[alife] A new thesis

Hideaki Suzuki hsuzuki at atr.jp
Mon Nov 1 21:56:27 PST 2004


Apologies if this has been cross-posted.

Dear everyone:

A new thesis is now available on-line.
This might be of some interest for people in this community.

http://www.nis.atr.jp/~hsuzuki/body/PhDthesisInfo_E.html

Cheers,

=============================================================
Hideaki Suzuki, Ph.D
Senior Researcher
ATR Network Informatics Laboratories
2-2-2 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0288 Japan
Tel:  +81 774-95-1063             Fax:  +81 774-95-2653
E-mail: hsuzuki at atr.jp
URL:    http://www.nis.atr.jp/~hsuzuki
=============================================================


Title:
"Design Optimization of Artificial Evolutionary Systems"

Abstract:

The performance of an artificial evolutionary system is
largely determined by the basic design prepared by a human
designer. This thesis describes a sequence of the author's
studies that aim at improving the design and implementing a
computational system able to evolve complex programs or
solutions in a life-like way.

The thesis first describes background theories on the
design in artificial life (alife). From the comparison to
the biological system, several design criteria on alife
systems are presented and representative alife systems are
assessed under the criteria. A mathematical theory for the
analysis on a creature genotype space is also described.

Next, the thesis proposes a machine language core memory
system, SeMar. SeMar is designed using a strong comparison
between computation and biochemical reactions. In imitation
of biological molecules, four kinds of data words are
prepared in the core. They are Membrane, Nutrient, DNA, and
Protein, and in the revised form of SeMar, all of the core
reactions are propelled by the parallel Protein execution.
The possibility of evolution of complex programs in SeMar
is discussed based on experimental results and design
criteria for an alife system.

Then, the thesis considers evolvability of artificial
evolutionary systems in general. Considering the relation
between evolvability and the fitness landscape, a measure
that parametrizes evolvability of alife systems is
proposed, and the design of an example alife system, a
string rewriting system, is numerically optimized in terms
of the maximization of the measure. Experimental results
show that numerical optimization by a computer can find a
far better design than that prepared by a human. In
addition, using the same system, the connectivity of viable
genotypes in the genotype space (evolvability) is examined
as a function of the measure, demonstrating strong
correlation between evolvability and the measure.

In the final part, the thesis proposes a new evolutionary
optimization algorithm named chemical genetic algorithms
(CGAs). The CGA focuses on an important factor of the
artificial evolutionary system design, translation.
Mimicking the biological translation in a living cell, the
CGA uses cellular structure with DNA and other smaller
molecules for translation as a selection unit, enabling
coevolution of DNA information and translation. Numerical
experiments reveal that the CGA can optimize the
translation, smooth the fitness landscape and enhance the
GA's evolvability, and as a consequence, have high
performance with a wide range of functional optimization
problems.





More information about the alife-announce mailing list