Expert Systems Knowledge Based Systems

Transcription

1 Expert Systems Knowledge Based Systems ES-1

2 Medical diagnosis» Disease identification Example Areas of Use ES-2

3 Example Areas of Use 2 Medical diagnosis» Disease identification Natural resource exploration» Analyzing geological data ES-3

4 Example Areas of Use Medical diagnosis» Disease identification Natural resource exploration» Analyzing geological data Customizing complex equipment» Computer systems ES-4

5 Properties Behaves like an expert in a narrow area ES-5

6 Properties 2 Behaves like an expert in a narrow area Has a knowledge base of the information in the area ES-6

7 Properties 3 Behaves like an expert in a narrow area Has a knowledge base of the information in the area Ability to explain its behaviour ES-7

8 Properties 4 Behaves like an expert in a narrow area Has a knowledge base of the information in the area Ability to explain its behaviour Ability to deal with uncertain data ES-8

9 Structure Knowledge base Inference Engine User Interface Shell ES-9

10 If then else rules The most popular form of knowledge representation Typical types of rules» If condition P holds then conclude C» If situation S exists then do action A» If conditions P and Q hold then conditions C1 and C2 cannot hold ES-10

11 If then else examples See Figures 15.2, 15.3 & 15.4 in Bratko ES-11

12 Figure 15.5 in Bratko Kitchen leak example» How do you read the graph? ES-12

13 Kitchen leak example 2 Figure 15.5 in Bratko» Can see how if then else rules can represent the graph on the left hand side» Note the use of AND / OR for inputs > Arc represents AND of inputs > No arc represents OR of inputs ES-13

14 Properties of if then else rules Moduarity» Each rule or group of rules encapsulates a part of the domain ES-14

15 Properties of if then else rules 2 Moduarity» Each rule or group of rules encapsulates a part of the domain Incrementabiity» Add / delete rules as needed ES-15

16 Properties of if then else rules 3 Moduarity» Each rule or group of rules encapsulates a part of the domain Incrementabiity» Add / delete rules as needed Modifiability» Can modify small parts of the knowledge as needed ES-16

17 Properties of if then else rules 4 Moduarity» Each rule or group of rules encapsulates a part of the domain Incrementabiity» Add / delete rules as needed Modifiability» Can modify small parts of the knowledge as needed Supports transparency» Relatively easy to explain and guide system s behaviour ES-17

18 Probabilistic behaviour Can extend rule syntax to include probability information» If condition A then conclude C with probability P» See Figure 15.3 in Bratko ES-18

19 Inference Engine» With if then else rules there are two ways of making inferences.» What are they? ES-19

20 Inference Engine 2» With if then else rules there are two ways of making inferences.» What are they? > Backward chaining > Forward chaining ES-20

21 Backward chaining» What is backward chaining? ES-21

22 Backward chaining 2» What is backward chaining? > The way Prolog works from conclusions to the base facts, the confirmed facts, the evidence > See pages ES-22

23 Problems with backward chaining» What are the problems with Prolog and backward chaining? ES-23

24 Problems with backward chaining 2» What are the problems with backward chaining? > Syntax is not suitable of people unfamiliar with Prolog > Cannot distinguish knowledge base from the rest of the system ES-24

25 Problems with backward chaining 3» What are the problems with backward chaining? > Syntax is not suitable of people unfamiliar with Prolog > Cannot distinguish knowledge base from the rest of the system» How can we overcome these problems? ES-25

26 Problems with backward chaining 4» What are the problems with backward chaining? > Syntax is not suitable of people unfamiliar with Prolog > Cannot distinguish knowledge base from the rest of the system» How can we overcome these problems? > Customize the syntax with new operators Bottom of page 349 ES-26

27 Problems with backward chaining 5» What do we need to do? ES-27

28 Problems with backward chaining 6» What do we need to do? > Build an inference engine for the new rules Figure 15.6 and program text ES-28

