Abstract
This article presents an approach to find multilayer neural networks with better performance by using the Design of Experiments method. Multilayer neural networks can approximate nonlinear functions by employing back-propagation learning algorithm. However, it is difficult to design the neural network with high performance because of an unknown relationship between the learning capability of the neural network and the design parameters such as its structure and conditions at learning. Our approach is to introduce the Design of Experiments method, which is based on an orthogonal array and analysis of variance, into design of multilayer neural networks in order to efficiently design them with a quantitative index.
