Adaptive Geno to Phenotypic Mappings:
Unlike most forms of Genetic Programming,
the Developmental paradigm assumes a model in which genotype and
phenotype spaces are explicitly separated through a mapping. The concept of a
mapping follows that established by Keller and Banzhaf in the sense that an encoding
is necessary from a base binary codon sequence. The approach adopted here,
however, is to make use of an explicitly symbiotic model of coevolution to co-evolve
independent populations of mappings and genotypes. In effect we are including in
the search the selection of the most appropriate instruction sets with respect to
the space of programs. Needless to say, this increases the size of the search space
at initialization, but potentially also enables one to reduce the size of the search
space as useless instructions are penalized.
- Probabilistic Cooperative Geno to Phenotypic Model:
A symbiotic coevolutionary model is introduced to efficiently search the space of mappings to
genotypes. Such a scheme, when used with a redundant mapping, provides
a framework for shrinking the initial large number of instructions comprising
the GP function set. This work establishes that such a model is actually faster
than canonical GP (constant search space) as applied to medical classification
benchmarks. Moreover, the solutions are also more transparent, on account
of the reduced instruction set and compact nature of the phenotype.
- Case of evolving recursive programs: GP solutions are typically not recursive.
That is to say, solutions are provided by instruction sets that do not permit any
form of jump, recursive reference, or loop instruction. In this work the Geno-
Phenotypic model of theme 4(a) is demonstrated to be much more capable
at discovering solutions to a series of recursive problems than the equivalent
canonical GP model.
- Wilson G., Heywood M.I. (2007) Learning Recursive Programs with Cooperative
Coevolution of Genetic Code Mapping and Genotype. Proceedings of the Genetic and
Evolutionary Computation Conference (GECCO),7, ACM Press. Vol. 1:1053-1060
- Dr. Wilson's PhD thesis develops the above results in tutorial form.