Upload
others
View
9
Download
0
Embed Size (px)
Citation preview
* *
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . * * * * * * * U S L / D B M S N A S A / R E C O N * * * * * * W O R K I N G P A P E R S E R I E S * * *
Rep0 r t N u m b e r
DWIS.NASA/RECON-lO
* *
The USL/DEEUIS NASA/RECON Working Paper Series contains a collection of reports representing r e s u l t s of activities being conducted by the Computer Science Department of the University of Southwestern Louisiana pursuant to the specifications of National Aeronautics and Space Administration Contract Number NASW-3846. The work on this contract is being performed jointly by the University of Southwestern Louisiana and Southern University.
For more information, contact:
Wayne D. Dominick
Editor USL/DWIS NASA/REWWorking Paper Series
Computer Science Department University of Southwestern Louisiana
P. 0. Box 44330 Lafayette, Louisiana 70504
(318) 231-6308
. . . . . . . . . . . . . . . . . . . . . . I DBMS.NASA/REW-lO I
https://ntrs.nasa.gov/search.jsp?R=19890005587 2020-04-26T20:20:33+00:00Z
K N W E D G E BASED SYSTEUS:
A CRITICAL SURVEY OF MAJOR COINCEPTS,
ISSUES AND TECHNIQUES:
VISUALS
Srinu Kavi
Center for Advanced Computer Studies University of Southwestern Louisiana
Lafayette, Louisiana 70504 P. 0. Box 44330
D e c e m b e r 11, 1984
K N a E D G E BASED S Y S T M :
ISSUES, AND TECHNIQUES A CRITICAL SURVEX OF MAJOR CONCEPTS,
SRINU KAVI FALL 1984
.
OBJECTIVES
- To examine various issues and concepts involved in KBSs involved in KBSs
- To examine various techniques used to build KBSs
- To examine (at least one) KBS in detail ( L e o 9 case study)
- To list and identify limitations and problems w i t h KBSs
- To suggest future areas of research
- T o p r o v i d e e x t e n s i v e r e f e r e n c e s
TABLE OF CONTENTS
I . INTRODUCTION
2. KN-EDGE-BASED SYSTEMS (KBSS)
Introduction Hypothetical KBS KBS Components
3. TECHNIQUES USED TO CONSTRUCT KBSs
Introduction M e t h o d s of Representing Knowledge Inference Engine W o r k s p a c e Representation T h e Interface
4. KBS BUILDING TOOLS AND LANGUAGES
General Purpose P r o g r a m i n g Languages Skeletal Systems General Purpose Representation Languages Computer-Aided Design Tools Case Studies
'.
5. APPLICATION CONSID-TIONS
Initial Considerations Technology Considerations Environmental Considerations
6. CONCLUSIONS
7. POTENTIAL FUTURE RESEARCH AREAS
APPEND I C I ES
A. CASE STUDY - MYCIN B. LIST OF KBSs C. FIFTH GENERATION PROJECT
REFERENCES
'.
CHARACTERISTICS OF KBSs
- Organization of Knowledge
- Performance
- Utility ( o r Usefulness)
- Understandability ( o r Explainability)
- Heuristics
- Uncertainity
- Flexibility
- Modularity
TECHNIQUES USED TO CONSTRUCT KBSs
Introduction
Origins of KBS Techniques Choices and Restrictions Knowledge Representation Problems Knowledge Representation Forms Knowledge Representation Unit Credibility Factors Procedural Versus Declarative
Representation
M e t h o d s of Representing KB
Finite-State Machine Programs Predicate Calculus Production R u l e s Semantic Networks F r a m e s
Inference Engine (IE)
Primary Functions of IE Some Definitions IE Strategies M e t h o d s of Implementing the IE Measures of Performance
W o r k s p a c e Representation
Introduction HEARSAY-Blackboard AND/OR Gr a p h Blackboard Versus AND/OR Gr a p h
T h e Interface
Functions of the Interface User Interface Expert Interface Knowledge Acquisition (KA) Process
i
Table 3-1 ORIGINS OF KBS TECHNIQUES (Based on [Barnett & Bernstein, 7 7 1 )
ARTIFICIAL INTELLIGENCE (AI)
Heuristic Search Inference and Deduction Pattern Matching Knowledge Representation and
Acquisition S y s t e m Organization
LANGUAGE PROCESSING
Parsing and Understanding Question and Response Generation Knowledge Representation and
Acquisition
THEDRY OF PROGRAMMING LANGUAGES
Formal Theory of Computational Power Control Structures D a t a Structures S y s t e m Organization Par s i ng
NgODELING A N D SIMULATION
Representation of Knowledge Control Structures Calculation of Approximations
DATA BASE MANAGEME"
Information Retrieval U p d a t i ng File Organization
SOFWARE ENGINEXRING
S y s t e m Organization Do cumen t a t i on Iterative System Development
APPLICATION AREAS
Domain-Specific Algorithms H u m a n Engineering
..
