Information Technology Reference
In-Depth Information
As time goes by, the case base becomes larger and larger. This will cause
great waste of the memory space and retrieval time. In this light, the system
should perform effective organization and management on the case base.
Retrieval, reuse, revise and retainion are four main processes in a CBR
cycle, and the reasoning process of CBR is also named as Four- R process.
5.10 Instance-Based Learning
Instance-based learning (IBL for short), is a family of inductive learning
methods closely linked to case-based learning (Aha,1991). The learning
algorithms in IBL simply store the already classified instances, and when a
new query instance is encountered, classify the instance into the same class as
the most similar related instance retrieved from the instance base. Rather than
complex index mechanisms, IBL uses the feature-value pairs as the primary
representation method. This approach also does no revisions on instances, yet
it is proved to be very useful.
In the researches of learning from examples or supervised learning, a
variety range of concept representation formalisms have been put forward,
including rules, decision tree, connectionist networks, etc. All these methods
predict the new instance based on the abstraction and generalization of the
training instances. Instance-based learning can be viewed as an extension of
the nearest neighbor method in that it uses the typical instances to denote the
corresponding concepts directly, rather than generalizing a set of abstractions
of the training instances. Its prediction about the new query is estimated under
the similarity assumption, i.e., the classification results on similar instances
are similar too. IBL retrieves a set of instances similar to the new instance,
and returns a result for the new query based on a systemic analysis of the
retrieved results. The nearest neighbor learning is incremental, and gains the
best prediction accuracy for instances whose feature values are continuous,
compared with other learning methods (Biberman,1994).
The k-nearest neighbor algorithm (k-NN for short) is the general form of
nearest neighbor method, where k is the number of most nearest neighbors.
Two key points in the application of k-NN are: How to retrieve some
instances similar to the instance form the instance base; and how to assess the
retrieved results to form the prediction value for the present example. The
former includes how to define the similarities among instances and the criteria
