Graph composition in a graph grammar-based method for automata network evolution

The dynamics of neural and other automata networks are defined to a large extent by their topologies. Artificial evolution constitutes a practical means by which an optimal topology can be determined. Constructing a grammar of good graphs and then deriving new graphs from this grammar can facilitate this process. The following paper presents a simple but novel method of evolving a hypergraph grammar for this purpose. Different strategies for composing graphs within this framework are evaluated on problems of symbolic regression, time series approximation, and neural networks. The results favour a selectively modular approach that connects nodes with the most similar, rather than identical, labels.