Lecture 5

Technologies to Manage
Knowledge: Artificial
Intelligence
Objectives

„ Introduce artificial intelligence as a facilitating

technology for knowledge management
„ Introduce knowledge as an important facet of
intelligent behavior
„ Introduce the early state space search
techniques
„ Introduce expertise in the context of knowledge
„ Introduce knowledge-based systems as a
modern evolution of the early state space search
techniques
Historical perspective

„ Definition of AI is elusive
‰ Repetitive tasks are algorithmic in nature
„ Humans excel in symbolic problem solving
„ Symbolic manipulation, basis of AI
“ AI is the science that provides computers with
the ability to represent and manipulate symbols
so they can be used to solve problems not easily
solved through algorithmic models”
Definition
„ Artificial Intelligence (AI) as the science that
“… encompasses computational techniques
for performing tasks that apparently require
intelligence when performed by humans.”
„ Idea of intelligent machines has been with us
with information age.
Knowledge and Intelligence
„ People are able to solve problems
intelligently
„ AI systems aim to mimic intelligent
problem solving
„ Problem-solving requires:
‰ Knowledge
‰ Inference (ability to manipulate, acquire, and
manage knowledge effectively and efficiently –
for recognition, reasoning, learning, etc.)
Knowledge and intelligence

„ How can inference best be implemented in AI

models, so as to manipulate, acquire, and
manage knowledge most effectively and
efficiently?
„ How can knowledge best be represented in
AI models, so as to facilitate effective and
efficient inference?
„ Is an inference model itself a kind of
knowledge?
Early AI research

„ Game playing (chess playing programs, Big Blue

defeated Boris Kasparov 1997
„ Machine translation efforts were not so successful
‰ No 1:1, context free correspondence between words in
different languages
‰ http://babelfish.altavista.com/
„ Eliza (Weizenbaum, 1965) used NL interface to act
as an artificial psychoanalyst.
‰ http://www-ai.ijs.si/eliza/eliza.html
„ Blocks World (Winograd)
Different types of knowledge

„ Natural language processing is more

challenging than game playing
Modern game playing models can beat even
world champions at chess (Big Blue)
But a child can process language better than any
AI model
Early history of AI

„ Game playing depends on inference

mechanisms (and typically use relatively
simple knowledge representations)
„ Language processing depends on how to
represent knowledge
AI and its component areas
Artificial Intelligence
Knowledge-based Systems Natural Language Processing
•Rule-based systems •NL understanding
• Classification •NL synthesis
• Diagnosis •Speech understanding
• Design •Speech synthesis
• Decision support
• Planning
• Scheduling Computer Vision
•Case-based Reasoning •Image processing
• Diagnostics •Image understanding
• Design
• Decision support
• Classification Machine Learning
•Constraint-based reasoning •Inductive learning
• Planning •Case-based learning
• Scheduling •Connectionist learning
•Model-based reasoning •Learning from analogy
• Monitoring •Explanation-based learning.
• Diagnostics •Data mining
• Design •Others.
AI and its component areas

Soft Programming Approaches

•Neural networks
•Uncertainty management Games
• Bayesian probability •Chess
• Certainty factors •Checkers
• Bayesian belief nets •Go
• Fuzzy logic •Backgammon
•Evolutionary Techniques
• Genetic algorithms
• Genetic programming Robotics
•Control
•Navigation and tactics
Human Behavior
Representation Automated Know. Acquisition
•Context-based Reasoning •Repertory grids
•Cognitively-inspired modeling •Conceptual maps
•Others
Search based systems
„ State space search
‰ Solution space to a problem. Includes actions, states,
beliefs representing the status of a problem.
‰ Path through the solution space between initial and goal
state
„ General knowledge employed in state space
searches was not sufficient to solve the difficult
problems
„ “Dumb search” within solution spaces uses only
general domain-independent knowledge – no
domain-specific knowledge
„ A way to expedite the state space search is to use
heuristics
Representing portion of a solution space
for car diagnosis
Searching a portion of the solution space
Heuristics

