Title: MPE and Partial Inversion in Lifted Probabilistic Variable Elimination
1MPE and Partial Inversion inLifted Probabilistic
Variable Elimination
- Rodrigo de Salvo Braz
- University of Illinois at
- Urbana-Champaign
with Eyal Amir and Dan Roth
2Lifted Probabilistic Inference
- We assume probabilistic statements such as8
Person, DiseaseP(sick(Person,Disease)
epidemics(Disease)) 0.3 - Typical approach is grounding.
- We seek to do inference at first-order level,
like it is done in logic. - Faster and more intelligible.
- Two contributions
- Partial inversion more general technique than
previous work (IJCAI '05) - MPE and Lifted assignments
3Representing structure
epidemic(measles)
epidemic(flu)
sick(mary,measles)
sick(mary,flu)
sick(bob,measles)
sick(bob,flu)
Poole (2003) named these parfactors, for
parameterized factors
Logical variable
epidemic(D)
Atom
sick(P,D)
4Parfactor
8 Person, Disease f(sick(Person,Disease),
epidemic(Disease))
5Parfactor
Person ? mary, Disease ? flu
8 Person, Disease f(sick(Person,Disease),
epidemic(Disease)), Person ? mary, Disease ? flu
6Joint Distribution
- As in propositional case, proportional to product
of all factors - But here, all factors means all instantiations
of all parfactors - P(...) ? ÕX f1(p(X)) ÕX,Y f2(p(X),q(X,Y))
7Inference task - Marginalization
- åq(X,Y) ÕX f1(p(X)) ÕX,Y f2(p(X),q(X,Y))
- Marginal on all random variables in p(X)
- summation over all assignments to all instances
of q(X,Y)
8The FOVE Algorithm
- First-Order Variable Elimination (FOVE) a
generalization of Variable Elimination in
propositional graphical models. - Eliminates classes of random variables at once.
9FOVE
epidemic(measles)
epidemic(D)
D ? measles
sick(mary,measles)
sick(mary, D)
D ? measles
hospital(mary)
10FOVE
epidemic(D)
D ? measles
sick(mary,measles)
sick(mary, D)
D ? measles
hospital(mary)
11FOVE
epidemic(D)
D ? measles
sick(mary, D)
D ? measles
hospital(mary)
12FOVE
D ? measles
sick(mary, D)
D ? measles
hospital(mary)
13FOVE
hospital(mary)
14Counting Elimination - A Combinatorial Approach
- åe(D) ÕD1?D2 f(e(D1),e(D2))
- åe(D) f(0,0)(0,0) in assignment
f(0,1)(0,1) in assignment
f(1,0)(1,0) in assignment
f(1,1)(1,1) in assignment - Let i be the number of e(D)s assigned 1
- å i Õv1,v2 f(v1,v2)(v1,v2) given i
- ? (number of assignments with D
e(D)1 i)
15Counting Elimination - Conditions
- It does not work oneliminating class epidemic
fromf(epidemic(D1, Region), epidemic(D2,
Region), donations). - In general, counting elimination does not apply
when atoms share logical variables. - Here, Region is shared between atoms.
16Partial Inversion
Provides a way of not sharing logical
variables åe(D,R) ÕD1?D2,R f( e(D1,R), e(D2,R), d
) ÕR åe(D,r) ÕD1?D2 f( e(D1,r), e(D2,r), d ) (R
is now bound, so not a variable anymore) ÕR f(
d ) f( d )R f( d )
17Partial Inversion, graphically
epidemic(D1,R)
Each instance a counting elimination problem
donations
D1 ? D2
epidemic(D2,R)
epidemic(D1,r1)
epidemic(D1,r10)
D1 ? D2
D1 ? D2
epidemic(D2,r1)
epidemic(D2,r10)
donations
18Another (not so partial) inversion
- åq(X,Y) ÕX,Y f(p(X),q(X,Y)) (expensive)
- ÕX,Y åq(X,Y) f(p(X),q(X,Y)) (propositional)
- ÕX,Y f'(p(X))
- ÕX f'Y(p(X))
- ÕX f''(p(X)) (marginal on p(X))
19Another (not so partial) inversion
p(X)
Each instance a propositional elimination problem
q(X,Y)
p(x1)
p(xn)
q(x1,y1)
q(xn,yn)
20Partial inversion conditions
f( friends(X,Y), friends(Y,X)) Cannot partially
invert on X,Y because friends(bob,mary) appears
in more than one instance of parfactor.
