Abstract
Genetic programming is the newest form of evolutionary
computation that was conceived in the late 1980's as a
possible means for automatic programming. Genetic
programming performs an evolutionary search in the
space of computer programs and selects the program that
solves a given task according to certain criteria. In
the first part of the dissertation we give an overview
of evolutionary computation and in particular genetic
programming. We raise key issues for genetic
programming: code growth, diversity, real world
applications.
In the second part we present our contribution to the
theory of genetic programming. We demonstrate two
methods for limiting the code growth. The first method
consists in applying an additional mutation operator
that simplifies the structure of a genetic program
without altering its behavior. The second method
applies multiobjective optimization for the objectives
of fitness and program size. We show that both methods
are successful in reducing code growth without
significant loss of accuracy. We then define a distance
metric for genetic programs and use it for applying the
fitness sharing technique. We propose a simple
diversity measure based on our metric and study the
effects of fitness sharing with the help of this
diversity measure.
In the third part we show the application of genetic
programming in two complex real world problems. The
first problem comes from mechanical engineering. Four
bar mechanisms play a very important role in practical
mechanism design. We describe our four bar mechanism
design system. We demonstrate how genetic programming
can be a vital component of a complex design system. We
integrate genetic programming with decision trees into
a powerful learning machine.
The second problem belongs to the decision support
domain of economics. The decision-makers have to make
many subjective decisions. Consequently, the final
decision is sensitive to even small changes in these
subjective values. We present our genetic programming
system that helps the decision-makers to arrive at
stable decisions. That is, for small variations in the
values of the involved variables, the final decision
remains unchanged.
Users
Please
log in to take part in the discussion (add own reviews or comments).