Neuroevolution based on reusable and hierarchical modular representation

Abstract

The framework of neuroevolution (NE) provides a way of finding appropriate structures as well as connection weights of artificial neural networks. However, the conventional NE approach of directly coding each connection weight by a gene is severely limited in its scalability and evolvability. In this study, we propose a novel indirect coding approach in which a phenotypical network develops from the genes encoding multiple subnetwork modules. Each gene encodes a subnetwork consisting of the input, hidden, and output nodes and connections between them. A connection can be a real weight or a pointer to another subnetwork. The structure of the network evolves by inserting new connection weights or subnetworks, merging two subnetworks as a higher-level subnetwork, or changing the existing connections. We investigated the evolutionary process of the network structure using the task of double pole balancing. We confirmed that the proposed method by the modular developmental rule can produce a wide variety of network architectures and that evolution can trim them down to the most appropriate ones required by the task. © 2009 Springer Berlin Heidelberg.