# Bushes; "Real" conference; GAIA stuff

[From Bill Powers (930126.0830)]

Rick Marken (930125.1500) --

RE: output and perception bushes

(reformatted:)

When people think of control theory, they think of set points
for scalars -- so if control theory is applied to arm lifting,
the tendency is to imagine that one control system should be
assigned to each scalar component of the vector output (each
muscle). But it is also possible to have a control system
controlling a scalar quantity that is a function of ALL
components of the output vector at once -- a single scalar
number, p, can vary as the sum of 5 muscles tensions, for
example.

This is a very important insight. Any high-level control system
necessarily makes use of many lower-level control systems at
once, and its perceptions are a function of the individual
control effects of all those systems, plus disturbances. Yet it's
still a scalar perception, not a vector. It can only get bigger
or smaller.

This brings to mind a passing comment by John Gardner, to the
effect that all the degrees of freedom of a hardware system have
to be specified. While this is true, it means only that all the
degrees of freedom ARE somehow specified. This does not mean they
are all controlled. You can hold out an arm and control its
position in the vertical direction but not the horizontal
direction. Someone pushing up or down will find that your arm
resists, but if they push sideways the arm will simply give way.
A mechanical system's state is always the consequence of all
forces acting on it, but not all of those forces necessarily come
from associated control systems.

So if a single control system's output affects 5 degrees of
freedom of the environment, and the controlled variable is some
weighted sum of those 5 degrees of freedom, the environment will
be controlled ONLY in that respect. The variables can all change
in uncountable ways, but the only CONTROLLED aspect of those
changes will be the one that keeps the value of a function of the
5 variables at a reference level. Actually the output can affect
far more than 5 degrees of freedom, and normally would. But only
the five that are sensed matter, and even for those, only the
particular function that creates the perception is under control.
The environment remains unspecified in all other respects, which
is to say that environmental disturbances are free to alter the
state of the environment in all other respects.

To get a 5-df environment completely under control, it's
necessary to have control in all 5 df. This can be accomplished
by 5 scalar control systems, each of which controls a function of
the 5 variables that is linearly independent of the other
functions. Any one of these control systems allows certain kinds
of variation in all 5 df, but all five systems together exert
complete control over the 5 variables. Just how they do this may
not be readily apparent to passers-by if all the control systems
act by affecting all of the variables.

One of the logical errors that's easy to make in learning control
theory is to suppose that because a particular state of the
environment is observed to be involved in control of a higher-
level perception, ONLY that state of the environment will result
when the perception is controlled.

I've seen this in linguistics. In a top-down model, some global
feature of a sentence is specified. This feature is then
exemplified by some element of a specific sentence at a lower
level. But why that sentence, and not a totally different one
that is also an example of the higher form? In fact the detailed
sentences used as examples vary all over the place, so there is
clearly no constraint on which sentence is to be used as an
example. This is a major problem for top-down models (at least as
far as implementing them is concerned).

A control-theoretic model of language production works the other
way: it selects sentences until one is found that can be
PERCEIVED as having a certain form matching a specified form.
None of the degrees of freedom of the sentence matter except the
combination that results in satisfying the reference-form; in all
other respects the sentence is free to vary. Any old words that
can be perceived as an NP will satisfy the requirement that an NP
appear in the final result in that position. But you can't go the
other way: you can't write an algorithm that will start with the
specification that a noun phrase be uttered, and come up with a
specific utterable noun phrase. The specification "NP" doesn't
care which noun phrase is found; therefore it can't specify ANY
noun phrase.

You linguists out there have heard me harp on this before. So far
you haven't dealt with the basic problem. It's part of the same
problem that Rick is talking about.

ยทยทยท

-------------------------------------------------------------
Cliff Joslyn (930125.1746) --

Are you trying to use CSG '93 as a "real" conference, with
submissions open to everyone, review of articles, proceedings?
If you're not big enough for that yet, do you aspire to be?

From my point of view that's what we're trying to avoid, if

possible. It may not be possible forever. At "real" conferences,
you have a couple of plenary sessions, but mostly all the
special-interest groups go off into parallel sessions and never
interact with anyone with different ideas. The review process
serves mainly to assure that only approved dogma is discussed,
and that people with good ideas but without specialized training
are never heard from. And those that are heard from get up with
set pieces to say, say them, and sit down, with scarcely a
comment or question. There isn't time -- got to get on to the
next paper. The result is that the audience (there are no
participants) hears a paper spoken at about 1/4 the speed that
they could have read it, provided they are still awake at the end
of the session. Worst of all, at the plenary sessions, what do
you get? You get the movers and shakers, the Important People,
the Leaders, who stand up and put on a two-hour show, mostly
nauseum for years. You don't get a graduate student standing
before the whole assembly sharing the excitement about an
experiment of his or her own devising that worked just
beautifully, or throwing out a cockeyed idea to be shot down. The
graduate student isn't important enough to merit a plenary
session.

