Bayesian Networks. Exact Inference by Variable Elimination. Emma Rollon and Javier Larrosa Q

Transcription

1 Bayesian Networks Exact Inference by Variable Elimination Emma Rollon and Javier Larrosa Q Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

2 Recall the most usual queries Let X = E Q Z, where E are the evidence variables, e are the observed values, Q are the variables of interest, Z are the rest of the variables. Probability of Evidence (Q = ): Prior Marginals (E = ): Posterior Marginals (E ): Pr(e) = Z Pr(e, Z) Pr(Q) = Z Pr(Q, Z) Pr(Q e) = Z Pr(Q, Z e) Most Probable Explanation (MPE) (E, Z = ): q = arg max Q Pr(Q e) Maximum a Posteriori Probability (MAP): q = arg max Q Pr(Q e) = arg max Q Z Pr(Q, Z e) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

3 In this session We are going to see a generic algorithm for inference (inference = query answer) It is called Variable Elimination because it eliminates one by one those variables which are irrelevant for the query. It relies on some basic operations on a class of functions known as factors. It uses an algorithmic technique called dynamic programming We will illustrate it for the computation of prior and posteriors marginals, P(Y) = Z P(Y, Z) P(Y e) = Z P(Y, Z e) It can be easily extended to other queries. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

4 Factors A factor f is a function over a set of variables {X 1, X 2,..., X k }, called scope, mapping each instantiation of these variables to a real number R. X Y Z f (X, Y, Z) Observation: Condition Probability Tables (CPTs) are factors. We define three operations on factors: Product: f (X, Y, Z) g(q, Y, R) Conditioning: f (X, y, Z) Marginalization: y Y f (X, y, Z) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

5 Factors: Product X Y f (X, Y ) Y Z g(y, Z) X Y Z f (X, Y ) g(y, Z) Factor multiplication is commutative and associative. It is time and space exponential in the number of variables in the resulting factor. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

6 Factors: Conditioning X Y f (X, Y ) Y f (0, Y ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

7 Factors: Marginalization X Y f (X, Y ) Y x X f (x, Y ) Factor marginalization is commutative. It is time exponential in the number of variables in the original factor, and space exponential in the number of variables in the resulting factor. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

8 Factors: Example Given the following BN: compute: Pr(S I ) Pr(I ), Pr(G i 1, D), G Pr(L G), L Pr(L G) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

9 The basic property: distributivity When dealing with integers, we know that: n n K i = K i i=1 i=1 The same happens when dealing with factors (e.g., D(X ) = {x 0, x 1 }): f (Z) g(x, Y ) = f (Z) g(x 0, Y ) + f (Z) g(x 1, Y ) = X = f (Z) (g(x 0, Y ) + g(x 1, Y )) = f (Z) X g(x, Y ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

10 VE: prior marginal on a chain We will illustrate VE in the simplest case (a chain with no evidence): Boolean variables X i The BN is a chain: P(X 1, X 2, X 3, X 4 ) = Pr(X 1 )Pr(X 2 X 1 )Pr(X 3 X 2 )Pr(X 4 X 3 ) We have no evidence (E = ) The only variable of interest is X 4 Thus, we want to compute: Pr(X 4 ) = Pr(X 1 )Pr(X 2 X 1 )Pr(X 3 X 2 )Pr(X 4 X 3 ) X 1 X 2 X 3 Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

11 VE: prior marginal on a chain Elimination of X 1 : Pr(X 4 ) = Pr(X 1 )Pr(X 2 X 1 )Pr(X 3 X 2 )Pr(X 4 X 3 ) = X 1 X 2 X 3 = Pr(X 1 )Pr(X 2 X 1 )Pr(X 3 X 2 )Pr(X 4 X 3 ) = X 2 X 3 X 1 = Pr(X 3 X 2 )Pr(X 4 X 3 )( Pr(X 1 )Pr(X 2 X 1 )) = X 2 X 3 X 1 = Pr(X 3 X 2 )Pr(X 4 X 3 )λ 1 (X 2 ) X 2 X 3 where λ 1 (X 2 ) = X 1 Pr(X 1 )Pr(X 2 X 1 ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

12 VE: prior marginal on a chain Elimination of X 2 : P(X 4 ) = Pr(X 3 X 2 )Pr(X 4 X 3 )λ 1 (X 2 ) X 2 X 3 = Pr(X 3 X 2 )Pr(X 4 X 3 )λ 1 (X 2 ) = X 3 X 2 = Pr(X 4 X 3 )( Pr(X 3 X 2 )λ 1 (X 2 )) = X 3 X 2 = X 3 Pr(X 4 X 3 )λ 2 (X 3 ) = where λ 2 (X 3 ) = X 2 Pr(X 3 X 2 )λ 1 (X 2 ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

