Information Technology Reference
In-Depth Information
1 Introduction
Evolutionary algorithms (EAs) are biologically inspired, randomized search
meta-heuristics. They unify the fundamental principles of biological evolution:
inheritance of genes, variation of genes in a population, translation of genotype
into phenotype and selection of the fittest in the sense of the Darwinian princi-
ple
survival of the fittest
[28]. In the sixties Holland, Rechenberg and Schwefel
translated this paradigm of evolution into a concept of algorithms which is called
evolutionary computation (EC). Today, this computational method has grown
to a rich and frequently used optimization method. It comprises several variants
of algorithms which are structurally similar, but specialized to certain search
domain characteristics.
1.1
Motivation
Today, EA research concentrates on application areas and theoretical questions.
A rather unexplained area is the question for parameter settings. EAs exhibit
different kinds of parameters, e.g. population sizes or mutation strengths. Their
success is bound to the choice of the right parameterizations. Parameters can be
tuned before
the optimization. Human experience is necessary to find the right
settings for particular situations. Sometimes it is useful to
control
the parame-
ters
during
the run. A famous example are mutation strengths. At the beginning
big mutations accelerate the search. Later on, smaller mutations make the con-
vergence to an optimum possible. Many control techniques are specialized and
try to integrate domain knowledge to control the parameters. But flexibility and
problem-independence is the spirit of EAs. The question arises: how can EAs
learn
parameters during the optimization automatically? Evolution strategies
gave an answer to this question decades ago: self-adaptation. Their mutation
strengths self-adaptation became very famous. The parameters are
learned
dur-
ing the optimization process.
This work concentrates on the famous concept of self-adaptation. What other
kinds of parameters can successfully be controlled automatically? How much
is evolution able to learn? First, we concentrate on mutation and propose to
