on the revision of probabilistic beliefs using uncertain evidence hei chan and adnan darwiche ucla...

16
On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Upload: timothy-morgan

Post on 16-Dec-2015

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

On the Revision of Probabilistic Beliefs using Uncertain Evidence

Hei Chan and Adnan Darwiche

UCLA

Presented by: Valerie Sessions

October 6, 2004

Page 2: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Overview

• Jeffrey’s Rule / Probability Kinematics

• Virtual Evidence Method

• Switching between methods

• Interpreting evidential statements

• Commutativity of Revisions

• Bounding Belief Change

Page 3: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Questions to Keep in Mind(1) How should one specify uncertain evidence?(2) How should one revise a probability

distribution?(3) How should one interpret informal evidential

statements?(4) Should, and do, iterated belief revisions

commute?(5) What guarantees can be offered on the amount

of belief change induced by a particular revision?

Page 4: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Probability Kinematics

• Two probability distributions disagree on probabilities for a set of events, but agree on how that event affects another event.

)(rP)Pr( gg cc

)|(rP)|Pr( gg cscs

Page 5: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Jeffrey’s Rule• Uses Probability Kinetics

• Given a probability distribution and some uncertain evidence bearing on this we have…

)Pr(

),Pr()(rP

1 g

gn

ii c

csqs

)(rP gg cq

Page 6: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Example 1

)Pr(

)(rP),Pr()|(rP

g

ggg c

ccscs

3.0

7.012.0

= 0.28

Page 7: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Virtual Evidence Method

• Given PR and new evidence n we have

)|Pr(),|Pr(

)|Pr(

AnBAn

An

n

jj

n

ii

A

ABnB

1

1

)Pr(

),Pr()|Pr(

Page 8: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Example 2

000369.0)|,Pr(

)989901.0*1()000005.0*1()009999.0*4()000095.0*4(

000095.0*4)|,Pr(

),Pr(

),Pr()|,Pr(

1

ba

ba

ba

baba

n

jj

a

Page 9: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Virtual Evidence -> Jeffrey’s Rule

Virtual Evidence

aaaa :)|Pr(:)|Pr(

To Jeffrey’s:

)|Pr()(rP

)|Pr()(rP

aqa

aqa

a

a

)Pr()Pr(

)Pr()|Pr(

aa

aa

aa

a

Page 10: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Jeffrey’s Rule -> Virtual Evidence

• Divide new Prob. by old Prob. for ratio

41

4

)Pr(

)(rP

a

aa

Page 11: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Virtual Evidence and Jeffrey’s Rule in Belief Networks

• Virtual Evidence was built for this

Burglary TRUE FALSETRUE 0.95 0.01FALSE 0.05 0.99

TRUE 0.0001FALSE 0.9999

P(B) P(A)

Alarm TRUE FALSETRUE 0.4 0.1FALSE 0.6 0.9

P(n|A)

1:4)|Pr(:)|Pr( aa For Jeffrey’s Rule -> Convert to Virtual Evidence and then put in belief network (cheat)

Page 12: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Interpreting Evidential Statements

• Looking at the evidence, I am willing to bet 2:1 that David is not the killer.

• Jeffrey’s Rule – “All things considered”– Pr'(killer) = 2/3

– Pr'(not killer) = 1/3

• Virtual Evidence – “Nothing else considered”– Pr(evidence|killer):Pr(evidence|not killer) = 2 : 1

Page 13: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Process for Mapping Evidence(1) One must adopt a formal method for specifying

evidence (Jeffrey’s Rule or Virtual Evidence)

(2) One must interpret the informal evidence statement as a formal piece of evidence using the method chosen

(3) One must apply a revision, by mapping the original probability distribution and formal piece of evidence into a new distribution, according to a belief revision principle

Page 14: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Commutativity of Iterated Revisions

• Jeffrey’s Rule is not commutative

• Wagner suggests Bayes Factors

)|Pr(

)|Pr()|(

ba

baba

Odd of a given b are defined by:

Bayes factor given by:

2

1

21

21

21

2121Pr,rP )Pr()Pr(

)(rP)(rP

):(

):():(

ee

ee

ee

eeeeF

Page 15: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Bounding Belief Change

• Chan and Darwiche present a distance measure to bind belief revisions

)Pr(

)(rPminln

)Pr(

)(rPmaxln)rP(Pr,

D

)rP(Pr,)rP(Pr,

)|(

)|(

DD eBA

BAe

Page 16: On the Revision of Probabilistic Beliefs using Uncertain Evidence Hei Chan and Adnan Darwiche UCLA Presented by: Valerie Sessions October 6, 2004

Bounding Belief Change

• Using these theorems with Jeffrey’s Rule and the Virtual Evidence MethodJeffrey’s Rule

)Pr(minln

)Pr(maxln)rP(Pr,

11i

in

ii

in

i e

q

e

qD

Virtual Evidence Method

i

n

ii

n

iD

11minlnmaxln))|Pr((Pr,