Some Special Purpose Knowledge Representations

1
Some Special Purpose Knowledge Representations

• Bob McKay
• School of Computer Science and Engineering
• College of Engineering
• Seoul National University

2
Outline

• Knowledge Representation for Knowledge
  Acquisition
• Ripple-Down Rules
• Model Based Knowledge Acquisition MORE
• Case-Based Reasoning
• Qualitative Reasoning QSIM

3
Ripple-Down Rules

• Based on work at UNSW (Kensington)
• Rules with exceptions
• Reverses usual conflict resolution (last matching
  conclusion is used)
• When the system comes to a conclusion with which
  the expert disagrees
• case is generalised as far as appropriate
• added to the knowledge-base as an exception rule
• System eventually consists of exceptions to
  exceptions to exceptions....

4
Ripple Down Rules
5
Key features

• Automatic rule placement
• And matching inference
• Expert identifies features that distinguish the
  case from
• A single past case
• A selection of past cases
• All seen cases

Show stored cases to the expert one by one

• Case by case development while in use

6
Problem Types

• Preston, Srinivasan
• Kang, Preston
• Preston, Ramadan
• Richards
• Beydoun Hoffman
• Kang, Ho, Wobcke
• Singh, Mishra, Sowmya
• Beckman, Hoffman
• Hoffman, Kang
• Cao, Martinez-Bejar
• Finlayson
• Hoffman

• Single Classification
• Multiple Classification
• Configuration
• Resource allocation
• Heuristic search
• Document management
• Image processing
• Genetic Algorithm
• Information extraction
• Ontology development
• Planning (soccer)
• Translation

7
Viewpoint

• No system is ever complete
• A robust case-by-case maintenance system is
  required
• Needs domain expert (user)
• Build the whole system using the maintenance
  system
• Applies to all systems ??????

8
Model-Based Elicitation

• We will look at an example system MORE
• Model-based systems have two knowledge
  representations
• The domain model
• Representation tailored for easy expression of
  domain
• With which the expert interacts
• The performance model
• Representation tailored for simple reasoning,
  explanation etc
• In which the actual reasoning occurs
• usually some form of logic rules

9
MORE Domain

• MORE deals with heuristic classification tasks
• Specifically, maintenance of drilling fluid in
  systems such as oil rigs

10
MORE Models
11
MORE Domain Models

• Symptoms
• Items user may observe during diagnosis, and seek
  to explain
• Attributes
• Serve to further discriminate symptoms
• eg rapid increase or decrease in some property
• Hypotheses
• Events that are possible causes of symptoms
• therefore serve as hypotheses
• Background Conditions
• Conditions which make symptoms more or less
  likely
• given the occurrence of a hypothetical cause
• Tests
• can be used to determine presence or absence of
  background conditions
• Test Conditions
• Conditions that have a bearing on the accuracy of
  the tests

12
MORE Performance Representation

• The MORE performance representation
• form the system actually reasons with
• maintained in the form of production rules
• They fall into three classes
• Diagnostic Rules
• heuristic mapping between symptoms and hypotheses
• Worked out from the causal links in the domain
  model
• Symptom Confidence Rules
• Give estimates of test reliability under
  different background conditions
• Hypothesis Expectancy Rules
• Give estimates of the prior probability of
  hypotheses under different background conditions.

13
MORE KA Strategies

• Differentiation
• Seek symptoms that distinguish between hypotheses
• for example, symptoms that have only one cause
• Applied when MORE discovers a pair of hypotheses,
  H1 and H2, that have no differentiating symptom
• MORE elicits a symptom from the expert, and adds
  it to the event model with appropriate links
• Frequency Conditionalisation
• Determine any background conditions that make a
  particular hypothesis more or less likely
• Applied when a family of rules (ie rules with the
  same conclusion) lacks any rules with either high
  or low, positive or negative, confidence factors

14
MORE KA Strategies

• Symptom Distinction
• Identify special properties that indicate the
  underlying cause
• Eg a rapid decrease in density suggests an influx
  of water, rather than shale contamination
• Applied by MORE when a family of rules contains
  no members with high positive confidence factors
• Symptom Conditionalisation
• Find out the conditions under which different
  symptoms might be expected to manifest
  themselves, given a particular disorder
• Such conditions set up expectations which may
  serve to rule out hypotheses if they are not
  confirmed
• Applied when there are no rules in a family that
  have a high negative confidence factor

