Training hidden markov models with the taguchi method

In some control systems structures, like predictive control, mathematical models for the control process must be derived. Those models can be obtained by a broad class of methods like parametric models applied to experimental data. In this context, and for systems with multiple operation regimes, the Hidden Markov model, due to its properties, is a convincing choice. However the parameter estimation of this type of models involves the optimization of a non-convex cost function. So the Baum-Welch method only can find sub-optimal parameters. This article shows that the use of the Taguchi method minimizes the training algorithm sensibility local minima