Step-by-step hidden Markov chain log-likelihood calculation

Specific Domains Statistics
1

I have a HMC for which the transition and observation matrices are time varying (and have some 0 elements).

I would like to calculate the log-likelihood (for Bayesian inference). Which library should I use? I looked at a few of them, but they seemed to have a high-level interface for constant matrices, and I need the low-level building blocks for updating, because of the time variation.

I need a numerically stable implementation, I can always program the algorithm from Rabiner’s tutorial, but would prefer not to do it if it has been done. Playing well with AutoDiff.jl would be a bonus (for NUTS/Hamiltonian MC).

2

Not sure about AutoDiff.jl, but our code follows the conventions of Distributions.jl and will work with ForwarDiff.jl:

https://github.com/jmxpearson/VinDsl.jl/blob/master/src/distributions/HMM.jl

3

something weird is going on with the logz member. you don’t set it in the constructor, but you use it later. maybe simply missing from new?

4

Thanks for that. Fixed now.