29 Forward chaining» What is forward chaining? ES-29

30 Forward chaining 2» What is forward chaining? > Work from the base facts, the confirmed facts, the evidence, to the conclusion ES-30

31 Forward chaining 3» What is forward chaining? > Work from the base facts, the confirmed facts, the evidence, to the conclusion» What do we need to do? ES-31

32 Forward chaining 4» What is forward chaining? > Work from the base facts, the confirmed facts, the evidence, to the conclusion» What do we need to do? > Build a forward chained inference engine Figure 15.7 and program text ES-32

33 Abstract view Forward vs backward chaining Backward chaining < input information à à derived information > Forward chaining See Figure 15.5 ES-33

34 Forward vs backward chaining 2 More concrete views» data à à goals» evidence à à hypotheses» findings, observations à à explanations, diagnosis» manifestations à à diagnoses, causes ES-34

35 Forward chaining? Backward chaining? Which is better ES-35

36 Depends upon the problem Which is better ES-36

37 Which is better 2 Depends upon the problem» Check if a hypothesis is true > Work backward ES-37

38 Which is better 3 Depends upon the problem» Check if a hypothesis is true > Work backward» Many hypotheses, cannot choose > Work forward ES-38

39 Which is better 3 Depends upon the problem» Check if a hypothesis is true > Work backward» Many hypotheses, cannot choose > Work forward Forward is better when» Accumulating evidence, data ES-39

40 Shape heuristic» When input information is sparse relative to derived information» Work forward or backward? ES-40

41 Shape heuristic 2» When input information is sparse relative to derived information» Work forward or backward? > Use forward chaining ES-41

42 Shape heuristic 3» When input information is sparse relative to derived information» Work forward or backward? > Use forward chaining» When input information is rich relative to derived information» Work forward or backward? ES-42

43 Shape heuristic 5» When input information is sparse relative to derived information» Work forward or backward? > Use forward chaining» When input information is rich relative to derived information» Work forward or backward? > Use backward chaining ES-43

44 Reality Do we work forward or backward? ES-44

45 Reality 2 As tasks get more complex» Better to interleave forward and backward chaining ES-45

46 Reality Example Doctor uses initial observations, evidence to form hypothesis ES-46

47 Reality Example 2 Doctor uses initial observations, evidence to form hypothesis» Forward direction ES-47

48 Reality Example 3 Doctor uses initial observations, evidence to form hypothesis» Forward direction Pursues most likely hypothesis ES-48

49 Reality Example 4 Doctor uses initial observations, evidence to form hypothesis» Forward direction Pursues most likely hypothesis» Backward direction to find if there is confirming evidence ES-49

50 Reality Example 5 Doctor uses initial observations, evidence to form hypothesis» Forward direction Pursues most likely hypothesis» Backward direction to find if there is confirming evidence Can lead to gathering more evidence ES-50

51 Reality Example 6 Doctor uses initial observations, evidence to form hypothesis» Forward direction Pursues most likely hypothesis» Backward direction to find if there is confirming evidence Can lead to gathering more evidence, need new hypothesis» Forward direction, again ES-51

52 Reality Example 2 Top page 353 ES-52

53 Generating explanations» There are two types of explanation What are they? ES-53

54 Generating explanations» There are two types of explanation What are they? > How > Why ES-54

55 Generating explanations how How did you find the answer?» What do you present? ES-55

56 Generating explanations how 2 How did you find the answer?» What do you present? > Typically present a path trace ES-56

57 Generating explanations how 3 How did you find the answer?» What do you present? > Typically present a path trace» There is a problem in the kitchen, which was concluded from the hall being wet and the bathroom dry > And» No water came from the outside, which was concluded from the window being closed ES-57

58 Proof tree The how answer is to print out the proof tree» Top page 354, Figure 15.8 > Given program text > Compare with Figure 15.6 backward chained interpreter ES-58

