An evolutionary approach to rule generation from neural networks

A method of extracting rules from neural networks formed using an evolutionary algorithm is presented. The evolutionary algorithm used here is a random optimization method (ROM). In particular, deterministic mutation (DM) is introduced in ROM. It is performed on the basis of the result of neural network learning. The DM procedure can evolve a candidate of a solution to increase a ROM fitness function in a deterministic manner. In the paper iris data are used as inputs. ROM are utilized to reduce the number of connection weights in the neural network. The network weights survived after the ROM training represent rules to perform pattern classification for the iris data. The rules are then extracted from the networks in which hidden units use signum and sigmoid functions to produce binary outputs. It enables us to extract simple logical functions from the network. Simulation results show this approach can generate a simple network structure and as a result simple rules.