Document: John von Neumann and the Evolutionary Growth of Complexity

6 Conclusion: Looking Forwards

By re-examining von Neumann's work in the light of his own description of the problem he was working on, I have tried to show that there is much more substance to it than has been generally recognised. However, as Popper [16] has emphasised, the scientific enterprise is intrinsically iterative: the solution of any given problem gives rise to a new problem situation, which is to say new problems to be addressed. It seems to me that von Neumann's work is particularly fruitful in this regard, as it poses a number of new and profound questions which need to be addressed by the (still) fledgling field of Artificial Life. Among these are:

Can we clarify (or even formalise) the notion of ``biological'' or ``adaptive'' complexity. This of course is not specifically a problem for Artificial Life, but rather underlies the entire discipline of evolutionary biology [11].
What is the precise significance of the self-reproductive capability of von Neumann's machines? Technically, based on the idea of a ``general constructive automaton'', growth of complexity per se could take place in isolation from self-reproduction. One simple scenario here would be to imagine being made to simply grind through the (countable) infinity of all description tapes, constructing every described automaton in turn. This would require only a trivial extension of the capabilities of . While this would fail in practice due to the essential fragility of von Neumann's automata (discussed further below) it is not clear whether there is any fundamental problem with this general idea. Nonetheless, it seems clear that if the growth of complexity is to involve Darwinian evolution, then self-reproduction surely is a necessary additional requirement.
The problem of identity or individuality. In the alife systems discussion above, the question of what constitutes a distinct individual ``machine'' or ``automaton'' is addressed in a completely ad hoc manner. In a real sense, these putative individuals exist as such only in the eye of the human observer. Whereas, the notion of a self-defining identity, which demarcates itself from its ambiance, is arguably essential to the very idea of a biological organism. Some early and provocative work on this question, including simulation models overtly reminiscent of von Neumann's CA, was carried out by Varela et al. [21]. However, there has been relatively little further development of this line since.
The evolutionary boot-strapping problem: in the von Neumann framework, at least, is already a very complicated entity. It certainly seems implausible that it could occur spontaneously or by chance. Similarly, in real biology, the modern (self-consistent!) genetic system could not have plausibly arisen by chance [5]. It seems that we must therefore assume that something like (or a full blown genetic system) must itself be the product of an extended evolutionary process. Of course, the problem with this--and a major part of von Neumann's own result--is that it seems that something like a genetic system is a pre-requisite to any such evolutionary process. As noted, a framework of Genetic Relativism, which envisages evolutionary modification of the genetic machinery itself, may contribute to resolving this problem.
Perhaps most importantly of all, what further conditions are required to enable an actual, as opposed to merely potential growth of complexity? It was well known even to von Neumann himself that his system would not in practice exhibit any evolutionary growth of complexity. The proximate reason is that, in his CA framework, all automata of any significant scale are extremely fragile: that is, they are very easily disrupted even by minimal perturbation from the external environment. The upshot is after the completion of even a single cycle of self-reproduction the parent and offspring would almost immediately perturb, and thus effectively destroy, each other. This can be avoided by ad hoc mechanisms to prevent all interaction (again, this was suggested by von Neumann, and since shown in practice by Langton and others). However, eliminating interaction eliminates the grist from the darwinian mill, and is thus a non-solution to the substantive problem (growth of complexity). Tierra offers a somewhat different approach, whereby the integrity of the automata (process images) is ``protected'' by the underlying operating system, while still allowing some limited, but significant, forms of interaction. This has been very fruitful in allowing the demonstration of some significant evolutionary phenomena. However, any serious claim to substantively model real biological organisms will inevitably have to confront their capacity for self maintenance and repair in the face of continuous perturbation and material exchange with their environments. This, in turn, is clearly also related to the earlier problem of biological individuality.

In conclusion, it seems to me that von Neumann's work in Artificial Life--properly understood--is as profound and important today as it was half a century ago; and that it should continue to provide structure, insight and inspiration to the field for many years to come.

Acknowledgements

John von Neumann and

5 Evolvability: A Closer

Document: John von Neumann and the Evolutionary Growth of Complexity

Barry.McMullin@dcu.ie