Cumulant-based parameter estimation using neural networks

A two-level, three-layer artificial neural network to estimate the MA (moving average) model parameters based on second- and third-order cumulant matching is developed. The first level is composed of some RAM units that are used to control the synaptic connectivities of the second-level neurons. The second level is composed of three layers of linear weighted-sum neurons in which the weight parameters of any neuron represent the MA parameter to be estimated. Each second- and third-order cumulant is viewed as a pattern the neural network needs to learn, and a steepest descent algorithm is proposed to train the neural network. The main advantage of this approach is that it uses a parallel architecture to represent the problem and a parallel to perform the estimation. A simulation is performed to demonstrate the performance of the neural network approach. Extension to ARMA (autoregressive moving-average) parameter estimation is discussed