21Summary of Partial Inversion
- More general than previousInversion Elimination.
- Generates Counting Elimination or Propositional
sub-problems. - Cannot be applied to entangled parfactors.
- Does not depend on domain size.
22Second contribution Lifted MPE
- In propositional case,MPE done by factors
containing MPE of eliminated variables.
C
A
B
D
23MPE
- In propositional case,MPE done by factors
containing MPE of eliminated variables.
B D f MPE
0 0 0.3 C1
0 1 0.2 C1
1 0 0.5 C0
1 1 0.9 C1
A
B
D
24MPE
- In propositional case,MPE done by factors
containing MPE of eliminated variables.
B f MPE
0 0.5 C1,D0
1 1.4 C1,D1
A
B
25MPE
- In propositional case,MPE done by factors
containing MPE of eliminated variables.
A f MPE(B,C,D)
0 0.9 B0,C1,D0
1 0.7 B1,C1,D1
A
26MPE
- In propositional case,MPE done by factors
containing MPE of eliminated variables.
f MPE
0.9 A0,B1,C1,D1
27MPE
- Same idea in First-order case
- But factors are quantified and so are
assignments
8 X, Y f(p(X), q(X,Y))
p(X) q(X,Y) f MPE
0 0 0.3 r(X,Y) 1
0 1 0.2 r(X,Y) 1
1 0 0.5 r(X,Y) 0
1 1 0.9 r(X,Y) 1
28MPE
8 X, Y f(p(X), q(X,Y))
p(X) q(X,Y) f MPE
0 0 0.3 r(X,Y) 1
0 1 0.9 r(X,Y) 1
1 0 0.5 r(X,Y) 0
1 1 0.3 r(X,Y) 1
Liftedassignments
- After Inversion Elimination of q(X,Y)
8 X f(p(X))
p(X) f MPE
0 0.05 8 Y q(X,Y) 1, r(X,Y) 1
1 0.02 8 Y q(X,Y) 0, r(X,Y) 1
29MPE
8 X f(p(X))
p(X) f MPE
0 0.05 8 Y q(X,Y) 1, r(X,Y) 1
1 0.02 8 Y q(X,Y) 0, r(X,Y) 1
- After Inversion Elimination of p(X)
f()
f MPE
0.009 8 X 8 Y p(X) 0, q(X,Y) 1, r(X,Y) 0
30MPE
8 D1, D2 f(e(D1), e(D2))
e(D1) e(D2) f MPE
0 0 0.3 r(D1,D2) 1
0 1 0.9 r(D1,D2) 1
1 0 0.5 r(D1,D2) 0
1 1 0.3 r(D1,D2) 1
f()
- After Counting Elimination of e
f MPE
0.05 938 D1,D2 e(D1)0, e(D2) 0, r(D1,D2) 1912 D1,D2 e(D1)0, e(D2) 1, r(D1,D2) 1915 D1,D2 e(D1)1, e(D2) 0, r(D1,D2) 0925 D1,D2 e(D1)1, e(D2) 1, r(D1,D2) 1
31Conclusions
- Partial InversionMore general algorithm,
subsumes Inversion elimination - Lifted Most Probable Explanation (MPE)
- same idea as in propositional VE, but with
- Lifted assignments
- describe sets of basic assignments
- universally quantified comes from Partial
Inversion - existentially quantified comes from Counting
elimination - Ultimate goal
- to perform lifted probabilistic inference in way
similar to logic inference without grounding and
at a higher level.
32(No Transcript)