Transcript Problem Solving Methods and Computer

Problem Solving Methods and
Computer-Aided Knowledge Acquisition
Goals and Achievements:
Tools applicable for construction of many systems
Structured design and elicitation for single system
Overview of this lecture
•
•
•
•
•
•
Limitations of Rule-based knowledge representation
Expert System classifications
Classification tasks
Example: Electronics repair knowledge
MORE: classification knowledge elicitation
Conclusions
Expertsystemen 10
2
Knowledge representation in Rules
• RULE as domain knowledge? • Strategic knowledge is often
IF
X is a rabbit
represented implicitly in
THEN X has four legs
Conflict Resolution (PRESS
Describes in fact how to infer
lecture 5).
a conclusion: operational.
• Implicit representations:
update / maintenance
• Mix Support, Strategic,
• Rule models and checkers
Structural knowledge:
(lecture 9): partial solution
IF
Radio is dead
THEN Put voltmeter on
Conclusion: Rules are
battery
• GOOD as a basis for
inferencing
• Context dependence:
• POOR as a general
IF pinguin THEN not fly
knowledge representation
IF bird
THEN fly
formalism
Expertsystemen 10
3
Strategy in Facade NLU
• Linguistic domain knowledge
is partitioned in
•
synonyms,
•
idiomatic expressions,
•
negations,
•
discourse (sub) acts
• Per class same treatment
(strategy)
•
id.expr: retract words
•
synonyms: high salience
• This structure of knowledge
is also linguistic knowledge!
• Elicitation: talk to expert in
familiar terms
Expertsystemen 10
4
Knowledge compilation
Tool
System
• Rules: the inferencing
“assembly language”
• Maintain knowledge at
more abstract level
• Compile knowledge into
inferencing rules
Domain
Knowledge Base
Compilation
Elicitation
Rule Base
Consultation
Inferencer
Expertsystemen 10
5
General Knowledge Compilation Tools?
• Design of tool is tightly
linked with roles of
knowledge and expert’s
approach
• Attempt: classify all
possible expert systems
into small number of
categories
• Ideal: make one tool per
category
• Expert Systems are too
much different!
• End with one compilation
tool per system
Expertsystemen 10
6
Classification by Hayes-Roth (1983)
of 10
systems
?
Ten categories of
• Interpretation
• Prediction
Diagnosis
and
• Diagnosis
treatment
of
• Design
illness
called
•ignorance
Planning
• Monitoring
• Debugging
• Repair
• Instruction
• Control
description of
the future?
Expert Systems:
description from observation
consequences from events
faults from symptoms configuration
configuration from constraints
of steps?
step sequence from goal
deviations from behavior
remedies from faults
the same?
remedies from faults
module sequence from feedback
steps from goals and observations
Expertsystemen 10
7
Clancey: Interpretation and Construction tasks
Interpretation
• Task involving some
working system
• Solution from enumerable
set
• Top-down inference
Construction
• Task of formation of a
working system
• Solution space implicitly
defined
• Bottom-up inference
Construction:
• Different Problem Solving
methods:
Backtracking,
Propose-and-Apply,
Propose-and-Revise,
Least-commitment …
• RIME/XCON, VT/SALT ..
• Lecture 14
Applies to subtasks
Expertsystemen 10
8
Interpretation
System as input output map
• Input unknown: Control
(what treatment is the best)
• System unknown: Identify
(what component is failed)
• Output unknown: Predict
(will the reactor explode)
Input
System
Output
Lectures
Student
Knowledge
Treatment
Patient
Life exp.
Bar control
Reactor
Pressure
If the solution space is an enumerable set:
Problem is to determine in what category
our instance belongs:
CLASSIFICATION
Expertsystemen 10
9
Heuristic Classification method
Clancey’s three steps:
• Data Abstraction:
20.6Volt: “Low voltage”
• Heuristic Match:
Low Voltage indicates Power
Supply problem
• Solution Refinement:
Continue within limited
search space
Abstract
data
Solution
class
Data
Solution
Systems with classification
as main or sub task:
• MYCIN: match data to preenumerated disease using
rules with CF
• SACON: suggest simulation
type for MARC software
• SOPHIE: Find faulty module
in circuit, faulty component
in module (measurements)
• COMPASS: diagnose
telephone switch (error
messages)
Expertsystemen 10
10
Heuristic and Hierarchical Classification?
Clancey 1985, Heuristic
• Choices may lead to
overlapping subspaces
• Difficult choices can be
postponed
Chandrasekaran 1986, Hierarchical
• Strict taxonomy of solutions: no
overlap
• Need confirmation of each step
because no correction possible
• Choose bird if it flies,
correct bat later
• Choose bird if it flies, lays eggs
and has feathers and bones.
Expertsystemen 10
11
Repair Knowledge and Repair Strategies
How to repair a circuit?
