A new algorithm to design optimal Hidden Markov Models

A new algorithm for the design of Hidden Markov Models (HMM) from observed symbol and bisymbol probabilities is presented. The algorithm provides a global optimum and makes use of linear vectorial models of sequences of probabilistic vectors estimated during an off-line learning process. Moreover, a method to enhance observed data so as to comply with the constraints of HMM strings is proposed. An optimal estimate of the number of states of the HMM for the given observed data is also provided.