Exploring Structural and Dynamical Factors in the Evolution of Artificial Neural Networks

Abstract

Real-world networks, such as the network of neurons in the brain, have structural and dynamical properties that distinguish them from artificially created networks. This thesis studies how a selection of these properties can be applied to the evolution of artificial neural networks (ANNs).

Artificial evolution is used to create networks that solve control and pattern recognition tasks. However, the main focus is not on the evolution of successful controllers. Rather, I have studied the characteristics of the search processes leading up to the evolved networks, in an attempt to describe how changes to the structure and dynamics of the evolved networks affect the search landscape and the evolutionary trajectory.

Network dynamics is studied by estimating the mutual information between the nodes of the evolved networks, and applying a set of measures inspired by recently proposed theories about the information-theoretic properties of the mammalian brain. The results indicate that the evolved networks’ ability to perform an unstable control task is correlated with the proposed informationtheoretic measures, but that this correlation is far from straightforward, and that more work is needed to achieve measures that can be used to actively guide the evolutionary search in a way that will improve the search efficiency over a wide range of tasks.

The study of the relationship between network structure and the evolution of ANNs for pattern recognition focuses on the degree of modularity of the evolved networks. An important difference from earlier studies is the distinction between genotypic modularity and network modularity. The main question is whether modular networks are “better” than non-modular networks, in some sense of the word. The results presented in this thesis work indicate that this is only the case under certain conditions. It is shown that a stochastic mapping from genome to network is an example of such a condition.