Giters

fmorenopino / HeterogeneousHMM

Discrete, Gaussian, and Heterogenous HMM models full implemented in Python. Missing data, Model Selection Criteria (AIC/BIC), and Semi-Supervised training supported. Easily extendable with other types of probablistic models.

Home Page:https://pyhhmm.readthedocs.io/en/latest/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Learning Rate for Multinomial HMM

tahaceritli opened this issue · comments

commented

Hi all,

Thanks for developing this tool! It is quite useful that it handles missing data unlike hmmlearn.

I was playing with the Multinomial HMM. I noticed that the learning rate was said to be 0 by default in the documentation, but it was set to 0.1 in the implementation.

I have set it to 0 so that it's aligned with the documentation and ran the example in the notebook. See Problem 3 in https://github.com/tahaceritli/HeterogeneousHMM/blob/master/examples/hmm_tutorials.ipynb. I compared it with the hmmlearn.MultinomiallHMM. The loglikelihoods are quite similar, but the estimated parameters are different. The parameters estimated via hmmlearn are very similar to the true model parameters, but I can't say the same for HeterogeneousHMM. I also tried with n_iter > 1; however, t didn't change the results.

I was wondering if you have any idea about why that might be the case.

Btw, the parameters of the true model were printed before rather than the parameters of the learned model in the example at https://github.com/fmorenopino/HeterogeneousHMM/blob/master/examples/hmm_tutorials.ipynb.

Best,
Taha

commented

Hi Taha,

I'd like to apologise for getting back to you so late. Thanks for contacting us with regards to this. I haven't had time yet to take a look at it, but I will make some during the upcoming week, and get back to you.

Best,
Emese