Competitive Learning Neural Networks For Rules Extraction

Clusters in the input-output problem space represent "patches" of data, which can be represented asrules. Neurons in competitive learning neural networks learn to represent centers of clusters. A weight vector w_jmay be viewed as a geometrical center of a cluster of data.

These characteristic of competitive learning neural networks can be used for the purpose of rules extraction, and finding fuzzy rules in particular. Figureshows a two-dimensional input output space for learning rules of the form of IF X is A, THEN Y is B, where A and B can be defined either as intervals for extracting interval rules or as fuzzylabels for extracting fuzzy rules. This clustering can be achieved in a competitive learning neuralnetwork.

The main requirement of learning rules through clustering in competitive learning algorithms is that the training data set should include a significant number of samples.

There are some characteristics of this type of learning rules:-

The set of extracted rules may only partially cover the whole input-output space.
If fuzzy quantization of the input and the output variables is used, then the set of fuzzy rules may overlap, that is, one input data vector may be covered by several fuzzy rules.
The number of rules can be controlled by a threshold of significance, being proportional to the number of data elements in a cluster, so only significant rules may be extracted.

After arough set of fuzzy rules is extracted, further refinement of these rules may be achieved by.

Using more fuzzy labels to quantize the input and output variables
Using another inference method one more suitable for the actual application task
Using a more precise form of rules by adding coefficients of importance and other parametersforrepresenting the uncertainty in the data set
Inserting the initial rules into a connectionist structure, with further training and consecutive rules extraction.

The approach presented here uses real data for training a competitive learning neural network. Fuzzyrules were then extracted based on fuzzy quantization.

SKEDSOFT