15
MORE KA Strategies

• Path Division
• Attempt to uncover intermediate events between a
  hypothesised disorder and an expected symptom
• which have a higher conditional probability of
  occurrence than the symptom itself
• If you don't observe that intermediate event, you
  have stronger evidence against the hypothesis
  than just failing to observe the symptom
• Applied when a rule family lacks any rule that
  associates a high negative confidence factor with
  failure to observe a symptom of that family's
  hypothesis
• MORE seeks an intermediate event that is caused
  by the hypothesis
• failure to observe it constitutes stronger
  evidence against the hypothesis

16
MORE KA Strategies

• Path Differentiation
• Try to split up causal pathways between disorder
  and symptom
• As in path division
• The motivation is to discover intermediate events
  that will allow us to differentiate between
  disorders that have similar symptoms.
• Applied when a symptom is linked to two different
  hypotheses
• MORE elicits from the expert an intermediate
  event that causes the symptom, and is caused by
  one hypothesis, but not the other
• Test Differentiation
• Determine the degree of confidence to be placed
  in test results
• Evidence is normally the result of tests with
  varying reliability
• Applied when a rule family lacks rules with
  either high or low positive or negative
  confidence factors
• Will ultimately generate symptom confidence rules

17
MORE KA Strategies

• Test Conditionalisation
• Determine the background conditions that affect
  the reliability of tests
• This information has a bearing on the
  significance of observations in particular cases
• Applied when a rule family lacks rules with
  either high or low positive or negative
  confidence factors
• Will ultimately generate symptom confidence rules

18
MORE Reanalysis

• Confidence Factors Trigger Reanalysis
• MORE has some expectations about the
  relationships between confidence factors
• When its expectations are violated, MORE consults
  with the user to try to rectify the rules
• Inference Path Lengths
• Suppose that a disorder D leads to symptom S1,
  and S1 leads to another symptom S2
• MORE expects that the confidence factor
  associated with the rule mapping S1 to D will be
  greater than or equal to that associated with the
  rule mapping S2 to D

19
MORE KA Strategies

• Diagnostic Significance
• Suppose that symptom S1 is only caused by D1,
  whereas S2 is caused by a number of disorders
  including D1
• MORE expects that the confidence factor
  associating S1 with D1 is greater than that
  associating S2 with D1
• Rule Families
• MORE has expectations regarding the relative
  values of confidence factors associated with
  rules in the same rule family
• For example, adding a symptom condition to a rule
  family that increases the conditional likelihood
  of the symptom should result in rules that have
  greater negative confidence factors than their
  constituent rules
• That is, the more we anticipate a particular
  symptom, given a particular hypothesis, the
  greater our shift towards disbelief in the
  hypothesis if that symptom is not observed

20
Case-Based Reasoning

• In KBS applications discussed previously
• human heuristic knowledge and experience are
  compiled and transformed into a condensed
  form
• However, human experience can be used directly
  and explicitly as well
• A motor mechanic noticed some blue smoke coming
  out of the muffler of a car. He recalled a
  previous case where an engine problem caused the
  same kind of smoke. So he decided to check the
  engine first.
• This is a typical example of case-based reasoning
  
• A stored, similar case is retrieved to derive a
  solution for the new case

21
Case-Based Reasoning

• A case-based reasoning system consists of three
  major components
• A library of historical cases
• A means of using the key elements of the present
  problem to find and retrieve the most similar
  case(s) from the library
• A means of modifying the proposed solution when
  the retrieved case is not identical to the
  current one
• The key issue in case-based reasoning is how to
  organise the library so that efficient search for
  a similar case can be achieved
• It is ideally suited for the situation where many
  well-documented problems and solutions exist
• Legal cases as precedents
• Predicting weather from previous similar
  situations

22
Components

• Representation
• Retrieval
• Matching engine retrieves cases similar to target
  case.
• Adaptation
• Remembering

23
Case-Based Reasoning Cycle
24
Breathalyser

• Example cases

• Duration is duration of drinking session.
• Perhaps elapsed time should be added as a case
  feature?

