Newsgroups: comp.ai.alife
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!Germany.EU.net!EU.net!uunet!tandem!oobii
From: oobii@cpd.Tandem.COM (oobii@cpd.tandem.com(Tim McFadden))
Subject: Computational Hardness and Expense of AL
Message-ID: <CzorFs.FsD@cpd.tandem.com>
Originator: oobii@runt2
Keywords: artificial life, computaitonal complexity, automata, evolution
Sender: news@cpd.tandem.com
Nntp-Posting-Host: runt2
Organization: Tandem Computers, Inc.
Date: Tue, 22 Nov 1994 20:26:15 GMT
Lines: 170
X-Disclaimer: This article is not the opinion of Tandem Computers, Inc.


Life is hard and expensive.
AL will be hard and expensive.

* this post is about 1000 words long *

The point of this post is to try to give some practical
estimation of the difficulties of producing ALife.

We argue:

1) AL is hard, because it must abstract from biological 
systems, which is hard because they are so wiggly and
complicated.

2) AL is expensive, because the abstraction is hard,
much detail of biological systems may have to be included,
which gets expensive, esp. in real-time. 

3) there are computational reasons to support the above.

It centers around the question, "how well can AL models
abstract biological life behavior?"   

This is important in the following sense.

Spose you are designing or changing an AL machine.

+---------------
|
|  Push AL machine on the stack
|  We have to talk about nanotech first, then we'll pop
|  the stack back up to AL.
|
|  Hypothesis:
|
|  Nanotechnology, as undertood now,  will not give 
| real-time
|  atom-by-atom control of living systems. Nor can it 
| operate  with arbitrarily complicated bio-macromolecules.
|
|  If you accept this or don't care, pop the stack now.
|
|  Argument:
|
|  Nanotechers, as in Drexler's book, "Nanotechnology",
|  have opened up a new field which will allow extensive 
|  atom-by-atom manipulation of matter. What has not been 
|  discovered is
|  how to keep large hydrocarbon molecules together while
|  nano-machines are doing atom-by-atom manipulation 
|  and/or how to coordinate the individual manipulators. 
|  In other words, 
|  the molecules themselves would be in control much of 
|  the time and not the nano-machines. Therefore the space 
|  of control is not that of all possible molecular 
|  arrangements. At this time,
|  the project of building any random, large moleclue at the
|  biochemistry level is science fiction. 
|
|  Anything is possible, but I'm not going to bother with
|  more than one layer of wild supposition at a time here.
|
|  Pop!
|
+------------------

Because we are using some material which will not be 
controlled at the nano-level, except for those chores which
don't need a lot of cooperation and molecular complexity,
we will be working at some higher level analog, which may
be OK in itself for some really useful chores.

That is, we have to practice some reverse reductionism.

+---------------------------
|
| Push
|
| Here *reductionism* means, roughly,  that cellular 
biology explains our
| body-minds, biochemistry explains cells, and physics 
| explains chemistry.Well sort of, there have been found 
| no rogue cells which don't obey
| the laws of physics, but we haven't even started coming 
| close to derivng all of animal behavior and human 
| experience from physics. This is a distinct problem; each
| layer seems to have its own rules, which don't break the
| rules of the layer below, but which are very hard to 
| derive from below, *reverse reductionism*.
|
| Pop
|
+--------------------------  

In the case where the number of possiblities for an 
operation on the analog material (which might even be 
software) become very large, or arbitrary, or must be 
decided withreal-time constraints, the issue of 
"how well the AL model abstracts" is important.

At some point, as in the case above, there might be a 
large number of possible combinations at the analog level.
Any one of these combinations may work under the AL model 
used, however, if the modeldoesn't capture the right 
biological details, the AL may die.It is at this point 
where the biological life behavior  might just  depend on 
the way some molecule does things or some evolutionary 
holdover hack. There is no law of nature  which sez that 
it neatly separates stuff into our reductionist levels.

Much death avoidance may be dependent on expensive 
complexity, at leastfrom having a large number of 
components. Real-time AL may be esp. expensive because of 
the parallel processing involved. Chaosexplains a lot of 
complex behavior, but not all of it.

There are also two computational reason to believe that AL 
will be hard and expensive.

First is Chaitin's theorem, which limits how much a given 
system can say about another, depending on its size. See
Rucker's "Mind Tools."

Second is Wolfram's Hypothesis (which I have extended from 
his comments on cellular automata):There are some systems 
(like celllar automata, for example) whose behavior cannot 
be predicted, all you can do is watch them evolve.

These apply directly to the case of an AL analog model. An 
AL model significantly smaller than a biological system 
may not be ableto represent the system well enough in the 
sense of Chaitin's theorem. To apply Wolfram's Hypothesis, 
there may be no way to find out whether a given design 
works other than testing it from allits possible 
configurations. Even biological evolution has not had time 
to try a small fraction of the combinations around the 
path getting to our current state.

If there is to be a direct analogy with biological life 
and AL, then the stakes are even higher. According to S. J.
Gould in his "Wonderful Life", there were more phyla a 
billion years ago than there are now. The major leaps in 
evolution occured whenhuge numbers of creatures were 
decimated, after which a much smaller gene pool was able to
expand into the empty eco-niches.All this confirms 
Schroedingers "What is life?" definition:a neat way to get 
from a high entropy state to a lower entropy state.
 
Given the fact that artificial intelligence has not 
produced much in the way of *abstract laws of intelligence*
in the last thirty or so years, there is little reason to 
hope that AL will progress any faster. Here *abstract laws 
of intelligence* refer to practical,everyday intelligence 
and not math/logic. Math/logic is not yet ready for 
real-time. Books like AL IV, however, show that we going to
learn a lot!

Just to show I'm paying my dues and not just complaining, 
I'm slowly
working on a hacked C version of Holland's broadcast 
language from his
"Adaptation .." book, Ch. 8. I hope GA will help AL.

thanks,
tim




