COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	∂22-Jan-79  1001	Greg Harris at CMU-10A 	Re: update 
C00007 ENDMK
C⊗;
∂22-Jan-79  1001	Greg Harris at CMU-10A 	Re: update 
Date: Monday, 22 Jan 1979 1259-EST
From: Greg Harris at CMU-10A
Subject: Re: update
To:   Doug Lenat <DBL at SU-AI>
Message-ID: <22Jan79 125904 GH30@CMU-10A>
In-Reply-To: Doug Lenat's message of 18 Jan 79 17:29

Doug,

Finding plausible generation methods by examining the DNA code sounds
like an incredibly long inference chain:  the generation, after all,
starts with the cross-combinatorial gene mixing which, when you think
about it, is incredible in its ability to preserve species traits at
all; then there is cell-specialization; and finally, the functionality
of the cell itself:  all the DNA is present, but only certain
"interpreters" are present, resulting in further selectivity in what
will be synthesized.  Certain privileged cells get to propose the next
iteration of cross-combination candidates.  Given that the "test" of
natural selection is based on long-term aggregate performance of this
entire cascade of selective generation, I see little hope in looking
for even longer-term information storage in, of all opaquely
interpreted places, the DNA code itself.

My alternate suggestion is that you try to find out what constrains the
variability over a single generation:  where in this code does it say
that eye color varies, but number of eyes does not?  Where is bilateral
symetry called for, and where excepted?  A bit of embryology might be
informative, since many of these things are fixed on the cellular scale
during gestation.  In particular, look for combinatorial choices that
cease to be combinatorial:  what is a dominant trait, expressed as the
behavior of "generation interpreters", and does this relate to any
possible mechanism that eventually transmits a combinatorial trait as
fixed forever after?

Instead of looking for a pun on the word "generation", why not regard
this as a challenge about the nature of interpreters:  is there some
kind of interpreter used in evolution that is unknown to us in computer
science, capable of gradual adaptive changes AND progressive species
splitting?  changing its tools while coming to depend on them?

To falsify the hypothesis that all the direction comes from the
environment, an AI approach might be to build a multiple interpreter
architecture that makes "superior" traits no longer optional but an
obligatory (or even just preferred) part of transmission to future
cycles, in a manner computationally analogous to evolution.  Your tasks
then are to draw enough but not too much inspiration from biology, and
to find some area of behavior on which an evolutionary architecture can
operate and make meaningful progress.  Oh, yes:  also, to conform to
cannons of taste in two fields at once.

You sounded serious, so I answered seriously.  Still, my feelings about
advice offered in this way are those of caution.  Maybe what Nature
does WOULD strike me as a total hack; but I expect otherwise.  Nature
is sophisticated, and frequently offers good advice of its own.

Greg