Studies in artificial evolution
Robert James Collins · 1992
We define artificial evolution as a particular class of genetic algorithms. Artificial evolution genetic algorithms require a clear separation between genotype and the information encoded in the genotype. The genotype is represented as a linear string, and the genetic operators of recombination and mutation operate randomly at the lowest level of organization of the string, without reference to any syntactic nor semantic structure that may be encoded there. We view the genotype as encoding a program; the fitness of the genotype is determined by decoding and executing the program, perhaps in an environment that is shared by the other members of the population. To achieve realistic population dynamics, we simulate large populations (at least tens of thousands of individuals in each generation). We apply artificial evolution to three classes of problems: the study of natural evolution, the evolution of complex behavior in artificial organisms, and function optimization. We simulate elaborations of Kirkpatrick's analytic model of sexual selection, exploring the effects of relaxing the simplifying assumptions (required for the analytic solution). We demonstrate that both the equilibrium and non-equilibrium dynamics are strongly affected by the details of the model. We also simulate the effect of host-parasite coevolution on the evolution of a recombination rate modifier gene, which is a test of the parasite hypothesis for the maintenance of sexual reproduction. Our results empirically demonstrate a strong correlation between the rate of parasite evolution and the equilibrium recombination rate in the host species. We also use artificial evolution to evolve foraging behavior in colonies of artificial ants. This study attacks the problem of representing a computer program both as a function that produces complex behavior and as a bitstring that is subject to a genetic algorithm. We introduce the connection descriptor artificial neural network (ANN) encoding scheme that places both the connection strengths and the connectivity pattern under genetic control, and use this encoding to evolve ant-like behavior. Our study concludes with the application of artificial evolution to function optimization. We perform a head-to-head comparison of conventional selection and mating schemes to those that involve spatial structure. We have found that spatial structure leads to much faster and robust discovery of optimal partitions.