NaNs in posterior as HMM state probability converges to 1.0

[mshin]

Hi, @Marty.

NaNs may be produced due to numerical instability when calculating hmm_hidden_state_prob.

In the post below, I had a similar chat with you. My guess is that the function hmm_hidden_state_prob may need to employ log_sum_exp to bypass the numerical instability issue. I also encountered NaNs when using hmm_hidden_state_prob and the problem was gone when custom filtered probabilities was used, which employed log_sum_exp.

Manual ordering of two simplexes for emission probabilities in multinomial HMM

Hi, @Marty. I encountered similar issue and I think I kind of figured out the answer. Basically, you can re-calculate filtered state probabilities using log_omega, Gamma, rho, which are required for hmm_marginal function. One thing to note here is that if you try to modify the hmm_gidden_state_prob function, the results produce NaNs at times, possibly because of under/overflow. I needed to use some parts of the code that I originally used, to avoid NaN issue. The implemented function is as …

I hope this helps!
Minho