A Research on the Relation Between Training Ambiguity and Generalization Capability

The classification result of an example matching to fuzzy IF-THEN rules is usually a possibility distribution, which can be measured by the ambiguity. This paper attempts to find the relation between the ambiguity on training set and the testing accuracy (which is usually called the generalization capability) and tries to give a new criterion to evaluate the generalization capability of fuzzy decision trees. Suppose that we first make use of the fuzzy decision tree to generate a set of fuzzy IF-THEN rules and then pay particular attention to the training ambiguity by matching training examples and testing examples to the generated IF-THEN rules. Our experiments show an interesting result, that is, with the precondition that the training accuracy does not decrease, the higher the ambiguity of the training set is, the higher the testing accuracy is. Some explanations and speculation about this experimental result are given