Abstract
Highly parallel ‘Self-X’ machines are inevitably coevolutionary since elements of a dynamic structural hierarchy always interact, an effect that will asymptotically dominate system behaviour at great scale. Here, we extend recent biologically-based models of machine cognition to a ‘farming’ paradigm for programming large, self-dynamic, coevolutionary devices. While broadly similar to the liquid state machines of Maass et al. these differ in that convergence is to an information source, a systematic dynamic behaviour pattern, rather than to a computed fixed ‘answer’. As the farming metaphor suggests, stabilising complex coevolutionary mechanisms appears as inherently difficult as programming them. One inference is that sufficiently large networks of (even dimly) cognitive devices will become emergently coevolutionary, and we do not presently understand how to program highly dynamic coevolutionary machines. This suggests the necessity of ‘second order’ evolutionary programming.
Acknowledgements
The author thanks two anonymous reviewers and the editor for comments useful in revision.