Case Based Reasoning

Case-Based Reasoning

Case-Based Reasoning uses the first two principles of Instance Based Learning:

  1. Lazy Learning
  2. Classify new instances based on those similar to them

But not the third
3. CBR doesn't represent instances as real-valued points in an n-dimensional Euclidean space.

In CBR, instances are typically represented using more rich symbolic descriptions, and the methods used to retrieve similar instances are correspondingly more elaborate.

For instance, we might represent an instance as the following (learner to create new mechanical designs):

Screen%20Shot%202012-04-15%20at%2010.32.20%20PM.png

As such they're more 'cases' than instances - quite different from each other, but with the same purpose.

CBR Cycle

  • RETRIEVE the most similar case or cases
  • REUSE the information and knowledge useful to solve the problem from the case
  • REVISE the proposed solution
  • RETAIN the parts of this experience likely to be useful for future problem solving

Similarity

The similarity metric different for Cases is quite different than the Euclidean distance for instances, due to their rich symbolic nature.

For instance we can use a match and mismatch function (counting the number of variables that match or don't as a proportion) to come up with something like:

(1)
\begin{align} score(Q,C) = \frac{match(Q,C) - mismatch(Q,C)}{|C|} \end{align}

Q is the query task, C is the case base task

Case Retrieval and Problem Solving

Multiple retrieved cases may be combined to form the solution to the new problem. This is similar to the k-NN approach, in that multiple similar cases are used to construct a response for the new query. However, the process for combining them is different - relying on knowledge-based reasoning rather than statistical methods.

There's a tight coupling between case retrieval and problem solving, so finding the instance is linked to solving the problem.

  • Functional qualitative relationships can be extended by rewrite rules to explore space of “related” specifications
  • May suggest novel combinations of structures for the design