I have described a method for evolving representations for GAs with the goal of improving their performance. The method is modeled after the process of biological genome evolution, in which newly created genes become stably incorporated in the genome only when they produce a fitness increase. This selection in the construction of the genome is expected to filter out genes that disrupt highly adapted traits, biasing the evolution of the genotype-phenotype map toward a modular structure. Such a modular structure in the representation used for a GA would reduce epistatic interactions between genes and confer greater evolvability. This predicted outcome is tested using Kauffman's NK landscape model.
In the NK model, by simply adding new genes with random effects to the genome and rejecting those that reduce fitness, genotype-phenotype maps evolve very little epistasis between genes. But the resulting adaptive landscapes are smoother, with higher peaks, than even low epistasis can account for.
These results suggest that much more efficient algorithms may be obtained if there is an opportunity to build up representations incrementally, gene by gene, and keep only those additions to the representation that produce a fitness increase.
A caveat should be made to the application of adaptive landscape theory to real organisms. The evolutionary processes that constructed their genomes may result in genotype-phenotype maps that are not well described by generic models. Constructional selection provides a novel mechanism to produce the low values of K hypothesized by Kauffman [18].