f I
MODEL I
EXPERT'S I KNauDGE
d
B m REPRESENTATICPJ ' J
, I , 0
I I
r * I ' w ,
EXPERT'S I
- mKsPAcE REPRESENTATION
b
REASONING 3 INFERENCE PRINCIPLES I ENGINE
c i
USER EXPECTATIONS
EXPLANATION I
I SYSTEM I 1
KNaWLEDGE REPRESENTATION METHODS
- Finite state machines
- Programs
- Predicate calculus
- Production rules
- Semantic networks I
- F r a m e s
Representation = Knowledge + Access [Newell, 821
PRODUCTION SYSTEM COMPONENTS
Three parts [Barr & Feigenbaum, 811:
- A R u l e Base: A collection of production rules.
- A W o r k s p a c e : A buffer like data structure.
- A n Interpreter: W h i c h controls the system activity.
INTEZtPFWTER TASKS
- Matching or Building a Conflict-Set
- Conflict-Resolution
- Action or Execution
CONFLICT RESOLUTION STRATEGIES
- Rule Order
- Rule Precedence
- Generality Order
- Data Order
- Regency Order
- Non-Deterministic
AN EXAMPLE
Automotive Repair Agency
T h e S y s t e m Contains
- Knowledge Base of production rules (Performance characteristics and Measurable attributes)
- A Database (Past problems, Repairs, and Service performed)
Rl IF fan belt tension is low THEN alternator output will be low [ . 5 ] l
and engine will overheat [ . 2 ] I
I
R2 IF alternator output is low THEN battery charge will be low [ . ' ? ' I I
R3 IF battery is low THEN car will be difficult to start [.5]
R4 IF automatic choke malfunctions OR automatic choke needs adjustment
THEN car will be difficult to start [ . 8 ]
R5 IF battery is out of warranty THEN battery charge m a y be low [ . 9 ]
Figure 3-9 PRODUCTION RULES FOR AUTOMDTIVE SYSTEMS KS
R6 IF coolant is lost OR coolant system pressure cannot be maintained
THEN engine w i l l overheat [ . 7 ]
R7 IF there is a high resistance short AND fuse is not blown
THEN battery charge will be low [ . 8 ]
R8 IF battery flui-d is low THEN battery will boil off fluid [.3]
R9 IF battery fluid is low THEN battery charge will be l o w [ . 4 ]
Figure 3-9. PRODUCTION RULES FOR AUTONDTIVE SYSTEMS KS (CONT'D)
OBSEXVATIONS AGENT D I FF I CULTY
Alternate Output Level Battery Charge Level Battery Fluid Level Choke Adjustment Choke Function Coolant Level Coolant S y s t e m Pressure Difficulty to Start Engine Temperature F a n Belt Tension F u s e Condition Short in Electric S y s t m Voltage Regulator Level W a r r a n t i e s
M e c h 4 M e c h 3 S r v R 2 M e c h 5 M e c h 5 S r v R 2 M e c h 5 c u s t I c u s t I M e c h 3 S r v R 2 M e c h 8 M e c h 4 Da t a B a s e 0
Fi g u r e 3-10. DATA GATHERING PROCEDURE FACT FILE
< CHOKE OUT DIFFICULT r To START OF ARlusTFEM
FUSES NOT f%=J
b v
I I ! I
F W I D Lcsrl
. _ ,-I SHORT HIGH 1
VOLTAGE REGULATOR
I I
I I
& HIGH ENGINE
AGENTS
MECH (5
MECH(5)
SRVR(2)
SRVR (2 1
MECH(4)
h I
BAlTERY BAm VOLTAGE
Low HIGH HIGH FLUID CHARGE REG, OUTPUT
b -
FIGURE 3-E, FRAGMENT OF GRAPH STRUCTURE
+
VIS IBI LIlY OF
- I ta I RECT 4 LIHITED
~ + INTERACTION - CHANNEL BEHAVIOR
RULES AS + EXPLANA- - PRIMITIVE * TIONS OF ACT IONS SOLLJT I ON
L *
+
MOWLARITY
OF BEHAVIOR i
1 > + + ' A
CUNFLIC r- RESOLUTION STRATEGY BILITY
-NS1- t7
- MODIFI- ABILITY
I
FORMAT MCH I NE READ- ABILITY
+
I
I
+
FIGC~RE 3-14, WCTFRISTICS OF PRODUCTION SYSTFMS
BASED ON cBARNEll& BERNSTEIN, '771
SEMANTIC N E m R K s
Semantic networks are used in
Psychological modeling of human memory
P r o g r a m i n g languages
Natural language understanding
Data base management systems
A SEMANTIC N E m R K ( o r NET) consists of nodes and links.
