Você está na página 1de 46

Knowledge Acquisition, Representation, and Reasoning

Prof. Rushen Chahal

Prof. Rushen Chahal

Page 1

Learning Objectives
Understand the nature of knowledge. Learn the knowledge engineering processes. Evaluate different approaches for knowledge acquisition. Examine the pros and cons of different approaches. Illustrate methods for knowledge verification and validation. Examine inference strategies. Understand certainty and uncertainty processing.

Prof. Rushen Chahal

11-2 Page 2

Development of a Real-Time Knowledge-Based System at Eli Lilly Vignette

Problems with fermentation process
Quality parameters difficult to control Many different employees doing same task High turnover

Expert system used to capture knowledge

Expertise available 24 hours a day

Knowledge engineers developed system by:

Knowledge elicitation
Interviewing experts and creating knowledge bases

Knowledge fusion
Fusing individual knowledge bases

Coding knowledge base Testing and evaluation of system

Prof. Rushen Chahal 11-3 Page 3

Knowledge Engineering
Process of acquiring knowledge from experts and building knowledge base
Narrow perspective
Knowledge acquisition, representation, validation, inference, maintenance

Broad perspective
Process of developing and maintaining intelligent system

Prof. Rushen Chahal

11-4 Page 4

Knowledge Engineering Process

Acquisition of knowledge
General knowledge or metaknowledge From experts, books, documents, sensors, files

Knowledge representation
Organized knowledge

Knowledge validation and verification Inferences

Software designed to pass statistical sample data to generalizations

Explanation and justification capabilities

Prof. Rushen Chahal 11-5 Page 5

Prof. Rushen Chahal

11-6 Page 6

Written, viewed, sensory, behavior


Acquired from
Human senses Machines

Prof. Rushen Chahal

11-7 Page 7

Surface level Input-output

Problem solving Difficult to collect, validate Interactions betwixt system components

Prof. Rushen Chahal

11-8 Page 8

Descriptive representation

How things work under different circumstances How to use declarative knowledge
Problem solving

Knowledge about knowledge

Prof. Rushen Chahal

11-9 Page 9

Knowledge Engineers
Professionals who elicit knowledge from experts
Empathetic, patient Broad range of understanding, capabilities

Integrate knowledge from various sources

Creates and edits code Operates tools

Build knowledge base

Validates information Trains users

Prof. Rushen Chahal

11-10 Page 10

Prof. Rushen Chahal

11-11 Page 11

Elicitation Methods
Based on interview Track reasoning process Observation

Build base with minimal help from knowledge engineer Allows execution of routine tasks with minimal expert input

Minimal input from both expert and knowledge engineer
Prof. Rushen Chahal 11-12 Page 12

Manual Methods
Goal-oriented Walk through

Complex domains Data unrelated and difficult to integrate


Prof. Rushen Chahal

11-13 Page 13

Manual Methods
Process tracking
Track reasoning processes

Protocol analysis
Document experts decision-making Think aloud process

Motor movements Eye movements
Prof. Rushen Chahal 11-14 Page 14

Manual Methods
Case analysis Critical incident User discussions Expert commentary Graphs and conceptual models Brainstorming Prototyping Multidimensional scaling for distance matrix Clustering of elements Iterative performance review

Prof. Rushen Chahal

11-15 Page 15

Semiautomatic Methods
Repertory grid analysis
Personal construct theory
Organized, perceptual model of experts knowledge Expert identifies domain objects and their attributes Expert determines characteristics and opposites for each attribute Expert distinguishes between objects, creating a grid

Expert transfer system

Computer program that elicits information from experts Rapid prototyping Used to determine sufficiency of available knowledge

Prof. Rushen Chahal

11-16 Page 16

Prof. Rushen Chahal

11-17 Page 17

Semiautomatic Methods, continued Computer based tools features:

Ability to add knowledge to base Ability to assess, refine knowledge Visual modeling for construction of domain Creation of decision trees and rules Ability to analyze information flows Integration tools

Prof. Rushen Chahal

11-18 Page 18

Automatic Methods
Data mining by computers Inductive learning from existing recognized cases Neural computing mimicking human brain Genetic algorithms using natural selection

Prof. Rushen Chahal

11-19 Page 19

Multiple Experts
Experts contribute individually Primary experts information reviewed by secondary experts Small group decision Panels for verification and validation

Consensus methods Analytic approaches Automation of process through software usage Decomposition
11-20 Page 20

Prof. Rushen Chahal

Automated Knowledge Acquisition Induction

Training set with known outcomes Creates rules for examples Assesses new cases

Limited application Builder can be expert
Saves time, money

Prof. Rushen Chahal

11-21 Page 21

Automated Knowledge Acquisition

Rules may be difficult to understand Experts needed to select attributes Algorithm-based search process produces fewer questions Rule-based classification problems Allows few attributes Many examples needed Examples must be cleansed Limited to certainties Examples may be insufficient
Prof. Rushen Chahal 11-22 Page 22

Automated Knowledge Acquisition

Interactive induction
Incrementally induced knowledge
General models
Object Network

Based on interaction with expert


Computer supported
Induction tables IF-THEN-ELSE rules
Prof. Rushen Chahal 11-23 Page 23

