51
Artificial Intelligence Learning from observations Chapter 18, AIMA

Artificial Intelligence

  • Upload
    ethel

  • View
    21

  • Download
    0

Embed Size (px)

DESCRIPTION

Artificial Intelligence. Learning from observations Chapter 18, AIMA. Two types of learning in AI. Deductive : Deduce rules/facts from already known rules/facts. (We have already dealt with this) Inductive : Learn new rules/facts from a data set D. - PowerPoint PPT Presentation

Citation preview

Page 1: Artificial Intelligence

Artificial Intelligence

Learning from observationsChapter 18, AIMA

Page 2: Artificial Intelligence

Two types of learning in AI

Deductive: Deduce rules/facts from already known rules/facts. (We have already dealt with this)

Inductive: Learn new rules/facts from a data set D.

CACBA

CAnyn Nn ...1)(),(xD

We will be dealing with the latter, inductive learning, now

Page 3: Artificial Intelligence

Two types of inductive learning

Supervised: The machine has access to a teacher who corrects it.

Unsupervised: No access to teacher. Instead, the machine must search for “order” and “structure” in the environment.

Page 4: Artificial Intelligence

Inductive learning - example A

• f(x) is the target function• An example is a pair [x, f(x)]• Learning task: find a hypothesis h such that h(x) f(x) given a

training set of examples D = {[xi, f(xi) ]}, i = 1,2,…,N

1)(,

0

0

1

0

1

0

1

1

1

xx f

1)(,

0

0

1

1

1

0

0

1

1

xx f

0)(,

0

1

0

1

1

0

0

1

1

xx f

Etc...

Inspired by a slide from V. Pavlovic

Page 5: Artificial Intelligence

Inductive learning – example B

• Construct h so that it agrees with f.• The hypothesis h is consistent if it agrees with f on all

observations.• Ockham’s razor: Select the simplest consistent hypothesis.

• How achieve good generalization?

Consistent linear fit Consistent 7th order polynomial fit

Inconsistent linear fit.Consistent 6th orderpolynomial fit.

Consistent sinusoidal fit

Page 6: Artificial Intelligence

Inductive learning – example C

x

y

Example from V. Pavlovic @ Rutgers

Page 7: Artificial Intelligence

Inductive learning – example C

x

y

Example from V. Pavlovic @ Rutgers

Page 8: Artificial Intelligence

Inductive learning – example C

x

y

Example from V. Pavlovic @ Rutgers

Page 9: Artificial Intelligence

Inductive learning – example C

x

y

Example from V. Pavlovic @ Rutgers

Sometimes a consistent hypothesis is worse than an inconsistent

Page 10: Artificial Intelligence

The idealized inductive learning problem

Our hypothesis space

f(x)

hopt(x) H

Error

Find appropriate hypothesis space H and findh(x) H with minimum “distance” to f(x) (“error”)

The learning problem is realizable if f(x) ∈ H.

Page 11: Artificial Intelligence

Data is never noise free and never available in infinite amounts, so we get variation in data and model.The generalization error is a function of both the training data and the hypothesis selection method.

Find appropriate hypothesis space H and minimize the expected distance to f(x) (“generalization error”)

{f(x)}

{hopt(x)}

Egen

The real inductive learning problem

Page 12: Artificial Intelligence

Hypothesis spaces (examples)

f(x) = 0.5 + x + x2 + 6x3123

1={a+bx}; 2={a+bx+cx2}; 3={a+bx+cx2+dx3};Linear; Quadratic; Cubic;

1 2 3

Page 13: Artificial Intelligence

Learning problems

• The hypothesis takes as input a set of attributes x and returns a ”decision” h(x) = the predicted (estimated) output value for the input x.

• Discrete valued function ⇒ classification• Continuous valued function ⇒ regression

Page 14: Artificial Intelligence

Classification

Order into one out of several classes

KD CX Input space Output (category) space

D

D

X

x

x

x

2

1

xK

K

C

c

c

c

0

1

0

2

1

c

Page 15: Artificial Intelligence

Example: Robot color vision

Classify the Lego pieces into red, blue, and yellow.Classify white balls, black sideboard, and green carpet.Input = pixel in image, output = category

Page 16: Artificial Intelligence

RegressionThe “fixed regressor model”

)()( xx gf

x Observed inputf(x) Observed outputg(x) True underlying function I.I.D noise process

with zero mean

Page 17: Artificial Intelligence

Example: Predict price for cotton futures

Input: Past historyof closing prices,and trading volume

Output: Predictedclosing price

Page 18: Artificial Intelligence

Decision trees

• “Divide and conquer”: Split data into smaller and smaller subsets.

