Expert Systems

1
Expert Systems

• Perhaps the most successful area of AI research
• Showed people that AI could solve real problems
  and not just toy problems
• Brought AI into the spotlight in the 70s and
  80s
• Currently, at least 2500 Expert Systems in use in
  the world (maybe as many as 12,500!)

2
Expert Systems

• Problem solvers which solve problems normally
  solved by human experts
• Requires two areas of expertise
• Expert Domain Knowledge
• Expert Procedural Knowledge
• This knowledge was usually very large (possibly
  too large for computers of that time or even
  today)

3
Why Expert Systems?

• Order-of-magnitude increase by expert in solving
  tasks
• Increased quality of work
• Reduced errors
• Reduced cost (?)
• Decreased personnel and training time
• Improved decisions
• Improved customer service

4
History of early systems

• Dendral was first (1965) - chemical analysis
• Macsyma (1967) - math reasoning
• Mycin (1974) - medical diagnosis
• Hearsay (1971) - speech recognition
• Harpy (1974) - speech recognition
• Puff (1975) - pulmonary disorders
• Prospector (1977) - mineral surveyor
• R1/XCON (1980) - configure VAX systems
• Internist (1982) - medical diagnosis

5
Other Notable (and more recent) Systems

• SOPHIE - tutorial on circuit diagnosis
• MOLGEN - design gene-cloning experiments
• CADUCEUS - internal medical diagnosis
• GUIDON - tutorial on bacterial infections
• REACTOR - diagnose reactor accidents
• PTRANS - DEC computer diagnosis
• YES/MVS - monitor IBM MVS OS
• ABEL - diagnose acid-base electrolytes
• REX - regression analysis expert

6
Learning systems and aids

• Systems that learn rules
• INDUCE/PLANT - infers plant disease rules
• RX - statistical analysis in medicine
• Meta-DENDRAL - infers new rules for DENDRAL
• Tools to aid in the development of KBS
• LEX - learns heuristics for selecting rules
• SEEK - collects statistics on rule performance
• TIRESIAS - rule debugging system

7
Types of Knowledge

• Rules
• Objects/Classes
• Decision Trees
• Cases
• Ability to acquire additional data
• Ability to perform Induction, Analogy
• Explanation
• Common Sense Knowledge

8
Choices of Representation

• Rules
• Frames/Objects/Attributes/Classes
• Hierarchies
• Propositions
• Constraints
• Procedures
• Certainty Factors/Probabilities/Fuzzy Variables

9
Choices of Inference and Control

• Chaining (Forward, Backward, Mixed)
• Resolution/Unification
• Inheritance
• Hypothetical Reasoning
• Means-Ends Analysis
• Constraint Propagation
• Heuristic Search

• Goal-directed
• Data-directed
• Message-passing
• Demons and Triggers
• Focus
• Metaplans
• Problem/Task Scheduling
• Blackboard
• Task-Specific Inference

10
Other Forms of Knowledge

• Conflict Resolution Knowledge
• Pattern Matching Algorithm
• Special-purpose Heuristics
• Meta-knowledge
• Explanation generating facility
• User Interface
• Knowledge Acquisition Facility

11
ES construction

• Knowledge Acquisition
• Prototyping
• Testing
• Enlarging
• Generalizing
• Documenting

12
Knowledge Acquisition

• Requirements Specification
• defining the goals of the system
• How to get the information from the expert into
  the system?
• interviewing expert
• reading literature
• rule induction
• What knowledge should be used?
• knowledge element identification
• knowledge classification system
• functional layout, flowchart

13
Questions about ESs

• Can they explain their conclusions?
• Can they support multiple beliefs?
• Do they have confidences?
• Can they learn?
• Do they use common sense?
• Do they use deep or shallow knowledge?
• Are they efficient?
• Can they share knowledge?

14
Problems with ESs

• Brittleness
• Lack of Generality
• Lack of Common or Temporal Sense
• Lack of focus
• Need for Search
• Knowledge Acquisition difficulties
• Validation

15
Why Brittleness?

• Shallow knowledge instead of deep knowledge
• Limited knowledge base
• Poor or no learning abilities
• No common sense knowledge
• No temporal or spatial knowledge
• Lack of graceful degradation
• Inconsistent rules

16
Why Lack of Focus

• Too search dependent
• Weak methods do not provide focus
• Meta-knowledge can be applied to provide focus,
  but when is meta-knowledge used?
• Lack of content-addressable memory
• Without learning, its hard to develop proper
  indexes for an enhanced search

17
Improving Knowledge Acquisition

• Used to be done by knowledge engineer
• Need the expert to enter knowledge
• KA requires immediate feedback as knowledge is
  being entered
• KA tools now available (KADS for instance)
• Theories on KA have developed over the last 2
  decades

18
Consider development times

• Dendral (1965) - 40 man-years
• MACSYMA (1967) - 40 man-years
• MYCIN (1971) - 20 man-years
• HEARSAY (1972) - 35 man-years
• HARPY (1974) - 20 man-years
• PUFF (1975) - 5 man-years
• PROSPECTOR (1977) - 20 man-years
• XCON (1980) - 7 man-years

19
More Recently

• Shift to Knowledge-Based Systems (is a Speech
  Rec. system an expert?)
• Use of Deep Knowledge
• Shift to Artificial Life Systems (that can
  learn on their own, build their own
  representations and knowledge bases)
• Shift to Intelligent Agents (systems that solve
  very specific problems, are possibly mobile and
  are easier to construct)
• Much ES work is done in other fields

20
ES Shells and Tools

• Previous ESs were all written in Lisp, C, Prolog,
  etc...
• With a Shell, control (procedural knowledge) is
  built-in, you only have to specify the domain
  knowledge
• Much easier to build an ES
• Examples EMYCIN, OPS5 (rule-based), SOAR
  (revision of OPS5 with chunking), KADS (Knowledge
  Acquisition), CSRL, DSPL, RA, Peirce (GT tools)

21
Emerging Capabilities

• Temporal, spatial and geometric models
• Deep and shallow models combined
• Ontologies
• Reasoning with very large KBS
• Distributed control, real-time control, Object
  brokers
• Task-specific software architectures
• Domain-specific software architectures
• Persistent, sharable knowledge, reusable KBS
• Automatic knowledge acquisition and learning
  techniques

22
Survey of ES/KBS area and year

• 1980 - 10
• 1981 - 15
• 1982 - 25
• 1983 - 70
• 1984 - 80
• 1985 - 60
• 1986 110
• 1987 - 310
• 1988 - 280
• 1989 - 240
• 1990 - 410
• 1991 - 425
• 1992 - 480

• Control
• Design
• Diagnosis
• Instruction
• Interpretation
• Monitoring
• Planning
• Prediction
• Prescription
• Selection
• Simulation

23
Specific Domains

• Agriculture
• Business
• Chemistry
• Communications
• Computer Systems
• Education
• Electronics
• Engineering
• Environment
• Geology
• Image Processing
• Info Management

• Law
• Manufacturing
• Mathematics
• Medicine
• Meteorology
• Military
• Mining
• Power Systems
• Science
• Space Technology
• Transportation

24
An Irony

• Expert (Knowledge-based) Systems has been the
  biggest success in AI
• Now, ES/KBS development is not considered AI at
  all
• Information Technology, Intelligent Agents,
  Decision Support Systems, Tutorial Programs, etc
  are now considered other fields even though they
  are all based on ES/KBS development whose
  foundations are AI