MAP estimation of HMMs using a prior that encourages 0s is an effective
way of learning structure (i.e., sparse transition matrices). See
@article{Brand99,
author = "M. Brand",
title = "Structure learning in conditional probability models via an
entropic prior and parameter extinction",
journal = "Neural Computation",
year = 1999,
volume = 11,
pages = "1155--1182"
}
Another classic approach is to build a tree that models the data
perfectly, and then merge states to simplify the model. See
@inproceedings{Stolcke92,
author = "A. Stolcke and S. M. Omohundro",
title = "Hidden Markov Model Induction by Bayesian Model Merging",
year = 1992,
booktitle = "NIPS-5"
}
HTH,
Kevin
