Experiments on Theory Refinement

Abstract

In this chapter, we apply the C-IL²P system to two problems of DNA classification, which have become benchmark data sets for testing the accuracy of machine learning systems. We compare the results obtained by different neural, symbolic and hybrid inductive learning systems. For example, the test-set performance of C-IL²P is at least as good as those of KBANN and Backpropagation, while C-IL²P’s training-set performance is considerably superior to KBANN and Backpropagation. We also apply C-IL²P to fault diagnosis, using a simplified version of a real power generator plant. In this application, we use the system extended with classical negation. We then compare C-IL²P with Backpropagation, using three different architectures. The results corroborate the importance of the background knowledge for learning in the presence of noisy data sets.