``Fly me to the moon
and let me play among the stars
Let me see what spring is like
on Jupiter and Mars ...''
Bart Howard, `Fly Me to the Moon'
I began this thesis by talking about life, and the possibility of recreating it on a computer. However, the lack of a precise and satisfactory definition of life led me to concentrate on more specific issues. In particular, most of this thesis has been concerned with open-ended evolution.
The approach to modelling open-ended evolution pioneered by Tom Ray with the Tierra platform has been fairly widely used, and its validity fairly widely accepted, within the artificial life community. However, a number of concerns have been voiced about Tierra; for example, the influence that specific yet fairly arbitrary design features exert upon the system's behaviour was mentioned in Sections 3.2.1 and 3.3.1. I suggested that by experimenting with a similar--but not identical--system, some light might be shed on such issues. I therefore decided to design, implement and experiment with such a system, called Cosmos (described in Chapter 4).
A wide range of experiments with Cosmos were reported and analysed in Chapters 5 and 6. The behaviour of the system was different to Tierra in various ways. For example, no parasitism or similar ecological phenomena were observed. This result was, in fact, expected, and is due to differences in the kinds of inter-organism interactions allowed in the two platforms.
The role of contingency (chance events), as opposed to general evolutionary principles or to specific design details, in determining the outcome of Cosmos experiments was investigated (Section 6.1). It was found that contingency did play an important role; in a series of 19 experiments run under identical conditions apart from the number used to seed the random number generator, each one performed significantly differently, on a number of measures, to at least a third of the other runs. I suggested that these results should be broadly applicable to similar platforms, although the increased ecological interactions in Tierra compared to Cosmos might change the situation to some extent.
Other results were reported, such as the emergence of `speciation' in runs where energy was distributed heterogeneously to the environment (Section 6.5.3). However, it was hard to escape the feeling that many results were due to fairly specific features of the system's design. On top of this, the parameter space was far too large to allow a full and systematic study of the platform's capabilities.
With the benefit of the experience gained with Cosmos, I took a step back in Chapter 7 to discuss some problems that I now perceive with the Tierra approach to modelling open-ended evolution. Specific problems include the facts that organisms have a `hard-wired' structure, and that their interactions with other organisms are very restricted. I also suggested that the fact that organisms are represented by self-reproduction algorithms might limit the system's evolvability, and that the lack of competition for matter and energy might restrict the potential for various ecological phenomena (such as self-maintaining organisations, food webs, etc.) to emerge in the system. Note that we do not know a priori whether this last shortcoming deprives the system of the capacity for open-ended evolution (although it may turn out that it does), but it does deprive it of the capacity for modelling many other processes associated with life. All of these shortcomings can ultimately be traced back to the lack of an adequate theoretical grounding to guide the design of such systems.
I then went on to analyse the process of self-reproduction in Tierra, and in a number of other systems, in terms of von Neumann's genetic self-reproduction architecture, A + B + C + D + φ[A + B + C + D] (Section 7.2). I suggested that the general distinction between implicit and explicit encoding of the self-reproduction process, which has been a preoccupation (in the context of attempts to avoid `trivial' self-reproduction) of many researchers working with models of self-reproduction recently, is not very enlightening when considering the capacity of the self-reproducers to partake in an open-ended evolutionary process (Section 7.2.3). However, I argued that it is enlightening to consider this distinction in the rather more specific context of the components of von Neumann's self-reproduction architecture.
In Section 7.2.3 I demonstrated that programs in Tierra-like platforms can be seen to conform to von Neumann's architecture, despite suggestions to the contrary in the literature. However, the distinction between genetic self-reproduction and reproduction by self-inspection is blurred in these systems, because the interpretation machinery (automaton A of von Neumann's architecture) is hard-wired into the operating system rather than being explicitly encoded on the individual self-reproducers. Furthermore, the process by which the program's instructions (the equivalent to the genetic tape in von Neumann's architecture) are copied is itself an example of self-inspection; indeed, von Neumann remarked upon this feature of his architecture himself. I argued that a program's instructions in a Tierra-like platform correspond to the tape φ[B + D] in terms of von Neumann's analysis.
After considering the process of DNA replication in von Neumann's terms, I next discussed the desirable properties of `proto-DNA', i.e. a class of object capable of acting as a seed for an open-ended evolutionary process. I concluded that such an object would correspond to the tape φ[D] only (Section 7.2.3). Unlike programs in Tierra-like platforms, I argued that the copying process B of the proto-DNA should be implicitly encoded in the environment (e.g. the operating system or the laws of physics), at least initially. The fact that B is explicitly encoded in Tierran programs means that it is susceptible to disruption by mutations and perturbations from the environment. This is why the interactions between programs in platforms such as this have to be restricted--in Tierra, for example, direct interaction between programs is restricted to the reading by one program of the instructions of its neighbours. If the copying process is implicit in the environment, far fewer restrictions need to be placed upon interactions between components in the system to ensure that the population survives and evolves. With fewer restrictions upon interactions, the system is able to evolve in more diverse ways.
