teaching; genes; HPCT and prediction

[From Bill Powers (960227.0600 MST)]

Lars Christian Smith (960226 21:20 CET) --

     It could be that it is changes in the brain that brings about
     changes in social status, but what is observed is usually larger
     body size. E.g. in the sea wrasse, if a male wrasse is caught, the
     largest female in the territory changes gender and becomes male.
     Visible changes in other species is often associated with dominance
     or victory in mock combat, which again is associated with body
     size. You may of course argue that it was a change in the
     organization of the brain that allowed the animal to catch more
     prey, eat more, and therefore become larger, but would you get any
     additional insight from this argument?

I'm not so much concerned with getting added insight as with finding a
believable model to explain the facts that are observed. It's very hard
for me to understand how a gene can know that it's inside the largest
female wrasse in the vicinity, and make an adjustment of gender on that
basis -- all in the lifetime of a single organism. Genes have no organs
that can sense the state of the external world, as far as I know. They
can have no knowledge of how the body and brain have developed as a
result of interactions with the current environment. The "Little Man in
the Gene" is no more plausible as an explanation than the "Little Man in
the Head."

What I _can_ believe is that genes may evolve which can construct an
organism that can learn from its interactions with the world, with an
organization that makes control processes especially easy to learn, that
can sense both external and internal variables that are important for
controlling other variables that affect the organism's survival, and so
on. I could accept the idea that genes are control systems concerned
with the immediate chemical environment, and that one result of their
operation is the building of more control systems at many levels. But to
go from the gene to the behavior of the whole organism in one step is,
to me, a vast oversimplification with no justification but laziness.

The phenomenon you cite, it seems to me, depends on the presence of
structures that can sense the external world, recognize aspects of it,
and take action to control what is sensed. The action can quite easily
extend to changes in biochemical functions and organ functions, which is
what a gender change must involve in organisms that can change gender.
For example, the female wrasse may sense the concentration of certain
pheremones in the water. As long as that concentration matches an
inherited reference level, the control system experiences no error and
takes no action. But if the concentration drops below a certain level
(when enough males disappear from the environment), the resulting error
operates a mechanism for producing the pheremone, and that mechanism,
when active, also converts the fish to functioning as a male. All the
female fish begin to respond in this way, but the largest one produces
the most pheremone first, and provides enough pheremone to correct the
error in _all_ the fish, so the other females do not complete the
transition, or go back to being male before the transition becomes
irreversible (you didn't say whether introducing a number of males would
cause a conversion of smaller males back into females).

I have proposed a mechanism like this to account for the specialization
of cells into organs during the initial growth of an embryo. All cells
contain genes that could make any type of cell. Suppose that the gene
contains reference signals for what, in the adult organism, we see as
the products of one of the organs. When one cell starts emitting a
specific product, the effect is to reduce the error signal relating to
that product in all cells. So the cell that starts emitting it first, or
the cell mechanically positioned so its products affect the most other
cells, shuts down production of that substance in the other cells while
simultaneously, by dividing, increasing its ability to produce that
substance. This unstable situation can end only with one cell becoming a
group of cells that takes over the function of producing one substance,
with similar genes in all other cells being shut off by the fact that
the required substance is already being provided somewhere else.

This is, of course, only a rough sketch of an idea, but it shows how
genes could produce cells which then become an organization that acts
back on the genes to construct a further layer of organization. The
adult organism results from a series of developments like this, with the
effects of genes becoming less specific with the addition of each new
layer of organization, and interactions with the current world becoming
a more and more important influence on the final organization of the
adult individual.


Shannon Williams (960226.17:00 CST) --

In replying to Rick, you said

     If a person is presented with a problem (he is having difficulty
     making some perception match its reference), and I present some
     partial solution to the problem, I want to be able to predict if my
     partial solution gives him enough information to solve his problem.

I think you are asking for an ability to predict that assumes a
uniformity among people that doesn't exist. Reorganization, at least as
I visualize it, will produce _any_ solution that works, not a particular
solution. Most problems have multiple solutions (except for problems
deliberately constructed so as to have only one). The most you can do is
to study the problem (rather than the person) to determine what the main
possible solutions are. But a person presented with that problem,
together with any number of hints, might end up with any one of the
solutions, or might end up with no solution (even though somebody else
might come up with one).

