Rule Extraction from Artiicial Neural Networks Trained on Elementary Number Classiication Tasks

Cascade-Correlation and BpTower networks are trained on pattern classiication tasks. The digits of four-digit integer numbers are sparsley coded and neural networks are trained to recognise the digit patterns of numbers divisible by ve, four and three. The performance of the decompositional rule extraction technique LAP is compared with that of the pedagogical technique RuleVI when extracting rules from trained artiicial neural networks. It turns out that the quality of the extracted rules depends on the combination of the neural network and the rule extraction method. The results are compared with results obtained by C4.5. Finally SHRUTI is applied to one of the rule sets.