A modified artificial fish swarm algorithm for the optimization of extreme learning machines

Abstract

Neural networks have been largely applied into many real world pattern classification problems. During the training phase, every neural network can suffer from generalization loss caused by overfitting, thereby the process of learning is highly biased. For this work we use Extreme Learning Machine which is an algorithm for training single hidden layer neural networks, and propose a novel swarm-based method for optimizing its weights and improving generalization performance. The algorithm presents the basic Artificial Fish Swarm Algorithm (AFSA) and some features from Differential Evolution (Crossover and Mutation) to improve the quality of the solutions during the search process. The results of the simulations demonstrated good generalization capacity from the best individuals obtained in the training phase. © 2012 Springer-Verlag.