Evolving Artificial Neural Networks through Topological Complexification
Thomas Doensig Jorgensen, Barry Haynes, Charlotte C. F. Norlund · 2009
Abstract—This paper describes a novel methodology for evolving artificial neural network topologies by intelligently adding connections and neurons. The neural networks are complexified and grown to optimise their neural complexity, which is a measure of the information-theoretic complexity of the network. Complexification of neural networks describes the process of increasing the neural complexity whilst increasing the structural complexity of the neural networks. This novel technique is tested in a robot control domain, simulating a racecar. It is shown, that the proposed methodology is a significant improvement over other more common and randomised growing techniques. The technique proposed here helps to discover a network topology that matches the complexity of the problem it is meant to solve. This results in networks which in some cases learn faster than some fixed structure networks.