First, the introduction of
He said before inference problem is known probability variables seek other variables, in view of the model is mainly the posterior probability of seeking hidden variables will be used. There are some hidden relationship between the variables is not so complicated, can be accurately calculated, although they are still in trouble, but whatever the outcome is calculable, concluded that this method is accurate, precise inference is relatively simple, do not write; there is really no calculated law, can only approximate estimation, the inferred approximate belief propagation and the main loop, the introduction of variation estimation variation distribution points, sampled by the analog sampling line with a distribution of the sample. Sample collection have direct sampling and indirect sampling, direct sampling if they can explain the probability distribution is relatively simple, there will mainly discuss indirect sampling, sampling is through a relatively simple distribution, and then add a number of conditions to process samples.
Second, the exact inference
(A) variable elimination method
(B) belief propagation method
Third, the approximate inference
(A) variation estimation
(B) based on the sampling inference
Fourth, the sampling method
(A) sampling theory
(B) sampling refuse
(C) the importance sampling
(D) Gibbs sampling
(E) Markov Monte Carlo sampling