Kapitel 5 Casebased Reasoning CBR

1
Kapitel 5 Case-based Reasoning (CBR)

• 5.1 Case-based Reasoning
• 5.1.1 CBR - basis
• 5.1.2 CBR scenario
• 5.1.3 CBR system - characteristics
• 5.2 CBR Knowledge management
• 5.2.1 CBR as a subsystem in the KM system
• 5.3 CBR applications
• 5.4 Conclusion

2
Sources, partially used for preparing this
chapter

• http//www.cs.auckland.ac.nz/ian/teaching.html
• Ian WatsonDept. of Computer ScienceUniversity
  of Auckland
• http//wwwagr.informatik.uni-kl.de/sections/educat
  ion/index.html
• Dr. rer. nat. Ralph Bergmann
• Dr. rer. nat. habil. Klaus-Dieter Althoff
• University of Kaiserslautern Department of
  Computer Science
• http//www.aic.nrl.navy.mil/aha/slides/
• David W. AhaNavy Center for Applied Research in
  AINaval Research Laboratory

Thanks to these authors!
3
5.1 Case-based Reasoning (CBR)5.1.1 CBR - basis

• CBR is a discipline in AI and is commonly
  described as
• a problem-solving method or reasoning model
• whose core processes revolve around the
• retrieval,
• reuse, and
• retention of previously encountered
  problem-solving episodes or cases.
• CBR is a methodology
• to model human reasoning and
• for building intelligent application systems
• CBR is about storing, sharing and using
  individual experience a Knowledge Management
  task

4
Definitions

• A case- based reasoner solves new problems by
  adapting solutions that were used to solve old
  problems (Reisbeck Schank 1986)
• CBR is both the ways people use cases to solve
  problems and the ways we can make machines use
  them. (Kolodner, 1993)
• CBR is a recent approach to problem solving and
  learning (Aamodt Plaza 1994)
• CBR is reasoning by remembering (Leake,1996)

5
A simple CBR model

• Solve new problems by selecting previous cases
  for similar problems and adapting them to current
  problem

new solution
solution
adaptation
solved by
new problem
problem
similarity
Case1
new case
Case-base
6
The CBR-cycle
(Aamodt Plaza, 1994)
7
The CBR-cycle

• Process-model
• 1. Retrieve
• - find all similar cases
• 2. Reuse
• - adapt to the new case-problem
• 3. Revise
• - evaluate solution for the new problem
• 4. Retain
• - store the acquired experience

8
CBR models human reasoning

• People use previous experience to solve the
  current problem
• a judge arguments with a similar case
• a technician remembers a similar fault of the
  same type of machine
• a mathematician tries to apply a similar proof
  for a new problem
• People share their experience to solve some
  (hard) problems
• A group of doctors makes diagnosis by exchanging
  their individual experience about symptoms of
  other patients

9
CBR-based intelligent application systems in a
nutshell

• store previous experience (cases) in memory
• to solve new problems
• retrieve similar experience (cases) about similar
  situations from the memory
• reuse the experience (cases) in the context of
  the new situation complete or partial reuse, or
  adapt according to differences
• to retain experience
• store new experience (cases) in memory (learning)

10
What is a Case

• several features describing a problem
• plus the solution or outcome
• cases can contain
• - text, numbers, symbols, plans, multimedia,
• cases are not distilled knowledge
• - unlike rule-base systems
• cases are records of real events
• and are excellent for justifying decisions

11
Two types of case features
metadata unindexed features Not predictive not
used for retrieval, they provide background
information to the users Provide valuable
contextual information and lessons learned
Drivinglicense 1024 Name John Doe Adress 12
Elm Street Photo Birthday 11.11.1955 Typ
A Issued at 12.12. 1975 Issued in Ohio Valid
until - Constraints glasses
data indexed features Predictive and used for
retrieval
Case vocabulary - the features that describe a
case
12
Case representation

• Representation depends on
• Requirements of domain and task
• Structure of available case data
• Flat feature-value list (like a database record)
  
• Simple case structure is sometimes sufficient
  for problem solving
• Easy to store and retrieve in a CBR system
• Suitable for shallow technical diagnosis,
  product recommendation, ...

Case 1 Name John Doe Address 12 Elm
Street Birthday 11.11.1955 Typ A Issued at
12.12.1975 ...
feature
value
13
5.1.2 CBR Scenario Example - Technical diagnosis

• Simple example Car Faults
• Symptoms are observed (e. g. engine doesnt
  start) and values are measured (e. g. battery
  voltage 6.3V)
• Goal Find the cause for the failure (e. g.
  battery empty) and repair strategy (e. g. charge
  battery)
• Case-Based Diagnosis
• A case describes a diagnostic situation and
  contains
• description of the symptoms
• description of the failure and the cause
• description of a repair strategy

14
CBR Scenario Example - Technical diagnosis
15
CBR Scenario Example - Technical diagnosis

• Each case describes one situation
• Cases are independent of each other

16
Solving a diagnostic problem
New problem

• Make several observations about new problem
• Not all features must be known
• The new problem is a case without the solution
  part

