Biol=Trivial? and Computation

Nicholas Gessler (gessler@UCLA.EDU)
Tue, 13 Aug 1996 11:30:15 -0700

I would agree that programming requires "some comprehension" of the
phenomena under question. An interesting question arises when we ask what
the source of that "comprehension" is or might be? Traditionally, the
source was the programmer, but this is increasingly less necessary.
"Comprehension" (of a most primitive kind) in a neural net comes from a
period of training where it is either rewarded or punished for its behavior.
Thought of in another way, this "comprehension" may be considered "feedback"
from the external world. If we further think of this "feedback" as a form
of natural (artificial) selection, then we can envision the quickly
developing fields of Evolutionary Computing (EC). "Genetic algorithms"
(GAs) have long been used to "adjust" neural nets and recently the GA field
has expanded to encompass "genetic programming" (GP). In GP, the programs
write themselves and are evaluated by an external fitness function. There
evolution can either be directed (user defined fitness function) or dynamic
(co-adaptationally defined implicit fitness determination).

I just participated in the first GP conference in Stanford and would like to
pass on a couple of observations. The analogies built upon analogies (some
would describe these as metaphors built upon metaphors) can be initially
confusing, even to practitioners of EC. One such confusion which I
critiqued at the conference was the expectation that "cultural algorithms"
should work more quickly than "genetic algorithms" in programming because
culture works more quickly than biology in the natural world. In the
natural world culture and biology work on different problems in different
media with vastly different generational periods. Translated into the
artificial world of computer media there is no a priori reason why the
problems or the generation times cannot be made equal or even reversed.
Consequently, this aspect of the analogy is no longer valid in EC. The more
valid analogy is between culture and biology in the natural world and
between software and hardware in the computer world. Both cultural and
genetic algorithms are software in EC. The next engineering task is the
invention of evolvable hardware (EH)...

One of the presenters gave a longing plea for the development of EH. To his
surprise, it has already been done. Another presenter reported on Evolvable
Hardware now commercially available in the form of Field Programmable Gate
Arrays (FPGAs). The connections between the logic gates are now
programmable by software making it possible to evolve circuits of hard-wired
logic. An anticipated 1000-fold increase in EC computing power is expected.

The take-home message for me is this: Considering evolution through natural
selection as a deep phenomenon underlaying many natural world processes is
an increasingly productive and prolific cognitive tool. In this deep sense,
evolution through natural selection may well be the source of
"comprehension" not only in the biological and cultural domains, but in
those of physics as well (no room to expound upon physics). Computational
approaches are relevant to the social sciences precisely because we are
computational engines (as Minsky has put it).

Nick Gessler

In response to:

>>Read writes,
>>> >A generative program can be written only if there is already comprehension
>>> >of the phenomena in question.
>>John McCreery writes:
>>> My memories of working as an AI programmer suggest a slightly different
>>> conclusion. The first iteration of a program requires *some* comprehension
>>> to get off the ground.
>>Danny Yee writes,
>>What about a neural net?
>John McCreery writes,
>Damn good question. I've heard it said that neural nets operate without
>presuppositions. I still don't quite understand how it works.

Nicholas Gessler
UCLA Anthropology
Computational Evolution and Ecology Group
Artificial Culture Site:
Phone/FAX 310-559-6661