Rule extraction from neural networks trained using evolutionary algorithms with deterministic mutation

Abstract

A method of extracting rules from neural networks trained using evolutionary algorithms (EAs) is presented. The EAs used are a genetic algorithm (GA) with deterministic mutation (DM) and a random optimization method (ROM) with DM. The DM is performed on the basis of the result of neural network learning. It can evolve chromosomes of individuals to increase their fitness functions in a deterministic manner. The EAs are utilized to reduce the number of neural network connections. The network connections surviving after training represent rules to perform pattern classification. The rules are then extracted from the network in which hidden units use signum functions to produce binary outputs. Simulation results show this method can generate a simple network structure and as a result simple rules for the iris data classification.