Robust nonlinear ARMA model parameter estimation using a stochastic recurrent neural network

We introduce a new approach for estimating linear and nonlinear stochastic ARMA model parameters, using recurrent neural networks. This new approach is a 2-step approach in which the parameters of the deterministic part of stochastic ARMA parameters are first estimated via a three-layer network and then re-estimated using the prediction error as one of the inputs to the networks. Using this simple two-step procedure, we obtain more robust model predictions than the deterministic network approach despite the presence of significant amounts of either dynamic or additive noise in the output signal. A comparison between the deterministic and stochastic approaches is made using renal blood pressure and flow data.