I am not familiar with the RHmm package, but in theory there should be
no problem with discrete multivariate observations. However, in general
in order to fit an HMM you need to specify a family of ***distributions***
for your observations (one distribution for each of the hidden states).
Usually the family is of parametric form with the values of the parameters
depending on the states.
This could be tricky, depending on the nature of your data. Without more
information it is impossible to say.
If the number of possible values of observations is reasonably small
relative to the total number of observations --- which seems to me to be
unlikely --- then you could fit a discrete non-parametric family. (You could
try using the package hmm.discnp.)
cheers,
Rolf Turner
On 09/03/13 02:47, Claus O'Rourke wrote:> Hi r-help,
>
> I have been using your RHmm package for some time and have recently
> had to try using the package for a new dataset.
>
> Basically I have a dataset with a number of discrete observation
> variables that change over time, and I would love to try modeling them
> using a HMM.
>
> Basically I was wondering if RHmm can be used to model a multivariate
> discrete HMM, i.e., the observations are a vector of discrete
> measurements? From what I see in the documentation and from playing
> around with examples, it seems like this may not be possible. My
> understand of the mathematics behind multivariate HMMs is limited, so
> I would appreciate any advance you might be able to give.
>
> Thanks for any help anyone can give.