13 VE: prior marginal on a chain Elimination of X 3 : P(X 4 ) = X 3 Pr(X 4 X 3 )λ 2 (X 3 ) = λ 3 (X 4 ) Question: What would be the time and space complexity of this execution? Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

14 VE: prior marginal on general networks Joint probability distribution: P(C, D, I, G, L, S, H, J) = P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )P(D C)P(C) Goal: P(G, L, S, H, J) Elimination order: C, D, I Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

15 VE: prior marginal on general networks Elimination of C: P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )P(D C)P(C) = I,D,C = P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )( P(D C)P(C)) I,D C = I,D P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )λ C (D) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

16 VE: prior marginal on general networks Elimination of D: P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )λ C (D) I,D = P(H G, J)P(J L, S)P(L G)P(S I )P(I )( P(G D, I )λ C (D)) = I D = I P(H G, J)P(J L, S)P(L G)P(S I )P(I )λ D (G, I ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

17 VE: prior marginal on general networks Elimination of I : P(H G, J)P(J L, S)P(L G)P(S I )P(I )λ D (G, I ) I P(H G, J)P(J L, S)P(L G)( P(S I )P(I )λ D (G, I )) = I P(H G, J)P(J L, S)P(L G)λ I (S, G) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

18 VE: posterior marginal on general networks Given the joint probability: P(C, D, I, G, L, S, H, J) = P(H G, J)P(J L, S)P(L G)P(S I )P(I )P(G D, I )P(D C)P(C) and evidence {I = i, H = h}, we want to compute: P(J, S, L i, h) = P(J, S, L, i, h) P(i, h) = C,D,G J,S,L,C,D,G P(C, D, i, G, L, S, h, J) P(C, D, i, G, L, S, h, J) using elimination order C, D, G. Trick Compute P(J, S, L, i, h) and normalize. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

19 VE: posterior marginal on general networks Elimination of C: P(h G, J)P(J L, S)P(L G)P(S i)p(i)p(g D, i)p(d C)P(C) = C,D,G P(h G, J)P(J L, S)P(L G)P(S i)p(i)p(g D, i)( D,G C P(h G, J)P(J L, S)P(L G)P(S i)p(i)p(g D, i)λ C (D) D,G P(D C)P(C)) = Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

20 VE: posterior marginal on general networks Elimination of D: P(h G, J)P(J L, S)P(L G)P(S i)p(i)p(g D, i)λ C (D) = D,G P(h G, J)P(J L, S)P(L G)P(S i)p(i)( P(G D, i)λ C (D)) = G D P(h G, J)P(J L, S)P(L G)P(S i)p(i)λ D (G) G Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

21 VE: posterior marginal on general networks Elimination of G: P(h G, J)P(J L, S)P(L G)P(S i)p(i)λ D (G) = G P(J L, S)P(S i)p(i)( P(h G, J)P(L G)λ D (G)) = G P(J L, S)P(S i)p(i)λ G (J, L) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

22 VE: Summary Condition all factors on evidence. Eliminate/marginalize each non-query variable X : Move factors mentioning X towards the right. Push-in the summation over X towards the right. Replace expression by λ function. Normalize to get distribution. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

23 VE: Exercise Given the following BN: Eliminate Letter and Difficulty from the network. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

24 VE: Some tricks Given a BN and a set of query nodes Q, and a set of evidence variables E: Pruning edges: the effect of conditioning is to remove any edge from a node in E to its parents. Pruning nodes: we can remove any leaf node (with its CPT) from the network as long as it does not belong to variables in Q E. Emma Rollon and Javier Larrosa Bayesian Networks Q / 25

25 Complexity and Elimination Order P(C, X 1,..., X n ) = P(C)P(X 1 C)P(X 2 C)... P(X n C) Eliminate X n first: P(C)P(X 1 C)P(X 2 C)... P(X n 1 C)( P(X n C)) = C,X 1,X 2,...X n 1 X n P(C)P(X 1 C)P(X 2 C)... P(X n 1 C)λ Xn (C) C,X 1,X 2,...,X n 1 Eliminate C first: X 1,X 2,...,X n ( P(C)P(X 1 C)P(X 2 C)... P(X n C)) C X 1,X 2,...,X n λ C (X 1, X 2,..., X n ) Emma Rollon and Javier Larrosa Bayesian Networks Q / 25