Anda di halaman 1dari 54

1

Week 10 Expert Systems

This week

Artificial Intelligence (AI) Expert Systems (ES)


Expertise, experts and transferring expertise Some knowledge elicitation techniques Knowledge representation Examples and use of ES Structure of ES Benefits Limitations

Contrast with DSS

Artificial Intelligence (AI)

Simulation of human intelligence by machines

Characteristics
learning
acquisition of information and rules for using the information

reasoning
using the rules to reach approximate or definite conclusions

self-correction
updating existing material in response to feedback

Applications include expert systems, tutoring systems, speech recognition, machine vision

Expert Systems (ES) - summary

Expert System
a decision-making computer package intended to
attain or surpass the levels of performance of a human expert in some highly specialised field
4

Expert Systems (ES) - summary

Expert System

expertise of a human expert is transferred to ES knowledge is stored and accessed as required makes inferences and arrives at a specific conclusion

differs from DSS: identifies best solution to a problem rather providing a range of possible solutions
advises the user

consists of

knowledge base inference engine user interface


5

Expertise

Expertise is the extensive, task-specific knowledge acquired from


training reading experience

Expertise

Includes

Theories Rules and procedures Rules (heuristics) about what to do in a given situation Global strategies for problem-solving Meta-knowledge (knowledge about knowledge) Facts
Turban (2001)

Recognise two main types


Facts Rules (procedures) as seen in prolog lecture.

Categorisation: novice to expert

Barfield (1986) distinguished between


experts non-experts
naive novice intermediate

Categorisation: novice to expert

Dreyfus (1987) suggested stages


Novice Advanced beginner Competence Proficiency

Expertise

Categorisation: novice to expert Dreyfuss stages 1. Novice

o o o o

skill acquisition

novice receives instruction


rules followed blindly lacks coherent sense of task

judges performance on use of rules


10

Categorisation: novice to expert Dreyfuss stages 2. Advanced beginner

performance becomes barely acceptable after experience of real situations recognises situations on perceived similarity to previous examples

11

Categorisation: novice to expert Dreyfuss stages 3. Competence

increased experience recognises similarities views decision-making in a hierarchical manner improves performance choice of plan is not a simple matter
12

Categorisation: novice to expert Dreyfuss stages 4. Proficiency

view altered by outcome of recent events certain features stand out - others ignored plan modified as necessary understanding the task becomes intuitive triggered naturally without explicit thought
13

Categorisation: novice to expert Dreyfuss stages 5. Expertise

proceeds without detached deliberation acts without conscious contemplation of options performance is fluid not interrupted by detached deliberation or analysis

14

Categorisation: novice to expert

Kolodner (1983) distinguishes between experts & novices


experts have more knowledge of domain
experts apply and use knowledge more effectively than novices

15

Categories of expertise

Clare (1989) categorised types of expert


Practitioner

Academic

Craftsman

Samurai
16

Categories of expertise

Practitioner
expertise based on organised principles and methods expertise gained from
theoretical knowledge developed from experience and practice

work environment may not offer fully coherent basis for development of expertise
such expertise is therefore rare

17

Categories of expertise
Academic
individual expected to
guide direct teach

others in field of own expertise

18

Categories of expertise

Craftsman
expertise gained in routine activities individual solves similar problems rote learning of methods and procedures

19

Categories of expertise

characterised by performance of expertise


performance is the key aspect final outcome is consequence of performance
"There is definitely a moment ... when there is loss of self. ... you are 'in the zone', when you are not aware of what you're doing...you are simply in the act of doing it." Michael Boyd "...it's similar with music in a way although you have to learn the notes first, you then have to forget the notes to play the music." Oliver Sacks
extracts from "More than merely the sum of their parts", public discussion at Columbia University, reported in The Sunday Times Culture Supplement, page 10-11, 23 November 2008

Samurai

20

Knowledge acquisition

Knowledge acquisition techniques extend and enhance those used in systems analysis, such as
repertory grid analysis based on Kellys Personal Construct Theory - every person:
has a personal view of the world forms theories tests hypotheses analyses results structured interviews triads

to extract knowledge of relationships and differences between elements

21

Stages of knowledge acquisition


Reformulations Redesigns
Refinements requirements Identify problem characteristics Find concepts to represent knowledge concepts Design structure to organise knowledge structure Formulate rules to embody knowledge rules Validate rules that organise knowledge

Identification

Conceptualisation

Formalisation

Implementation

Testing

Buchanan, Barstow, Bechtal, Bennett, Clancey, Kulikowski, Mitchell and Waterman, Constructing an Expert System: in Building Expert Systems (1983), ed. Hayes-Roth F, Waterman DA and Lenat DB.
22

Experts

Expert activities
Recognising and formulating the problem Solving problem quickly and correctly Explaining the solution Learning from experience Restructuring knowledge Breaking rules if necessary Determining relevance Being aware of limitations
Turban (2001)
23

