Evolution of the individual control hierarchy

CSGnet Archive CSG2003
1

[From Bill Powers (2003.02.04.1916 MST)]

This is Runkel on 4 Feb 03.

How is it that I have seen on the CSGnet no comment on the article in
Scientific American for February 2003 (vol. 288 no. 2), page 52 ff.,
"Evolving Inventions" by Koza, Keane, and Streeter?

Sir, I am delighted to say that I can immediately answer your question. I
don't read Scientic American. That's not quite as good an answer as Mark
Twain's, but it will have to do.

The article is about constructing electrical circuits by growing them,
so to speak, over a series of trials (corresponding to generations),
selecting those variations according to a standard set by the grower
(the engineer). In the organism, the standard could be zero disturbance
to the reorganizing system.

This is called adaptive neural networks, or something like that.
Self-organizing systems is another name. All the examples of it that I have
seen (not many) cheat: the programmer knows what the goal is, and allows
the deeveloping organization to survive if it simply changes itself in the
right direction (without actually achieving the desired goal). A model
cockroach, for example, is allowed to survive if it makes a move toward
food. As a model of evolution, of course, this is not good, because real
evolution requires that you actually survive, and doesn't give credit for
partial answers.

It strikes me that this evolutionary procedure could "grow" the levels
of the control hierarchy in the individual. That is, maybe overall
inner conflict (let's say) is minimized not by logical combinations of
inputs and outputs such as we easily conceive, but by sheer brute
selection processes among millions of circuits and trials. Could that
process leave behind the sort of encompassing control levels that Wm.
Powers conceives?

That's what I call reorganization, driven by intrinsic error and carried
out by random variation and selective retention.

Are Koza, Keane, and Streeter unwittingly running models of the growth
of the control hierarchy?

See my comment to Bill Williams about what to give credit for.

Best,

Bill P.

2

[From Bill Powers (2003.02.07.0919 MST)]

Phil Runkel (2003.02.07) --

>I did not make it clear enough. The article is not mere speculation.

Their selective retention process or "genetic programming" has
apparently constructed some number of circuits that do work as intended.

The problem I was expressing was that in other demonstrations of this sort
that I have seen, the goal was known in advance, and the programmer in
effect let his program know whether it was getting closer or moving farther
away from the goal. As you know from my World Futures article, I have no
objection to this as a model of evolution -- but it's not what most other
people call "natural selection." It is very strongly guided evolution.

Natural selection is like this. You're walking along the street, and
suddenly someone walking beside you gives you a painful clout in the ear.
"What was that for?" you complain. No answer. So you go on walking, and POW
there's another clout in the ear. "Jeez, what am I doing wrong?" you ask.
No answer. So you try walking in a different direction: POW. You try
skipping. POW. You try limping. POW. By this time you can hardly stand.
POW. You're down. POW POW. You're losing consciousness. POWPOWPOWPOWPOW.
You're dead.

Evidently you were not fit to survive. But what did you learn? What did
your descendants learn? Nothing. The next person this is tried on might get
a little farther, but before learning what the trouble is, POWPOWPOW and he
is dead, too. You eliminate wrong moves, but does that leave only right
moves, or does it just reveal an endless supply of more wrong moves?

It would take rather a long time to figure out that the person walking down
the street will keep hitting you until you pull out a pocket watch and say
"Oh, my stars and whiskers!"

On the other hand, the person walking with you might relax the rules
somewhat, and reward you for making any move more or less in the right
direction by letting you survive a little longer. In this way that person
can guide you toward the result he wants, as in a game of hot and cold.
When that kind of guidance is allowed, progress and arrival at the
goal-state is essentially assured. B. F. Skinner called this method
"shaping," although he used systematic selection from systematic variations.

The examples of "evolutionary" models that I have seen have all involved
this kind of guidance, built into the program. Something in the program
knows what is wanted and can see whether the changes are leading toward it
or away from it. With that kind of perceptual information available, as we
know from our experiments with the E. coli model, random variations of the
output will eventually achieve the goal, through selective retention of
changes in the right direction.

