No Output for Long Observation Sequences

[Mohamad Aboufoul]

Hello,

I have been trying to build an unsupervised HMM with 3 hidden states and 10 emissions. I currently have 81 sequences with lengths ranging from ~500 to ~1700 observations. I used a KMeansLearner followed by a BaumWelchLearner to build the model.

My issue is that my output has no start, transition, nor emission probabilities. I have no runtime nor compile time errors, but I get an output like this (with box characters in place of the question mark in diamond characters):

State 1

  Pi: NaN

  Aij: � � �

  Opdf: Integer distribution --- � � � � � � � � � �

When I cut down the lengths of each sequence to 1/7th their original length, however, I get an actual output like this:

State 1

  Pi: 0.34567901234567905

  Aij: 0.11 0.475 0.415

  Opdf: Integer distribution --- 0 0.375 0.557 0.068 0 0 0 0 0 0

I was wondering if there was some limit on the number of observations Jahmm can handle when building Hidden Markov Models. There appears to be a tradeoff as I tried seeing how many full sequences I could use and still get an actual output (which ended up being 1-2). Is there a way of changing this so that Jahmm can handle larger sequences?

Any input is much appreciated. Thank you!