Tri-state neural networks for time series prediction

Inspired by the physiology of natural neurons which exhibit a third state in their action potential known as hyperpolarization, the tri-state neuron utilizing the double sigmoid nonlinearity was developed to produce faster converging solutions with no loss in performance of the converged multilayer perceptron artificial neural network. The authors apply the traditional bi-state network and the tri-state network to the one-step forward prediction problem. The performance of the neural network predictor structures is examined for three different time series models: a linear auto-regressive model, a nonlinear Volterra model, and a deterministic chaotic model. Results illustrate the increased convergence speed attainable from the tri-state network for both the linear model and Volterra nonlinear model without a loss in performance when tested. Results from the chaotic model are mixed but, in some respects, the tri-state network is superior