25
Case Representation

• The knowledge engineering task is focused on
  deciding how to represent cases
• what features best characterise cases
• i.e. predictive features
• may require expert analysis
• e.g. for image classification the bitmap may need
  to be converted to an edge map.
• e.g. height and weight may not be useful in
  themselves for classifying apples and pears,but
  height/weight ratio is.

26
Case retrieval

• Based on some similarity measure.
• e.g number of matching features
• e.g. distance measure based on difference between
  numeric features
• Indexes may be used to speed the retrieval

27
Case indexing - Example
28
k-Decision Tree
29
Case Adaptation

• Breathalyser
• if actual consumption is 2 more than in retrieved
  case add 0.5 to blood alcohol count.
• Property Valuation
• for extra bedroom add x to price
• More complex adaptation may be needed where
  solutions are plans or designs, rather than
  single values.

30
CBR Advantages Disadvantages

• Advantages
• allows the reasoner to propose solutions quickly
• allows the reasoner to propose solutions in
  domains not completely understood by the reasoner
  
• cases are useful in interpreting open-ended and
  ill-defined concepts
• lends itself to analogical reasoning
• knowledge acquisition process is considerably
  simplified
• Disadvantages
• solutions may be biased towards stored cases
• solutions might be used blindly without proper
  validation, especially by novices

31
CBR Applications

• Early Software Cost Estimation
• ww.cs.tcd.ie/publications/tech-reports/
  reports.99/TCD-CS-1999-36.pdf
• Weather Prediction
• www.mcs.vuw.ac.nz/comp/Publications/
  CS-TR-93-7.abs.html
• www.aic.nrl.navy.mil/papers/2001/
  AIC-01-003/ws5/ws5toc1.pdf
• Travel Check-in
• http//www.check-in.com/
• Building Design
• http//nathan.gmd.de/projects/fabel/fabel.html
• And dozens of other applications
• http//www.cbr-web.org/CBR-Web/?infoprojectsmenu
  pp

32
QSIM A Qualitative Reasoning Language

• An example of a very narrowly focussed domain
  language
• Attempts to use physical constraints to describe
  what qualitative states are possible for a
  physical system being simulated
• Parameters
• Qualitative states (QS) are defined as a set of
  parameters
• each of which is either
• Increasing
• decreasing
• steady
• Critical Points
• Each physical parameter is represented as a
  function of time that has a finite number of
  critical points
• A critical point is where the physical parameter
  changes value
• Landmark Values
• A totally ordered set of landmark values
  represents all the values of the physical
  parameter at its critical points

33
QSIM A Qualitative Reasoning Language

• Time
• Time is represented by a set of distinguished
  time points in the time line
• Only two types of time elements can be defined
• the time at a point
• the time interval between two points
• Initial State
• The initial state is defined by the qualitative
  values of various physical parameters at a
  particular time point or time interval
• Qualitative Behaviour
• The basis of the qualitative simulation is the
  determination of all the possible combinations
  of values for all the physical parameters at a
  point or interval in time
• A sequence of qualitative states describes a
  qualitative behaviour of the system

34
QSIM A Qualitative Reasoning Language

• Constraints
• The number of possible combinations of values
  for all the physical parameters is huge
• Constraints of the model must be used to reduce
  the search space
• If more than one qualitative change remains, then
  the current qualitative state has multiple
  successors
• The simulation will produce a tree
• Objective
• The objective of qualitative simulation is to
  determine the next feasible qualitative state (or
  states) of the physical parameters
• If the current QS is defined at a point,
  then the next QS would be defined at an
  interval starting from this point
• If the current QS is defined at an interval, then
  the next QS would be defined at a point that is
  the end of the interval

35
QSIM A Qualitative Reasoning Language

• Transitions
• A physical parameter is restricted to a
  predetermined set of possible transitions from
  one QS to the next
• There are two types of transitions
• p-transitions (from point to interval)
• i-transitions (from interval to point)

36
QSIM Example

• We describe a qualitative simulation of a ball
  being thrown upwards
• The system is determined by three physical
  parameters
• acceleration
• velocity
• height

