Hidden Markov Problem...

Hi everyone,

I hope I'm not all the way wrong here with my problem.
I am researching face recognition here at the university of ilmenau,
and I want/have to do it with hidden-markov-models.

I implemented the HMM-code using mainly the rabiner-tutorial,
and it seems to work. (discrete HMM work definitely, continuous
I am not THAT sure, does anyone know a way to "test" this,
to verify that the algorithm is correct?)

When I do the training with my observation, created with
wavelet decomposition of picture blocks, after some iterations
of the EM-algorithm I get an error because the values of the
gaussian distributions go zero. I calculate these values with:

dens= 1/sqrt((2.*pi).^D.*norm(Sigma)) * exp((-0.5)*vec'*inv(Sigma)*vec)

vec being O-mu.
I found out that they go zero, because the exp-term is negative,
and e^ a negative number is zero, so the whole thing is zero.

The problem is, I don't know why that happens.
Can anyone give me a hint, what could be wrong or what I could do?
If you need code, just tell me, I'll show you. :)

Kox

[ comp.ai is moderated. To submit, just post and be patient, or if ]
[ that fails mail your article to <comp-ai@xxxxxxxxxxxxxxxxxx>, and ]
[ ask your news administrator to fix the problems with your system. ]
.