On the Artificial Evolution of Neural Graph Grammars
Martin H. Luerssen, David M W Powers · 2003
Artificial neural networks and other connectionist models of computation are frequently credited with biological plausibility. Since biological systems are products of Darwinian evolution, network optimisation by artificial evolutions has considerable appeal. However, the computational expense of this can become prohibitive unless an emphasis is placed on modularity and reuse. Gene expression holds the answer to this. Genes are translated into proteins that self-organize into phenotypic traits such as the brain, with feedback loops controlling the further expression of genes. In this paper we present a generalization of this mechanism, a context-free graph grammar that describes a graph of finite-state automata. The graph is generated by replacing hyperedges with subgraphs of automata and other hyperedges according to a set of hypergraph productions. These automata need not be homogeneous, e.g. they may correspond to different types of neurons, reflecting the diversity of neurons in the brain. Desirable hypergraph productions are retrieved from a population of productions, which evolve by mutation of existing productions and subsequent selection against a user-defined criterion.