Abstracting Biological Development to Evolve Large-Scale Artificial Neural Networks

• Kenneth O. Stanley, Evolutionary Complexity Research Group, University of Central Florida School of Electrical Engineering and Computer Science, USA

Kenneth O. Stanley is an Assistant Professor in the School of Computer Science at the University of Central Florida. He graduated with a B.S.E. in Computer Science Engineering and a minor in Cognitive Science from the University of Pennsylvania in 1997 and received a Ph.D. in 2004 at the University of Texas at Austin. He is the inventor of the Neuroevolution of Augmenting Topologies (NEAT) algorithm for evolving increasingly complex artificial neural networks. His main research contributions are in neuroevolution (i.e. evolving neural networks), generative and developmental systems, coevolution, machine learning for video games, and interactive evolution. He has won best paper awards for his work on NEAT (at the 2002 Genetic and Evolutionary Computation Conference) and for his work on NERO (at the IEEE 2005 Symposium on Computational Intelligence and Games), in addition to winning the Independent Games Festival Student Showcase Award and supervising the team that produced AAAI's 2007 Best Student Video Award winner.

In his talk, Stanley will focus on recent work in indirectly encoding neural networks through an abstraction of biological development called Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT). The primary conceptual focus will be on the critical role of abstraction in biologically-inspired systems. In particular, because DNA encodes complexity on a massive scale, interest has grown in recent years in the powerful mapping between genotype and phenotype in nature. Thus, the question, what is the right level of abstraction for the development of the embryo in nature? The conclusion is that it is possible to capture the essential properties of developmental encoding at a much higher level of abstraction than previously thought. This new developmentally-motivated encoding, called Compositional Pattern Producing Networks (CPPNs), is then extended to represent large-scale neural networks by exploiting a surprisingly simple geometric trick. The result is a novel ability to evolve working networks with millions of connections that exploit geometric regularities in several problem domains. Finally, Stanley will conclude with thoughts on prospects for open-ended evolution and unbounded innovation.