But it's not evolution, or rather Natural Selection as normally conceived.

Mary and I are doing well; we actually eat quite a bit of food in addition
to the pills. But let's not bore the youngsters with the favorite topic of
Aged Persons.

Best,

Bill P.

3

This is Runkel on 4 Feb 03.

Good morning. If you read this in the afternoon, just erase the word
"morning" and type "afternoon" in its place. If you read it in the
evening, erase the word "afternoon" and type "evening" in its place. If
you do not read it until the next morning, just leave it as it is.

How is it that I have seen on the CSGnet no comment on the article in
Scientific American for February 2003 (vol. 288 no. 2), page 52 ff.,
"Evolving Inventions" by Koza, Keane, and Streeter?

The article is about constructing electrical circuits by growing them,
so to speak, over a series of trials (corresponding to generations),
selecting those variations according to a standard set by the grower
(the engineer). In the organism, the standard could be zero disturbance
to the reorganizing system.

It strikes me that this evolutionary procedure could "grow" the levels
of the control hierarchy in the individual. That is, maybe overall
inner conflict (let's say) is minimized not by logical combinations of
inputs and outputs such as we easily conceive, but by sheer brute
selection processes among millions of circuits and trials. Could that
process leave behind the sort of encompassing control levels that Wm.
Powers conceives?

Are Koza, Keane, and Streeter unwittingly running models of the growth
of the control hierarchy?

Hey, Cziko!

��Phil Runkel

4

Bill:

I did not make it clear enough. The article is not mere speculation.
Their selective retention process or "genetic programming" has
apparently constructed some number of circuits that do work as intended.

Part of your reply was: "That's what I call reorganization, driven by
intrinsic error and carried out by random variation and selective
retention."

Yes, me too. That is the core of what I was inquiring about. It seems
to me that is what those authors claim to be doing. I am not sure
whether the combinations they tried were generated randomly, but they
used a lot of combinations that were winnowed by whether they reduced
overall error.

You referred to your comment to Bill Williams. Yes, I am familiar with
those overweening claims, and I looked out for them in the article. I
think it possible that the authors have avoided those errors. They do
insert the engineer's reference signal, but in this case I think their
reference is only somewhat more specific than your intrinsic reference
signals. Their signals are certainly SOME more specific than yours,
because they set a criterion that the resulting circuit is to perform a
particular function; for example, a "negative feedback amplifier" (they
note that Black produced one in 1927, and the one "evolved" by their
"genetic programming" turned out very much like his). Other examples:
"second-derivative controller," "sorting network," "voltage-current
converter," "tunable integrated active filter" and analog computational
circuits for the square, cube, square root, cube root, logarithm, and
Gaussian functions. But those inserted criteria seem to me to be
somewhat like we could expect to be generated by the successive layers
of the hierarchy to reduce conflicts produced by the actions called
forth at the lower levels.

I am wondering whether the experimental work of these authors could be
helpful in putting more specificity onto a model of the growth of the
hierarchy. They are not, by the way, appealing to averages over
individuals, human or modeled. And no tables or other "knowledge" are
inserted into the "genetic programming." They do specify subfunctions
in the circuits, subfunctions that can be selected and retained or not.

You have wondered what the rules for combination could look like above
the first couple of levels. I am asking whether the evolutionary
selection and retention might generate the "rules," those being somewhat
different in every individual, though with recognizable similarities,
the similarities being due partly to what culture makes available and
partly to what the inherited intrinsic signals make available or permit.

As to the Scientific American, I am not surprised at your disdain. It
seems to me that the worthwhile articles steadily become fewer. At
each notice of time for renewal, I wonder harder whether I want to
cancel. I suppose I am suffering from nostalgia.

I hope you and Mary are OK. We are both kept functioning by various
pills. They do not taste as good as food, but they are less trouble to
prepare. And they are more fun than surgery.

��Phil R.