Case-Based Reasoning - Advanced Artificial Intelligence - page 186

cases retrieved influence the quality of the current problem solving. The
process can be divided into three phases: feature identification, tentative
matching, and final selection.
Feature Identification is to identify the relevant features on the analysis of
the problem. The features can be obtained by the following method: (1)
extract the features of the question directly from its description. For example,
the system can extract some keywords from the problem description that is in
natural languages. The extracted keywords can be viewed as features of the
problem. (2) elicit the features after a understanding of the problem, such as
the feature elicitation in image analyzing and processing. (3) obtain the
feature from users by the human-machine interaction according to the
requirements of context or knowledge model. The system utilizes users'
answers heuristics to constrain and direct the search and make the retrieved
case more accurate.
Tentative Matching means to find a group of candidate from the case
base to the current problem. It relies on the indices of the features fixed in
Feature Identification. Retrieval of cases from the case base must be equipped
with the ability to perform partial matches and construct a partial order on the
similarities among cases, since in general there is no existing case that exactly
matches the new case. The similarities can be computed in terms of syntax
structures without the utilization of domain knowledge, or be estimated on the
basis of a deep analysis and thorough understand. Some well-known concrete
methods are: nearest neighbor, induction, knowledge guided induction and
template retrieval; we can also give weights to the features reflecting their
importance.
Final Selection is to choose one or several cases with the most similarity
to the target case from the tentative matches. It is closely related to the
domain knowledge. First, explanations can be made by the knowledge
engineers, or be computed based on the knowledge model. Then the system
evaluates the explanations, and arranges the candidates into a queue according
to some criterion. The one that receives the highest rating becomes the best
match, for example the most relevant one or the one with the most rational
explanations.
A normal course of case retrieval is sketched by the left picture of Fig.
5.4. The input is the target case, i.e., the present case. It comprises the current
scene and reasoning goal. The scene is refined after an analysis on it. If there
are cases similar to the target case in the case base, then the indices of the