Copy Link
Add to Bookmark
Report
Alife Digest Number 017
Artificial Life Digest, Number 17
Friday, April 20th 1990
Issue's Topics:
multiple genotypes per phenotype
Science article
emergence and "alive" simulations
binary or other coding within GA's
----------------------------------------------------------------------
Date: Tue, 17 Apr 90 09:43:02 PDT
From: xdos!speedy!doug@apple.com (Doug Merritt)
Subject: multiple genotypes per phenotype
I'm surprised to hear any controversy on the subject of multiple
genotypes per phenotype; I had thought the evidence was overwhelming
on this, even prior to the multi-sequence/same folded protein argument.
Consider first the classic biological view of phenotype -- a massively
complex macroscopic phenomonen, seperated by many levels of scale
from the microscopic genotype. One might run across two individuals
who appear macroscopically identical, but in general one would assume
they are genetically distinct. Or the structure of the human eye, said
to be extremely similar to that of the octopus. Yet given the extreme
distance to a common evolutionary ancestor, one should assume the
genetic encoding for this structure is different in the two species,
in the absence of evidence to the contrary.
In an AL context, consider the experiments I did with breeding Tic Tac Toe
players around 1985 (first with finite state machines as "bodies", second
with simple neural nets). The smaller the gene-sequence, the longer it
took to create "good" players. The genotype was a search space, and
seemingly paradoxically, the larger it was, the faster good solutions
were found (in units of number of generations). The longer gene sequences
had many possible ways of creating good players, so some subset of them
were found relatively quickly. The smaller sequences had very few ways,
which amounted to a search for the global optimum for the sequence, which
of course took a long time.
In this example, I'm claiming that the phenotype is the external appearance
of the critter, which was its playing behavior. And of course there are
many different possible state machines or neural nets which could give
rise to identical playing behavior. And thus many different "genotypes"
per "phenotype".
BTW it was interesting that replacing the finite state machine
mechanism with a neural net and rerunning Genesis showed no real difference
in external characteristics (e.g. improvement curve over N generations).
I was startled to see the rapidity with which the system discovered and
took advantage of a bug in my implementation to prosper without actually
playing the game fairly (i.e. the discovery of cheating), and when I fixed
that, next they conspired to win by collusion due to a problem in my
tournament system (i.e. rackateering). But I digress.
Doug
Doug Merritt {pyramid,apple}!xdos!doug
Member, Crusaders for a Better Tomorrow Professional Wildeyed Visionary
------------------------------
Date: Thu, 19 Apr 90 22:00:18 EDT
From: RAY <ray@vax1.udel.edu>
Subject: Science article
April 19, 1990
Just in case anyone missed it, there was an article in Science recently
that dealt somewhat with AL. It talked a lot about the work of Stuart
Kauffman.
Waldrop, M. Mitchell. 1990. Spontaneous order evolution, and life.
Science 247: 1543--1545.
It appears that there will be an article in the Sunday April 22
(Earth Day) issue of the ``Globe and Mail'', Canada's equivalent of the
New York Times and the Wall Street Journal combined, ``the serious one
that is distributed nation wide''. The author, Steven Strauss seems to
be very much intrigued by AL and will probably give a thoughtful treatment.
Tom Ray
University of Delaware
School of Life & Health Sciences
Newark, Delaware 19716
ray@vax1.acs.udel.edu
302-451-2753
------------------------------
Date: Thu, 19 Apr 90 03:36:23 edt
From: "Peter Cariani" <peterc@chaos.cs.brandeis.edu>
Subject: emergence and "alive" simulations
I'd like to submit a couple of basic questions to the ALife list for general
discussion and clarification:
1) What distinguishes an "emergent computation" from a nonemergent one?
2) What distinguishes an "artificially alive" simulation from one that is not?
3) Do these distinctions depend upon the formal behavior of the computations
carried out or are they dependent upon the interpretations given them by an
external observer/programmer?
I think until we come up with a clear definition of "life" and the means of
distinguishing living organization/behaviors from nonliving ones, artificial
life will be seen as a specialized branch of computer programming rather
than an autonomous field of inquiry with deep implications for the rest of
science.
The best definitions thus far have come from theoretical biology, and have
stressed recurrent self-production network relations: Rashevsky's
relational biology, Rosen's M,R systems, Pattee's semantic closure,
Varela & Maturana's autopoiesis, and Fleischaker's operational definition
of life. I would also include von Neumann's self-reproducing kinematic
automaton (as long as it's understood that Neumann's cellular automaton
was conceived as a <formal description> of some aspects of a physical
self-reproducing device).
