first-order probabilistic inference rodrigo de salvo braz
DESCRIPTION
3 Slides online Just search for “Rodrigo de Salvo Braz” and check at the presentations page. (www.cs.berkeley.edu/~braz)TRANSCRIPT
![Page 1: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/1.jpg)
First-OrderProbabilistic Inference
Rodrigo de Salvo Braz
![Page 2: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/2.jpg)
2
A remarkFeel free to ask clarification questions!
![Page 3: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/3.jpg)
3
Slides online Just search for
“Rodrigo de Salvo Braz” and check at the presentations page.(www.cs.berkeley.edu/~braz)
![Page 4: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/4.jpg)
4
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 5: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/5.jpg)
5
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 6: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/6.jpg)
6
How to solvecommonsense
reasoning
How to solveNatural
LanguageProcessing
How to solvePlanning
How to solveVision
DomainKnowledge
Inferenceand
Learning
LanguageKnowledge
Inferenceand
Learning
Domain &Planning
Knowledge
Inferenceand
Learning
Objects &Optics
Knowledge
Inferenceand
Learning
Abstracting Inference and Learning
Commonsense reasoning
Natural LanguageProcessing
Planning Vision ...
Inferenceand
Learning
AI Problems
![Page 7: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/7.jpg)
7
Inference and LearningWhat capabilities should it have? Inference
andLearningIt should represent and use:
•predicates (relations) and quantification over objects: 8 X tank(X) Ç plane(X) 8 X tank(X) ) color(X)=greencolor(o121) = red
•varying degrees of uncertainty: 8 X { tank(X) ) color(X)=green }
since it may hold often but not absolutely.•essential for language, vision, pattern recognition,
commonsense reasoning etc•modal knowledge, utilities, and more.
![Page 8: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/8.jpg)
8
Abstracting Inference and Learning
Domain knowledge:
tank(X) Ç plane(X){ tank(X) ) color(X)=green}color(o121) = red...
Commonsense reasoningLanguage knowledge:
{ Sentence(S) Æ Verb(S,”attacks”) Æ Subject(S,X) Æ Object(S,Y) ) attacking(X,Y) }...
Natural Language Processing
Inferenceand
Learning
Sharing inference and learning module
![Page 9: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/9.jpg)
9
Abstracting Inference and Learning
Domain knowledge:
tank(X) Ç plane(X){ tank(X) ) color(X)=green}color(o121) = red...
Commonsense reasoning WITH Natural Language ProcessingLanguage knowledge:
{ Sentence(S) Æ Verb(S,”attacks”) Æ Subject(S,X) Æ Object(S,Y) ) attacking(X,Y) }...
Inferenceand
Learning
Joint solution made simpler
Verb(s13,”attacks”), Subject(s13,o121)...
![Page 10: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/10.jpg)
10
Standard solutions fall short Logic
Has objects and properties; but statements are absolute;
Graphical models and machine learning Have varying degrees of uncertainty; but propositional; treating objects and
quantification is awkward.
![Page 11: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/11.jpg)
11
Standard solutions fall shortGraphical models and objects
tank(X) Ç plane(X){ tank(X) )color(X)=green }{ next(X,Y) Æ tank(X) ) tank(Y) }
color(o121)=red color(o135)=green
color(121)
tank(o121) plane(o121)
next(o121, o135)
tank(o135) plane(o135)
color(o135)
•Transformation (propositionalization) not part of graphical model framework;
•Graphical model have only random variables, not objects;
•Different data creates distinct, formally unrelated model.
![Page 12: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/12.jpg)
12
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 13: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/13.jpg)
13
First-Order Probabilistic Inference Many languages have been proposed
Knowledge-based model construction (Breese, 1992) Probabilistic Logic Programming (Ng & Subrahmanian, 1992) Probabilistic Logic Programming (Ngo and Haddawy, 1995) Probabilistic Relational Models (Friedman et al., 1999) Relational Dependency Networks (Neville & Jensen, 2004) Bayesian Logic (BLOG) (Milch & Russell, 2004) Bayesian Logic Programs (Kersting & DeRaedt, 2001) DAPER (Heckerman, Meek & Koller, 2007) Markov Logic Networks (Richardson & Domingos, 2004)
“Smart” propositionalization is the main method.
