for loop missing in the backward algorithm

[davidbp]

In the current version of the pseudocode of backward algorithm we have

where c_l is not specified. Should it be something like this?

This is relevant. I have create an issue in the lxmls_toolkit since:

• the emission probabilities are outside the logsum in the forward algorithm
• the emission probabilities are inside the logsum in the backward algorithm

Maybe this detail about the logsum is already correct. I am sure anyway that the loop over the different states is needed.

[ramon-astudillo]

The code is right. Have a look at Eq 25 in http://www.gao.ece.ufl.edu/eel6825/rabiner_hmm.pdf

for a more general view on why this is like this see

http://www.stat.columbia.edu/~liam/teaching/neurostat-fall14/papers/hmm/minka-lds-techreport.pdf

[davidbp]

THe code might be right but what about the pseudocode of the algorithm? In the code there is a double for in both the forward and the backward algorithms but the pseudocode given just has a for in the backward algorithm.

[ramon-astudillo]

Ok, that part seems to be wrong, yes. So the parenthesis are ok, but there is indeed a missing for. Could you please ad it?.

[davidbp]

Corrected with pull request