what's the differenece between continuousHMM and discreateHMM

[daniellocust]

Hello, Nickgillian!

I'm studying the HMM, and read the code of continuousHiddenmarkovmodel.cpp. In the train_(...) function, i didn't find the code to implement Baum-Welch algorithm, but the algorithm is implemented in DiscreteHiddenMarkovModel.cpp. So, i wanna to know what difference is between this two kind of algorithm, and why continuousHMM need not to implement Baum-Welch algorithm? Thanks!

[nickgillian]

Hi daniellocust,

Your analysis of the two HMM algorithms (discrete and continuous) is correct. The two algorithms use very different algorithms:

• the discrete HMM algorithm uses the classic HMM algorithm, part of which is the Baum Welch algorithm for training the state transition and obverstation matrices, see my thesis for more info on this.
• the continuous HMM algorithm is roughly based on a version of this HMM algorithm by Bevilacqua et. al. If you read this paper you should see the main algorithm behind the continuous HMM.

At some point, I would like to revisit both algorithms and improve them, as neither implementation is optimal, so if you're interested in helping to improve the core algorithms behind either/both then let me know.

[daniellocust]

Hi Nickgillian, Thank you very much!