Research on an improved GMDH-type neural network based on reconstruction of phase space

An improved GMDH-type neural network based on the reconstruction of phase space and its application to complex systems are proposed. The structure of the conventional GMDH neural network is fixed, and it is difficult to obtain the suitable threshold value and the output of the network is best in local. The new model of the GMDH-type neural network is improved in two ways: one is that GA and selection coefficients are adopted to get the best structure of the network and the best solution; the other is that the mean of the fitness function is used as the threshold value, in this way the necessary data are reserved and the needless terms are adequately omitted. The improved GMDH-type neural network proposed is fit for the prediction of chaotic time series and the result of the simulation demonstrates that the new network makes better performance than the conventional one.