FITTING HIDDEN MARKOV MODELS IN R

depmixS4 provides a framework for specifying and fitting hidden Markov models. The observation densities use an interface to the glm distributions, most of which are now implemented. The vignette that accompanies the package has a table with available response distributions Besides these, observations can be modelled using the multinomial distribution with identity or logistic link function. The latter provides functionality for multinomial logistic responses with covariates. The transition matrix and the initial state probabilities are also modeled as multinomial logistics (or multinomials with identity link, which is the default when no covariates are present) with the possibility of including covariates.

Optimization is by default done using the EM algorithm. When (linear) constraints are included, package Rsolnp is used for optimization (there is also support for using Rdonlp2 as optimizer, see USING RDONLP2 below). New response distributions can be added by extending the response-class and writing appropriate methods for it (dens, and getpars and setpars); an example of this is provided on the ?makeDepmix help page. depmixS4 also fits latent class and mixture models, see ?mix for an example.

The latest development versions of depmixS4 (and depmix) can be found at: https://r-forge.r-project.org/projects/depmix/

FOR DEPMIX USERS

depmixS4 is a completely new implementation of the depmix package using S4 classes. Model specification now uses formulae and family objects, familiar from the lm and glm functions. Moreover, the transition matrix and the initial state probabilities (as well as multinomial responses) are now modeled as multinomial logistics with a baseline. Specification of linear constraints uses the same mechanism as was used in depmix, with the only difference that constraints are passed as arguments to the fit function rather than the model specification function. See the help files for further details. NOTE: most of the functionality of depmix is now also provided in depmixS4; in the future therefor I may stop updating depmix.

USING RDONLP2

Optimization of models with (general) linear (in-)equality constraint is done using Rsolnp (available from CRAN). Optionally the Rdonlp2 package can be used; Rdonlp2 was written by Ryuichi Tamura(ry.tamura @ gmail.com), and can currently be installed using:

install.packages(“Rdonlp2”, repos= c(“http://R-Forge.R-project.org", getOption(“repos”)))

Note the licence information which says, among other things: “The free use of donlp2 and parts of it is restricted for research purposes …”