Hidden Markov Model (Underflow Problem)

[Alwardi]

Hi,

I am working on building a Class Hidden Markov Model consists of submodels which are connected as a whole model. The whole model is trained by Baum-Welch algorithm. I use training sequences which are DNA sequences of length 5000 to 10000 nucleic acids (samples). Because the training sequences are long I am facing an underflow problem in the training of the model. I used the LOG function in the calculations and I used scaled HMM but this did not solve the problem. Does any one have a better solution for this problem??

 

Thanks,

Saj