![Page 14: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/14.jpg)
14
Propositionalization Bayesian Logic Programs
Prolog-like statements such as
rescue(Battalion) | in(Soldier,Battalion), wounded(Soldier).
wounded(Soldier) | in(Soldier,Battalion),
attacked(Battalion).
associated with CPTs and combination rules.
|instead of
:-
![Page 15: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/15.jpg)
15
Propositionalizationrescue(Battalion) | in(Soldier,Battalion),
wounded(Soldier). (plus CPT, cr)BN grounded from and-or tree:
rescue(bat13)
in(john,bat13)
wounded(john)
in(peter,bat13)
wounded(peter)
...
...
...
...
...
![Page 16: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/16.jpg)
16
PropositionalizationMulti-Entity Bayesian Networks (Laskey, 2004)
in(Soldier,Battalion) wounded(Soldier)
rescue(Battalion)
in(Soldier,Battalion) attacked(Battalion)
wounded(Soldier)
![Page 17: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/17.jpg)
17
PropositionalizationMulti-Entity Bayesian Networks (Laskey, 2004)
in(Soldier,Battalion) wounded(Soldier)
rescue(Battalion)
in(john,bat13) wounded(john)
rescue(bat13)
in(peter,bat13) wounded(peter)
rescue(bat13)
...
first-order fragments
instantiated
![Page 18: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/18.jpg)
18
PropositionalizationMulti-Entity Bayesian Networks (Laskey, 2004)
in(john,bat13) wounded(john)
rescue(bat13)
in(peter,bat13) wounded(peter)
...
in(john,bat13) wounded(john)
rescue(bat13)
in(peter,bat13) wounded(peter)
rescue(bat13)
...
![Page 19: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/19.jpg)
19
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 20: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/20.jpg)
20
Lifted Inference
rescue(bat13)
rescue(Battalion) ( in(Soldier,Battalion) Æ wounded(Soldier){ wounded(Soldier) ( in(Soldier,Battalion) Æ attacked(Battalion) }
wounded(Soldier)
attacked(bat13)
in(Soldier, bat13)
Unbound, general variable
•Faster•More compact•More intuitive•Higher level – more
information/structure available for optimization
![Page 21: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/21.jpg)
21
Bayesian Networks (directed)
epidemic
P(sick_john, sick_mary, sick_bob, epidemic)= P(sick_john | epidemic) * P(sick_mary | epidemic) * P(sick_bob | epidemic) * P(epidemic)
sick_john sick_mary sick_bob
P(sick_bob | epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
![Page 22: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/22.jpg)
22
Factor Networks (undirected)
epidemic_france
P(epi_france, epi_belgium, epi_uk, epi_germany)/ (epi_france, epi_belgium) * (epi_france, epi_uk) * (epi_france, epi_germany) * (epi_belgium, epi_germany)
epidemic_belgium epidemic_germany
epidemic_uk
factor,or potential
function
![Page 23: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/23.jpg)
23
Bayesian Nets as Factor Networks
epidemic
P(sick_john, sick_mary, sick_bob, epidemic)/ P(sick_john | epidemic) * P(sick_mary | epidemic) * P(sick_bob | epidemic) * P(epidemic)
sick_john sick_mary sick_bob
P(sick_bob | epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
![Page 24: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/24.jpg)
24
Inference: Marginalization
epidemic
P(sick_john) / epidemic sick_mary sick_bob P(sick_john | epidemic) * P(sick_mary | epidemic) * P(sick_bob | epidemic) * P(epidemic)
sick_john sick_mary sick_bob
P(sick_bob | epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
![Page 25: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/25.jpg)
25
epidemic
sick_john sick_mary sick_bob
P(sick_bob | epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
P(sick_john) / epidemic P(sick_john | epidemic) * P(epidemic)
* sick_mary P(sick_mary | epidemic)
* sick_bob P(sick_bob | epidemic)
Inference: Variable Elimination (VE)
![Page 26: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/26.jpg)
26
epidemic
sick_john sick_mary
(epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
P(sick_john) / epidemic P(sick_john | epidemic) * P(epidemic)
* sick_mary P(sick_mary | epidemic) * (epidemic)
Inference: Variable Elimination (VE)