Evaluation, Validation, Verification

Dynamic activities
Assess systems overall value

Compares systems performance to experts Concordance and differences

Building and implementing system correctly Can be automated
Prof. Rushen Chahal 11-24 Page 24

Prof. Rushen Chahal

11-25 Page 25

Production Rules
IF-THEN Independent part, combined with other pieces, to produce better result Model of human behavior Examples
IF condition, THEN conclusion Conclusion, IF condition If condition, THEN conclusion1 (OR) ELSE conclusion2
Prof. Rushen Chahal 11-26 Page 26

Artificial Intelligence Rules

Knowledge rules
Declares facts and relationships Stored in knowledge base

Given facts, advises how to proceed Part of inference engines Metarules

Prof. Rushen Chahal

11-27 Page 27

Artificial Intelligence Rules

Easy to understand, modify, maintain Explanations are easy to get. Rules are independent. Modification and maintenance are relatively easy. Uncertainty is easily combined with rules.

Huge numbers may be required Designers may force knowledge into rule-based entities Systems may have search limitations; difficulties in evaluation

Prof. Rushen Chahal

11-28 Page 28

Semantic Networks
Graphical depictions Nodes and links Hierarchical relationships between concepts Reflects inheritance
Prof. Rushen Chahal 11-29 Page 29

All knowledge about object Hierarchical structure allows for inheritance Allows for diagnosis of knowledge independence Object-oriented programming
Knowledge organized by characteristics and attributes
Slots Subslots/facets

Parents are general attributes Instantiated to children

Often combined with production rules

Prof. Rushen Chahal 11-30 Page 30

Decision tables Decision trees

Knowledge Relationship Representations

Spreadsheet format All possible attributes compared to conclusions Nodes and links Knowledge diagramming

Computational logic
True/false statement

Predicate logic
Variable functions applied to components of statements

Prof. Rushen Chahal

11-31 Page 31

Prof. Rushen Chahal

11-32 Page 32

Reasoning Programs
Inference Engine
Algorithms Directs search of knowledge base
Forward chaining
Data driven Start with information, draw conclusions

Backward chaining
Goal driven Start with expectations, seek supporting evidence

Inference/goal tree
Schematic view of inference process
AND/OR/NOT nodes Answers why and how

Rule interpreter
Prof. Rushen Chahal 11-33 Page 33

Explanation Facility
Makes system more understandable Exposes shortcomings Explains situations that the user did not anticipate Satisfies users psychological and social needs Clarifies underlying assumptions Conducts sensitivity analysis

Why How Journalism based
Who, what, where, when, why, how Why not Prof. Rushen Chahal 11-34 Page 34

Generating Explanations
Static explanation
Preinsertion of text

Dynamic explanation
Reconstruction by rule evaluation

Tracing records or line of reasoning Justification based on empirical associations Strategic use of metaknowledge
Prof. Rushen Chahal 11-35 Page 35

Widespread Important component Representation
Numeric scale
1 to 100

Graphical presentation
Bars, pie charts

Symbolic scales
Very likely to very unlikely
Prof. Rushen Chahal 11-36 Page 36

Probability Ratio
Degree of confidence in conclusion Chance of occurrence of event

Bayes Theory
Subjective probability for propositions
Imprecise Combines values

Belief functions Creates boundaries for assignments of probabilities
Assumes statistical independence

Prof. Rushen Chahal

11-37 Page 37

Certainty factors
Belief in event based on evidence Belief and disbelief independent and not combinable Certainty factors may be combined into one rule Rules may be combined

Prof. Rushen Chahal

11-38 Page 38

Expert System Development

Project initialization Systems analysis and design Prototyping System development Implementation Postimplementation

Prof. Rushen Chahal

11-39 Page 39

Project Initialization
Identify problems Determine functional requirements Evaluate solutions Verify and justify requirements Conduct feasibility study and cost-benefit analysis Determine management issues Select team Project approval
Prof. Rushen Chahal 11-40 Page 40

Systems Analysis and Design

Create conceptual system design Determine development strategy
In house, outsource, mixed

Determine knowledge sources Obtain cooperation of experts Select development environment

Expert system shells Programming languages Hybrids with tools
General or domain specific shells Domain specific tools
Prof. Rushen Chahal 11-41 Page 41

Rapid production Demonstration prototype
Small system or part of system Iterative Each iteration tested by users Additional rules applied to later iterations

Prof. Rushen Chahal

11-42 Page 42

System Development
Development strategies formalized Knowledge base developed Interfaces created System evaluated and improved

Prof. Rushen Chahal

11-43 Page 43

Implementation Adoption strategies formulated System installed All parts of system must be fully documented and security mechanisms employed Field testing if it stands alone; otherwise, must be integrated User approval
Prof. Rushen Chahal 11-44 Page 44

Operation of system Maintenance plans
Review, revision of rules Data integrity checks Linking to databases

Upgrading and expansion Periodic evaluation and testing

Prof. Rushen Chahal 11-45 Page 45

Facilitates knowledge acquisition and distribution Problems with use of informal knowledge Open knowledge source

Prof. Rushen Chahal

11-46 Page 46