TEMP (WARM-BLOODED MAMMAL) I SA (DOG, MAMMAL) I SA ( CAT, MAMMAL 1
LOC (MARY 's, F IDO) LOC (F I REHOUSE, BOWSER) LOC (BOB'S, PUFF) COLOR(TAN,FIDO) COLOR(TAN,BOWSER) COLOR(BLACK, PUFF)
BETWEEN(MARY'S,FIREHOUSE,BOB'S)
I SA (F fDOI DOG) I SA (BOWSER, DOG) I SA (PUFF, CAT)
S I Z E ( 4 0 L B , F I D O ) SIZE(14LB,BOWSER) S IZE(gLB,PUFF)
MAMMAL
P U W 3 C K 1 - -
rDM-nl I nnncn TEMpY \ I S A
PUFF
\ DhWPCD FIDO
MARY'S ~ O L B TAN 1 4 ~ ~ FIREHOUSE 4 LB BLACK BOB'S
Fi r s t Ru 1 e
PUFF is a cat and CAT is a MAMMAL; therefore, PUFF is a MAMUAL.
Second Rule
SIZE(4,PUFF) & SIZE(14,BCMEER) & 4 < 14 = > SMALLER(PUFF, B m E R )
Third Rule
ISA(FID0, DOG) & ISA(DOG, MAlvlMAL) = > ISA(FID0, MAlvlMAL)
ISA(FID0, MAMMAL) & TEMP(WmBLOODED, MANMAL) = >
TEMP(FID0, W m B L O O D E D )
- W h a t does a node (object) really mean?
- Is there a unique w a y to represent an idea?
- H o w is the passage of time to be represented?
- H o w does one represent things that are not facts about the world but rather ideas or beliefs?
- W h a t are the rules about inheritance of properties in networks?
FRAMES
Frame Characteristics
- Description
- Instantiation
- Prediction or Expectation
- Justification
- Variation
- Correction
- Perturbation
- Transformation
I dog FRAME I SA m a m a 1 2 kind breed 3 color SUBSET.OF {tan brown black
w h i t e rust]
4 F R O M c o l o r Q F k i n d
5 leggedness O . . . &
7 state d u 1 t . OR 6 weight ' 0 , ~ ~ Q E ~
puppy i f age < I 8 age ' 0 , now birth- *
9 bi r thday date IO n ame string
11 END dog
Figure 3-16. EXAMPLE FRAME DEFINITIONS [Barnett & Bernstein, 7'71
I boxer FRAME ISA 2 color
3 size 4 tai 1 5 ears 6 temperment 7 COMPLA I NT S
8 END
breed OF dog ONELOF { t a n
40.. . 6 0 bobbed OR long
p l a y f u l IF weight > I00 THEN ASSUME (great dane) boxer
brown brindle)
bobbed OR f l o p p y
F i g u r e 3-16. EXAMPLE FRAME DEFINITIONS (CONT'D) [Barnett & Bernstein, 771
L W - LEVEL I WORMAT I ON
OBJECT - 654
color = tan ears = bobbed leggedness = 4
temperment = m e a n size = 40 - 45
TRIAL IDENTIFICATION
[OBJECT 654 ISA dog
kind boxer WITH [ c o l o r tan size 40 - 45 tail ASSUMED bobbed ears bobbed temperment EXCEPTIONAL
me a n ] color tan leggedness 4
state ASSUMED adult] weight 40 - 45
F i g u r e 3-17. INEXACT MATCH BY A FRAME SYSTEM [Barnett & Bernstein, 771
