More Probabilistic Models Introduction to Artificial Intelligence COS302 Michael L. Littman Fall 2001.

More Probabilistic ModelsMore Probabilistic Models

Introduction toIntroduction toArtificial IntelligenceArtificial Intelligence

COS302COS302

Michael L. LittmanMichael L. Littman

Fall 2001Fall 2001

AdministrationAdministration

2/3, 1/3 split for exams2/3, 1/3 split for exams

Last HW due WednesdayLast HW due Wednesday

Wrap up WednesdayWrap up Wednesday

Sample exam questions later…Sample exam questions later…

Example analogies, share, etc.Example analogies, share, etc.

TopicsTopics

Goal: Try to practice what we know Goal: Try to practice what we know about probabilistic modelsabout probabilistic models

• Segmentation: most likely Segmentation: most likely sequence of wordssequence of words

• EM for segmentationEM for segmentation• Belief net representationBelief net representation• EM for learning probabilitiesEM for learning probabilities

SegmentationSegmentation

Add spaces:Add spaces:

bothearthandsaturnspinbothearthandsaturnspin

Applications:Applications:• no spaces in speechno spaces in speech• no spaces in Chineseno spaces in Chinese• postscript or OCR to textpostscript or OCR to text

So Many Choices…So Many Choices…

Bothearthandsaturnspin.Bothearthandsaturnspin.B O T H E A R T H A N D S A T U R N S P I N.B O T H E A R T H A N D S A T U R N S P I N.

Bo-the-art hands at Urn’s Pin.Bo-the-art hands at Urn’s Pin.

Bot heart? Ha! N D S a turns pi N.Bot heart? Ha! N D S a turns pi N.

Both Earth and Saturn spin.Both Earth and Saturn spin.

……so little time. How to choose?so little time. How to choose?

Probabilistic ApproachProbabilistic Approach

Standard spiel:Standard spiel:

1.1. Choose a generative modelChoose a generative model

2.2. Estimate parametersEstimate parameters

3.3. Find most likely sequenceFind most likely sequence

Generative ModelGenerative Model

Choices:Choices:• unigram unigram Pr(w)Pr(w)• bigram bigram Pr(w|w’)Pr(w|w’)• trigram trigram Pr(w|w’,w’’)Pr(w|w’,w’’)• tag-based HMM tag-based HMM Pr(t|t’,t’’), Pr(w|t)Pr(t|t’,t’’), Pr(w|t)• probabilistic context-free grammar probabilistic context-free grammar

Pr(X Y|Z), Pr(w|Z)Pr(X Y|Z), Pr(w|Z)

Estimate ParametersEstimate Parameters

For English, can count word For English, can count word frequencies in text sample:frequencies in text sample:

Pr(w) = count(w)/sumPr(w) = count(w)/sumw w count(w)count(w)

For Chinese, could get someone to For Chinese, could get someone to segment, or use EM (next).segment, or use EM (next).

Search AlgorithmSearch Algorithm

gotothestoregotothestore

Compute the maximum probability Compute the maximum probability sequence of words.sequence of words.

pp0 0 = 1= 1

ppj j = max= maxi<j i<j ppj-i j-i Pr(wPr(wi:ji:j))

pp5 5 = max(p= max(p0 0 Pr(Pr(gototgotot), p), p1 1 Pr(Pr(otototot), ), pp22 Pr(Pr(tottot), p), p33 Pr( Pr(otot), p), p44 Pr( Pr(tt))))

Get to point i, use one word to get to j.Get to point i, use one word to get to j.

Unigrams Probs via EMUnigrams Probs via EM

g g 0.010.01 go go 0.780.78 got got 0.210.21 goto goto 0.610.61

o o 0.020.02 t t 0.040.04 to to 0.760.76 tot tot 0.740.74

o o 0.020.02 t t 0.040.04 the the 0.830.83 thes thes 0.040.04

h h 0.030.03 he he 0.220.22 hes hes 0.160.16 hest hest 0.190.19

e e 0.050.05 es es 0.090.09

s s 0.040.04 store store 0.810.81 t t 0.040.04 to to 0.700.70 tore tore 0.070.07

o o 0.020.02 or or 0.650.65 ore ore 0.090.09

r r 0.010.01 re re 0.120.12 e e 0.050.05

EM for SegmentationEM for Segmentation

Pick unigram probabilitiesPick unigram probabilitiesRepeat until probability doesn’t Repeat until probability doesn’t

improve muchimprove much1.1. Fractionally label (like forward-Fractionally label (like forward-

backward)backward)2.2. Use fractional counts to Use fractional counts to

reestimate unigram probabilitiesreestimate unigram probabilities

Probability DistributionProbability Distribution

Represent probability distribution on Represent probability distribution on a bit sequence.a bit sequence.

A BA B Pr( Pr(ABAB))

0 00 0 .06.06

0 10 1 .24.24

1 01 0 .14.14

1 11 1 .56.56

Conditional Probs.Conditional Probs.

