FOR BETTER PERFORMANCE USE THE RDONLP2 PACKAGE (SEE INSTRUCTIONS BELOW).
depmix is a package for fitting multigroup mixtures of latent/hidden Markov
models for arbitrary length multivariate timeseries data of mixed
categorical and continuous variables. This includes as special cases the
following models: finite mixtures and latent class models (T=1), the latent
Markov model for univariate and multivariate timeseries, and mixtures of
the latter. Moreover, it includes the possibility of specifying general
linear constraints between parameters.
The possible response distributions are the multinomial and the gaussian
(normal). Analytic gradients are implemented. Standard errors are
computed based on a finite differences approximation of the Hessian using
the analytic gradients.
Parameter estimation is done by direct optimization of the likelihood using
reparametrization (for the linear equality constraints) and a penalty
function for violoation of inequality constraints. The use of the Rdonlp2
package for optimization is recommended as it deals much better with
general linear constraints; furthermore, there is optional support for
using NPSOL.
USING RDONLP2
Optimization of these models works much better using Rdonlp2. To get this
to work do the following:
1) Install the Rdonlp2 package (from http://arumat.net/Rdonlp2/, the
licencse says, among other things, "The free use of donlp2 and parts of it
is restricted for research purposes.")
2) Change the default optimization option in the function fitdmm, ie the
method argument, to "donlp", ie set method="donlp".
3) Install depmix.