• Splits usually on a single variable

x1 > ?

yes no

x2 > ? x2 > ?

yes yesno no

Page 19: Artificial Intelligence

The wait@restaurant decision tree

This is our true function.Can we learn this tree from examples?

Page 20: Artificial Intelligence

Inductive learning of decision tree

• Simplest: Construct a decision tree with one leaf for every example = memory based learning.Not very good generalization.

• Advanced: Split on each variable so that the purity of each split increases (i.e. either only yes or only no)

• Purity measured,e.g, with entropy

Page 21: Artificial Intelligence

Inductive learning of decision tree

• Simplest: Construct a decision tree with one leaf for every example = memory based learning.Not very good generalization.

• Advanced: Split on each variable so that the purity of each split increases (i.e. either only yes or only no)

• Purity measured,e.g, with entropy

Page 22: Artificial Intelligence

Inductive learning of decision tree

• Simplest: Construct a decision tree with one leaf for every example = memory based learning.Not very good generalization.

• Advanced: Split on each variable so that the purity of each split increases (i.e. either only yes or only no)

• Purity measured,e.g, with entropy

)](ln[)()](ln[)(Entropy noPnoPyesPyesP

i

ii vPvP )(ln)(EntropyGeneral form:

Page 23: Artificial Intelligence

The entropy is maximal whenall possibilities are equallylikely.

The goal of the decision treeis to decrease the entropy ineach node.

Entropy is zero in a pure ”yes”node (or pure ”no” node).

The second law of thermodynamics:Elements in a closed system tend to seek their most probable distribution; in a closed system entropy always increases

Entropy is a measure of ”order” in asystem.

Page 24: Artificial Intelligence

Decision tree learning algorithm

• Create pure nodes whenever possible• If pure nodes are not possible, choose the

split that leads to the largest decrease in entropy.

Page 25: Artificial Intelligence

Decision tree learning example

10 attributes:1. Alternate: Is there a suitable alternative restaurant

nearby? {yes,no}2. Bar: Is there a bar to wait in? {yes,no}3. Fri/Sat: Is it Friday or Saturday? {yes,no}4. Hungry: Are you hungry? {yes,no}5. Patrons: How many are seated in the restaurant? {none,

some, full}6. Price: Price level {$,$$,$$$}7. Raining: Is it raining? {yes,no}8. Reservation: Did you make a reservation? {yes,no}9. Type: Type of food {French,Italian,Thai,Burger}10. Wait: {0-10 min, 10-30 min, 30-60 min, >60 min}

Page 26: Artificial Intelligence

Decision tree learning example

T = True, F = False 6 True,6 False 30.012

6ln126

126ln12

6Entropy

Page 27: Artificial Intelligence

Decision tree learning example

30.063ln6

36

3ln63

12

66

3ln63

63ln6

312

6Entropy

Alternate?

3 T, 3 F 3 T, 3 F

Yes No

Entropy decrease = 0.30 – 0.30 = 0

Page 28: Artificial Intelligence

Decision tree learning example

30.063ln6

36

3ln63

12

66

3ln63

63ln6

312

6Entropy

Bar?

3 T, 3 F 3 T, 3 F

Yes No

Entropy decrease = 0.30 – 0.30 = 0

Page 29: Artificial Intelligence

Decision tree learning example

29.073ln7

37

4ln74

12

75

3ln53

52ln5

212

5Entropy

Sat/Fri?

2 T, 3 F 4 T, 3 F

Yes No

Entropy decrease = 0.30 – 0.29 = 0.01

Page 30: Artificial Intelligence

Decision tree learning example

24.054ln5

45

1ln51

12

57

2ln72

75ln7

512

7Entropy

Hungry?

5 T, 2 F 1 T, 4 F

Yes No

Entropy decrease = 0.30 – 0.24 = 0.06

Page 31: Artificial Intelligence

Decision tree learning example

30.084ln8

48

4ln84

12

84

2ln42

42ln4

212

4Entropy

Raining?

2 T, 2 F 4 T, 4 F

Yes No

Entropy decrease = 0.30 – 0.30 = 0

Page 32: Artificial Intelligence

Decision tree learning example

29.074ln7

47

3ln73

12

75

2ln52

53ln5

312

5Entropy

Reservation?

3 T, 2 F 3 T, 4 F

Yes No

Entropy decrease = 0.30 – 0.29 = 0.01

Page 33: Artificial Intelligence

Decision tree learning example

14.06

4ln64

62ln6

212

6

40ln4

04

4ln44

12

42

2ln22

20ln2

012

2Entropy

Patrons?

