Chaotic signal emulation using a recurrent time delay neural network

The authors describe a method for training a dispersive neural network to imitate a chaotic signal without using any knowledge of how the signal was generated. In a dispersive network, each connection has both an adaptable time delay and an adaptable weight. The network was first trained as a feedforward signal predictor and then connected recurrently for signal synthesis. The authors evaluate the performance of a network with twenty hidden nodes, using the Mackey-Glass (1977) chaotic time series as a training signal, and then compare it to a similar network without internal time delays. The fidelity of the synthesized signal is investigated for progressively longer training times, and for networks trained with and without momentum.<>