the idea
of PCT," or something. In fact, your suggested outline sounds pretty
good to me.
···
------------------------------------------------------------------------
Best,
Bill P.