The success of our last eight CSG meetings has always rested on
fragile grounds. The basic theme is participation by everybody in
everything. All sessions are plenary, about 26 hours of them.
Speakers are strongly discouraged from reading papers, but
encouraged to bring enough copies to pass around so people can
read them in the off time. The ideal presentation consists of
someone giving a short (ten-minute) talk ABOUT the paper they
want to present, and then opening up the floor for discussion,
assuming that everyone who is interested has read the copy
distributed on the first day. The time allotted to a speaker or
group of speakers is one hour, usually, of which most of the time
is spent in discussion with all the participants in the room.
People will get up and take the podium or the blackboard to make
a point, and then sit down. Conversations go in all directions.
The speaker is not necessarily at the center.

Sometimes pandemonium results. Mostly it doesn't. I think that
one reason it runs so well is that we don't "accept papers" or
"review submissions." Instead, during the first session
(Wednesday evening) we haul out a blackboard and schedule the
meeting. The time blocks are drawn out on a calendar, and the
chairman says "OK, who wants to talk about what, and when, and
how long?" There is the whole time pie in front of everyone,
clearly finite. As it gets sliced, it becomes apparent that not
everyone can talk as long as he or she pleases, and that room
must be left for latecomers and accomodations must be made for
people who have to leave early, and so on. Often the schedule is
revised during the meeting to reshuffle speakers, bring up new
topics that have caught everyone's interest, and so on. The whole
initial scheduling process takes less than half an hour.

The interesting thing about all of this is that everyone wants to
hear what the others have to say. It's assumed that the people
present understand control theory well enough not to be snowed by
the discussions, but we make no distinctions in that regard.
Somehow we have never had any serious problem with an ignorant
newcomer going off on a tangent -- few people have tried that,
and it's quite clear that the group won't put up with it anyway.
Everyone is aware that this brief time together once a year is
precious and that it's important to hear from as many people as
possible, whether they're standing up in front at the podium or
just holding forth from a corner of the room. There is great
mutual respect shown, and even newcomers catch on fast, both as
to the implied obligations and as to their implied right to speak
up. There is absolutely no prestige in being a speaker; the
speaker basically kicks off a discussion on a new subject. One
year it turned out that I didn't speak at all (from the position
in front of the group, that is). Most years I don't have much to
present in any formal way, although I may have some remarks for
guests at the beginning.

One reason for the success of the meetings, I think, is that we
have adopted the Gordon Research Conference format. Morning
sessions run from 9:00 AM to lunch, afternoons are totally free
(no sessions), and there is an evening session from 7:00 to 9:00
or whenever we get thrown out of the room. The bull sessions then
continue through a large part of the night. You have no idea how
welcome that afternoon off is (well, you do, Cliff, as you've
been to Gordon conferences). Of course people congregate in the
computer room for demonstrations, or in the auditorium for
impromptu tutorials or expansions on talks for people who want to
hear more details, or on the tennis court, or the hiking paths,
or down the hill in Durango. None of that is scheduled; you do
what you please in the afternoon. That's when the special
interests can get together to talk shop.

I don't know what will happen to this way of running the meetings
when more than about 50 people attend (the largest attendance so
far, of real participants not counting guests, has been about
30). The Gordon Conferences invite about 100 people, and that is
really straining the concept of having all plenary sessions. What
the Gordon people have done is to run many conferences, limiting
each to 100. Maybe that's what we'll have to do. I don't want to
think about it; the way we do it now is so great that I hate to
contemplate the end of it that will surely come. Unless someone
comes up with an absolutely brilliant idea.

The one thing I DON'T want to see is a "scientific meeting." I've
been to enough of those to have had my fill of boredom.

Hope you'll decide to come to the '93 meeting. Just register when
the announcement comes out -- that's all the qualification you
need. You certainly know enough about control theory to find the
meeting interesting, and to contribute.

As to cross-posting, I'm of two minds. When people ask me about
coming, I tell them that of course they're welcome, but they
won't get much out of it unless they've done their homework on
PCT. This isn't a meeting you attend to find a job or increase
--------------------------------------------------------------
Greg Williams (930126) --

RE GAIA:

Good on George Williams. I had a heck of a time with Lovelock --
he's one of those who adamantly rejects the concept of a "set
point". His Daisyworld is characterized by two equations, one
showing the dependence of temperature on the number of daisies,
and the other showing the dependence of daisy proliferation on
temperature. Plots of these equations intersect at a point, which
is what Lovelock claims is all that is necessary for control of
the variables.

