Extraction of semantic rules from trained multilayer neural networks

Methods to extract logical rules from trained multilayer neural networks are described. A general and theoretical method of logical rules extraction is proposed. To each logical unit of a trained neural network, a truth table and a logical rule are associated. Before extracting rules, it is necessary to make certain that the operation done by the unit is a logical one. Two approaches fulfilling this criterion are proposed. The first consists of forcing the outputs of the units to be in {1,1}. An algorithm based on backpropagation (BP) is proposed. The second approach consists of forcing each unit to realize only an elementary logical operation, namely, AND, OR or NOT, but not a combination of them. Three algorithms are proposed in this case. Simulation results for these methods are reported