17
Solving a problem Computation of similarity
Compare the new problem with each case select
most similar case(s)

• Similarity is the most important concept in CBR
• Purpose of similarity is to select cases that
• - can be adapted easily to solve the current
  problem and/or
• - have (nearly) the same solution to the current
  problem
• Basic assumption
• similar problems have similar solutions

18
Similarity

• Degree of similarity utility or reusability
  of the solution
• Goal of similarity modelling/computation
• - provide a good approximation
• - easy to compute
• Different approaches
• nearest neighbour
• ML classifiers (ID3, C4.5)
• Statistical clustering
• fuzzy sets/logic

19
A similarity measure - Nearest Neighbour

• The most widely used technology in CBR
• - it is provided by the majority of CBR tools
• Algorithm
• 1. The similarity of the problem (T - target)
  case to a source case (S) in the case-library is
  determined by similarity function (f)
• for each (i) case attribute
• 2. This similarity measure may be multiplied by a
  weighting factor (w)
• 3. Then the sum of the similarity of all
  attributes (n -the number of attributes in each
  case) is calculated to provide a measure of the
  similarity of that case in the library to the
  target case
• This can be represented by the equation

20
Similarity Example Technical Diagnosis

• Similarity is assessed for each feature

• Similarity depends on the feature value
• Features can have different weights
  (importance)
• - High importance Problem, Battery_ voltage,
  State_ of_ light
• - Low importance Make, Model, Year, Colour

21
Compare new problem with case 1
Normalisation factor
22
Compare new problem with case 2
23
Reuse the solution of Case 1
24
Store the new experience
If the diagnosis is correct store new case in
the case base
25
5.1.3 CBR system - characteristics

• Complex knowledge acquisition is avoided
• Simple maintenance of the knowledge in the
  system
• High quality of retrieved solutions
• High efficiency by problem-solving
• Better reuse of existing data/knowledge
• User acceptance

26
CBR is Intuitive

• this is how we routinely make decisions
• experts rely on their experience
• novices use rules
• people learn by acquiring new cases
• experts decision making is dynamic

27
CBR is Simple

• easily understood by users
• this increases acceptance
• simple to implement
• all you need are cases
• and simple software
• consequently CBR is
• easy to sell to management
• easy to sell to users

28
Problem-solving or decision-making tasks of CBR
CBR
Analytic tasks
Synthetic tasks
forecasting
diagnosis
design
planning
classification
decision support
configuration
29
Analytic tasks

• Charakterisierung
• Schwerpunkt der Anwendung liegt in der Analyse
  einer vorliegenden Situation
• Situation muß hierzu i.d.R. einer Klasse
  zugeordnet werden.
• Anzahl der Klassen ist i.d.R. fest vorgegeben
• Je nach Art der Aufgabe kommen weitere
  Problemlöseschritte hinzu
• Fallbasierte Systeme für analytische Aufgaben
• Fälle typischerweise (Situation, Klasse)
• Schwerpunkt liegt beim Retrieval
• Lösungsanpassung ist oft nicht erforderlich
• Viele Anwendungen bereits im täglichen Einsatz

30
Synthetic tasks

• Charakterisierung
• Schwerpunkt der Anwendung liegt im Zusammensetzen
  einer komplexen Lösung aus einzelnen
  Bestandteilen
• Problembeschreibung ist i.d.R. eine Anforderung
  an die Lösung
• Lösungsraum ist i.d.R. unendlich groß
• Fallbasierte Systeme für synthetische Aufgaben
• Fälle typischerweise
• (Problem, Lösung) oder
• (Problem, Lösungsweg)
• Schwerpunkt liegt bei der Lösungsanpassung
• Häufig viel allgemeines (zusätzliches) Wissen
  erforderlich
• Viele Forschungsprototypen
• Entwicklungsaufwand größer als für analytische
  Aufgaben
• Anwendungen an der Schwelle zum täglichen Einsatz

31
When to Apply CBR?

• when a domain model is difficult or impossible to
  elicit
• when the system will require constant maintenance
• when records of previously successful solutions
  exist
• or when similar problems show up often

32
CBR system - applicability
Knowledge management
Knowledge-based systems
Cognitive science
CBR
Machine learning
Information retrieval
Neural networks
Database
Statistic
33
Who Uses CBR?

• American Express - credit card risk assessment
• Microsoft help desks
• Barclaycard - fraud watch
• General Electric train diagnostics, plastic
  fabrication
• British Airways plane maintenance
• Daimler Chrysler software support
• Analog component selection
• NASA space shuttle support
• Swiss Bank - investment management
• Deloitte Touche fraud detection

34
5.2 CBR KM - Where is the Knowledge in CBR
(I)

• knowledge about domain is in the case-base
• knowledge (problem-solving) in a CBR system is
  in 4 containers Richter 95
• - case representation
• - retrieval algorithm
• - similarity measures
• - adaptation methods
• knowledge (problem-solving) can be moved
  between containers

35
CBR Knowledge Management

• Framework for comparison (I)
• Goal is the same
• to capture and reuse experience or knowledge (to
  improve and support the overall business strategy
  of an organisation)

