likelihood in EM algorithm -
regarding em algorithm. if understand correctly, likelihood obtained in e-step (for instance in baum–welch algorithm likelihood can obtained form forward-backward procedure). however, final step in each em-iteration m-step. means likelihood computed in step k, "belongs" parameters of step (k-1). so, or missing something?
the em algorithm aims at, given model, finding parameters of model maximize likelihood of data.
the focus on parameters , not on likelihood value.
from understanding, e-step meant compute expected value of log-likelihood function. in practise, consists in computing values needed m-step, using current parameters values. in m-step, parameters maximize likelihood (or lower bound). can subsequently compute likelihood value new parameters.
are working hmms? if yes, can specify quantities computed @ each step.
hope can help! :)
Comments
Post a Comment