Including Phenotype Information in Mutation to Evolve Artificial Neural Networks

This paper addresses artificial neural network (ANN) evolution by presenting a mutation approach based on a novel self-adaptation strategy. The proposed strategy involves the phenotype information, incorporated in a value, called the network weight (NW), which depends on a total number of hidden layers and an average number of neurons in hidden layers. The inclusion of the phenotype information determines the increment of the mutation step size and the average percentage of successful mutations, which is achieved by means of adaptation to characteristics and the complexity of ANN architectures. The NW operator is combined with the genotype information, included in the dynamic component and represented by the fitness of a particular chromosome. These two components in the mutation approach drive the evolution of chromosomes according to characteristics of an ANN "internal" architecture and a fitness of a particular chromosome simultaneously.