Automatic Change of Representation in Genetic Algorithms

In the areas of Genetic Algorithms, Artificial Life and Animats, genetic material is often represented as a fixed size sequence of genes with alleles of 0 and 1. This is in accord with the ‘principle of meaningful building blocks’. The principle suggests that epistatically related genes should be positioned very close to one another. However, in situations in which gene dependency information cannot be determined a priori, a Genetic Algorithm that uses a static, list chromosome structure will often not work. The problem of determining gene dependencies is itself a search problem, and seems well suited for the application of a Genetic Algorithm. In this paper, we propose a self-organizing Genetic Algorithm, and, after describing four different chromosome representations, show that the best one for a Genetic Algorithm to use to coevolve the organization and contents (gene dependencies and values) of a chromosome is a hierarchy.