Learning in (hidden) Markov-models using th imprecise Dirichlet model