Next: Current version and changes
Up: Introduction
Previous: Introduction
H2M is a set of MATLAB/OCTAVE functions that implement the EM algorithm
[1], [2] in the case of mixture models or hidden Markov
models with multivariate Gaussian state-conditional distribution. More
specifically, three special cases have been considered
- Gaussian mixture models.
- Ergodic (or fully connected) Gaussian hidden Markov models.
- Left-right Gaussian hidden Markov models.
In fact, the case 2 and 3 above do not significantly differ except for the fact
that in the case of a left-right HMM, one needs to estimate the parameters from
multiple observation sequences. In all three cases, it is possible to use
either diagonal or full covariance matrices for the state-conditional
distributions.
The H2M/cnt extension (added in version 1.6) handles similar models but for
scalar count (discrete valued positive) data. Three cases have been considered
- Mixture of Poisson distributions.
- Hidden Markov models with Poisson state conditional distribution.
- Hidden Markov models with Negative binomial state conditional
distribution.
Compared to the main H2M functions, only the case of ergodic models (ie.
models that can be trained from a single long observation sequence rather than
from multiple sequences) has been considered.
Olivier Cappé, Aug 24 2001