Case-based reasoning

1
Case-based reasoning

• ISYS 370
• R. Weber

2
CBR applicationsCCBRconversational
CBRhttp//www.egain.com/pages/Level2.asp?Section
ID4PageID4http//support.lucasarts.com/yoda/s
tart.htm
3
Deployed CBR applications (i)

• PROFIT valuates residential properties to
  evaluate mortgage packages for a division of GE
  Mortgages. Values of a property change with
  market conditions, so estimates have to be
  updated constantly according to real estate
  transactions, which validate the estimations.
• CARMA is designed to provide expert advice on
  handling rangeland grasshopper infestations.
  CARMA has reused its expertise combined with
  model-based methods to devise policies on pest
  management and the development of industry
  strategies.

4
Deployed CBR applications (ii)

• General Motors has developed an organizational
  CBR system to support the goals of dimensional
  management, an area in the manufacturing of
  mechanical structures (e.g., vehicle bodies) that
  enforces quality control by reducing
  manufacturing variations that occur in fractions
  of millimeters.
• Western Air is an Australian distributor of heat
  and air conditioning systems they have chosen to
  use a web-based CBR application 20 to guarantee
  a competitive advantage that also poses an entry
  barrier to competition. They guarantee the
  precision of the specifications of each new
  system and the accuracy of the quotes by relying
  in knowledge captured in previous installations.

5
Deployed CBR applications (iii)

• Dublet recommends apartments for rental in
  Dublin, Ireland, based on a description of the
  users preferences. It employs information
  extraction from the web (of apartments for rent)
  to create cases dynamically and retrieves units
  that match the users preference. Dublet performs
  knowledge synthesis (creation) and extends the
  power of knowledge distribution of the CBR system
  by being operational in cell phones.
• PTV combines case-based (content-based)
  personalization with collaborative filtering to
  recommend shows to watch on digital television.

6
Deployed CBR applications (iv)

• NEC has developed SignFinder, which is a system
  that detects variations in the case bases
  generated automatically from customer calls. When
  they detect variations on the content of typical
  customers requests, they can discover knowledge
  about defects on their products faster than with
  any other method.

7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
Further reading

• Riesbeck Schank (1989) Inside case-based
  reasoning
• Kolodner (1993) Case-based reasoning
• Aamodt Plaza (1994) AICom paper (todays
  reading)
• Leake (1996) Leake, David. (1996). Case-Based
  Reasoning Experiences, Lessons, and Future
  Directions.
• Watson (1997) Applying Case-Based Reasoning
  techniques for enterprise systems.

12
Introduction

• from a knowledge representation concept (i.e.
  scripts, MOPS)
• role of understanding in solving problems
• CBR assumptions
• similar problems have similar solutions
• problems recur (Leake, 1996)

13
Definitions

• From Riesbeck Schank (1989), "A case-based
  reasoner solves new problems by adapting
  solutions that were used to solve old problems".
• Case-Based Reasoning systems mimic the human act
  of reminding a previous episode to solve a given
  problem due to the recognition of their
  affinities (Weber, 98).
• Case-based reasoning is a methodology that reuses
  previous episodes to approach new situations.
  When faced with a new situation, the goal is to
  retrieve a similar previous one and reuse its
  strategy (Weber, 02).

14
CBR methodology
Task?
AI TaskDiagnosisPrescriptionInterpretation-ad
viceRecommendationAnalysis-predictionScheduleP
lanning
15
CBR methodology
Task?
case base
16
CBR methodology
situation assessment
case base
17
CBR methodology
RETRIEVE
case base
RETAIN
REUSE
REVISE
18

• Knowledge in case-based reasoning systems
• by Richter, M. M., The Knowledge Contained in
  Similarity Measures Some remarks on the invited
  talk given at ICCBR'95 in Sesimbra, Portugal,
  October 25, 1995. Online http//www.cbr-web.org/
  documents/Richtericcbr95remarks.html

19
Case representation

• case problem symptoms A, B, C
• case solution disease 1
• case outcome confirmed

