Complexity in artificial life

Abstract

In this work we have defined and implemented an abstract evolutionary platform based on the chemostat, for the study of evolutionary complexity in populations of artificial organisms. An artificial organism consists of a virtual CPU and an assembler program (``genotype''), which self-replicates under perturbation and computes functions (``phenotype''). By defining a merit function over the computable functions, differential reproductive success is applied to the populations, such that they evolve to the implicit fitness landscape. Evolutionary complexity in an evolving finite population is defined as the product of diversity and average genotype size, and the work focuses on factors that may create such complexity. In particular, recombination and an approach to open-ended evolution are presented. Besides studying complexity, genetic recombination is also introduced and compared with mutation for evolutionary success in complicated fitness landscapes. The thesis is that mutual relations between fitness and environment are a major source of evolutionary complexity.