HPCT isn't really concerned with finding solutions to problems. It's
more concerned with defining the types of solutions that are likely to
be found. For example, if an organism knows how to head for a goal like
a piece of food, a certain type of directional control is implied. But
if an obstacle appears between the organism and the food, the same
organism will simply butt its head against the obstacle and never reach
the food. Another type of control is necessary, of a higher level, which
can operate by _changing_ the immediate goal of movement. Instead of
heading directly toward the food, the higher system constructs a path
around the obstacle by varying the immediate directional goal, and thus
gets to the food anyway. Yet directional control is still resistant to
disturbances that tend to alter the direction away from the immediate
intended direction, so control is going on at both levels. All the
levels in HPCT are supposed to have this nature; they use existing
levels as a means of controlling a higher-level version of the world,
while the lower-level systems still act as autonomous control systems
relative to the goal they are being given.

So far the only mechanism in PCT or HPCT that has been proposed by me
for solving problems is random reorganization. I know that there must be
other mechanisms, such as the ones involved in adaptive neural networks,
but so far nobody has merged those ideas into either PCT or HPCT in the
form of a working model. Hans Blom has come the closest to doing this,
with his Kalman Filter approach, and Martin Taylor has proposed
hierarchies with local reorganization in them. I have tried out other
means of local reorganization. I presume that you're working on still
another approach. But so far all these ideas are only preliminary
sketches of a whole system.
Chris Cherpas (960226.1842 PT) --

It helps to know more about what you're doing! I love it.

     A big question I'm trying to resolve now is how far to go with the
     PCT hierarchy (e.g., Robertson & Powers, 1990). Should I actually
     try starting with controlling intensity (necessarily in some
     sensory modality, though), then move gradually up through
     sensation, configuration, etc.?

With computers you're pretty much limited to vision and sound; it's hard
to test kinesthetic controlled variables such as force and position,
because you need special equipment to introduce disturbances. Taste and
smell are even harder. But sight and sound cover a lot of territory.

As to the levels, I always start gnawing my lips when people take them
as given. Maybe they have some usefulness, but they are UNTESTED! I
understand that you're trying to develop applications, but wouldn't it
be possible to devote some parts of your learning programs to gathering
data? We really need some research on the idea of levels of control.
What classes of variables do people have to control in order to control
other classes of variables? What controlled variables have to be changed
in order to maintain what other variables against disturbances? What
perceptions can't exist unless what other perceptions already exist?
These were the questions I was trying to answer in developing the
definitions of levels, but the source of information was just subjective
examination of my own and other people's behavior in informal settings.

Perhaps you could use my levels as a starting point, but I have to tell
you that Rick and Tom and I and others have had great difficulties in
proving experimentally that multiple levels of control even exist. The
main difficulty is that you can always collapse a multiple-level control
system model into a single system with equivalent properties. I haven't
been able to think of a clear experimental way to show the separate and
simultaneous existence of different levels of control. We have some
anatomical evidence of clear levels of control, but only at the bottom
one or two levels. Information about what higher centers in the brain do
is almost nonexistent; it's still mostly guesswork.

Maybe your field offers a way. I presume that children generally learn
the levels from the bottom up (with a lot of backing and filling). There
should be a stage of development where the child is more skilled at one
level than at a higher one. If so, you could measure the ability to
control at the different levels, and see where the systems become
unskilled (or even better, where they are absent). But this implies that
you know what the levels really are, and that they're of similar types
in everyone. It's pretty clear that a child has to be able to control
the position of a pencil-point in order to control the appearance of a
drawn shape, and has to be able to recreate drawn shapes in order to
write specific letters -- but how do we prove this?

I can't answer these basic questions. Research is needed. If you use the
levels exactly as I have defined them, you'll be taking a chance; I
can't say how much of a chance.

You evidently have a LOT of experience in this field. You will probably
get farther toward answering these questions than I could do. I hope
you'll try.
Rick Marken (960226.2130) --

A brilliant analysis of the problem of distinguishing performance from
learning. When you use all of your horsepower, the result is impressive.

You point out one of the questions that AIers have been vague about. Are
we trying to design SOME way of carrying out intelligent-looking
behavior, or are we trying to model the HUMAN way of doing this? The two
problems are not equivalent, although solving the first can give us
possibly useful ideas about the second (but can also throw us off the
Best to all,

Bill P.