GDA and Case Based Reasoning

The most major issue in CBR (case-based reasoning) is retrieval of appropriate cases. There are several big problems that must be solved to realize such retrieval, which comprise the indexing problem.

First, a computer must be able to recognize that a case is applicable to the current problem. This is the similarity-assessment problem. Sometimes cases that fit best to the present situation do not look very similar to the problem on the surface. For example, chess and football are dissimilar in the concrete features of the games, but have many things in common in terms of situations and strategies such as a fork. Cases must hence be compared at more abstract levels of representation for retrieval.

To make this happen we must discover the right level of abstract ways for case representation --- indexing vocabulary. The CBR research community has been pursuing to find a generally applicable indexing vocabulary, a theory of the representational content of indices that crosses domains, such as the Universal Index Frame (UIF) (Schank and Osgood, 1990). Nevertheless, those endeavors have been limited to the particular problem domains, and have failed to span across diversity of domains.

Globally available tags and documents annotated with them can accelerate the research. By investigating the applicability of the tags that are originally developed for other purposes such as translation to case retrieval, we can easily evaluate the effective form of information for case-based problem solving. We can then probably transform annotated documents into another format more appropriate to case retrieval. If automatic transformation is too hard, we can propose more specific tags for CBR at the right level of abstraction.

For accurate case retrieval, we must also enable the computer to elaborate. Sometimes what is known about a current problem is in such a raw form for comparison that additional features of the problem need to be derived from it. This is known as the situation-assessment problem. For example, to predict a winner of a battle, the most informative feature is the ratio of defender strength and offender strength. Cases need to be retrieved based on the similarity of the ratio (a derived feature) rather than that of individual strength values (surface features). Such useful information is rarely available in raw cases but must be derived from the facts of the problem. The issue here is to come up with a way of elaborating situation descriptions and generating derived features for cases in an efficient way.

An interesting research direction regarding this problem is case-based tag derivation. Based on the accumulated cases of manually generating domain and task-specific tags from the descriptions annotated with generic tags, we might be able to induce some patterns or rules to transform information with the generic tags to that annotated with the derived tags in the problem solving situations at hand. Exploiting those induced patterns, if case descriptions annotated with (generic) tags are available, users can enjoy more accurate case retrieval without spending extra time in elaborating case features by themselves.

GDA Home Page