„ You try to go from A to B

„ Possible search strategies to achieve this:
‰ Random search
‰ Systematic blind search
‰ Heuristic search (or directed search)
„ A heuristic is a “rule of thumb” that tries to
guess which search directions are more
promising
Example of using heuristic knowledge
You notice a backfiring noise in your car
„ You remember a similar situation several months
earlier
„ The mechanic at that time told you the problem was
a loose vacuum hose connection in the engine (and
showed you the hose)
‰ You stop the car, discover the hose is loose, reconnect it
‰ You restart the car, notice the problem is gone
You did not:
‰ analyze the operation of the engine
‰ know how an internal combustion engine operates
‰ have a solution space in mind, did not search a solution
space
When to use heuristics

„ Heuristics often provide the correct solution

to a problem
„ Failure to do so is because:
‰ The number of possibilities to be examined is too
large,
‰ The algorithmic evaluation function applied to
each possible answer to determine correctness is
too complex, or
‰ The evaluation function is unknown and must be
approximated
Algorithmic vs heuristics
„ You need the contractor to give you an estimate to build a house
before you leave town tonight
The contractor typically prices by doing a detailed cost estimate:
„ Call a building supply warehouse for material prices
„ Evaluate quotations from subcontractors on cost of labor
„ Determine appropriate contractors’ fees
„ Allow a reasonable contingency figure, etc
BUT this takes too much time to finish by tonight!
„ The contractor instead takes a heuristic approach:
‰ Find another home of similar size that he recently built

‰ Normalize for the exact size

‰ Adjust for large differences

„ Number of bathrooms
„ Luxury kitchen fittings
Takes only 30 minutes
Knowledge Based Systems and KM

„ Knowledge-based systems are an excellent platform

for capturing, sharing, and applying knowledge (of
certain kinds).
„ Knowledge-based systems were designed primarily
for the purpose of being able to apply knowledge
automatically and solve problems
„ In many KM contexts, we just want a tool to support
knowledge capture, discovery, and/or sharing – we
may not need (or desire automatic application)
KBS

„ The ability of knowledge-based systems to

apply knowledge can be useful for knowledge
sharing:
‰ Demonstrate the system on example problems
‰ Have the system indicate what heuristic
knowledge was used to solve the problems
„ An excellent way to help humans internalize
knowledge.
Definition of KBS

„ A computerised system that uses domain knowledge to

arrive at a solution to a problem within that domain.
This solution is essentially the same as one concluded
by an expert, when confronted with the same problem

1. The use of highly specific domain knowledge

2. The heuristic nature of the knowledge employed
instead of algorithmic
3. The separation of the knowledge of how it is used.
Use of high specific domain knowledge

„ Initially pioneered in the late 1960s to early

1970s in the DENDRAL and Meta-DENDRAL
systems [Lindsay 1980]
„ DENDRAL infers the molecular structure of
unknown compounds from mass spectral and
nuclear magnetic response data
„ Meta-DENDRAL assists in the determination
of the dependence of mass spectrometric
fragmentation on sub-structural features
Separating knowledge from making
inferences
„ Note there are two kinds of knowledge
involved in automatically making inferences:
‰ The general knowledge of how to infer something
(anything): general inference methods
‰ The specific knowledge about that something
‰ Emphasizes that a declarative (rather than
procedural) representation should be used for the
specific knowledge.
Inferencing

„ Allows general inference engines to be

developed
‰ Can be re-used in any domain
‰ Simply encode domain-specific knowledge using
the declarative representation language, and then
run the general inference engine
„ See CLIPS system (NASA)
http://www.ghg.net/clips/WhatIsCLIPS.html
MYCIN

„ These three key differences from

conventional software were first merged in
Stanford’s MYCIN system in the early 1970s
[Shortliffe 1976; Buchanan 1984]:
‰ The use of highly specific domain knowledge.
‰ The heuristic nature of the knowledge employed,
instead of algorithmic
‰ The separation of the knowledge from how it is
used.
MYCIN

„ MYCIN diagnoses and specifies treatments

for blood disorders through a Q&A session
with a physician:
‰ asks questions about the signs and symptoms of
the patient
‰ requests certain laboratory tests as appropriate
‰ recommends a drug treatment after the set of
possible infections has been sufficiently narrowed
MYCIN

