2

I am running a Baum-Welch HMM algorithm (in R). The sequence vector contains a series of observations which have been gathered from a dataset where the data has 17 states.

I can successfully run the HMM algorithm and it converges without any problems when I set pseudo count to 1e-09, below that the algorithm fails.

My question relates to the Baum-Welch algorithm and local minima. When I run the algorithm and obtain the estimates for the emission matrix and transition matrix, I then use these as inputs to calculate the posterior probability of a sequence.

That is all fine except that when I redo all of the above (in a new R session) I get completely different posterior probability estimates. Is this because the BW algorithm is getting stuck in different local minima each time?

Raphael
  • 73,212
  • 30
  • 182
  • 400
user3745220
  • 21
  • 1

1 Answers1

2

It is impossible to locate the issue for sure without inspecting the implementation and results in more detail than is ontopic here. For instance, numeric algorithms always have potential issues related to the precision of the used number format, especially when dealing with (very small) probabilities.

That said, two notes.

  • Baum-Welch does only find local optima.
  • Insufficient training data can cause parameter estimations to vary wildly.

So I recommend you try more training data if possible. Otherwise, just run the algorithm many times and pick the best model (according to the likelihood of the training data according to the resp. model) you find.

Raphael
  • 73,212
  • 30
  • 182
  • 400