Pr(Pr(AA||~B~B) = .14/(.14+.06) = .7) = .14/(.14+.06) = .7

Pr(Pr(AA||BB) = .56/(.56+.24) = .7) = .56/(.56+.24) = .7

Pr(Pr(BB||~A~A) = .24/(.24+.06) = .8) = .24/(.24+.06) = .8

Pr(Pr(BB||AA) = .56/(.56+.14) = .8) = .56/(.56+.14) = .8

So, Pr(So, Pr(ABAB)=Pr()=Pr(AA)Pr()Pr(BB))

Graphical ModelGraphical Model

Pick a value for A.Pick a value for A.

Pick a value for B.Pick a value for B.

Independent influence: kind of Independent influence: kind of and/or-ish.and/or-ish.

AA BB.7.7 .8.8

Probability DistributionProbability Distribution

A BA B Pr( Pr(ABAB))

0 00 0 .08.08

0 10 1 .42.42

1 01 0 .32.32

1 11 1 .18.18

Dependent influence: Dependent influence: kind kind of xor-ish.of xor-ish.

Conditional Probs.Conditional Probs.

Pr(Pr(AA||~B~B) = .32/(.32+.08) = .8) = .32/(.32+.08) = .8

Pr(Pr(AA||BB) = .18/(.18+.42) = .3) = .18/(.18+.42) = .3

Pr(Pr(BB||~A~A) = .42/(.42+.08) = .84) = .42/(.42+.08) = .84

Pr(Pr(BB||AA) = .18/(.18+.32) = .36) = .18/(.18+.32) = .36

So, a bit more complex.So, a bit more complex.

Graphical ModelGraphical Model

Pick a value for B.Pick a value for B.

Pick a value for A, based on B.Pick a value for A, based on B.

AA

BB

B Pr(A|B)B Pr(A|B)0 .80 .81 .31 .3

.6.6

CPT: ConditionalCPT: ConditionalProbability TableProbability Table

General FormGeneral Form

Acyclic graph; each node a var.Acyclic graph; each node a var.Node with k in edges; size 2Node with k in edges; size 2kk CPT. CPT.

NN

PP11

PP1 1 PP2 2 … P… Pk k Pr(Pr(NN|P|P11 P P2 2 … P… Pkk))0 0 0 p0 0 0 p00…000…0

… …1 1 1 p1 1 1 p11…111…1

PP22 PPkk……

Belief NetworkBelief Network

Bayesian network, Bayes net, etc.Bayesian network, Bayes net, etc.

Represents a prob. distribution over Represents a prob. distribution over 22n n values with O(2values with O(2kk) entries, where ) entries, where k is the largest indegreek is the largest indegree

Can be applied to variables with Can be applied to variables with values beyond just {0, 1}. Kind of values beyond just {0, 1}. Kind of like a CSP.like a CSP.

What Can You Do?What Can You Do?

Belief net inference: Pr(Belief net inference: Pr(NN||EE11,~E,~E22,E,E33, , ……).).

Polytime algorithms exist if Polytime algorithms exist if undirected version of DAG is undirected version of DAG is acyclic (singly connected)acyclic (singly connected)

NP-hard if multiply connected.NP-hard if multiply connected.

Example BNsExample BNs

EE

BB

CC

DD

AA

EE

BB

CC

DD

AA

singlysinglymultiplymultiply

Popular BNPopular BN

WW

CC

XXVV YY ZZ

Recognize this?Recognize this?

BN ApplicationsBN Applications

Diagnosing diseasesDiagnosing diseases

Decoding noisy messages from deep Decoding noisy messages from deep space probesspace probes

Reasoning about geneticsReasoning about genetics

Understanding consumer purchasing Understanding consumer purchasing patternspatterns

Annoying users of WindowsAnnoying users of Windows

Parameter LearningParameter Learning

A B C D EA B C D E

0 0 1 0 10 0 1 0 1 Pr( Pr(BB||~A~A)?)?

0 0 1 1 10 0 1 1 1

1 1 1 0 11 1 1 0 1

0 1 0 0 10 1 0 0 1

1 0 1 0 11 0 1 0 1

0 0 1 1 00 0 1 1 0

0 0 1 1 10 0 1 1 1

EE

BB

CC

DD

AA

1/51/5

Hidden VariableHidden Variable

A B C D EA B C D E

0 0 1 0 10 0 1 0 1 Pr( Pr(BB||~A~A)?)?

0 0 1 1 10 0 1 1 1

1 1 1 0 11 1 1 0 1

0 1 0 0 10 1 0 0 1

1 0 1 0 11 0 1 0 1

0 0 1 1 00 0 1 1 0

0 0 1 1 10 0 1 1 1

EE

BB

CC

DD

AA

What to LearnWhat to Learn

Segmentation problemSegmentation problemAlgorithm for finding the most likely Algorithm for finding the most likely

segmentationsegmentationHow EM might be used for parameter How EM might be used for parameter

learninglearningBelief network representationBelief network representationHow EM might be used for parameter How EM might be used for parameter

learninglearning