59 Proof tree 2 The how answer is to print out the proof tree» Top page 354, Figure 15.8 > Given program text > Compare with Figure 15.6 backward chained interpreter The main work is printing the result in a readable format ES-59

60 Generating explanations why The why explanation is required during the reasoning process» What do you present? ES-60

61 Generating explanations why 2 The why explanation is required during the reasoning process» What do you present? > The system asks the user for information ES-61

62 Generating explanations why 3 The why explanation is required during the reasoning process» The system asks the user for information» The user may ask why ES-62

63 Generating explanations why 4 The why explanation is required during the reasoning process» The system asks the user for information» The user may ask why» The system then gives an explanation ES-63

64 Generating explanations why 5 The why explanation is required during the reasoning process» The system asks the user for information» The user may ask why» The system then gives an explanation > Pages ES-64

65 Generating explanations why 5 The why explanation is required during the reasoning process» The system asks the user for information» The user may ask why» The system then gives an explanation > Pages > Figure 15.9 and program text ES-65

66 On introducing uncertainty Chapter 15 introduces an ad hoc way of dealing with probabilities ES-66

67 On introducing uncertainty 2 Chapter 15 introduces an ad hoc way of dealing with probabilities» We will not look at these methods ES-67

68 On introducing uncertainty 3 Chapter 15 introduces an ad hoc way of dealing with probabilities» We will not look at these methods» We defer to Chapter 16 where we handle probabilities in a proper mathematical way ES-68

69 On introducing uncertainty 4 Chapter 15 introduces an ad hoc way of dealing with probabilities» We will not look at these methods» We defer to Chapter 16 where we handle probabilities in a proper mathematical way > Use Bayesian networks ES-69

70 On introducing uncertainty 5 Chapter 15 introduces an ad hoc way of dealing with probabilities» We will not look at these methods» We defer to Chapter 16 where we handle probabilities in a proper mathematical way > Use Bayesian networks > A sound and correct way of dealing with probability and uncertainty ES-70

71 On introducing uncertainty 6 Chapter 15 introduces an ad hoc way of dealing with probabilities» We will not look at these methods» We defer to Chapter 16 where we handle probabilities in a proper mathematical way > Use Bayesian networks > A sound and correct way of dealing with probability and uncertainty > Modern approach ES-71

72 Semantic networks & frames Structure facts so as to elicit information ES-72

73 Semantic networks & frames 2 Structure facts so as to elicit information Introduce concepts that were adapted by object-oriented programming ES-73

74 Semantic networks & frames 3 Structure facts so as to elicit information Introduce concepts that were adapted by object-oriented programming Amounts to adopting a particular style of programming and organizing a program ES-74

75 Semantic networks & frames 4 Structure facts so as to elicit information Introduce concepts that were adapted by object-oriented programming Amounts to adopting a particular style of programming and organizing a program» Requires discipline > The programming environment does not directly support the style ES-75

76 Semantic networks Consist of» Entities ES-76

77 Semantic networks 2 Consist of» Entities» Relations between Entities ES-77

78 Semantic networks 3 Consist of» Entities» Relations between Entities» Some similarity with Entity-Relation model in databases ES-78

79 Semantic networks 4 Consists of» Entities» Relations between Entities» Some similarity with Entity-Relation model in databases» A graph representation is used ES-79

80 Page 361 Bratko, E4 Example semantic network ES-80

81 Example semantic network How would the graph be represented in Prolog? ES-81

82 Semantic method representation The graph is represented in Prolog as a set of facts» Functor a relation name» Arguments are the entities at the head and tail of a relation moving_method (kiwi, walk) colour ( kiwi, brown) is_a ( Kim, kiwi) active_at ( kiwi, night) ES-82

83 Inheritance representation How would inheritance be represented? ES-83

84 Inheritance representation Inheritance can be represented as a custom rule for each relationship moving_method ( X, Method ) :- is_a ( X, SuperX ), moving_method ( SuperX, Method). ES-84

