Introduction - Self-Adaptive Heuristics for Evolutionary Computation - page 2

control the bias of mutations self-adaptively. Usually, mutations are unbiased,
because evolution is not supposed to prefer any direction. But it turned out that
a bias is useful, in particular on ridge functions and in constrained domains.
Furthermore, we equip inversion mutation with self-adaptation. Inversion muta-
tion changes permutations in combinatorial search domains. In the case of the
traveling salesman problem it swaps two edges of random cities. But how many
swaps are useful, e.g. at the beginning of the search? We let self-adaptation learn
the optimal number of swaps in each generation.
Crossover is the other important genetic operator for EAs. It combines the
genetic material of two parents. Self-adaptation for crossover has been neglected
in the past. We try to investigate whether it is possible to learn automatically
which parts of the genetic material to take from each parent. Another interest-
ing question concerns the beginning of self-adaptation: what about the limits
of the original mutation strength control of evolution strategies? Experiments
and theoretical analysis reveal that their control may fail, in particular at the
boundary of infeasible search domain regions. We propose constraint handling
heuristics to overcome these limitations.

1.2 A Survey of This Topic

The following section gives an overview of this topic and summarizes each chap-
ter. Figure 1.1 gives a graphical survey of the structure of this work.

Chapter II: An Introduction to Evolutionary Computation

The field of EC belongs to the biologically inspired methods of computational in-
telligence. Evolutionary algorithms model the principle of the famous Darwinian
[28] concept of evolution. Various kinds of algorithmic variants exist becoming
more and more similar to each other in recent years. Chapter 2 presents the
basic principles of the evolutionary computation field, defining basic terms and
reaching from practical aspects to a survey of theoretical approaches. Besides a
short description of the different types of evolutionary algorithms, the chapter
points out the similarities between evolutionary algorithms and particle swarm
optimization algorithms.

Chapter III: An Extended Taxonomy of Parameter Setting Techniques and
Definitions of Self-Adaptation

After the design of the problem representation and the evaluation function, the
EA has to be specified. Besides the chose of the genetic operators, adequate
parameters for the features of the EA have to be set. The parameter values de-
termine the effectivity and the eciency of the heuristic. Appropriate operators
have to be chosen as well as appropriate initial parameterization of the strategy
parameters. Furthermore, it can be useful to change these parameters during
the run of the EA. In chapter 3 we introduce and extension of Eiben's [37] pa-
rameter control taxonomy. A survey of the various kinds of control techniques