![Page 27: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/27.jpg)
27
epidemic
sick_john sick_mary
(epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_mary | epidemic)
P(sick_john) / epidemic P(sick_john | epidemic) * P(epidemic) * (epidemic)
* sick_mary P(sick_mary | epidemic)
Inference: Variable Elimination (VE)
![Page 28: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/28.jpg)
28
epidemic
sick_john
(epidemic)
P(epidemic)
P(sick_john | epidemic)
P(sick_john) / epidemic P(sick_john | epidemic) * P(epidemic) * (epidemic) * (epidemic)
(epidemic)
Inference: Variable Elimination (VE)
![Page 29: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/29.jpg)
29
sick_john
(sick_john)
P(sick_john) / (sick_john)
Inference: Variable Elimination (VE)
![Page 30: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/30.jpg)
30
A Factor Network
sick(mary,measles)
hospital(mary)
epidemic(measles) epidemic(flu)
sick(mary,flu)
…
… sick(bob,measles)
hospital(bob)
sick(bob,flu)……
…
… …
… …
![Page 31: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/31.jpg)
31
First-order Representation
sick(P,D)
hospital(P)
epidemic(D)
Atoms representa set of random
variables
Logical Variablesparameterize random
variables
Parfactors, for parameterized
factors
Contraints to logical
variables
P john
![Page 32: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/32.jpg)
32
Semantics
sick(mary,measles)
hospital(mary)
epidemic(measles) epidemic(flu)
sick(mary,flu)
…
… sick(bob,measles)
hospital(bob)
sick(bob,flu)……
…
… …
… …sick(mary,measles), epidemic(measles))
![Page 33: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/33.jpg)
33
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 34: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/34.jpg)
34
Inversion Elimination (IE)
sick(P, D)
epidemic(D)
D measles
D measles
IE
…sick(john, flu)
epidemic(flu)
sick(mary, rubella)
epidemic(rubella)…
Grounding
…sick(john, flu)
epidemic(flu)
sick(mary, rubella)
epidemic(rubella)…
VE
Abstraction
epidemic(D)
D measles
D measles
sick(P, D)
sick(P,D) (epidemic(D),sick(P,D))
![Page 35: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/35.jpg)
35
Inversion Elimination (IE) Proposed by Poole (2003); Lacked formalization; Did not identify cases in which it is
incorrect; Formalized and identified correctness
conditions (with Dan Roth and Eyal Amir).
![Page 36: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/36.jpg)
36
Requires eliminated RVs to occur in separate instances of parfactor
…sick(mary, flu)
epidemic(flu)
sick(mary, rubella)
epidemic(rubella)…
sick(mary, D)
epidemic(D)
D measles
InversionElimination
correct
Inversion Elimination - Limitations
![Page 37: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/37.jpg)
37
IE not correct
Requires eliminated RVs to occur in separate instances of parfactor
Inversion Elimination - Limitations
epidemic(D2)
epidemic(D1)
D1 D2 month
epidemic(measles)
epidemic(flu)
epidemic(rubella)
…month
epidemic(D2)
epidemic(D1)
D1 D2 month
![Page 38: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/38.jpg)
38
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 39: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/39.jpg)
39
Counting Elimination Need to consider joint assignments; Exponential number of those; But actually, potential depends on
histogram of values in assignment only: 00101 the same as 11000;
Polynomial number of assignments instead.
epidemic(D2)
epidemic(D1)
D1 D2 month
epidemic(measles)
epidemic(flu)
epidemic(rubella)
…month
![Page 40: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/40.jpg)
40
A Simple Experiment
![Page 41: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/41.jpg)
41
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 42: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/42.jpg)
42
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 43: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/43.jpg)
43
BLOG (Bayesian LOGic) Milch & Russell (2004); A probabilistic logic language; Current inference is propositional
sampling; Special characteristics:
Open Universe; Expressive language.