„ Mycin represented its knowledge as a set of

IF-THEN rules with certainty factors.
IF the infection is pimary-bacteremia
AND the site of the culture is one of the
sterile sites
AND the suspected portal of entry is the
gastrointestinal tract
THEN there is suggestive evidence (0.7) that
infection is bacteroid.
Pioneering KBS

„ Medicine: CASNET, INTERNIST, PUFF,

TEIRESIAS
„ Geology: PROSPECTOR
‰ SRI [Duda 1978]
‰ Elicited, preserved, and applied geologic formation
knowledge to assist in mineral exploration
„ Manufacturing: XCON
„ Developed by DEC & CMU [McDermott 1982]
‰ Elicited, preserved, and applied the knowledge of human
configurators of computer systems to automate and
duplicate their functions
Other examples

„ Manufacturing: COOKER
‰ Texas Instruments for Campbell Soup
‰ Assists in the maintenance of soup-making equipment
‰ Captured and preserved the knowledge of a highly
experienced employee about to retire
„ Credit: AUTHORIZER’S ASSISTANT
‰ Inference Corp & American Express [Leonard-Barton
1988]
‰ Elicited, preserved, and applied human knowledge in
handling applications for AmEx credit cards.
‰ Takes information from multiple databases and issues
approval/denial of large purchase requests from merchants
Types of Knowledge in KBS –car example

„ General support simple knowledge (known to all

mechanics):
‰ Can test battery (in two different ways), inspect cables,
inspect starter switch etc.
„ Specific simple knowledge (known to only the
mechanic working on the particular car):
‰ The outcome of each test or inspection for that particular
car
„ Specific tactical or strategic knowledge (known only
to expert mechanics):
‰ What order should the inspections most efficiently be done
in, in all possible situations for all cars
Types of KBS knowledge (chess game)

„ Knowledge about the rules of the game

‰ General, simple support domain knowledge; embedded in
operators of the solution space
‰ This knowledge tells the player which moves can be made
„ Knowledge of how best to move the pieces to defeat
the opponent
‰ Complex, specific tactical knowledge, based on
experience.
‰ This knowledge tells the player which move should be
made
KBS and Expert Systems

„ KBS and ES synonymous with rule-based

systems
‰ Heuristics best represented with if-then rules
„ KBS excel at representing associational
knowledge
‰ this reflects heuristic ability and elicited through
the knowledge engineering process
‰ System’s knowledge can associate observations
(inputs) with specific causes (outputs) (if-then
rules)
KBS

„ KBS are poor in solving motor skills expertise

‰ Human performance knowledge is tacit
‰ This expert performance is physical, not cognitive
‰ Robotic capabilities for processing sensory data is not
sufficiently developed to navigate the physical world like
humans
„ RoboCup competition (Kitano, 1998)
„ Expert systems are knowledge-based systems
where:
„ the specific knowledge is at the level equivalent to a human
expert,
„ the inference engine is able to make use of the knowledge as
efficiently and effectively as a human expert
Advantages of KBS

„ Wide distribution of scarce expertise

‰ Very few experts, ability to reproduce expert’s knowledge
„ Ease of modification
‰ Separating K from the reasoning mechanism eases
modification; important in heuristic programming with
frequent changes
„ Consistency of answers
‰ different human experts often present dissimilar answers
to the same problem; KBS are consistent in problem
solving, uniform answers
Advantages

„ Perpetual accessibility
„ Preservation of expertise
„ Solution of problems involving incomplete
data
‰ KBS can solve incomplete or inexact data
„ Explanation of solution
‰ Partly because of heuristic nature, KBS track
knowledge used to generate solutions.
Explanations assist user by clarifying results,
tutoring.
Disadvantages

„ Answers may not always be correct

‰ Costly errors, monitoring system that protects
expensive equipment from malfunctions
„ Limits not always recognised
‰ KBS do not know when they do not know!
‰ Always try to deduce a solution; misleading or
incorrect answers generated
„ Lack of common sense
‰ Impossible to represent it in KBS
‰ Everything has to be explicitly said (in a rule)
Conclusions

‰ The concept of expertise in the context of

knowledge
‰ The state space search methods comprising early
AI work
‰ The difference between these and the modern
knowledge-based systems
‰ How knowledge-based systems can be used to
manage knowledge.