Evolutionary design of artificial neural networks with different nodes

Abstract

Evolutionary design of artificial neural networks (ANNs) offers a very promising and automatic alternative to designing ANNs manually. The advantage of evolutionary design over the manual design is their adaptability to a dynamic environment. Most research in evolving ANNs only deals with the topological structure of ANNs and little has been done on the evolution of both topological structures and node transfer functions. The paper presents a new automatic method to design general neural networks (GNNs) with different nodes. GNNs combine generalisation capabilities of distributed neural networks (DNNs) and computational efficiency of local neural networks (LNNs). We use an evolutionary programming (EP) algorithm with new mutation operators which are very effective for evolving GNN architectures and weights simultaneously. Our EP algorithm allows GNNs to grow as well as shrink during the evolutionary process. Our experiment results show the effectiveness and accuracy of evolved GNNs.