The major rift between the theoretical biologists
and the artificial life simulators lies in whether simulated behaviors can
be separated from the material substrate of the computing devices (platonic
idealism vs aristotelian hylomorphism). The assumption that this can be
done involves the premise that all state transitions are deterministic
(and therefore exhaustively describable by state-transition rules), but
the implementation of perception (or "measurement") operations requires
apparently contingent transitions (which cannot be described by rules on
that level of description). In other words, "computations" can be abstracted
from their physical substrates (due to their apparently determinate behavior),
but "measurements" cannot be so abstracted (due to their apparently indeter-
minate behavior, relative to the observer). I would argue (along with Pattee)
that "measurements" are every bit as essential to the definition of life
as "computations," hence the arguments over whether computer simulations
can really implement (as opposed to represent) "measurement" operations. The
logically-necessary must be complemented by the empirically-contingent.
I was surprised that there was so little of the basic discussion regarding
the definition of life at the two conferences (outside of the poster by
Fernandez, Moreno, & Etxeberria at AL II)--in the rush to simulate we
shouldn't forget the questions we seek to answer by simulating. The appearance
of lifelike motion can be so seductive and all-consuming....
--peter cariani peterc@chaos.cs.brandeis.edu 37 Paul Gore St, Boston, MA 02130
tel 617-524-0781
------------------------------
Date: 19 April 1990, 14:54:35 SET
From: UIN005 at DDOHRZ11
Subject: binary or other coding within GA's
There are good reasons, I guess, why GAs and evolution strategies (ESs)
coexist since more than 20 years. The former are well known to a broad
community in the West, the latter more frequently used in the East (as
seen from a mid Atlantic position). I would like to make some remarks
on the latter only here.
ESs were used at first in experimental optimization tasks like shape
design of slender bodies in a flow, nozzle design and the like.
In these cases it was quite 'natural' to program mutations as normally
or binomially distributed changes with zero mean and given - later on
self-adapted - variance(s). The standard deviations may be looked upon
as 'mean' step sizes of the mutations. If one looks into nature such
deviations (the apple normally does not fall far off from the tree)
from parental positions describes quite well what can be observed on
the level of the phenotypes. ESs therefore did not look for the lower
genetic level of mutations. The genetic code as well as the so-called
epigenetic apparatus are somehow smoothing the discrete steps of
genetic variation.
In case of discrete optimization problems - better: problems which are
formulated in a way which operates on the basis of discrete decision
variables - it obviously is useful to operate with discrete, may be even
with binary coded, variables as well as mutations. But,if the principle
that 'smaller' (in any useful sense) variations are more frequent than
larger ones is omitted totally, then the creeping random search process
tends towards a blind random search process. The latter has a better
chance to approach the global extremum under many local extrema, but at
a very high cost. Whether or not a binary encoding is appropriate,
depends on the structure of the problem to be solved. It does not seem
to be useful to convert every continuous-variable optimization problem
into a combinatorial one (since the former are easier to solve, in
pinciple at least). The opposite kind of reformulation may sometimes
be of more help, instead (like the rubber bend approach to TSP).
Much more interesting than the way of problem codification for me is
the question of organizing the self-adaptation of the strategy parameters
(or rules) - in case of ESs these are the variances and even covariances
of the mutability density distribution. My observations show that
diversity of the population (population principle) - may be enhanced by
annidation, recombination (sexual propagation), forgetting (limited
life span or reproduction capacity of individuals), as well as soft
selection are of utmost importance in that kind of groping-in-the-dark
search games. To maintain on-line learning-by-doing in a changing
environment (e.g. process optimum control) even polyploidy must be taken
into account.
Hans-Paul Schwefel
University of Dortmund
April 19th, 1990
------------------------------
End of ALife Digest
********************************
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=---=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
= Artificial Life Distribution List =
= =
= All submissions for distribution to: alife@iuvax.cs.indiana.edu =
= All list subscriber additions, deletions, or administrative details to: =
= alife-request@iuvax.cs.indiana.edu =
= All software, tech reports to Alife depository through =
= anonymous ftp at iuvax.cs.indiana.edu in ~ftp/pub/alife =
= =
= List maintainers: Elisabeth Freeman, Eric Freeman, Marek Lugowski =
= Artificial Life Research Group, Indiana University =
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=---=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
End of Alife Digest
********************************
