Chaotic Data and Model Building

The criterion hidden Markov models are trained to minimize is essentially an entropy. Here, such models are trained to fit numerically generated data from the Rossler system and estimates of the KS entropy h μ are derived from Lyapunov exponent calculations. The values attained for the training objective function are found to be inferior to the values that the h μ estimates suggest are possible. The nature of the models produced by extensive training is found to be sensitive to the choices of initial model parameters.