[Paleopsych] genetics as an intelligent system
Werbos, Dr. Paul J.
paul.werbos at verizon.net
Thu Nov 25 15:54:51 UTC 2004
Having spent all of about 5 minutes of real thinking about the
questions Greg raises... enough thoughts pop into the mind that
I doubt I have time to type them all.
First -- one of the reasons why the establishment may find it difficult to
address the questions is that they are very limited in this case in the degree
of mathematical abstraction they use. It's a kind of qualitative limitation
in how mathematical thinking is used...
The neuroscience establishment (which I know much better) has been struggling
with similar limitations... maybe a bit harder and a bit more successfully
It is interesting to ask: now that we have learned a lot about intelligent
systems in GENERAL..
and now that some of us have a reasonable first-order idea of how this maps
into the brain..
what about the genetic system?
Forgive me for using a new term which sounds a bit pretentious --
The prefix "meta" has been badly misused lately, but in this case -- what
be a good single word to refer to the idea of a genetic system which
"learns to learn"?
Part of Greg's message is that we need to understand metagenetics in order
any sense at all of 97 percent of the human genome. That's a big step, a
and an important one. That idea has existed in some form for a long time,
give it a snazzy new one-word version and focus more attention on it is
still a good step.
But is there more going on here?
A natural way to interpret "metagenetics"... is to think of ... a kind of
second-order system which is
still designed to perform the same basic functions people think about in
or evolutionary computing: maximizing some kind of fitness function U(w) as
a function of a set
of weights or parameters w. (Parameters could be anything from body
to behavioral response characteristics .. to anything...) A sophisticated
way to explore the space
of possible .. genotypes. Back in 1999
(at a plenary talk at CEC99, the IEEE Conference on Evolutionary
Computing), I challenged
people to send me proposals to address a more interesting computational task:
to design systems which LEARN to do stochastic search to maximize U(w,X),
where w is as before,
and X is a set of observed variables available to enhance performance. I
have reiterated this in many
talks and tutorials... I call this task "Brain-Like Stochastic search."
It's very important in
engineering, for example; if we use evolutionary search to find the best
possible chip design
for some task.... it would be good to represent DIFFERENT chip design tasks
by a vector X,
and then use a system which learns to do better on chip design task in general.
For now, it's enough of a challenge to treat X as "exogenous," but someday
one could advance to
Now: one COULD follow up on Greg's questions by asking whether we can model
the genetic system
as one which implements "Brain-Like Stochastic Search" with dynamic X. We
may ask: to what
extent does this richer functional interpretation become essential to
understanding the basics
of what we really see with the genome?
Now -- a certain degree of "stockpiling" can be important even in that
But another question occurs to me today: would it make any sense to go even
and evaluate the possibility of a still higher level of intelligence in the
In brains, evolutionary computing is certainly far from enough, in any form.
(And I suppose I know a few key things about Edelman's work that Edelman
In a word -- TIME. Optimizing results INTO THE FUTURE, with anticipation or
(both explicit and implicit), is absolutely central to how brains work.
Could there be anything like THAT in the genetic system? I wonder...
Various types of memory are essential in brains. There are many levels of
stockpiling in brains.
Could any of THAT be transferrable to the genetic case?
This morning I was thinking more about Einstein than about Greg... but I
suppose such thoughts would be
off-topic on this list. Oh, well.
Best of luck,
More information about the paleopsych