• Repair shop??
• 200 electrical components
• one or more faulty
• Knowledge about properties
of each component
• Knowledge about interaction
Strategy 1:
• Test/replace each component
in some order
Strategy 2:
• Employ structural grouping
of components
Expertsystemen 10
12
Grouping of system components
Planning/Analysis phase:
• Distinguish logical subunits
of circuitry
• Characterise behavior that
differentiates between faults
in subunits
• For each subunit, list normal
values for measurements
• For each measurement, give
components to determine it
Consultation:
• Run behavioral tests until
faulty subunit is found
• Measure in faulty unit
• For deviating measurements,
check suspect components
• Replace defective component
• Repeat until radio plays
Domain independent
Problem Solving
Strategy that can be
coded into Elicitation
Tool
Expertsystemen 10
13
MORE Domain Models
• Hypotheses
We want to select from one of
the things that can be wrong
• Symptoms
Selection is based on these
observations (attributes)
• Conditions
Influences on the likelyhood of
hypothesis and symptoms
• Tests
Find out if a condition arises
Expertsystemen 10
H1
H2
H3
S1
S2
S3
H4
H5
S4
14
Confidence Factors, Measure of (Dis) Belief
• MORE generates Diagnostic Rules for
H1
Hypo – Symp associations:
IF
S1
THEN H1
WITH (mb, md)
• Diagnostic rule: MB Positive and MD
Negative Confidence Factor
• MB is high if
• H1 is only/most likely explanation for S1
• Prior probability for S1 is low
S1
Pr(S1)
Pr(H1 -> S1)
Pr(H1)
• MD is high if
• S1 is a very likely consequence of H1
• XS based on CF, not probability
Expertsystemen 10
15
Conditions and Tests
MORE Background conditions:
• “Condensator problems are
more likely if the radio was
stored humid”
• “Resistor problems are more
likely if the radio was badly
ventilated”
MORE Tests:
• Humid storage gives
moisture patches
• Bad ventilation overheats
rectifier and output
Expertsystemen 10
16
Symptom and Hypothesis Rules
Symptom Confidence Rule:
• Rank importance of observed
syptoms
• Use prior probability and
background conditions
• Use reliability induced by
tests
Hypothesis Expectancy Rule:
• Rank probabilities of
hypotheses
• Use prior probabilities and
background conditions
Expertsystemen 10
Clancey’s heuristic
classification:
Abstract
data
Solution
class
Data
Solution
SCR
DiaR
HER
17
Knowledge Elicitation in MORE
Long before MORE:
• Give me a Rule …
• I’ll add it to the program
• Test exhaustively
MORE
• Tell
• Tell
• Tell
Before MORE:
• Give me a Rule
• I’ll check if it looks familiar
• I’ll add it to the program
• Test
• I’ll ask you questions until I
think I know enough
• I’ll convert the knowledge to
rules for you
Rule level
Knowledge elicitation:
me the Hypotheses
me their probabilities
me about Symptoms
Abstract level
Expertsystemen 10
18
Knowledge Elicitation Steps of MORE
Questions that MORE may ask
the Expert:
Questions are guided by
MORE’s state of the model:
• Differentiation:
What S differentiates
between H1 and H2?
• Apply when:
H1 and H2 have no
Differentiating Symptom
• Frequency Conditionalization: • Apply when:
What BC influences the
S has no rules with high mb
probability of S?
and md
• Symptom distinction:
Refine S to distinguish H1
from H2
• Apply when:
S has no rules with high mb
Expertsystemen 10
19
MORE: Knowledge driven knowledge elicitation
• MORE was good for
building MUD;
• otherwise insufficently
general!
Domain
Knowledge Base
Compilation
• MORE contains problem
solving knowledge
• MORE collects domain
knowledge from the
Human Expert
• MORE compiles PSM
plus Domain knowledge
into rules
• MORE uses PSM
knowledge to guide
elicitation
Rule
Base
Elicitation
Feedback
• Reason using cost
of test and repair
Expertsystemen 10
20
MUD
• Drilling fluid used in oil excavation
• Lubrication, cooling, waste removal,
information stream
• Drill interruptions are costly
• Carefully continuously examine mud
temperature, viscosity, composition
• MUD was developed for the quick
treatment of mud problems
• MORE was developed for the quick
treatment of MUD problems
Expertsystemen 10
21
Similar approaches
Construction systems: Interpretation systems:
• VT and SALT:
• PUFF and CENTAUR:
Propose and Revise
Hierarchical Hypothesize and Test
(w/o single fault assumption,
• XCON and RIME:
resembles construction)
Propose and Apply
• TEST and TDE:
Lectures 14 (and 15)
Abstract HHaT in tree of
hypotheses
Expertsystemen 10
22