Neuromodels of analytic dynamic systems

In contrast to recent work aimed at using neural networks for relatively ‘long term’ prediction of time series, this paper examines how neural networks designed for short term prediction can form very good approximation models, valid over a large region of the phase space, after having been trained on as few as 500 pointsfrom a single trajectory of the underlying dynamic system. This is illustrated using four dynamic systems of increasing complexity, including a simple chaotic system and a more realistic system describing rigid body rotation using orthogonal torques under a chaotic control regime.