The evolutionary adaptability of biological systems is not simply a matter of good variation-selection search procedures. The structure-function relations on which these procedures act must be evolution-friendly. Systems with multiple weak (or graded) interactions and high component redundancy satisfy this requirement. We here describe how these features can be represented in a pattern processing system, called the cytomatrix neuron. Computer experiments show that the malleable structure-function relations significantly enhance the responsiveness of the pattern processing capabilities to selective pressures and that increasing the effective dimensionality of the processor, by increasing the number of structural features subject to mutation, also increases rate of adaptation, quality of solution, and controllability of generalization in response to the structure of the pattern set used for training. The relative autonomy of biological systems is contingent on evolutionary adaptability. Nontrivial capabilities could neither arise nor persist in the absence of adequate adaptability. For the same reason evolution-friendly structure-function relations are key to artificial autonomy. Tradeoffs must be considered, however. Simulating structure-function malleability on top of a base programmable machine is computationally costly. Our extensive experimental work with the cytomatrix processor shows that useful adaptive pattern processing can nevertheless be achieved despite these computational costs.
展开▼