#map-inference

Question 1: Single choice

easyAI-checkedProbabilistic & Graphical Models

You want $\operatorname*{arg\,max}_{\mathbf{x}} p(x_1, \ldots, x_5)$ for an unnormalised model $p \propto \prod_k \phi_k$ by max-sum message passing. Do you need the normalising constant $Z$ ?

Yes; the max-sum messages are only valid for normalised factors1
Yes; it is needed to turn the log messages into probabilities2
No; scaling $p$ by a constant changes neither the arg max nor the backtracking3
Only if the factors have different scales4

Timed test