Host: Japan SOciety for Fuzzy Theory and intelligent informatics
Co-host: The Korea Fuzzy Logic and Intelligent Systems Society, IEEE Computational Intelligence Society, The International Fuzzy Systems Association, 21th Century COE Program "Creation of Agent-Based Social Systems Sciences"
Neural networks can approximate arbitrary nonlinear functions. Neural networks are often designed by trial-and-error. Though the design methods such as brute-force approaches, network construction and pruning have been proposed, a generic approach in which other conditions such as learning parameters are taken into account is rarely met. This paper presents an efficient approach to decide multiple design parameters of multilayer neural networks by using the Design of Experiment, whose features are efficient experiments by an orthogonal array and quantitative analysis by analysis of variance. We pick up not only the number of hidden nodes but also initial conditions and learning parameters as the design parameters. In this paper, our approach is applied to 3-layer feedforward neural networks (FNNs) and 5-layer FNNs. We also show that the approximation accuracy of multilayer neural network can increase by picking up much more parameters.