37
QSIM Example Transition Table

• p-transition From QS(f,ti)
  To QS(f,ti, ti1)
• P1 ltlj,stdgt
  ltlj,stdgt
• P2 ltlj,stdgt
  lt(lj, lj1),incgt
• P3 ltlj,stdgt
  lt(lj-1, lj),decgt
• P4 ltlj,incgt
  lt(lj, lj1),incgt
• P5 lt(lj, lj1),incgt
  lt(lj, lj1),incgt
• P6 ltlj,decgt
  lt(lj-1, lj),decgt
• P7 lt(lj, lj1),decgt
  lt(lj, lj1),decgt
• i-transition From QS(f,ti,
  ti1) To QS(f,ti1)
  
• I1 ltlj,stdgt
  ltlj,stdgt
• I2 lt(lj, lj1),incgt
  lt lj1,stdgt
• I3 lt(lj, lj1),incgt
  lt lj1,incgt
• I4 lt(lj, lj1),incgt
  lt(lj, lj1),incgt
• I5 lt(lj, lj1),decgt
  ltlj,stdgt
• I6 lt(lj, lj1),decgt
  ltlj,decgt
• I7 lt(lj, lj1),decgt
  lt(lj, lj1),decgt
• I8 lt(lj, lj1),incgt
  ltl,stdgt
• I9 lt(lj, lj1),decgt
  ltl,stdgt

38
QSIM Example

• The first state at time t0 is when the ball is
  initially thrown upwards
• The second QS can be described as an interval
• QS(A t0 t1) (g std), where
• A is acceleration
• g is a landmark value
• std means steady
• QS(V t0 t1) (0? dec), where
• V is velocity which is decreasing (dec)
• The qualitative value is somewhere between zero
  and the initial velocity
• which could be any positive value
• QS(Y t0 t1) (0 ? inc), where
• Y is height which is increasing (inc)
• The qualitative value is between the initial
  height (0) and the maximum height
• which could be any positive value

39
QSIM Example

• The time interval t0 t1 is when the ball is
  moving up
• The next objective is to compute the next QS,
  QS(F t1)
• The following steps are involved in computing the
  next QS.
• Determine all possible i-transitions from the
  current QS using the transition table. For
  example
• A I1
• V I5 I6 I7 I9
• Y I4 I8 (note Y ? is impossible, ruling
  out I2, I3)
• Apply various constraints to filter out
  infeasible transitions
• I4 for Y is incompatible with I5, I6, I9 for V
• Based on the relationships between velocity and
  position
• I8 for Y is incompatible with I5, I7, I9 for V
• Based on the relationships between velocity and
  position

40
QSIM Example

• After this step, two combinations remain
• A I1 V I7 Y I4
• A I1 V I6 Y I8
• The first is the same as the current QS
• thus filtered out
• The second will be the next QS
• That is
• QS(A t1) (g std)
• QS(V t1) (0 dec)
• QS(Y t1) (Ynew std)

41
QSIM Example

• Now to obtain QS(F,t2)
• Determine all possible p-transitions
• A P1, P2, P3
• V P6
• Y P1, P2, P3
• Physical constraints rule out P1, P2 for Y
• Since V is zero and decreasing
• Physical constraints rule out P2, P3 for A
• The new state is
• QS(A t2) (g std)
• QS(V t2) (-inf0 dec)
• QS(Y t2) (0Ynew dec)

42
QSIM Applications

• http//www.cs.utexas.edu/users/qr/qsim-users.html
  QSIM-apps
• Structural Design
• Classification Analysis of Materials
• Chemical Hazard Identification
• Modelling Chemical Reactions
• Genetic Regulatory Networks
• Verifying Controller Behaviour
• Air Conditioner Faults
• Etc

43
Summary

• Representation for Knowledge Acquisition
• Ripple-Down approach provides maintainable
  knowledge
• Separation of Domain and Performance
  representation in MORE
• Domain model supports knowledge acquisition
• Performance representation supports reasoning,
  explanation
• Case-Based Reasoning
• Reasoning with reduced explicit knowledge
  modelling
• Reasoning for Special Purposes
• QSIM for Qualitative Simulation of the Real World

44
Summary

• Its clear that people dont follow probabilistic
  laws in reasoning
• Response of probabilists people are just wrong
• Alternative response people have learnt
  effective methods to reason with complex
  dependencies
• Dempster-Shafer theory
• Distinguishes between uncertainty and lack of
  knowledge
• Consistent, systematic treatment of lack of
  knowledge
• Fuzzy logic
• Deals with fuzzy concepts rather than probability
• Emphasises efficient computation