Of course he's right -- but I couldn't convince him that the set
point is inherent in his system, which is really a control
system. Oh, well.

It occurs to me, now that you bring up the subject, that there is
a way of getting a high-gain control system out of a lot of low-
gain systems operating in parallel. An example is the real human
arm control system. No one control loop can have much effect,
operating through only one small muscle fiber. But when you put a
lot of such systems together, all sensing the same force (by
different paths) and all contributing a little to the same output
force, the result is multiplication of the loop gain by the
number of systems in parallel.

For GAIA-like concepts, you get a different version of this
effect. Suppose that plants have tight control over, say, the
oxygen concentration in the immediate vicinity of their leaves. I
don't know how tight this control actually is, but let's pretend
that it's perfect. Each plant keeps the oxygen tension at the
surface of its own leaves at exactly the level it wants.

So a disturbance of that oxygen tension will be completely
canceled by variations in the plant's oxygen output. But consider
the oxygen tension 1 millimeter away from the leaf. Here, the
oxygen tension is affected by disturbances, too, but is less
affected by variations in the plant's oxygen output because of
diffusion in three dimensions. So the oxygen tension 1 mm from
the leaf would be less controlled than that right at the surface
of the leaf.

We can see, then, that there is a field of control surrounding
the plant, with oxygen tension showing less and less resistance
to disturbance as we sample the air farther and farther away from
the plant. For a packet of air at sufficient distance from the
plant, we would see no resistance to disturbance at all.

Now, start increasing the number of plants. As the density of
plants gets greater, the mean distance of any arbitrary packet of
air at ground level from any plant decreases, and the amount of
stabilization of oxygen tension increases. To put it another way,
the net apparent loop gain for controlling the oxygen tension of
any arbitrary packet of air increases as the density of plant
distribution increases. At some packing, I would guess that
control of all packets of air from ground level to some
considerable altitude would be essentially as good as it is at
the surface of each plant's leaves, and the loop gain would be
about the same as that for a single plant controlling oxygen
tension at the surface of its own leaves.

All of this leads to an interesting and disturbing conclusion. If
this sort of GAIA-control exists, then the loop gain for
controlling essential environmental variables depends on the
number of plants. The problem is that we see very little effect
on the behavior of a control system over very large changes in
loop gain, as long as the minimum gain is high enough. So as
disturbances increase, the opposing outputs of the individual
control systems also increase to cancel the effects, and we see
little change in the controlled variable. It looks as though the
disturbances are ineffective, and it continues to look that way
even through the number of control systems acting in parallel is
drastically decreasing.

At some point, however, the loop gain will fall far enough to
begin allowing the disturbances to have significantly larger
effects. The unfortunate thing is that this will occur only when
the whole composite control system is right at the point of
complete failure. The fact that control is involved conceals the
progressive failure of the system until there is no margin of
safety left at all. At that point, a disturbance that could
easily have been counteracted before now can drive the controlled
variable far enough from its community-wide mean set point to
result in immediate disaster.

Just think in terms of deforestation. As the forests of the world
are burned and cut down, their contribution to control of CO2 and
O2 levels decreases. The composite loop gain, however, is still
high enough that our pouring of CO2 into the air still has only
very small measured effects. However, the loop gain is dropping
at the same time that the disturbance is increasing. Sooner or
later control will begin to fade away, and when the gain gets
down to 5 or 10, the growing disturbance will rapidly have more
and more effect. Just plot G/(1+G) as G decreases from 100 to 1.
By the time we start to see large effects, it will be much too
late to do anything about them. And the effect of any large but
still probable pulse of disturbance may already exceed the
capacity of the GAIAn control systems.
--------------------------------------------------------------
Bruce Nevin (930126.0831) --

... control systems of a given order (e.g. cells) are
necessarily oblivious to the means by which they implement ECSs
and their perceptual signals, hierarchies of perceptual
control, etc. in control systems of the next higher order (e.g.
animals and plants). The same principle applies to the Gaia
hypothesis.

This is a nice principle. It probably has an essential
relationship to the concept of mobile awareness, and the idea
that you are never aware OF the level of control FROM which you
view the lower levels of experience. Control systems at any level
-- which you have extended down as far as they will go -- are
necessarily oblivious to higher systems. The only indication that
a higher system might exist is in the sense that some states of
perceptions just seem "right." Is this the origin of the idea of
conscience, or superego?
-----------------------------------------------------------
Bob Clark (930124 or so) --

Best to all,

Bill P.