Created by W.Langdon from gp-bibliography.bib Revision:1.6297
Autopoiesis self-organization (Heylighen 2002)(Maturana and Varela 1980) of a system denotes its self-creation and self-preservation. (Self-production is the literal meaning of autopoiesis) The required self-organization and the resulting performance of current artificial systems appear insufficient in a practical environment, where an ideal system would autonomously identify and approach problems, possibly producing similarly independent subsystems that represent problem solutions.
For informatics, we call this objective autopoietic programming, assuming its feasibility as a working hypothesis. We follow a straightforward approach, advancing an instance of current Machine Learning toward perfect self-organization, and discuss limitations that are due to impenetrable barriers inherent to present programming paradigms. To the end of the approach, chapter 2 discusses the autopoietic process called natural evolution (Darwin 1859; Ayala and Valentine 1979) from which self-organising systems emerge. (Example: an ecosystem, an individual organism.) Therefore, artificial evolution (Alliot, Lutton, Ronald, Schoenauer, and Snyers 1996), man-made implementations of evolutionary principles, approaches our supreme objective of artificial, fully self-organizing systems. For informatics, our present realm of interest, we thus focus on Evolutionary Algorithms (EA) (Baeck, Fogel, and Michalewicz 1997), i.e., probabilistic, iterative direct search methods that are inspired by biological evolution.
Regarding autopoietic programming, an EA called Genetic Programming (GP) (Koza 1992; Banzhaf, Nordin, Keller, and Francone 1998) offers itself, because such algorithms produce algorithms. However, a GP user faces undesirable properties typical of all current semi-automatic problem solvers, such as costly manual creation, maintenance, and problem-specific adaptation, the last being particularly critical since practical environments usually come with incomplete problem knowledge. To ameliorate the situation and to boost system performance, self-adaptation, in the sense of automatic specialization by enriching the problem model of a GP run, is desirable and approaches autopoiesis.
Ontogeny, a.k.a. development, is the history of structural changes of a system. In the realm of biological systems (Meinhardt 1982), we meet endogenous development that is essential to a system self-organization. System-inherent genotypic information, emerging during phylogeny in nature, guides such ontogeny that builds phenotypic structure which, in turn, exhibits behaviour.
Chapters 2,3,4 propose a basic formal model of a non-trivial genotype-phenotype mapping for search algorithms. The model as well as natural ontogenic phenomena suggest the design of beneficial mappings that leads to our GP-framework that we call Developmental Genetic Programming (DGP), a subset of developmental Genetic Programming that itself is a relatively small class of GP approaches that emphasize ontogenic aspects. (In the years following the coining of DGP in 1998, the term developmental Genetic Programming gained popularity in the community as a token for all ontogenic approaches.) Given the trivial mapping (identity), the framework collapses into an instance of the vast majority of common Genetic Programming approaches.
Chapters 5,6,7 design toy and practical problems for thought experiments and experiments on the framework, and they evaluate the empirical outcome. In a dynamic environment, autopoiesis of a system requires the latter structural components to stay in flux. Since these elements carry the function of the system, including its autopoiesis, the concept of self-adapting ontogeny imposes itself.
Chapter 8 shifts the focus within artificial ontogeny toward the phenotypic level. In our framework, a repairing method is the only essential component of ontogeny that is solely concerned with phenotypes. Deleting repair is a particularly interesting flavour of this component. Therefore, the chapter considers this repair type, dealing with the phenotypic level only.
Chapter 9 summarizes technical results, and Chapter 10 discusses conclusions on exploiting the limited autopoiesis of current search algorithms and suggests an escape, inspired by adaptive DGP, to fully self-organizing computation.",
Genetic Programming entries for Robert E Keller