IDENTIFICATION OF NEURAL NETWORK PREDICTOR BY MEANS OF PREDICTION COMPLEXITY

Based on information entropy and mutual information, we proposed the definition of nonlinear partial autocorrelation. The concept is the generalization of partial autocorrelation. By means of it, we could get the quantitative method to measure the intrinsic prediction complexity of time series. The complexity is determined by the irreducible dependence between current quantities of time series and high order historical quantities, and indicated by the attenuation trend of nonlinear partial autocorrelation. In according to the attenuation trend, in principle, researchers could implement nonlinear model identification, e.g., identification of neural networks. Computer simulations perfectly supported our idea.