Rule generation from a rotation-invariant neural pattern recognition system

A method of extracting rules from a rotation-invariant neural pattern recognition system formed using a genetic algorithm (GA) is presented. In particular, deterministic mutation (DM) is utilized to improve its convergence properties. It is performed on the basis of the result of neural network structure learning. DM can evolve chromosomes of individuals to increase their fitness functions in a deterministic manner. In this paper, coin data are used as inputs. The coins used are a Japanese 500-yen coin and a South Korean 500-won coin, which are very similar. GA is utilized to reduce the number of connection weights in the neural network. The network weights surviving after training represent rules to perform pattern classification for the coin data. The rules are then extracted from the network. Furthermore, the network has a procedure to substitute signum units for hidden sigmoid ones in examining its recognition accuracy. It enables us to easily extract rules. Simulation results show that this approach can generate a simple network structure and, as a result, simple rules for coin data classification