CBR: Part 2

1
CBR Part 2

• Detailed CBR Process

2
Introduction

• CBR is a machine learner that solves a new
  problem by looking at similar solved problems.
• Adapting their solutions instead of generating a
  new solution from scratch.
• Advantages of CBR systems
• They can propose quick solutions by reusing prior
  solutions.
• They can solve problems in domains that are not
  completely understood.
• They can evaluate solutions when no algorithmic
  method is available for evaluation.
• They may focus its reasoning on important parts
  of the problem.
• They can be warned of potential problems that
  have occurred in the past.

3
The CBR Process

• It consists of the following steps
• Solve a new problem or new case.
• Extract problem features.
• Retrieve similar old cases to the new arrived
  problem based on a similarity metric.
• Determine the best matches of the retrieved cases
  to the new problem.
• Extract the differences between the new and best
  matching cases.
• For each difference, make adjustments to the case
  solution using rules or other algorithms.
• Check the proposed solution using rules and fine
  tune it.
• Evaluate it based on external feedback.
• Store the new case and its quality

4
CBR Process Diagram
5
Representing and Indexing Cases

• What is a case?
• A case is a contextualized piece of knowledge
  representing an experience that teaches a lesson
  fundamental to achieving the goals of the
  reasoner
• A case represents specific knowledge tied to a
  context
• Two major functional parts of a case
• - Its content ( the lesson(s) it teaches )
• - Its context ( the context in which it can teach
  its lesson(s) and described by its indexes )

6
Case Content

• Three major parts to the content of any case
• Problem/situation description
• Description of the problem or situation being
  understood and solved.
• Any environment features that has some effect on
  the new solution.
• Solution
• Store the new solution and a description of how a
  solution was derived (trace).
• It may include an explanation of each part of the
  solution.
• Outcome
• Record the results of applying the new solution
  (success/failure).
• If possible, store how and why it succeeded or
  failed.
• Evaluate the effectiveness of repeating or
  adapting an old solution to a new problem.

7
Indexing

• Indexes are combinations of important descriptors
  of a case.
• Those than can distinguish it from other cases.
• Indexes are assigned to new cases added to the
  case-base.
• In CBR we are not interested in balancing the
  index structure.
• We are more interested that indexes can
  distinguish cases from each other.

8
The Retrieval Process

• Retrieving cases and adapting solutions are the
  main problems in CBR.
• Less number of cases is needed in the case-base
  when adaptation works efficiently.
• During retrieval, CBR does situation assessment
  and uses matching and retrieval algorithms.
• Situation assessment identifies features in the
  same index vocabulary as the cases.
• Matching algorithms know how to
• compare 2 items and score their degree of match,
  and
• distinguish which indexed features are more
  important to focus on.
• Retrieval algorithms know how to search the case
  library
• Matching algorithms are applied to cases
  identified by retrieval algorithms.

9
Retrieval process (cont.)

• Suppose the N features F1, F2, .. FN, were
  extracted from a new problem (case).
• Let f1, f2, fN be the values for these features.
• Let c1, c2, , cN be the feature values for a
  case.
• One measure of similarity between a new and old
  case would be (a k-NN-like measure)
• Where wi are weights (that sum to 1) over the N
  features, and ß is a scaling parameter to avoid
  each term to exceed 1.
• Perfect match results in s equal to 1. Otherwise,
  s varies between 0 and 1.

10
Retrieval process (cont.)

• Updates to the case library uses similar
  algorithms as those used for retrieval.
• It starts with situation assessment.
• It looks for a place to insert the case instead
  of a place to retrieve a similar case.
• It inserts the case and reorganizes the index
  structure as necessary.
• Adaptation
• No old case is exactly the same as a new case.
• New things may be added to an old solution.
• Something may be deleted from it.
• Something may be replaced or transformed from it.

11
Forms of Adaptation

• Null adaptation Use solution as is.
• It happens when a large case-base exist.
• User adaptation User selects the modifications.
• It uses the user knowledge rather than using a
  knowledge base.
• Structural adaptation Modify retrieved case with
  some knowledge base.
• Adjust parameters for different operational
  environments.
• Substitute steps in the retrieved case using
  substitution rules.
• Apply other (ad hoc) methods to the retrieved
  case.

12
Adaptation methods (Ten)

• Substitution methods
• Reinstantiation (substitute old objects with new
  ones)
• Parameter adjustment (adjust parameters of an old
  solution).
• It relies on specialized heuristics.
• Local search (using auxiliary knowledge)
• Replace inappropriate objects for new situation.
• Query memory (search case memory)
• Replace objects satisfying some description.
• Specialized search (search the case or auxiliary
  memory)
• Using specialized search heuristics.
• Case-based substitution (use other cases to
  suggest substitutions).

13
Adaptation methods (cont.)

• Transformation methods
• transform an old solution that will work for the
  new situation.
• Two transformation methods
• Common-sense transformation
• Use of common-sense heuristics to delete,
  replace, or add components to a new solution.
• Model-guided repair (transformation guided by a
  causal knowledge)

14
Adaptation methods (cont.)

• Special-purpose adaptation and repair
• It carries out domain-specific and
  structure-modifying adaptations not covered by
  other methods.
• Heuristics that eliminate redundancies and repair
  faulty solutions.
• Heuristics implemented as critics using a
  rule-based production system.
• Derivational repair
• It reuses the methods for deriving a solution or
  solution piece.

15
Learning

• Effective learning can happen as a byproduct of
  the reasoning and memory update.
• A reasoner becomes more efficient over time.
• A reasoner learns as it reasons.
• The steps required to solve a problem is learning
  that is also stored.
• Inductive formation of generalization is one way
  of learning. Other ways are
• Through the accumulation of new cases, and
• Through the assignment and unassignment of
  indexes.

16
Learning (cont.)

• Generalization is formed when
• several cases are classified the same way, or
• all predict the same solution.
• A CBR system can be enhanced with learning
  capabilities
• By controlling which cases to store so that it
  makes retrieval more efficient.
• By learning which solutions or elaborations are
  costly to carry out.
• By using cases to enhance general knowledge.

17
Building CBRs

• Three implementations of CBRs
• Autonomous systems
• They do retrieval, adaptation, and evaluation
  steps by themselves.
• Human-machine systems
• Adaptation and evaluation usually performed by
  human
• Embedded systems
• A CBR system embedded in a larger system (e.g., a
  training system).