HMMTRAIN and HMMESTIMATE not working

Hello,

I am working on Discrete HMM. I have 1000 observations (2 observed states and 3 hidden states). My model parameters are initialized using rand generators. I use HMMVITERBI method to get the most likely hood hidden states. So i wl be having 1000 observed states and 1000 hidden states. And then I use HMMESTIMATE to train the model and get the appropriate transition and emission matrix. After getting that, I predict the next Hidden state from the new transition matrix (applying argmax) and also next observed state,

Now I do this prediction for next N number of days. My question is once I predict the next observation and next hidden state, I iterate back to HMMTRAIN with 1001 observation and again calculate new transition and emission matrix ..then again predict the next states.. now in this process, there is no significant change in the transition matrix across iteration. (coz adding one extra value is not making a big change). Coz of that I am getting the same output state for the whole N days..

Please help.. AM i missing something ?

Answer 1

[engstudent]

hi i am working in speech recognition using hmm and i need to know what is the step to build hmm andhow to train hmm in matlab7.9 ??????? i found in matlab hmmestimat,i have seq but i dont know what is the states? my Q is if i need to generat the states by my self????or this is wrong plz answer my question............

plz help me .....................

Answer 2

[engstudent]

hi i am working in speech recognition using hmm and i need to know what is the step to build hmm andhow to train hmm in matlab7.9 ??????? i found in matlab hmmestimat,i have seq but i dont know what is the states? my Q is if i need to generat the states by my self????or this is wrong plz answer my question............

plz help me .....................

Answer 3

[Felipe]

It depends in what are you defining by states in your problem. The training sequence takes as input a sequence of observed data. Take into account that the HMM works with hidden states (normally Q in literature) and Observations (normally O). There is a nice work about the Baum-Welch algorithm used in HMMTRAIN called: "A gentle tutorial of the EM Algorithm and its Application to Parameter Estimation for Gaussian Mixture and Hidden Markov Models" by Jeff A. Bilmes.

Good luck!