Time series generation by multilayer networks

The properties of time series, generated by continuous valued multilayer networks consisting of one or two hidden layers, are studied analytically. The time series is generated by using past output values to determine the next input vector. The main results for the generic asymptotic behavior are (a) The attractor dimension is only a function of the number of hidden units in the first hidden layer; (b) the analytical solution for the time series generated by the networks mirrors the structure of the network itself.