Transferring expertise

Expert

Knowledge engineer Development (including knowledge acquisition)

Expert system
Consultation

User

Transfer expertise from expert to ES to non-expert (user) Knowledge acquisition Knowledge representation Knowledge inferencing Knowledge transfer to user Turban (2001)
24

Structured interviews

Expert is interviewed
identifies important items in the domain identifies associated attributes creates a scale of characteristics (opposites) places items on scale solutions are placed on grid (table) ratings from scale may be reviewed

25

Repertory Grids (RGA)

Generated following structured interview


shows relative position of elements based on scoring in structured interview
Attributes Trait Opposite Choc mints Muesli bar Carrot Cabbage Availability High (5) Low (1) 4 5 3 3 Calories High (5) Low (1) 5 4 3 1 Vitamins High (5) Low (1) 1 3 4 5
26

Triads

Groups of three items are classified


why two are alike, and the third different combine in many different ways Consider:
Margaret Thatcher Ronald Reagan Ed Miliband

27

Triads example
left-wing
Ed Miliband

deceased
Ronald Reagan

female
Margaret Thatcher

Margaret Thatcher

Ronald Reagan

Margaret Thatcher

Ed Miliband

Ronald Reagan

Ed Miliband

right-wing

living

male
28

Knowledge representation

Knowledge must be stored and represented for use after identification


Semantic networks Association lists (alists) Search trees Frames Facts Rules

29

Semantic networks

Items linked
is-a relationships uses inheritance
employee

has-a

is-a manager

has-a

office

id-card

is 45 years old

works-in

Jim

sales
30

Association lists
Association list has attribute.value pairs: ((A . excellent)(B . good)(C . average) (D . weak)(E . borderline)(R . resit))

this alist shows the meaning of grades extracting the meaning of grade A will give excellent first attribute is used as key to extract value clumsy for large amounts of data
31

Search trees

Data can be stored on the tree Tree can be searched to find data

Jim

Pat

Mary

Here, tree shows an organisation structure


Liz Alan Ben Dora
32

Search trees

Depth first search


from start point

Jim

continues until downward path exhausted then next downward path is tried
Liz until all paths taken or item is found

Pat

Mary

Alan

Ben

Dora
33

Search trees

Breadth first search


from start point

Jim

considers all options at next level down


Pat then moves to subsequent level until all levels visited or Liz item is found Alan
Ben

Mary

Dora
34

Search trees

If tree becomes too large it is impossible to search the entire tree (time constraints) Unproductive paths must be pruned If searching for Marys staff, ignore staff working Liz for Pat Pat

Jim

Mary

Alan

Ben

Dora
35

Frames

Similar to a blank form (standard template) stereotypes used, other values specified

Frame: SALARY rate of pay: tax paid to date:

NI deductions: pension contributions:

unit ( per annum) unit (month, year) calculate ((TAX-PAID) (tax)) unit (calculate (NI-CON)) unit (calculate (PENSION))

36

Facts

Data that is known to be true prolog can store data as facts in database

manages(jim, pat). manages(jim, mary). manages(pat, liz). manages(pat, alan). manages(mary, ben). manages(mary, dora).

storing the details of the tree structure as an organisation chart

37

Rules

Production rules
format:
pattern action typically in the form of ifthen rules in prolog: predicate is true if components satisfied
sibling(X,Y) :- parents(X, M, F), parents(Y, M, F). X is a sibling of Y if X has mother M and father F and Y has mother M and father F
38

Examples of ES (1)

MYCIN
o medical diagnosis, identifying treatment for blood disorders o symptoms entered o ES compares with known symptoms presents diagnosis prompts for further information o each rule has a certainty factor

Prospector
o o o o Evaluates mineral potential of geological area Assessing risk of investments uses probabilistic reasoning (based on Bayes theorem) can explain why it needs answers to specific questions
39

Examples of ES (2)
XCON
o o o o configures VAX computer systems for DEC rule-based expert system probabilistic information not necessary exact statement of requirements in each case

CLUES
Countrywides Loan Underwriting Expert System
o o o o

developed to ensure rapid, consistent, high-quality loan decisions CLUES has 400 rules refined until it agreed with 95% of human expert decisions human underwriter still examines all rejected loans productivity increased from 6 or 7 applications/day to 16/day

40

(Laudon & Laudon, 2004)

Examples of ES (3)

Nestle Foods Corporation pension fund status


Maintains up-to-date knowledge base to advise on impact of regulation changes and conformance with new standards

Pingtung Teachers' College (Taiwan)


What-if? Analysis for retirement planning: calculating pension benefits

Credit analysis systems


Assessment of risk with respect to a customer's credit record

Automated help desk solutions


Rule-based support for small businesses to help technicians track issues and solve problems
41

(Turban et al, 2011)