85 Inheritance representation Inheritance can be represented as a custom rule for each relationship moving_method ( X, Method ) :- is_a ( X, SuperX ), moving_method ( SuperX, Method). What is the problem with this? ES-85

86 Inheritance representation 2 Inheritance can be represented as a custom rule for each relationship moving_method ( X, Method ) :- is_a ( X, SuperX ), moving_method ( SuperX, Method). Cumbersome to use extensively ES-86

87 Inheritance representation 3 Inheritance can be represented as a custom rule for each relationship moving_method ( X, Method ) :- is_a ( X, SuperX ), moving_method ( SuperX, Method). Cumbersome to use extensively» Need a separate rule for each relation ES-87

88 Inheritance representation 3 Inheritance can be represented as a custom rule for each relationship moving_method ( X, Method ) :- is_a ( X, SuperX ), moving_method ( SuperX, Method). Cumbersome to use extensively» Need a separate rule for each relation How can we solve the problem? ES-88

89 Inheritance representation 2 A better way is to have a general rule for is_a based on facts» Argument to fact is a compound term > relation_name ( Arg1, Arg2) fact ( Fact ) :- Fact,!. fact ( Fact ) :- Fact =.. [ Relation, Arg1, Arg2 ], is_a ( Arg1, SuperArg ), SuperFact =.. [ Relation, SuperArg, Arg2 ], fact ( SuperFact ). ES-89

90 Frames A frame is data structure whose components are slots ES-90

91 Frames 2 A frame is data structure whose components are slots Slots have a name and a value ES-91

92 Frames 3 A frame is data structure whose components are slots Slots have a name and a value FRAME: bird a_kind_of: animal moving_method: fly active_at: daylight ES-92

93 Frame representation How can a frame be represented in Prolog? FRAME: bird a_kind_of: animal moving_method: fly active_at: daylight ES-93

94 Frame representation in Prolog The frame name is the functor for a set of predicates. The arguments of the predicate are» The slot name» The slot value bird ( a_kind_of, animal ). bird ( moving_method, fly ). bird ( active_at, daylight ). FRAME: bird a_kind_of: animal moving_method: fly active_at: daylight ES-94

95 Slot values Slots have a name and a value» The value can be > What? ES-95

96 Slot values 2 Slots have a name and a value» The value can be > Simple values ES-96

97 Slot values 3 Slots have a name and a value» The value can be > Simple values > References to other frames ES-97

98 Slot values 4 Slots have a name and a value» The value can be > Simple values > References to other frames > Procedures to compute the slot value ES-98

99 Slot values 5 Slots have a name and a value» The value can be > Simple values > References to other frames > Procedures to compute the slot value > Unfilled ES-99

100 Slot values 6 Slots have a name and a value» The value can be > Simple values > References to other frames > Procedures to compute the slot value > Unfilled How would they be filled? ES-100

101 Slot values 7 Slots have a name and a value» The value can be > Simple values > References to other frames > Procedures to compute the slot value > Unfilled They are filled by inference ES-101

102 Frame inheritance How is inheritance represented? FRAME: albatross a_kind_of: bird color: fly size: 115 ES-102

103 Frame inheritance 2 Inheritance is represented by designated slot names, and the value being the name of a frame FRAME: bird a_kind_of: animal moving_method: fly active_at: daylight FRAME: albatross a_kind_of: bird color: fly size: 115 ES-103

104 Frame instance How is a frame instance designated? FRAME: Albert instance_of: albatross ES-104

105 Frame instance 2 An instance is represented by designated slot names, and the value being the name of a frame FRAME: Albert instance_of: albatross FRAME: albatross a_kind_of: bird color: fly size: 115 ES-105

106 Can override slot values Frame instance 3 FRAME: Albert instance_of: albatross size: 120 FRAME: albatross a_kind_of: bird color: fly size: 115 ES-106

107 Frame example Figure Program text pages ES-107