Generation of unpredictable time series by a neural network

Statistical properties of the time series were studied. The time series was generated by perceptron using Hebb learning rule. The product of learning rate and amplification was termed as control parameter, when continuous transfer function was used. The connection to Berasconi model and the suppression of autocorrelation function were discussed. Results showed exponential growth in the typical length and transient of system cycles with the system size.