![Page 44: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/44.jpg)
44
type Battalion;#Battalion ~ Uniform(1,10);
random Boolean Large(Battalion bat) ~ Bernoulli(0.6);
random NaturalNum Region(Battalion bat) ~ TabularCPD[[0.3, 0.4, 0.2, 0.1]]();
type Soldier;#Soldier(BatallionOf = Battalion bat)
if Large(bat) then ~ Poisson(1500); else if Region(bat) = 2 then = 300; else ~ Poisson(500);
query Average( {#Soldier(b) for Battalion b} );
BLOG (Bayesian LOGic)Open Universe
Expressive language
![Page 45: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/45.jpg)
45
Inference in BLOGrandom Boolean Attacked(bat) ~ Bernoulli(0.7);
random Boolean Wounded(Soldier sol)if Attacked(BattalionOf(sol)) then if Experienced(sol) then ~ Bernoulli(0.1); else ~ Bernoulli(0.4); else = False;
random Boolean Experienced(Soldier sol) ~ Bernoulli(0.1);
random Boolean Rescue(Battalion bat)= exists Soldier sol BattalionOf(sol) = bat & Wounded(sol);
guaranteed Battalion bat13;query Rescue(bat13);
![Page 46: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/46.jpg)
46
Inference in BLOG - Sampling
Rescue(bat13)
#Soldier(bat13)
Wounded(soldier1)
Wounded(soldier73)
Attacked(bat13)
...
Sampled false
Open Universe
false
false
false
![Page 47: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/47.jpg)
47
Inference in BLOG - Sampling
Rescue(bat13)
#Soldier(bat13)
Wounded(soldier1)
Wounded(soldier49)
Attacked(bat13)
...
Sampled true
Open Universe
true Experienced(soldier1)
truetrue
![Page 48: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/48.jpg)
48
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 49: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/49.jpg)
49
Temporal models in BLOG Expressive enough to directly write temporal
modelstype Aircraft;#Aircraft ~ Poisson[3];
random Real Position(Aircraft a, NaturalNum t) if t = 0 then ~ UniformReal[-10, 10]() else ~ Gaussian(Position(a, Pred(t)), 2);type Blip;// num of blips from aircraft a is 0 or 1#Blip(Source = Aircraft a, Time = NaturalNum t) ~ TabularCPD[[0.2, 0.8]]()
// num false alarms has Poisson distribution.#Blip(Time = NaturalNum t) ~ Poisson[2];
random Real ApparentPos(Blip b, NaturalNum t) ~ Gaussian(Position(Source(b), Pred(t)), 2));
![Page 50: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/50.jpg)
50
Temporal models in BLOG Inference algorithm does not use
temporal structure: Markov property: state depends on
previous state only; evidence and query come for successive
time steps; Dynamic BLOG (DBLOG); analogous to Dynamic Bayesian
networks (DBNs).
![Page 51: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/51.jpg)
51
DBLOG Syntax Only change in syntax is a special Timestep
type, with same semantics as NaturalNum:type Aircraft;#Aircraft ~ Poisson[3];
random Real Position(Aircraft a, Timestep t) if t = @0 then ~ UniformReal[-10, 10]() else ~ Gaussian(Position(a, Prev(t)), 2);type Blip;// num of blips from aircraft a is 0 or 1#Blip(Source = Aircraft a, Time = Timestep t) ~ TabularCPD[[0.2, 0.8]]()
// num false alarms has Poisson distribution.#Blip(Time = Timestep t) ~ Poisson[2];
random Real ApparentPos(Blip b, Timestep t) ~ Gaussian(Position(Source(b), Prev(t)), 2));
![Page 52: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/52.jpg)
52
Similar idea to DBN Particle Filtering
DBLOG Particle Filtering
evidence at t
State samples for t - 1
Samples weighed by evidence
resampling
State samples for t
Likelihoodweighting
![Page 53: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/53.jpg)
53
Weighting by evidence in DBNs
evidence
state
t-1 t
![Page 54: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/54.jpg)
54
Weighting evidence in DBLOG
evidence
state Position(a1,@1)
Position(a2,@1)
#Aircraft
t-1 t
obs {Blip b : Source(b) = a1} = {B1};
obs ApparentPos(B1, @2) = 3.2;ApparentPos(B1,@2)
#Blip(a1,@2)
Position(a1,@2)B1
Lazy instantiation
![Page 55: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/55.jpg)
55
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 56: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/56.jpg)
56
Retro-instantiation in DBLOG
Position(a1,@1)
Position(a2,@1)
#Aircraft
ApparentPos(B1,@2)
#Blip(a1,@2)
Position(a1,@2)B1
ApparentPos(B8,@9)
#Blip(a2,@9)
Position(a2,@9)
B8
Position(a2,@2)
...
