Re: some hidden markov understanding questions...

Hi Kox,

1. Is there always an "explanation" for the hidden states?

I believe sometimes the hidden states do not correspond to any known
variable, states, etc., and that one can only call them "latent state
1", "latent state 2" etc.

2. Is there any way to "test" my HMM-algorithm,

Yes. Like this:

a. Choose an arbitrary but plausible set of values for the parameters
of the model.

b. Based on those parameter values, generate a random sequence of
observed states (pseudo-data).

c. Input the pseudo-data to your B-W estimation algorithm. The
algorithm should give parameter estimates close to the values you used
to generate the data. The longer the pseudo-data sequence, the closer
the estimates should be to the values you specified.

For your other questions, you might ask the HMM experts in the Yahoo
group BayesNetToolbox:

http://groups.yahoo.com/group/BayesNetToolbox/

They are working on HMM software for Matlab.

Hope this helps.
--
John Uebersax PhD

.