Rule Extraction from Dynamic Cell Structure Neural Networks by Modified LREX Algorithm-A Comparative Study
Mansour Sheikhan, Amir Hossein Khalili · 2009
Abstract: The semantic of neural networks is not explicit and they are considered as black box systems. There are many researches investigating the area of rule extraction by neural networks. In this paper, the eclectic approach of rule extraction from a dynamic cell structure (DCS) neural network is investigated. To do this, a modified version of LERX algorithm is used for rule generation. Empirical results show that the DCS performs better than other self-organizing maps, e.g. Kohonen and competitive neural networks in generation of effective rules. Accuracy of classification, separability of Voronoi regions and mean squared error (MSE) in regions are the measures in which DCS performs better than other networks. Key words: Rule extraction • neural networks • dynamic cell structure