Improving the Performance of a Rule Induction System Using Genetic Algorithms
H. Vafaie, Kenneth Alan De Jong · 2001
Concept acquisition is a form of inductive learning that induces general descriptions of concepts from specific instances of a given concept. AQ15 is a conceptual inductive learning program that uses feature-based examples of concepts to generate rules which describe the underlying concept. This program has been successfully applied to a wide variety of domains. We have been exploring the use of AQ15 on a difficult class of image processing problems, namely, to infer descriptions of textures from noisy examples. In this context, we find that AQ15 produces rules that are sub-optimal from two view points: 1) we need to minimize the number of features actually used for classification; and 2) we need to achieve high recognition rates with noisy data. Our multistrategy approach is to apply genetic algorithms to select the best feature subset to use in conjunction with AQ15 to achieve our goals. The proposed approach has been implemented and applied to an initial set of randomly selected texture data. Our results are encouraging and indicate significant advantages to a multistrategy approach in this domain. 1.