By suggesting that proto-DNA should be of the form φ[D], I am claiming that a suitable scenario for the start of an open-ended evolutionary process entails a population of strings of arbitrary information, which can all potentially be reproduced via implicit properties of the environment in which they exist. However, in order for there to be any selection between different strings in such a scenario (and therefore any evolution), individual strings must have specific phenotypic properties associated with them. In this way, my arguments suggest a shift of focus away from the process of self-reproduction per se, and towards a more careful consideration of phenotypic action. At the beginning of an evolutionary process we would not expect any complicated interpretation machinery to be available to decode the proto-DNA, so the environment should be able to implicitly determine the phenotypic properties (e.g. catalytic activity) of specific strings--in other words, A is implicit in the environment.
An inspection of the behaviour of existing artificial evolutionary models leaves one with the impression that they are only capable of evolving variations on a limited number of themes--evolutionary innovations usually have a ``more of the same'' quality rather than being fundamentally novel. Howard Pattee's analysis of the question of how fundamentally new symbolic information can arise in a physical system (see Section 7.1.3) suggests that truly open-ended evolution requires that the genotype, phenotypic and interpretation machinery should all be explicitly represented within the (artificial) physical world--this is his condition of semantic closure [Pattee 95b]. Therefore, even though I have argued that proto-DNA should initially be interpreted implicitly, it is vital that it also has the potential to evolve explicit interpretation machinery, A' (together with the necessary control machinery C'). This would require at least that certain processes within the system could become associated with specific sections of information on the proto-DNA strings (representing the birth of the explicit genotype-phenotype distinction). More work is required to further investigate how this potential for the proto-DNA may be assured. (A more explicitly-encoded copying process B' may also evolve once the proto-DNA is being explicitly interpreted, as indeed seems to have happened during the evolution of life on Earth.) In Section 7.2.3 I also argued that the representation of the organism should be fully embedded in the evolutionary `arena of competition' in order not to further restrict the evolutionary potential of the system.
Another important issue regarding the evolution of fundamental novelty concerns how organisms can evolve measuring instruments to probe new aspects of their environment (Section 3.1.2). In Section 7.3.2 I suggested that the common evolutionary principle that new functions are generally formed by organs with arise as modifications to pre-existing organs (e.g. [Maynard Smith 86] p.46), together with a move towards modelling components with multiple phenotypic properties (in different modalities), may help here. Much more work is required on this topic, however.
Finally, I returned to the issue of modelling life in general, as opposed to specifically modelling open-ended evolution. In Sections 7.1.4 and 7.3.2 I observed that many of the more interesting ecological and evolutionary phenomena in the biosphere arise because organisms are able to interact in much richer ways than are allowed in most artificial life models. My discussion of proto-DNA included consideration of how these restrictions may be relaxed whilst maintaining the robustness of the self-reproduction process. However, a further issue in the context of modelling life concerns the distinction between logical and material models. Biological organisms are embedded in a material world, and therefore represent useful resources of matter and energy for potential use by other organisms. Without a material grounding (i.e. a system where organisms are composed of structural units which are, at their lowest level, conserved, and which are in limited supply), it is doubtful whether any selection pressure can exist for organisms to evolve properties such as self-maintenance. Also, it is only with such a material grounding that ecological phenomena such as food webs and trophic levels can be realised. If we wish to allow artificial life models the capacity to evolve in these ways, it is therefore likely that we would have to use a material model of some sort.
In Section 7.3.3 I noted that Bedau's picture of life as supple adaptation raises some intriguing issues. For example, it suggests the important ecological processes we commonly associate with life (i.e. precisely the processes that necessitate the use of a material model) might actually be necessary features of any system that is capable of open-ended evolution. Further work on this hypothesis could therefore provide us with a clearer idea of the relationship between the evolutionary and ecological aspects of life discussed in Chapter 2.
In short, I have argued that future progress in the synthetic modelling of open-ended evolution in general, and of the evolution of living systems in particular, requires explicit theoretical consideration not only of individual self-reproducers, but also of the ways in which they interact, and of the environments in which they exist (including consideration of spatial structure, of the way in which organisms form part of the environment experienced by other organisms, and of the degree of implicit or explicit encoding of processes). In the long term, the discipline requires a unifying paradigm--a general picture of individuals, interactions and environments--in terms of which more specific questions can be framed. I have suggested that the paradigm proposed by Waddington in [Waddington 69] represents a useful starting point, and would also provide a valuable connection between artificial life and more traditional theoretical biology. Further development of the theoretical issues identified in Chapter 7 should lead to useful contributions to our understanding of biological life, and to the development of computer-based systems with greatly improved evolutionary potential.