Integration of ConstraintBased Reasoning and CaseBased Reasoning

1
Integration of Constraint-Based Reasoning and
Case-Based Reasoning

• Mohammed H. Sqalli
• Eugene C. Freuder
• University of New Hampshire
• msqalli,ecf_at_cs.unh.edu

2
Motivation

• CSP used for the adaptation process in CBR
• Solve a problem when a complete knowledge of the
  domain is difficult to get (Weigel et al. 1998)
• Achieve domain independence in adaptation (Purvis
  Pu 1995)
• Make solution space easier to explore (Smith
  Faltings 1995)
• CBR completes the CSP model (Purvis 1998, Torasso
  1998, Sqalli Freuder 1998)
• CBR corrects the CSP model (Sqalli Freuder 1998)

3
Taxonomy

• Branting 1998

Empirical (examplars)
Analytic (models)
Social system behavior
Natural system behavior
Artifact behavior
Law
Physics

• Sqalli Freuder 1998

Complex
Simple
Complete
Incomplete/Incorrect
Physical systems
Interoperability testing
Planning
Physics
4
Categorization of Modeling (Sqalli Freuder 1998)
5
Domains of Application

• Diagnosis
• Configuration
• Planning
• Design

6
Good experiences

• CADSYN (Maher Zhang 1991-93) design
  constraints are used for adaptation
• JULIA (Hinrichs 1992) a case-based meal planning
  system with a constraint propagator
• CADRE (Hua Faltings 1993) Constraints used to
  reduce the adaptation space
• COMPOSER (Pu Purvis 1995) solves problems
  using CSP for adaptation
• CHARADE (Avesini, Perini Ricci 1993-94)
  decision making in environmental emergencies
• IDIOM (Smith Faltings 1995) CSP for adaptation

7
Bad Experiences

• It is hard to find in the literature such
  experiences since published papers usually
  include the successes and not the failures
• There is one example showing that CSP/CBR
  integration may not be the best alternative
• Nutritional menus CSP/CBR may not be the best
  way of solving this problem, because of the
  monotony of the solutions it provides. A CBR/RBR
  system seems to be a more suitable for these
  kinds of applications (Marling, Petot Sterling
  1998)

8
Drawbacks

• Integration tend to be domain oriented and may be
  applied to a limited domain theory (CBR
  limitation)
• Overhead of switching from one reasoning method
  to the other
• Time and Space limitations of each reasoning mode

9
Advantages (CBR enhances CSP)CSPMaster,
CBRSlave

• CSP solving efficiency improved when starting
  from a case rather than from nothing
• Fill values of CSP problem (Purvis Pu 1995)
• Reduce search space (Huang Miles 1996)
• Solve large CSPs characterized by heavy searches
  (Huang Miles 1996)
• CBR learning component (Sqalli Freuder 1998)
• Update the CSP model. Effectiveness of the model
  increases (Sqalli Freuder 1998)
• CBR used to solve DCSP (Purvis 1998)

10
Advantages (CSP enhances CBR)CBRMaster,
CSPSlave

• CBR adaptation process formulated as CSP (Purvis
  Pu 1995)
• Constraint-Based Adaptation for compensating
  incomplete cases. Cross-checking cases with
  constraints (Lee et al. 1997)
• Add generic knowledge (CSP/ RBR) to cases
  (Bartsch-Sporl 1995)
• Constraint-Based retrieval (Bilgic Fox 1996)
• Exploit the concept of interchangeability in CSP
  (Weigel, Faltings, Torrens 1998)
• Reduce number of cases used (Sqalli Freuder
  1998)

11
Trade-off

• Overhead of using two modes of reasoning vs.
  limitations of each mode
• Integration CBR/CSP can have advantages or may
  add more work
• Adaptability criterion (Purvis 1998)
• Updating models not for all domains
• CBR/CSP Space vs. Time
• Balanced integration of CBR/CSP

12
What can we learn from other integrations ?

• Approximate Model-Based Adaptation. Cases
  compensate for incompleteness in the model.
  Models compensate for insufficient case coverage
  CARMA (Branting 1998)
• CBR contributes new links into the causality
  model (Karamouzis Feyock 1992)
• Best scenario MBR small number of cases
  (Torasso1998)
• CBR accounts for errors in the model ADAPtER
  (Portinale Torasso 1995)
• CBR used as a form of caching to speedup later
  problem solving (Van Someren et al. 1997)
• Unifying two modes (voting) better than combining
  them (Domingos 1998)

13
What did we learn ?

• CSP provides a domain-independent representation
  of a task (adaptation in CBR)
• CBR is useful for incomplete domains. Model is
  either difficult or impossible to get
• CSP provides a rich representation of a task
• CSP provides many advanced algorithms to deal
  with hard problems
• CBR provides a very useful learning component