Literal and ProRulext: algorithms for rule extraction of ANNs

摘要

Artificial neural networks (ANN) present excellent capacity for generalization. Besides, they are applied to the most diverse human knowledge domains. However, since they represent knowledge in its topology, weight values and bias, explaining clearly how an ANN has obtained its outputs is not a trivial task for human experts. Usually such deficiency can be minimized through the "if/then" rule extraction from the trained network. Thus, this work presents two algorithms for the propositional rule extraction from trained ANNs: literal and ProRulext. Among other advantages, these methods can be applied to trained networks for pattern classification and time series forecast, obtaining rules that are compact, comprehensible and faithful to the networks from which they have been extracted, also at a lower computational cost compared to NeuroLinear.