1 / 33

Inference in Bayesian Networks

Inference in Bayesian Networks. Pr: Evidence: Pr(e) Posterior marginals: Pr(x|e) for every X MPE: Most probable instantiation: Instantiation y such that Pr(y|e) is maximal (Y = E) MAP: Maximum a posteriori hypothesis: Intantiation y such that Pr(y|e) is maximal (Y is subset of E).

diep
Download Presentation

Inference in Bayesian Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Inference in Bayesian Networks A. Darwiche

  2. Pr: Evidence: Pr(e) Posterior marginals: Pr(x|e) for every X MPE: Most probable instantiation: Instantiation y such that Pr(y|e) is maximal (Y = E) MAP: Maximum a posteriori hypothesis: Intantiation y such that Pr(y|e) is maximal (Y is subset of E) Query Types A. Darwiche

  3. Pr: Posterior Marginals Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  4. Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio ok on yes no ok off yes no Diagnosis Scenario .001 .090 A. Darwiche

  5. MPE: Most Probable Explanation Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  6. MPE: Most Probable Explanation Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  7. MAP: Maximum a Posteriori Hypothesis Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  8. MAP: Maximum a Posteriori Hypothesis Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  9. MAP: Maximum a Posteriori Hypothesis Battery Age Alternator Fan Belt Charge Delivered Battery Fuel Pump Fuel Line Starter Distributor Gas Battery Power Spark Plugs Gas Gauge Engine Start Lights Engine Turn Over Radio A. Darwiche

  10. Probability of Evidence A. Darwiche

  11. + * * A B ØB A B + + A ØA Factoring true false .9 true .3 true .1 true false true .8 false .7 * * * * * * false false .2 .3 λa λb .7 .1 .9 .8 .2 λ~a λ~b false A. Darwiche

  12. Notation • A binary variable X: • is variable with two values (true, false) • x is short notation for X=true • ~x is short notation for X=false • If X is a variable with parents Y and Z, then: represents the probability Pr(X=x | Y=y, Z=y) • If X is a binary variable with parents Y and Z (also binary), then: represents the probability Pr(X=true | Y=false, Z=true) A. Darwiche

  13. An instantiation is a set of variables with their values: X=true,Y=false, Z=true is an instantiation A=a, B=b, C=c is an instantiation x, ~y, z is short notation for the instantiationX=true, Y=false, Z=true a,b,c is short notation for the instantiation A=a, B=b, C=c Two instantiations are inconsistent iff they assign different values to the same variable: x,~y,z and x,y,z are inconsistent x,~y,z and a,b,c are consistent Notation A. Darwiche

  14. A B Pr true true .03 true false .27 false true .56 false false false .14 false Joint Probability Distribution Pr(a) = .03 + .27 = .3 A. Darwiche

  15. A B Pr true true .03 true false .27 false true .56 false false false .14 false Joint Probability Distribution Pr(~b) = .27 + .14 = .41 A. Darwiche

  16. λa λb …are called evidence indicators Evidence Indicators A B Pr λaλb .03 .03 true true .27 λaλ~b .27 true false λ~aλb .56 .56 false true false λ~aλ~b .14 .14 false false false F = .03λaλb + .27λaλ~b + .56λ~aλb + .14λ~a λ~b F is called the polynomial of the given probability distribution A. Darwiche

  17. To compute the probability of instantiation e: Evaluate polynomial F while replacing each indicator -by 1 if the instantiation is consistent with the indicator;-by 0 if the instantiation is inconsistent with the indicator Examples: Indicator λa is consistent with instantiation a,~b,c Indicator λb is inconsistent with instantiation a,~b,c Indicator λd is consistent with instantiation a,~b,c Indicator λ~d is consistent with instantiation a,~b,c Computing Probabilities A. Darwiche

  18. Computing Probabilities F = .03λaλb + .27λaλ~b + .56λ~aλb + .14λ~a λ~b To compute the probability of instantiation a, ~b: F(a,~b) = .03*1*0+ .27*1*1+ .56*0*0+ .14*0*1 = .27 To compute the probability of instantiation ~a: F(~a) = .03*0*1+ .27*0*1+ .56*1*1+ .14*1*1 = .70 A. Darwiche