20
Case acquisition/authoring

• cases are acquired from real experiences
• cases are created from categories of real
  experiences (prototypes)
• cases are authored by an expert
• cases are learned by data analysis
• cases are searched in patterns
• cases are converted (extracted) from text
• cases are learned from text

21
Similarity

• The key to its success is expertise to determine
  what makes a case similar to another. For
  example, if you have a common cold and your
  spouse has the flu, you will be able to recognize
  these two conditions are similar. But only a
  physician can determine whether two infirmities
  are similar so that the same treatment can be
  applied. It is expert knowledge that tells when a
  case is similar to another in the context of a
  CBR system.
• Similarity function is a knowledge representation
  formalism to measure similarity between two cases

22
Retrieval

• similarity functions measure similarity
• all cases (or a selected portion) are compared to
  the target (problem) case
• cases are retrieved when their similarity is
  above a pre-defined threshold
• this threshold determines the point from which
  cases are considered similar

23
Adaptation

• All features that describe a case and are not
  used for retrieval can potentially be adapted

24
Adaptation methods

• substitution
• reinstantiation replacement based on a role
• parameter adjustment (proportional)
• local search (taxonomy)
• query memory
• case-based substitution alternatives in cases
• transformation transform by changing features
  either by substitution or deletion
• common-sense transformation
• model-guided repair

25
Learning

• learning by incorporating new cases to the case
  base
• learning by adding cases that are adaptations
  from retrieved cases

26
CBR and AI tasks (i)

• interpretive
• past cases are used as references to categorize
  and classify new cases
• interpretation, diagnosis
• problem-solving
• past cases are used to provide a solution to be
  applied to new cases
• design, planning, explanation

27
CBR and AI tasks (ii)

• Mundane
• prediction-advice
• composition
• understanding
• reading
• planning
• walking
• uncertainty
• creativity

• Both
• interpretation
• classification
• categorization
• discovery
• control
• monitoring
• learning
• planning
• analysis
• explanation

• Expert
• diagnosis-troubleshooting
• prescription
• configuration
• design
• scheduling
• retrieval
• mediation
• argumentation
• recommendation

28
vocational counseling

• diagnosing headaches

29
Advantages of CBR systems (i)

• Knowledge acquisition and representation There
  is no need to explicit acquire and represent all
  the knowledge the system can use.
• CBR systems can avoid mistakes
• Common sense knowledge that would have to be
  represented explicitly is implicitly stated in
  cases.
• Not easily formalizable tasks such as in some
  medical domains, prototypical descriptions
  represent more easily a body of knowledge.

30
Advantages of CBR systems (ii)

• Creativity - Case solutions can be combined into
  new ones and cases can also be used in a
  different level of abstraction providing
  innovative solutions.
• Learning - can be done without human
  interference CBR systems can become robust and
  provide better solutions. Users feedback is
  easily incorporated in the revise phase.
• Degradation -CBR systems can recognize when no
  answer exists to a problem by simply defining a
  threshold from which a solution is no longer
  acceptable. In decomposable problem domains, a
  solution can be created from the combination of
  partial solutions.

31
Advantages of CBR systems (iii)

• (shared with ES and other AI methods)
• Permanence - CBR do not forget unless you program
  it to.
• Breadth - One CBR system can entail knowledge
  learned from an unlimited number of human
  experts.
• Reproducibility - Many copies of a CBR system.

32
current issues

• case authoring
• case base maintenance
• methods for distributed case bases

33
Building (shells), using, maintaining

• Shells/tools
• http//www.cbr-web.org/CBR-Web/?infotoolsmenupt
  
• Esteem examples, NISTP CBR Shell examples
• Using
• Laypeople, experts
• Maintaining
• Automatically learning new cases
• Cases are real or created
• Manually adding new cases

34
CBR and grounds for computer understanding

• Ability to represent knowledge and reason with
  it.
• Perceive equivalences and analogies between two
  different representations of the same
  entity/situation.
• Learning and reorganizing new knowledge.
• From Peter Jackson (1998) Introduction to Expert
  systems. Addison-Wesley third edition. Chapter 2,
  page 27.