Exploiting complexity in evolutionary search using neural networks

We describe the use of feedforward neural networks to measure the complexity of cellular automata (CA) rules displaying emergent computation. Cranny and Bossomaier (1999) have conjectured that all CA rules capable of emergent computation must possess a great deal of intrinsic structure, implying that each lookup table is far from a random bit-string. We use neural networks to validate this assertion, and then show how the structure thus revealed can be used to both classify all known examples of emergent computation and constrain the search space for future searches for emergent computation