Neural network chaotic system identification

This research focuses on system identification using alternative neural network architectures, where the dynamics of the underlying system are chaotic. A neural network is trained on the measured input-output data of the actual system. The actual system model examined is based on the Lorenz equations for the chaotic attractor of the same name. For comparison, a quadratic system is also studied. In this research, both standard feedforward network architectures and recurrent network topologies proved inadequate to successfully identify the chaotic system, although the quadratic system was readily identified by both architectures. A radial basis network architecture was able to capture the qualitative behavior of the chaotic system on predictive data sets. However, the absolute error in the estimates remained high. Subsampling either produced no substantive improvement or led to impractical retina sizes.