Probabilistic Partial Evaluation: Exploiting rule structure in probabilistic inference

Transcription

1 Probabilistic Partial Evaluation: Exploiting rule structure in probabilistic inference David Poole University of British Columbia 1

2 Overview Belief Networks Variable Elimination Algorithm Parent Contexts & Structured Representations Structure-preserving inference Conclusion 2

3 Belief (Bayesian) Networks n P(x 1,...,x n ) = P(x i x i 1...x 1 ) = i=1 n i=1 P(x i π xi ) π xi are parents of x i : set of variables such that the predecessors are independent of x i given its parents. 3

4 Variable Elimination Algorithm Given: Bayesian Network, Query variable, Observations, Elimination ordering on remaining variables 1. set observed variables 2. sum out variables according to elimination ordering 3. renormalize 4

5 Summing Out a Variable... h c d e f g a b Sum out e: P(a c, d, e) P(b e, f, g) P(e h) P(a, b c, d, f, g, h) 5

6 Structured Probability Tables P(a c, d, e) P(b e, f, g) t d e c t f t f p 1 p 2 p 3 p 4 f f e g p 5 p 6 p 7 p 8 p 2 = P(a = t d = t e = f ) 6

7 Eliminating e, preserving structure We only need to consider a & b together when d = true f = true. In this context c & g are irrelevant. In all other contexts we can consider a & b separately. When d = false f = false, e is irrelevant. In this context the probabilities shouldn t be affected by eliminating e. 7

8 Contextual Independence Given a set of variables C,acontext on C is an assignment of one value to each variable in C. Suppose X, Y and C are disjoint sets of variables. X and Y are contextually independent given context c val(c) if P(X Y=y 1 C=c) = P(X Y=y 2 C=c) for all y 1, y 2 val(y) such that P(y 1 c) >0and P(y 2 c) >0. 8

9 Parent Contexts A parent context for variable x i is a context c for a subset of the predecessors for x i such that x i is contextually independent of the other predecessors given c. For variable x i & assignment x i 1 =v i 1,...,x 1 =v 1 of values to its preceding variables, there is a parent context π v i 1...v 1 x i. P(x 1 =v 1,...,x n =v n ) n = P(x i =v n x i 1 =v i 1,...,x 1 =v 1 ) = i=1 n i=1 P(x i =v i π v i 1...v 1 x i ) 9

10 Idea behind probabilistic partial evaluation Maintain rules that are statements of probabilities in contexts. When eliminating a variable, you can ignore all rules that don t involve that variable. This wins when a variable is only in few parent contexts. Eliminating a variable looks like resolution! 10

11 Rule-based representation of our example a d e : p 1 b f e : p 5 a d e : p 2 b f e : p 6 a d c : p 3 b f g : p 7 a b c : p 4 b f g : p 8 e h : p 9 e h : p 10 11

12 Eliminating e a d e : p 1 b f e : p 5 a d e : p 2 b f e : p 6 a d c : p 3 b f g : p 7 a b c : p 4 b f g : p 8 e h : p 9 e h : p 10 unaffected by eliminating e 12

13 Variable partial evaluation If we are eliminating e, and have rules: x y e : p 1 x y e : p 2 e z : p 3 no other rules compatible with y contain e in the body y & z are compatible contexts, we create the rule: x y z : p 1 p 3 + p 2 (1 p 3 ) 13

14 Splitting Rules A rule a b : p 1 can be split on variable d, forming rules: a b d : p 1 a b d : p 1 14

15 Why Split? If there are different contexts for a given e and for a given e, you need to split the contexts to make them directly comparable: a b c e : p1 a b e : p 1 a b c e : p 1 a b c e : p 2 a b c e : p 3 15

16 Combining Heads Rules a c : p 1 b c : p 2 where a and b refer to different variables, can be combined producing: a b c : p 1 p 2 Thus in the context with a, b, and c all true, the latter rule can be used instead of the first two. 16

17 Splitting Compatible Bodies a d e : p 1 b f e : p 5 a d f e : p 1 b d f e : p 5 a d f e : p 1 b d f e : p 5 a d e : p 2 b f e : p 6 a d f e : p 2 b d f e : p 6 a d f e : p 2 b d f e : p 6 e h : p 9 e h : p 10 17

18 Combining Rules a d f e : p 1 b d f e : p 5 a d f e : p 1 b d f e : p 5 a d f e : p 2 b d f e : p 6 a d f e : p 2 b d f e : p 6 e h : p 9 e h : p 10 18

19 Result of eliminating e The resultant rules encode the probabilities of {a, b} in the contexts: d f h, d f h For all other contexts we consider a and b separately. The resulting number of rules is 24. Tree structured probability for P(a, b c, d, f, g, h, i) has 72 leaves. (Same as number of rules if a and b are combined in all contexts). VE has a table of size

20 Evidence We can set the values of all evidence variables before summing out the remaining non-query variables. Suppose e 1 =o 1... e s =o s is observed: Remove any rule that contains e i =o i, where o i = o i in the body. Remove any term e i =o i in the body of a rule. Replace any e i =o i, where o i = o i, in the head of a rule false. Replace any e i =o i in the head of a rule by true. In rule heads, use true a a, and false a false. 20

21 Conclusions New notion of parent context rule-based representation for Bayesian networks. New algorithm for probabilistic inference that preserves rule-structure. Exploits more structure than tree-based representations of conditional probability. Allows for finer-grained approximation than in a Bayesian network. 21