Relevant issues from the biological literature are discussed in Chapter 2. The distinction between evolutionary and ecological views of life is emphasised, and attempts to produce definitions of life from each of these views, as well as hybrid definitions, are discussed. Similar distinctions in origin-of-life models are described. Current debates concerning various preconceptions about life and evolution are then discussed. These include debates over the notion of progress in evolution, and in particular the common assumption that evolutionary processes necessarily lead to an increase in the complexity of organisms. This leads to a discussion of the term `complexity', and of the various issues it encompasses. The nature of the major evolutionary transitions that have occurred in the history of life on Earth is described. Current debates in the biological literature over the relative importance of general principles versus contingency (`historical accidents') in determining the course of evolution are also discussed. At the end of the chapter, definitions are provided for the term `open-ended evolution' and for various sorts of complexity, but an adequately precise and satisfactory definition of life is still lacking. In this light, a decision is made to frame the research questions addressed in the rest of the thesis in terms of more specific concepts. In particular, the experimental work will concentrate on the issue of open-ended evolution.
A review of the relevant artificial life literature is presented in Chapter 3, together with an introduction to the more important issues and debates of the subject. The distinction between weak and strong artificial life is discussed, as are some of the theoretical and practical hurdles that the subject faces. Various views of the relationship between artificial life and theoretical biology are also presented. The majority of the chapter is devoted to a description of previous work with artificial life models. Particular attention is given to studies addressing the issues of self-reproduction and open-ended evolution, including work by John von Neumann, Nils Barricelli, John Holland and Tom Ray. More briefly, a number of models addressing issues of self-organisation and the origin of life are also discussed. Having devoted some time to describing the literature, I end the chapter by talking about the various methodology and design issues for artificial life platforms that have been highlighted by the preceding review.
In Chapter 4 the platform used for the experimental portion of the thesis is introduced. This platform, called Cosmos, is based upon Ray's Tierra system [Ray 91]. Reasons for experimenting with such a system are discussed at the start of the chapter. As the approach pioneered by Ray with Tierra is fairly widely used in the artificial life community, and its validity is fairly widely accepted, I claim that it is a valuable endeavour to investigate such models more thoroughly. Although Cosmos shares the general Tierra approach of modelling evolving individuals as self-reproducing computer programs, there are a number of important differences in the design of the two systems. These include: a closer analogy between individual programs (or more accurately, processes) in Cosmos and biological cells, including features such as regulation of the genome; the introduction of the notion of energy (potential CPU-time) as a commodity that cells must collect, store and use to pay for the execution of instructions; the introduction of a two-dimensional discrete spatial environment in which the programs can move and interact; and finally, the potential for parallel programs (which are considered analogous to multicellular organisms) to evolve. (Tierra has also been modified to allow for parallel programs, e.g. [Thearling & Ray 96], but the implementation details are different.) These and other differences between Cosmos and Tierra are discussed throughout the chapter, and summarised in the final section.
The results of a wide variety of Cosmos runs are presented, analysed and discussed in Chapters 5 and 6. These two chapters together constitute possibly the fullest and most systematic investigation of a Tierra-like system to have been reported.
Chapter 5 is devoted to the detailed analysis of a single Cosmos run. The various measures and visualisation techniques used are first described, including a collection of measures suggested by Mark Bedau and colleagues (e.g. [Bedau et al. 98]). The results of the run were not spectacular; programs generally increased in length due to the acquisition of more instructions to collect energy from the environment (which thereby increased their chances of survival). No parallel programs emerged, neither were any parasites, or similar ecological phenomena, observed. Reasons for this behaviour are discussed. The lack of emergence of ecological phenomena reinforces the suspicion that their appearance in runs of Tierra is due to some fairly specific details of the system's design rather than to any more general principles. The phylogenetic tree of the significant types of program to have emerged during the run is reconstructed, and suggests that chance events may have played a significant role in determining the outcome of the run. This issue is investigated in more detail in the following chapter.
Chapter 6 describes the results of a wide variety of Cosmos runs, designed to explore the system's parameter space. This issue of contingency is investigated in detail, in a group of 19 experiments in which the system is run under identical conditions except for the number used as a seed for the random number generator. It is found that each run performed significantly differently, in a number of measures, to at least a third of the other runs. This result is used in a calculation which suggests that the system should be run at least nine times in future experiments to be sure of observing a variety of results due purely to contingency rather than any other factors. A range of other experiments are also reported. Some of the more interesting results from these include: the runs can generally by categorised into one of two classes proposed by Bedau and colleagues [Bedau et al. 98], where one class represents active evolutionary behaviour and the other represents a degenerative state where the system's evolutionary potential is effectively lost; and, in one set of experiments (reported in Section 6.5.3) `speciation' is observed, i.e. the coexistence in the population of programs of different lengths, in different areas of the environment. A gradient in the energy distribution to the environment is required to achieve this, although no physical boundaries are necessary. Some experiments with a sexually reproducing program are also described, although asexual varieties quickly invaded the population and displaced the sexual programs in all of the reported runs. A summary and discussion of all of the results is presented at the end of the chapter.
In Chapter 7, with the benefit of the practical experience gained by designing, implementing and using Cosmos, I step back and reconsider the scientific value of the approach. A number of areas are identified in which Cosmos and similar Tierra-like systems are lacking, from both theoretical and methodological points of view. The major problem with using these systems for scientific purposes is the inadequacy of the theoretical grounding upon which their design is based. Various other problems are discussed, including the fact that these systems still do not seem to have the ability to evolve fundamentally new types of innovations. These discussions lead on to a detailed analysis of various issues involved in the concept of self-reproduction, particularly within an evolutionary context. The reproduction process employed by the programs in these systems is analysed in terms of von Neumann's genetic self-reproduction architecture [von Neumann 66], and compared to the replication processes employed by von Neumann's self-reproducing automata, by DNA, and by a hypothetical `proto-DNA' structure which has properties making it suitable for acting as a seed for an open-ended evolutionary process. It is argued that the explicit encoding of the self-reproduction process on programs in Tierra-like systems is a liability in terms of open-ended evolution, and that these systems would have greater evolutionary potential if the copying process was, initially at least, implicitly encoded in the operating system. In this situation, the focus of the design of such systems then shifts to the consideration of the sorts of phenotypic properties the reproducers should possess. The domain of interaction of the phenotypes should be within the evolving system itself, to allow for intrinsic rather than extrinsic selection (e.g. [Packard 88]), but beyond this there would seem to be few restrictions. Some ways in which the study of open-ended evolution in artificial life models may be improved are discussed in Section 7.3, where a number of open questions are also listed. A paradigm for studying biological evolution, introduced 30 years ago by C.H. Waddington [Waddington 69], is described, and it is argued that this represents a useful starting place for an improved synthetic approach. As an example of a future research direction, I show how the important question of how organisms evolve fundamentally new measuring instruments can be addressed within Waddington's framework (see Section 7.3.2). In the final section of Chapter 7 I return to the issue of studying life as opposed to open-ended evolution. I suggest that artificial life models of open-ended evolution and of other processes associated with the biosphere can usefully contribute to a broader understanding of the nature of life, but only if the questions being addressed are expressed in terms of precise concepts.
A summary of the thesis message is presented in Chapter 8.
Further details of the design and implementation of Cosmos are given in Appendix A, and further details of the runs reported in Chapters 5 and 6 can be found in Appendix B. A glossary of the biological terms used in the thesis is included after these appendices.