# hw3sol - 10708 Graphical Models Homework 3 Solutions 1...

This preview shows pages 1–5. Sign up to view the full content.

10708 Graphical Models: Homework 3 Solutions 1 Triangulation 1. The moralized Bayes net (ﬁgure 1) is produced by marrying the parents and dropping direction on DAG edges. A B C D E F G H Figure 1: Moralized Bayes Net 2. A,B,G,H,D,E,F,C 3. C,A,B,G,H,D,E,F 4. The clique trees are in ﬁgures 2 and 3. 2 Clique Tree Factorization 2.1 Let us prove that if the clique tree is in the state π ( C i ) = P ( C i ) , μ ij ( S ij ) = P ( S ij ) , (1) 1

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
ABC EFH 0 CEF 1 DEC 1 2 DEG 0 1 1 2 1 2 Figure 2: Clique Tree for part 2 ABCDEF DEG 2 EFH 2 1 Figure 3: Clique Tree for part 3 then a new message for any edge and any direction i j will be equal to μ ij and will not change the potential of C j . By deﬁnition of belief propagation, δ i j = X C i \ S ij π i = X C i \ S ij P ( C i ) = P ( S ij ) = μ ij (2) so the edge belief for edge i - j does not change, and because π new j = π j δ i j μ ij = π j × 1 = π j (3) the belief of clique j does not change either. 2.2 Consider an edge i - j , where clique C i is a leaf in the clique tree. Let W i = C i \ S ij . By the chain rule, P ( X ) = P ( W i | X \ W i ) P ( X \ W i ) (4) and by the independence property from question 4, P ( W i | X \ W i ) = P ( W i | S ij ) = P ( W i , S ij ) P ( S ij ) = P ( C i ) P ( S ij ) . (5) 2
Thus, P ( X ) = P ( C i ) P ( S ij ) P ( X \ W i ) . (6) Now apply the same reasoning to any leaf of C \ C i , where C is the original clique tree, until there are no more cliques (each clique will become a leaf at some point in time, when all but one of its neighbors are eliminated). Each clique C j will contribute a factor P ( C j ) P ( S jk ) (7) where j - k is one of the edges involving clique C j , to result in P ( X ) = Q i P ( C i ) Q ( i,j ) E P ( S ij ) (8) (each P ( S ij ) will be contributed only once, when the ﬁrst of two corresponding cliques is eliminated). 3

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
3) Variable Elimination in Clique Trees Q: Consider a chain graphical model with the structure X 1 - X 2 - - X n , where X i takes on one of d possible assignments. You can form the following clique tree for this GM: C 1 - C 2 - - C n - 1 , where Scope @ C i D = 8 X i , X i + 1 < . You can assume that this clique tree has already been calibrated. Using this clique tree, we can directly obtain
This is the end of the preview. Sign up to access the rest of the document.

## This note was uploaded on 05/25/2008 for the course MACHINE LE 10708 taught by Professor Carlosgustin during the Fall '07 term at Carnegie Mellon.

### Page1 / 8

hw3sol - 10708 Graphical Models Homework 3 Solutions 1...

This preview shows document pages 1 - 5. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online