Download presentation

Presentation is loading. Please wait.

Published byMikayla Ludwick Modified over 3 years ago

1
Probabilistic Inference Reading: Chapter 13 Next time: How should we define artificial intelligence? Reading for next time (see Links, Reading for Retrospective Class): Turing paper Mind, Brain and Behavior, John Searle Prepare discussion points by midnight, wed night (see end of slides)

2
2 Transition to empirical AI Add in Ability to infer new facts from old Ability to generalize Ability to learn based on past observation Key: Observation of the world Best decision given what is known

3
3 Overview of Probabilistic Inference Some terminology Inference by enumeration Bayesian Networks

4
4

5
5

6
6

7
7

8
8

9
9 Probability Basics Sample space Atomic event Probability model An event A

10
10

11
11 Random Variables Random variable Probability for a random variable

12
12

13
13

14
14

15
15

16
16

17
17 Logical Propositions and Probability Proposition = event (set of sample points) Given Boolean random variables A and B: Event a = set of sample points where A(ω)=true Event ⌐ a=set of sample points where A(ω)=false Event aΛb=points where A(ω)=true and B(ω)=true Often the sample space is the Cartesian product of the range of variables Proposition=disjunction of atomic events in which it is true (aVb) = ( ⌐ aΛb)V(aΛ ⌐ b)V(aΛb) P(aVb)= P( ⌐ aΛb)+P(aΛ ⌐ b)+P(aΛb)

18
18

19
19

20
20

21
21

22
22

23
23

24
24

25
25 Axioms of Probability All probabilities are between 0 and 1 Necessarily true propositions have probability 1. Necessarily false propositions have probability 0 The probability of a disjunction is P(aVb)=P(a)+P(b)-P(aΛb) P( ⌐ a)=1-p(a)

26
26 The definitions imply that certain logically related events must have related probabilities P(aVb)= P(a)+P(b)-P(aΛb)

27
27 Prior Probability Prior or unconditional probabilities of propositions P(female=true)=.5 corresponds to belief prior to arrival of any new evidence Probability distribution gives values for all possible assignments P(color) = (color = green, color=blue, color=purple) P(color)= (normalized: sums to 1) Joint probability distribution for a set of r.v.s gives the probability of every atomic event on those r.v.s (i.e., every sample point) P(color,gender) = a 3X2 matrix

28
28

29
29

30
30

31
31

32
32

33
33

34
34 Inference by enumeration Start with the joint distribution

35
35 Inference by enumeration P(HasTeeth)=.06+.12+.02=.2

36
36 Inference by enumeration P(HasTeethVColor=Green)=.06+.12+.02+.24=.4 4

37
37 Conditional Probability Conditional or posterior probabilities E.g., P(PlayerWins|HostOpenDoor=1 and PlayerPickDoor2 and Door1=goat) =.5 If we know more (e.g., HostOpenDoor=3 and door3-goat): P(PlayerWins)=1 Note: the less specific belief remains valid after more evidence arrives, but is not always useful New evidence may be irrelevant, allowing simplification: P(PlayerWins|California- earthquake)=P(PlayerWins)=.3

38
38 Conditional Probability A general version holds for joint distributions: P(PlayerWins,HostOpensDoor1)=P(PlayerWins|HostOpensDoor1)*P(Ho stOpensDoor1)

39
39 Inference by enumeration Compute conditional probabilities: P( ⌐Hasteeth|color=green)= P(⌐HasteethΛcolor=green) P(color=green) 0.8 = 0.24 0.06+.24

40
40 Normalization Denominator can be viewed as normalization constraint α P( ⌐Hasteeth|color=green ) = α P( ⌐Hasteeth|color=green ) =α[P( ⌐Hasteeth,color=green, female )+ P( ⌐Hasteeth,color=green, ⌐ female)] =α[ + ]=α = Compute distribution on query variable by fixing evidence variables and summing over hidden variables

41
41 Inference by enumeration

42
42 Independence A and B are independent iff P(A|B)=P(A) or P(B|A)=P(B) or P(A,B)=P(A)P(B) 32 entries reduced to 12; for n independent biased coins, 2 n -> n Absolute independence powerful but rare Any domain is large with hundreds of variables none of which are independent

43
43

44
44 Conditional Independence If I have length <=.2, the probability that I am female doesn’t depend on whether or not I have teeth: P(female|length<=.2,hasteeth)=P(female|h asteeth) The same independence holds if I am >.2 P(male|length>.2,hasteeth)=P(male|length>.2) Gender is conditionally independent of hasteeth given length

45
45 In most cases, the use of conditional independence reduces the size of the representation of the joint distribution from exponential in n to linear in n Conditional independence is our most basic and robust form of knowledge about uncertain environments

46
46 Next Class: Turing Paper A discussion class Graduate students and non-degree students: Anyone beyond a bachelor’s: Prepare a short statement on the paper. Can be your reaction, your position, a place where you disagree, an explication of a point. Undergraduates: Be prepared with questions for the graduate students All: Submit your statement or your question by midnight Wed night. All statements and questions will be printed and distributed in class on Wednesday.

Similar presentations

OK

An Introduction to Artificial Intelligence Chapter 13 &14.1-14.2: Uncertainty & Bayesian Networks Ramin Halavati

An Introduction to Artificial Intelligence Chapter 13 &14.1-14.2: Uncertainty & Bayesian Networks Ramin Halavati

© 2018 SlidePlayer.com Inc.

All rights reserved.

Ads by Google

File type ppt on cybercrime examples Presentations ppt online shopping Free download ppt on revolt of 1857 Ppt on conservation of energy resources Ppt on next generation 2-stroke engine animation Ppt on applied operations research va Download ppt on indus valley civilization for kids Ppt on biodegradable and non biodegradable items Ppt on business development planning Ppt on diodes and transistors biasing