Summary and Conclusion - Self-Adaptive Heuristics for Evolutionary Computation - page 131

8 Summary and Conclusion

8.1 Contributions of This Topic

Evolution created manifold creatures since its beginning with the first organisms,
the prokaryotes, three to four billion years ago. Nowadays, computer scientists
translate the principles of inheritance, variation and natural selection into al-
gorithmic concepts known as EC. They mainly perform randomized search in
the domain space. Among its most famous variants are GAs, ES, EP and GP
growing more and more together. In the past many heuristic extensions have
been proposed for EAs. The contribution of research in EC is comparable to
old-fashioned AI research: the reduction of exponentially large search spaces by
appropriate representations, operators and heuristics. The success of the search
is often a result of the knowledge the practitioner integrates into the heuristics,
a symptom of the no free lunch theorem. Sometimes the practitioner should be
aware that classic search techniques may be faster. But most classical techniques
demand domain knowledge to be applicable or require mathematical features like
steadiness or the existence of the first or second derivative, which is not the case
for many practical problems. Nevertheless, randomized search has grown to a
strong, practicable and well-understood search technique, which is used all over
the world in research and engineering. EAs have successfully been applied to
domains like control, pattern recognition or automatic design. In the following,
we summarize the main results and research contributions of this topic.

• Taxonomy of parameter adaptation. The success of EAs partly de-

pends on proper parameter settings. Various kinds of parameter adaptation
techniques exist. We proposed an extended taxonomy for parameter setting,
which is based on the early taxonomy of Eiben [37]. At first, the methods can
be classified into tuning before and controlling during the optimization run.
The latter can furthermore be classified into deterministic, adaptive, self-
adaptive and meta-evolutionary approaches. Deterministic parameter con-
trol means changing the parameters in terms of the number of generations.
Heuristic rules are the basis for adaptive parameter control.