INFERENCE ENGINE CONTROL STRATEGIES
- Forward chaining
- Backward chaining
- Chain both ways
- Middle t e r m chaining
- Fixed directionality
- Variable directionality
- Hybrid strategy
- Breadth-first
- Depth-first
&t
3 2 N
/ \ J \ \ 1 7
rl 9 0 76 I4 22
F m CHAINING
A 5 2 7 9
St 5 3
10 / $1 5 3
- -33, / =. 26 40
2080 121
CHAIN BOTH WAYS
METHODS OF IMPLEMENTING THE IE
Search Methods
Simulation Methods
Pattern Matching
SEARCH SYSTEM COMPONENTS
F i v e m a j o r components
Select
Expand
E v a 1 ua t e
Prune
Termi na t e
EVALUATION FUNCTION
"The purpose of a n evaluation function is to provide a m e a n s for ranking those nodes (activities) that are candidates for expansion to determine w h i c h one is most likely to be on the best path to the g o a l " [Nilsson, '711.
A' - AN OPTIMAL SEARCH ALGORITHM
In A * , the evaluation function, f'(x) is the cost of a solution path constrained to go through node x ; f ' should be -.
m- I f'(n(i)) = E K(n(j), n(j+l)) l<=i<=m
J = 1
f'(n) = f ' ( s t a r t , n ) + f'(n, goal)
f'(n) = g(n) + h(n)
f(n) = g ' ( n ) + h ' ( n ) .
MEASURES OF PERFORMANCE
Search Techniques
Penetrance
P = LIT L length of the derived path
T total number of nodes f r o m initial to goal
Branching Factor
WRKSPACE REPRESENTATION
- Plan
- Agenda
- History
- Solution S e t
WW METHODS
- HEARSAY Blackboard
- AND/OR Graph
HEARSAY BLACKBOARD
A data structure
- Hypotheses and support criteria stored
- Intermediary between KSs and IE
-LEVELS-
CONCEPTUAL
PHRASAL
LEXICAL
SYLLABIC
SURFACE PHONEMIC
PHONETIC
1.-
SECMEMAL
PARAFlETRIC
-KNOWLEDGE SOURCES-
SEMANTIC WORD HYPOTHESIZER
SYNTACTIC PARSER
SYNTACTIC WORD HYPOTHESIZER
IWNW HYPOTHESIZER
WORD CANDIDATE GENERATOR
P"OL0GICAL RULE APPLIW
PWNE-PHONWE W"IZER
PHONE SYNTHES I ZER
SEGMENT PHONE SYNCHRONIZER
PARAMETER SEGMENT SWROkIIZER
SEGMENTER CLASSIFIER
FIGURE 3-28, HFARSAY I I 1 Fvfl S OF RE-TIOU' wasoURCEs BASED ON mrw, ET AL, '801
USER INPUT
Parsing Strategies
- Backtracking Versus Parallel Processing
- T o p Down Versus B o t t o m U p Processing
- Choosing How to Expand or Combine
I
~
- Multiple Knowledge Sources ~
TEMPLATE MATCHING
E . g . Y ELIZA, SIR, STUDENT
$1 x(i) ( I W A R E } NOT $2
UlHAT IF x ( i ) WEEW $ ( 2 ) ?
"Today's temperature is not hot"
"What i f temperature w e r e hot?"