  19. A B Pr .03=.3*.1 true true A B .27=.3*.9 true false ØB A B A ØA true false .9 56=.7*.8 false true true .3 true .1 true false true .8 false .7 false false .2 false .14=.7*.2 false false false A. Darwiche

  20. A B ØB A B A ØA θ b|a true true θa θ~b|a true true false θ~a θ b|~a false true false θ ~b|~a false false A B Pr θa θ b|a true true θa θ~b|a true false θ~a θ b|~a false true false θ~a θ ~b|~a false false false A. Darwiche

  21. A B ØB A B A ØA θ b|a true true θa θ~b|a true true false θ~a θ b|~a false true false θ ~b|~a false false A B Pr λaλbθa θ b|a true true λaλ~bθa θ~b|a true false λ~aλbθ~a θ b|~a false true false λ~aλ~bθ~a θ ~b|~a false false false A. Darwiche

  22. A B ØB A B A ØA θ b|a true true θa θ~b|a true true false θ~a θ b|~a false true false θ ~b|~a false false A B Pr λaλbθa θ b|a true true λaλ~bθa θ~b|a true false λ~aλbθ~a θ b|~a false true false λ~aλ~bθ~a θ ~b|~a false false λa λb θa θb|a+ λa λ~b θa θ~b|a+ λ~aλb θ~a θb|~a+ λ~a λ~b θ~a θ~b|~a A. Darwiche

  23. B A ØB A B A ØA true true θ b|a true θa true false θ~b|a false θ~a false true θ b|~a false false θ ~b|~a false The Polynomial of a Bayesian Network λa λb θa θb|a+ λa λ~b θa θ~b|a+ λ~aλb θ~a θb|~a+ λ~a λ~b θ~a θ~b|~a A. Darwiche

  24. C B D A The Polynomial of a Bayesian Network F = λa λb λc λd θa θb|a θc|a θd|bc + λa λb λc λ~d θa θb|a θc|a θ~d|bc + …. A. Darwiche

  25. + Factoring * * + + * * * * * * θa λa λb θab θa~b θ~ab λ~b θ~a~b λ~a θ~a Arithmetic Circuit λa λb θa θb|a+ λa λ~b θa θ~b|a+ λ~aλb θ~a θb|~a+ λ~a λ~b θ~a θ~b|~a A. Darwiche

  26. Arithmetic Circuit Pr(a) .3 + * * .3 0 + + 1 1 * * * * * .3 .1 .9 .8 .2 0 * θa λa λb 1 1 1 0 .3 .1 .9 .8 .2 .7 θab θa~b θ~ab λ~b θ~a~b λ~a θ~a A. Darwiche

  27. + * * A B ØB A B + + A ØA Factoring true false .9 true .3 true .1 true false true .8 false .7 * * * * * * false false .2 θa λa λb θ~a θb|a θ~b|a θb|~a θ~b|~a λ~a λ~b false A. Darwiche

  28. T S1 S2 S3 … Sn Factoring the Polynomial of a Bayesian Network A. Darwiche

  29. Sophisticated Platform (desktop) Eval Eval Eval compiler A. Circuit A. Circuit A. Circuit Primitive Platforms (embedded) Embedding Probabilistic Reasoning Systems A. Darwiche

  30. TreeWidth(Measures connectivity of Networks) Higher treewidth A. Darwiche

  31. TreeWidth(Measures connectivity of Networks) Multiply-connected networks Singly-connected network (polytree) A. Darwiche

  32. The treewidth of a polytree is m, where m is the maximum number of parents that any node If each node has at most one parent, the polytree is called a tree The treewidth of a tree is 1 Treewidth A. Darwiche

  33. Given a Bayesian network N with: Number of nodes: n Treewith: w We can generate an arithmetic circuit for N: In O(n 2w) time In O(n 2w) space It is easy to do inference on polytrees Treewidth A. Darwiche

More Related