2 F

4 T

None Full

Entropy decrease = 0.30 – 0.14 = 0.16

2 T, 4 FSome

Page 34: Artificial Intelligence

Decision tree learning example

23.04

3ln43

41ln4

112

4

20ln2

02

2ln22

12

26

3ln63

63ln6

312

6Entropy

Price

3 T, 3 F

2 T

$ $$$

Entropy decrease = 0.30 – 0.23 = 0.07

1 T, 3 F$$

Page 35: Artificial Intelligence

Decision tree learning example

30.04

2ln42

42ln4

212

44

2ln42

42ln4

212

4

21ln2

12

1ln21

12

22

1ln21

21ln2

112

2Entropy

Type

1 T, 1 F

1 T, 1 F

French Burger

Entropy decrease = 0.30 – 0.30 = 0

2 T, 2 FItalian

2 T, 2 F

Thai

Page 36: Artificial Intelligence

Decision tree learning example

24.02

2ln22

20ln2

012

22

1ln21

21ln2

112

2

21ln2

12

1ln21

12

26

2ln62

64ln6

412

6Entropy

Est. waitingtime

4 T, 2 F

1 T, 1 F

0-10 > 60

Entropy decrease = 0.30 – 0.24 = 0.06

2 F10-30

1 T, 1 F

30-60

Page 37: Artificial Intelligence

Decision tree learning example

Patrons?

2 F

4 T

None Full

Largest entropy decrease (0.16)achieved by splitting on Patrons.

2 T, 4 FSome

X? Continue like this, making new splits, always purifying nodes.

Page 38: Artificial Intelligence

Decision tree learning example

Induced tree (from examples)

Page 39: Artificial Intelligence

Decision tree learning example

True tree

Page 40: Artificial Intelligence

Decision tree learning example

Induced tree (from examples)

Cannot make it more complexthan what the data supports.

Page 41: Artificial Intelligence

How do we know it is correct?

How do we know that h f ? (Hume's Problem of Induction)– Try h on a new test set of examples

(cross validation)

...and assume the ”principle of uniformity”, i.e. the result we get on this test data should be indicative of results on future data. Causality is constant.

Inspired by a slide by V. Pavlovic

Page 42: Artificial Intelligence

Learning curve for the decision tree algorithm on 100 randomlygenerated examples in the restaurant domain.The graph summarizes 20 trials.

Page 43: Artificial Intelligence

Cross-validationUse a “validation set”.

Dtrain

Dval

Eval

valgen EE

Split your data set into twoparts, one for training yourmodel and the other for validating your model.The error on the validation data is called “validation error”(Eval)

Page 44: Artificial Intelligence

K-Fold Cross-validationMore accurate than using only one validation set.

Dtrain

Dval

Dtrain

Dtrain

Dtrain

Dval

Dval

Eval(1) Eval(2) Eval(3)

K

kvalvalgen kE

KEE

1

)(1

Page 45: Artificial Intelligence

PAC

• Any hypothesis that is consistent with a sufficiently large set of training (and test) examples is unlikely to be seriously wrong; it is probably approximately correct (PAC).

• What is the relationship between the generalization error and the number of samples needed to achieve this generalization error?

Page 46: Artificial Intelligence

instance space X

f

h

f and h disagree

The error

X = the set of all possible examples (instance space).D = the distribution of these examples.H = the hypothesis space (h H).N = the number of training data.

Image adapted from F. Hoffmann @ KTH

] fromdrawn |)()([)(error DfhPh xxx

Page 47: Artificial Intelligence

Probability for bad hypothesis

Suppose we have a bad hypothesis h with error(h) > .What is the probability that it is consistent with N

samples?

• Probability for being inconsistent with one sample = error(h) > .

• Probability for being consistent with one sample = 1 – error(h) < 1 – .

• Probability for being consistent with N independently drawn samples < (1 – )N.

Page 48: Artificial Intelligence

Probability for bad hypothesis

What is the probability that the set Hbad of bad

hypotheses with error(h) > contains a consistent hypothesis?

NNhhP )1()1())(error consistent ( bad HH

Page 49: Artificial Intelligence

Probability for bad hypothesis

What is the probability that the set Hbad of bad

hypotheses with error(h) > contains a consistent hypothesis?

NNhhP )1()1())(error consistent ( bad HH

If we want this to be less than some constant , then

ln)1ln(ln1 NN HH

Page 50: Artificial Intelligence

Probability for bad hypothesis

What is the probability that the set Hbad of bad

hypotheses with error(h) > contains a consistent hypothesis?

NNhhP )1()1())(error consistent ( bad HH

If we want this to be less than some constant , then

)ln(|)ln(|

)1ln(

)ln(|)ln(|

HH

N

Don’t expect to learn very well if H is large

Page 51: Artificial Intelligence

How make learning work?

• Use simple hypotheses– Always start with the simple ones first

• Constrain H with priors– Do we know something about the domain?– Do we have reasonable a priori beliefs on

parameters?

• Use many observations– Easy to say...

• Cross-validation...