RECURSIVE TRANSITION NEWRKS
E . g . Y "The little boy in the swimsuit k i c k e d the red b a l l "
NP: T h e little boy in t h e swimsuit
PP: in the swimsuit
NP: t h e swimsuit
V e r b : kicked
NP: t h e red ball
S: PP
NP <VERB> NP
NP:
PP:
NP
FIGURE 3-32, A RECURSIVF TRANSITION NETWORK BASED ON [BARR 8 FEIGENBAUN, ‘811
DIFFICULTIES IN EXPLANATIONS k
- 7
% r
Explanations
- Must be in terms of
Knowledge chunks P r o b l e m parameters Inference rules
- Must be translated to human understanding
METHODS OF PROVIDING EXPLANATIONS
- W o r k s p a c e Representation
- Using Knowledge Source(s)
- Re-solve the P r o b l e m
DIFFICULTIES IN K A
Representational mismatch
Verbalization by the expert (Protocol study)
Limitations on current technology
K A bottleneck
KBS BUILDING TOOLS AND LANGUAGES
- General purpose programning languages I
- Skeletal systems
- General purpose representation languages
- Computer-aided design tools for KBSs I
C a s e S t u d i e s
- EMYCIN
- HEARSAY-I11
- AGE
INITIAL CONSIDERATIONS
Task suitability
Availability of expert
KA process
Agreement w i t h the domain theory
E x p e r t ' s model
E x p e r t ' s principles of reasoning
Intermediate levels of abstraction
General versus domain s p e c i f i c knowledge
End users
Unanticipated support
Cost versus benefits
TECHNOLOGY CONSIDERATIONS
Building the prototype system
Chunk size
Representation of knowledge
Inference engine
M e t a knowledge
Procedural knowledge
Addition of knowledge by the users
Extensibility
Knowledge representation tools
Design of tools
K N a E D G E REPRESENTATION TOOLS
Generality
Appropriateness
Accessibility
Explanation/Interaction
P r o b l e m characteristics versus T o o l features
DESIGN OF TOOLS FOR BUILDING KBSs
- Generality
- Completeness
- High-level representation language
- Explanation/interaction facilities
- D a t a representation
- Control structure
E N V I R 0 " T A L CONSIDERATIONS
- Interactive KBSs
Interactive development
Local operating environment
CONCLUSIONS
- W i d e spectrum of application areas
- Highly successful
- S o m e systems are being used routinely (DENDRAL, MYCIN, RI, PROSPECTOR)
- Not yet corrmonly understood ( F e w "data points")
- M a j o r motivations
Replication/Distribution expertise
Union of expertise
D o cumen t at i on
- Building ESs expensive and time-consuming ($1-2 m i l l i o n ; 5 person-years with tools)
- General level of accomplishment high
- Number of unresolved issues
- Difficulties and potential risk
- High expectations/misunderstandings
..
POTENTIAL FUTURE RESEARCH AREAS
Knowledge acquisition
Representation theory
Cornparision of techniques
KBS building tools
Explanation
Evaluation
Parallel processing
Learning f r o m e x p e r i e n c e
Management of knowledge
Abstraction and hierarchies
Technological innovations
U n i f o r m terminology
1 1. R e p o n No. 2. Government Accession No. 3. Recipient’s Catalog No.
/ ~ ~ f i s - 6 5. Report Date
/# -8LL 4. Title and Subtitle
USL/NGT-19-010-900:, KNOWLEDGE BASED SYSTEMS: A CRITICAL SURVEY OF MAJOR CONCEPTS, ISSUES, AND TECHNIQUES: VISUALS
~
7. Author(s) SRINU KAVI
9. Performing Organization Name and Address
University of Southwestern Louisiana The Center for Advanced Computer Studies P.O. Box 44330 Lafayette, LA 70504-4330
12. Sponsoring Agency Name and Address
8. Performing Organization Report No.
10. Work Unit No.
11. Contract or Grant No.
NGT-19-010-900 13. Type of Repon and Period Covered
FINAL; 07/01/85 - 12/31/87 14. Sponsoring Agency Code
15. Supplementary Notes
16. Abstract
This Working Paper Series entry represents a collection of presentation visuals associated with the companion report entitled “Knowledge Based Systems: A Critical Survey of Major Concepts, Issues, and Techniques”, USL/DBMS NASA/RECON Working Paper Series report number DBMS .NASA/RECO N-9.
17. Key Words (Suggested by Authods) 1
Knowledge-Based Systems, Informa- tion Storage and Retrieval Systems
This report represents one of the 72 attachment reports to the University of Southwestern Louisiana’s Final Report on NASA Grant NGT-19-01G900. Accordingly, appropriate care should be taken in using this report out of the context of the full Final Report.
18. Distribution Statement
19. Security Classif. (of this report) 20. Security Classif. (of this page) 21. NO. of Pages
Unclassified . Unclassified 6 4
22. Rice’