A rule merging technique for handling noise in inductive learning

Abstract

Inductive learning algorithms are used for extracting IF-THEN rules from examples. The main weakness of most existing algorithms is their poor ability to handle data containing noise. This problem is even more severe when inductive learning techniques are applied to real engineering data. The paper presents a new pruning technique that improves significantly the performance of the RULES family of inductive learning algorithms. The technique is designed for RULES-5, the latest algorithm in the family, but could readily be applied to rule sets created by other algorithms.