|
Abstract : |
Most inductive learning systems generate complete and consistent descriptions. In order to achieve completeness and consistency in the presence of noise or iraprecision, one may generate overly complex and detailed descriptions. Such descriptions, however, may not perform well in future cases and suffer the disadvantage of excessive complexity. This is the well known phenomenon of overfitting. In this paper, a rule optimization method called SG-TRUNC is described and evaluated experimentally. SG-TRUNC improves previous TRUNC methods and has been implemented in a more efficient way. In the method, an optimized description is obtained through a sequence of generalization and/or specialization operations performed on a complete and consistent concept description. The operations applied always simplify a description. The method has been implemented in AQ16 that has been applied ' to two domains: a designed testing problem and "'multiplexer " Fll. The Exper/mental results have shown that both simplicity and performance improvements can be gained in the domains where noise is present., |
