Chapter 10: Indexing and Retrieval

1
Chapter 10 Indexing and Retrieval

• Next, we discuss the relationship between
  indexing and retrieval
• Weve seen three ways to implement indexes
• Label some parts of a case as being most
  predictive
• Place indexes on the internal nodes of a
  hierarchical memory structure
• Consider the importance weights in the nearest
  neighbor metric as indexes
• You start using indexes when you get a new
  problem to solve
• During situation assessment, you try to assign
  the new case the indexes it would have, if it
  were already stored

2
Assigning Indexes to a New Case

• The first step is context setting
• Here, you determine all the indexes you can by
  looking at the new case
• You may be able to use a checklist for this
• Certain features contained in the case may always
  be indexes
• Certain computations may always provide indexes
• Example Dividing number of home runs by number
  of at bats
• You may need heuristics to help find the indexes
• CASEY indexes on the diagnostic states of a
  patient, and that is not something already given
  in a new case. So, it infers possible diagnostic
  states and then retrieves cases with those
  diagnostic states to see which match best

3
Context Refinement

• Context setting may be sufficient
• Sometimes, you need context refinement to
  continue finding indexes during and/or after
  retrieval
• Context refinement can be seen in traversing the
  nodes of a discrimination network
• As you traverse the tree from the top, you
  compute the indexes needed to answer the
  questions at each node
• One advantage of this approach is that you dont
  waste time computing indexes in parts of the tree
  you dont traverse
• Context refinement may also be useful if the
  differences among cases you already have can
  suggest indexes to consider
• An example comes from diagnosing car problems

4
(No Transcript)
5
(No Transcript)
6
Redefining the Context

• Theres another time you may do situation
  assessment
• If the cases retrieved during context setting and
  context refinement dont solve the problem, you
  may need to redefine the context
• You have a failure, and there may be different
  reasons for it
• Perhaps a good solution exists, but you didnt
  find it because you didnt assess the situation
  properly the first time
• Perhaps there really isnt a close enough case in
  the case base
• Perhaps you found a close enough case but didnt
  adapt it correctly
• In the first situation, redefining the context
  can help

7
How Index Implementation Impacts Retrieval Methods

• If you implement indexes as questions or labels
  on the internal nodes of discrimination networks,
  then you retrieve cases by following paths in
  memory that the indexes point to
• If you implement indexes as labels in cases, you
  may use a two-step retrieval, first finding a
  subset of partially matching cases and then
  seeing which of these match best on the labeled
  features
• If you implement indexes as importance weights,
  you retrieve cases that score highest on the
  nearest neighbor metric
• Note that more than one kind of index may be used
  in a system
• If a system performs different tasks, cases that
  are useful for performing different tasks may use
  different indexing schemes
• Different index types may also be used in
  combination
• For example, a hierarchical memory scheme could
  return a subset of cases that are then ranked
  using nearest neighbor matching

8
Knowing When You Have a Good Enough Case

• Its important to know when you have a good
  enough case to solve the current problem
• We want to find the best case we have, but if the
  case base is small, or the problem is unusual,
  even the best case may not be good enough
• Example Consider the car repair case base. Say
  that instead of cases 2 and 3, the closest match
  was for the same kind of car with similar mileage
  that had a flat tire. This wouldnt help with an
  engine not starting at all. Its much better for
  the system to say, I dont know, than to give a
  ridiculous answer like, Change the tire and then
  see if the engine starts.
• When using the nearest neighbor approach, it
  often helps to set a numeric threshhold on the
  similarity score. If no case exceeds that score,
  then no solution is suggested. There isnt a
  magic cutoff score, but you can experiment to see
  how high it has to be for your own system to give
  sensible answers

9
Efficiency, Accuracy and Flexibility

• Efficiency refers to how fast we can retrieve
  cases
• Accuracy refers to always finding the best case
  we have
• Flexibility refers to being able to use cases in
  ways you hadnt originally planned for them to be
  used
• These are all good things, but they can conflict
  with each other, so sometimes you need to make
  tradeoffs among them
• For example, to achieve flexibility, you use a
  large number of descriptive indexes. Using more
  indexes makes retrieval take longer, because you
  may consider more cases and consider each one at
  greater length
• When you design a new system, you can make
  choices about indexing and retrieval based on
  your priorities for efficiency, accuracy and
  flexibility