It was great up until "renormalize," which is tossed out without explanation.

Clear, consise, to the point. Tutorial still holds up 9 years later.

Landed on this after bumping head first on many books, articles and videos. This helped, so thanks.

Thanks! One way to do it is as a search problem. Goal states are states where all variables have been put into the ordering. The state is the variables eliminated thus far. An action is the choice of next variable to eliminate. The cost could be, for example, the size of the factor generated during the elimination, or (a little more intricate) the max(0, size of factor generated - largest factor generated thus far). A uniform cost (or A*) search would allow you find the optimal ordering.

You nailed it. Please explain renormalize in details.

Video is great but I don't know how to renormalize and I am crying now.

UNITO: project IaLab, leave a like.

For a query of the type P(Q_1, Q_2, ..., Q_m | e_1, e_2, ..., e_n) we call the Q_i variables the query variables, the e_i variables the evidence variables, and the remaining variables in the Bayes' net are the hidden variables. Variable elimination proceeds by eliminating one hidden variable at a time. When all hidden variables have been consumed, then all the remaining factors need to be multiplied together to obtain P(Q_1, Q_2, ..., Q_m, e_1, e_2, ..., e_n).

Many thanks and it really helps :). When you say Re normalize to find for eg. P(U|+z), this implies P(+z,U)/P(+z). Is that right?

thank you for your great tutorial <3

Does ordering of elimination matter? If not, then how did you think of the order?

Could you please explain how to "renormalise" with an example? Thanks in advance

How to choose the elimination ordering ??

good even after 11 years

what if query was P(!U | +z)?

When you join the factors you aren't supposed to get a joint probability every time.