Fuzzy Adaptation for Particle Swarm Optimization for Modular Neural Networks Applied to Iris Recognition

In this paper a new Modular Neural Network (MNN) optimization is proposed, where a particle swarm optimization with a fuzzy dynamic parameter adaptation designs optimal MNNs architectures. This design consists in to find the number of hidden layers for each sub module with their respective number of neurons, learning method, error goal and the percentage of data used for the training phase. The proposed method is applied to pattern recognition based on the iris biometrics and has as objective function to minimize the error of recognition. The proposed fuzzy adaptation seeks to avoid stagnation of error of recognition during iterations updating some PSO parameters such as w, C-1 and C-2.