School of Informatics, University of Edinburgh, Edinburgh EH8 9YL, UNITED KINGDOM.
This work addresses the question: What are the basic design considerations for creating a synthetic model of the evolution of living systems (i.e. an `artificial life' system)? It can also be viewed as an attempt to elucidate the logical structure (in a very general sense) of biological evolution. However, with no adequate definition of life, the experimental portion of the work concentrates on more specific issues, and primarily on the issue of open-ended evolution. An artificial evolutionary system called Cosmos, which provides a virtual operating system capable of simulating the parallel processing and evolution of a population of several thousand self-reproducing computer programs, is introduced. Cosmos is related to Ray's established Tierra system, but there are a number of significant differences. A wide variety of experiments with Cosmos, which were designed to investigate its evolutionary dynamics, are reported. An analysis of the results is presented, with particular attention given to the role of contingency in determining the outcome of the runs. The results of this work, and consideration of the existing literature on artificial evolutionary systems, leads to the conclusion that artificial life models such as this are lacking on a number of theoretical and methodological grounds. It is emphasised that explicit theoretical considerations should guide the design of such models, if they are to be of scientific value. An analysis of various issues relating to self-reproduction, especially in the context of evolution, is presented, including some extensions to von Neumann's analysis of self-reproduction. This suggests ways in which the evolutionary potential of such models might be improved. In particular, a shift of focus is recommended towards a more careful consideration of the phenotypic capabilities of the reproducing individuals. Phenotypic capabilities fundamentally involve interactions with the environment (both abiotic and biotic), and it is further argued that the theoretical grounding upon which these models should be based must include consideration of the kind of environments and the kind of interactions required for open-ended evolution. A number of useful future research directions are identified. Finally, the relevance of such work to the original goal of modelling the evolution of living systems (as opposed to the more general goal of modelling open-ended evolution) is discussed. It is suggested that the study of open-ended evolution can lead us to a better understanding of the essential properties of life, but only if the questions being asked in these studies are phrased appropriately.