- author = "Una-May O'Reilly and Franz Oppacher",
- title = "An Experimental Perspective on Genetic Programming",
- booktitle = "Parallel Problem Solving from Nature 2",
- year = "1992",
- editor = "R Manner and B Manderick",
- pages = "331--340",
- address = "Brussels, Belgium",
- month = sep # " 28 - 30",
- publisher = "Elsevier Science",
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/ppsn92.ps.gz",
- size = "10 pages",
- abstract = "Genetic Programming (GP) has recently been introduced by John R. Koza as a method for genetically breeding populations of computer programs to solve problems. We believe GP to constitute a significant extension of the Genetic Algorithm (GA) research paradigm primarily because it generalizes the genetic search techniques: instead of looking for a solution to a specific instance of a problem, GP attempts to evolve a program capable of computing the solutions for any instance of the problem. We have implemented a genetic programming environment, GP*, that is capable of duplicating Koza`s experiments. In this paper we describe a specific GP experiment on the evolution of programs to sort vectors, and discuss the issues that must be addressed in any application of GP: the design of fitness functions and test suites, and the selection of program terminals and functions. Our observations point to several previously unnoticed shortcomings of the GP approach. We hypothesize that these shortcomings are due to the fact that GP only uses a hierarchical representation but does not construct its solutions in an explicitly hierarchical manner.",
-
notes = "Critical of Koza's GP (nb non-ADF) {"}We conclude that
GP in its current form is heirarchical only with
respect to its representation and not with resepect to
its process of constructing solutions. This limits the
ability of GP to evolve complex programs from simple,
general functions, and makes the algorithm stongly
dependant on initial human design
decisions.{"}
Proposes SPECIALISE and DECOMPOSE operators, like encapsulate and expand, but applied infrequently and depending upon how the GP is going. SPECIALISE would look for common code in better programs and convert them to functions which cannot be disrupted by crossover.
However: ``Regarding the specialize and decompose operators, we abandoned them after very preliminary work''.
References Ken De Jong ICGA-87
",