Examples of ES (4)
Areas for ES applications Finance insurance evaluation, credit analysis, tax planning, fraud detection and prevention, financial report analysis, financial planning, performance evaluation
Data processing Equipment selection and maintenance, network management
Marketing Customer relationship management, market analysis, product planning Human resources HR planning, performance evaluation, staff schedules, pension management Manufacturing 42 Production planning, quality management, product design, equipment repair (Turban et al, 2011)

Structure of ES

Main components
o Knowledge base o Inference engine o User interface

Other components
o Knowledge acquisition sub-system o Blackboard (workspace) o Explanation facility (justifier)

Human components
o o o

Domain expert Knowledge engineer User


43

Knowledge base

The software that represents the knowledge

Facts: what is known about the problem area (theories, details of the problem situation, etc.) Rules: rules of inference, logical deductions, if then etc.

Knowledge base
Knowledge acquisition Facts and Rules User
44

Inference engine

Inference engine

The brain of ES, control structure, rule interpreter Enables reasoning based on
knowledge base (facts and rules in problem domain) blackboard (input about the particular problem to be solved) forward chaining (data driven) follows the rules to arrive at a conclusion backward chaining (goal driven) starts with hypothesis to be confirmed or disproved

Inference engine
Interpreter Scheduler Consistency enforcer Knowledge base Blackboard (workspace)

45

Human components (1)

Domain expert
qualities already discussed

may be more than one expert


Problems arise if they disagree! Multiple experts may make fewer mistakes than single expert Wider domain than single expert Enhanced quality through synergy of experts ...but... Some experts may dominate discussion Time lost group meetings 46 Lack of confidentiality might present problems (Turban, 2011)

Human components (2)

Knowledge engineer

helps expert structure knowledge of domain builds ES / knowledge base in short supply

needs some knowledge of domain needs greater knowledge of development environment

47

Human components (3)

Different categories of users


non-expert
uses

ES as a consultant or advisor

student
uses

ES as an instructor

ES

builder

uses

ES to improve the knowledge base uses ES as a partner


expert
uses

the ES as a colleague or assistant e.g. for a second opinion

48

Benefits of ES

Financial savings
fewer human experts needed

Improved quality of decisions / answers / solutions


ES consistent, fewer mistakes, do not forget rules

Supports various approaches to decision-making


Use for training Experts time not wasted on trivial tasks Scarce expertise is preserved May be used in hazardous environments
49

Limitations of ES (1)

Knowledge not always readily available Expertise hard to extract from people Approach of experts may be different, but correct Difficult to model judgment under time pressure
50

Experts may not be able to explain what they do


ES work well only in a domain of knowledge

Experts may not be able to check answers are reasonable

Limitations of ES (2)

Experts vocabulary (jargon) often not well understood by others outside field of interest Help often required from knowledge engineers: rare and expensive - make ES expensive Low level of trust by end-users Knowledge transfer may be subject to bias
Turban (2001)
51

ES compared to DSS
Expert Systems

DSS

highly specialised in nature


extend scarce resource (expertise) allow non-experts access to knowledge of experts

extend range, capability and effectiveness of managers' decision-making processes


applied where analytic aids of value but manager's judgement essential do not attempt to automate the decision process, or impose solutions

aim to provide specific solution require specialist development skills and tools

52

References and suggested reading

Barfield, W, 1986, Expert-Novice Differences for Software:

Implications for Problem-Solving and Knowledge Acquisition,

Behaviour and Information Technology, Vol. 5, No. 1, p. 15-29 Chaffey, D. (ed.), 2003, Business Information Systems, 2nd edition, FT Prentice Hall Expert Systems: chapter 6, pages 255-257 Clare, J, 1989, Knowledge Elicitation for Financial Dealers p.237246 in Diaper D (editor) Knowledge Elicitation: Principles, Techniques and Applications, Ellis Horwood Dreyfus, HL, 1987, Misrepresenting Human Intelligence p. 41-54 in Born RP (editor) Artificial Intelligence: The Case Against, Croom Helm Turban E, Sharda R and Delen D, 2011, Decision Support and 53 Business Intelligence Systems, 9th edition, Pearson

References and suggested reading

Kolodner, JL, 1983, Towards an Understanding of the Role of Experience in the Evolution from Novice to Expert, IJMMS 19, p497-518 Laudon, K. & Laudon, J., 2004, Management Information Systems, 8th ed., Pearson Prentice Hall Expert Systems: chapter 10, pages 327-333 Turban E. & Aronson J.E., 2001, Decision Support Systems and Intelligent Systems (6th edition), Prentice Hall Business Publishing Expert Systems: chapter 10, pages 407 - 428 Whiteley, D., 2004, Introduction to Information Systems, Palgrave Expert Systems: chapter 11, pages 161-168 http://www.enquirewithin.co.nz/ (accessed 14 November 2006) 54

Anda mungkin juga menyukai