36
CBR Knowledge Management

• Framework for comparison (II)
• CBR vs. KM
• less emphasis on knowledge creation aspects
• cases are assumed to exist
• very simple knowledge representation
• feature-value pairs (suitable for databases)
• strong support for knowledge retrieval and
  adaptation
• contextual retrieval is supported
• suitable for processing noisy data
• similarity is the most important concept in CBR
• applicable only in domains in which similar
  problems occur
• modelling background knowledge in cases not
  possible

37
CBR Knowledge Management

• Framework for comparison (III)
• KM vs. CBR
• strong support for domain description (background
  knowledge)
• e.g. ontology-based KM
• (business) process-oriented
• strong support for a global management process in
  the organisation
• knowledge management covers also non-IT aspects
• more emphasis on knowledge creating aspects
• e.g. metadata generation

38
5.3 CBR Applications
Knowledge about a problem-domain is contained in
the documents Documents are cases. There are
three approaches to refer to the knowledge
contained in the cases - conversational -
textual - structural
39
CBR Applications

• Conversational approach Conversational CBR
• To each case is associated one (or more)
  questions-tree
• - questions-tree corresponds to case-features
• Characteristics
• User is navigated to the related solution
  (case, document) by the set of questions
• - question are organised in the form of decision
  tree
• Case-base has to be developed and maintained by
  expert
• The number of question should be optimal (not
  too much)
• Used for shallow diagnosis, product selection,
  planning.
• Main money earner for CBR

40
CBR Applications - Conversational CBR
Example HP customer care
1. Question
Problem description
http//h20015.www2.hp.com/en/siteHome.jhtmljsessi
onidN0L0ZM0KYXNB5QEXGR3UOSQ?regccuslcenpage
typesitehome
41
CBR Applications - Conversational CBR
Example HP customer care
2. Question
Case feature
42
CBR Applications - Conversational CBR
Example HP customer care
3. Question
Case feature
43
CBR Applications - Conversational CBR
Example HP customer care
Suggestion
44
CBR Applications - Conversational CBR
Example HP customer care
Feedback
45
CBR Applications - Conversational CBR

• Discussion (I)
• Used question-tree structure

46
CBR Applications - Conversational CBR

• Discussion (II)
• retrieve phase (CBR cycles) does not use any
  similarity measure
• reuse phase has no possibility for adaptation of
  retrieved solution
• revise phase is not supported
• - user could not explain why the solution is
  wrong
• - Similar system with user-feedback Philips
  Product Catalog Search http//www.homeandbody.phil
  ips.com /indexie.asp
• retain phase only in off-line mode (new
  documents new cases)
• Microsoft uses this style of CBR in Office
  Assistant (Decision-Theoretic CBR)

47
CBR Applications

• Textual approach
• Cases are represented as
• Documents short document description
• Problem description is in the form of free text
• Characteristics
• Case-base is created very easily
• Suitable in the case of a lot of documents (FAQ,
  manuals)
• Retrieval is based on string-comparison
• could be improved using Information Retrieval
  methods
• Quality is measured by
• recall the number of retrieved cases
• precision the number of relevant cases (for
  given problem)

48
CBR Applications textual approach
SIMATIC Knowledge Manager - customer support
system for industrial automation systems
Problem description
A CBR tool
relevance
www.ad.siemens.de/simatic-cs
49
CBR Applications textual approach
Questions to add more feature-value pairs gt
Better precision
50
CBR Applications textual approach
less results which are more relevant for the user
51
CBR Applications - textual approach

• Discussion
• retrieve phase (CBR cycles) is realized in
  Information Retrieval algorithm
• similarity measure has to be incorporated in that
  string-comparison
• reuse phase has no possibility for adaptation of
  retrieved solution
• revise phase is not supported
• - no feedback from user
• retain phase only in off-line mode (new
  documents new cases)

52
CBR Applications

• Structural approach
• Cases are represented as
• Documents ontology-based document description
  (ontology-based metadata)
• Problem description is in the form of
  ontology-based queries
• Characteristics
• Requires formally represented background
  knowledge about problem domain -
  domain ontology
• Case-base has to be developed and maintained by
  expert
• Retrieval is based on ontology-based querying
• - similarity could be defined (e.g. in hierarchy)

53
5.4 Conclusion

• Case-Based Reasoning (CBR) technology is
  increasingly employed for some processes in the
  Knowledge Management
• knowledge access
• contextual retrieval is supported
• users are able to articulate exactly what they
  want
• knowledge sharing it could be empowered by
  relevance-based retrieval, case reuse, and
  learning
• Knowledge Management opens itself as an
  application area for CBR of high current interest
• certain problem areas from KM have already been
  successfully addressed by using CBR technology,
  such as
• product experience bases,
• help desk systems, and
• user profiling and product recommendations for
  E-Commerce.

54
Information sources on the Web

• ai-cbr
• members mailing list, features, news,
  bibliography, software, etc
• www.ai-cbr.org (http//ai-cbr.cs.auckland.ac.nz/)
• the cbr web
• news, projects, publications, etc
• www.cbr-web.org