obs {Blip b : Source(b) = a2} = {B8};
obs ApparentPos(B8,@2) = 2.1;
...
...
![Page 57: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/57.jpg)
57
Coping with retro-instantiation Analyse possible queries and
evidence (provided by user) andpre-instantiate random variables that will be necessary later;
Use mixing time to decide when to stop sampling back.
![Page 58: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/58.jpg)
58
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 59: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/59.jpg)
59
Improving BLOG inference Rejection sampling
evidence
evidence
= ?
We count samples that match the evidence.
![Page 60: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/60.jpg)
60
Improving BLOG inference Likelihood weighting
evidence
We count samples using evidence likelihood as weight.
![Page 61: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/61.jpg)
61
Improving BLOG inference
{Happy(p) : Person p}=
{true, true, false}
Happy(john)= true
Happy(mary)= false
Happy(peter)=false
Evidence is deterministic function, so likelihood is always 0 or 1.So likelihood weighting is no better than rejection sampling.
![Page 62: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/62.jpg)
62
Improving BLOG inference
{Salary(p) : Person p}=
{90,000, 75,000, 30,000}
Salary(john)= 100,000
Salary(mary)= 80,000
Salary(peter)=500,000
The more unlikely the evidence, the worse it gets.
![Page 63: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/63.jpg)
63
Improving BLOG inference
{ApparentPos(b) : Blip b}=
{1.2, 0.5, 5.1}
ApparentPos(b1) = 2.1
ApparentPos(b2) = 3.3
ApparentPos(b3) = 3.5
Doesn’t work at all for continuous evidence.
![Page 64: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/64.jpg)
64
Structure-dependent automatic importance sampling
{ApparentPos(b) : Blip b}=
{1.2, 0.5, 5.1}
ApparentPos(b1) ApparentPos(b2) ApparentPos(b3)
Position(Source(b1)) = 1.5
Position(Source(b2)) = 3.9
Position(Source(b3)) = 4.0
= 1.2 = 5.1 = 0.5 High-level language allows algorithm to automatically apply better sampling schemes
![Page 65: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/65.jpg)
65
Outline A goal: First-order Probabilistic Inference
An ultimate goal Propositionalization
Lifted Inference Inversion Elimination Counting Elimination
DBLOG BLOG background DBLOG Retro-instantiation Improving BLOG inference
Future Directions and Conclusion
![Page 66: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/66.jpg)
66
For the future Lifted inference works on symmetric,
indistinguishable objects only; Adapt approximations for dealing with
similar objects; Parameterized queries:
P(sick(X, measles)) = ? Function symbols:
(diabetes(X), diabetes(father(X))) Equality (paramodulation); Learning.
![Page 67: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/67.jpg)
67
Conclusion Expressive, first-order probabilistic
representations in inference too! Lifted inference equivalent to
grounded inference; much faster, yet yielding the same exact answer;
DBLOG brings techniques for temporal processes reasoning to first-order probabilistic reasoning.
![Page 68: First-Order Probabilistic Inference Rodrigo de Salvo Braz](https://reader035.vdocuments.net/reader035/viewer/2022062412/5a4d1b0a7f8b9ab059989ee7/html5/thumbnails/68.jpg)
68