Lecture 6

# How hard is inference in practice exact inference in

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: lative error ρ is NP-hard. • Thm: Computing P (Xi|e) with absolute error for any ∈ (0, 0.5) is NP-hard. • But: special cases may have error bounds. \$ & ¨ ¨  ... ©  £ \$ £ ¡ \$ & § & ¦ & & ; • And: heuristics often work well. How hard is inference in practice? Exact inference in Gaussian models takes O (N 3) time • We will show later that exact inference can always be done in time O (N v W ), where WG is the tree-width of the graph (to be deﬁned later) and v = maxi |Xi| is the max number of values (states) each node can take. • For Gaussian graphical models, exact inference is O (N 3) no matter what the graph structure is! • The NP-hardness proof shows that, in the worst case, we have WG ∼ N . • But for many models used in practice, we have WG ∼ constant. • Also, for Gaussian graphical models, exact inference is O (N 3) no matter what the graph structure is! • c.f., linear programming easier than integer programming. • Lecture 3: Any undirected graphical model in which potentials have the form ψij = exp(Xi − µi)Σ−1(Xj − µj ) ij can be converted to a joint Gaussian distribution. • Book chap 4: any directed graphical model in which CPDs have the form p(Xi|Xπi ) = N (Xi; W Xπi + µi, Σi) can be converted to a joint Gaussian distribution. • Exact inference in a Gaussian graphical model = matrix inversion. Variable elimination algorithm Working right to left (peeling) Coherence Difficulty Intelligence Grade P (J ) = Job L G H I D L S G H I D L S G H I D φJ (J, L, S ) τ3 (G,S ) L L S G H G S τ4 (G,J ) φJ (J, L, S ) L S φL (L, G)τ4(G, J )τ3(G, S ) G τ5 (J,L,S ) = φJ (J, L, S )τ5(J, L, S ) L S τ6 (J,L) C φH (H, G, J ) H φS (S, I )φI (I ) I φ( G, I,...
View Full Document

## This document was uploaded on 03/28/2014 for the course CS 532 at UBC.

Ask a homework question - tutors are online