Ensembles of evolutionary Extreme Learning Machines through differential evolution and Fitness Sharing

Extreme Learning Machine (ELM) is a single-hidden-layer feedforward neural network which has been applied into many real world pattern classification problems. Recently, ELMs have been built in an automatic way through evolutionary algorithms. Most works, nonetheless, do not uses all population obtained, but choose only one individual in the last generation. In an attempt to improve performance, an ensemble is a more promising choice because a pool of classifiers might produce higher accuracy than merely using the information from only one classifier among them. One of the most important factors for optimum accuracy is the diversity of the classifier pool. In this work, an enhanced Differential Evolution incorporating sharing function method is used to generate a pool of ELMs. Fitness Sharing that shares resources if the distance between the individuals is smaller than the sharing radius is a representative specification method, which produces diverse results than standard evolutionary algorithms that converge to only one solution. Experimental results on 14 well known benchmark classification tasks suggest that our method can generate ensembles that are more effective than ensembles solely through DE and traditional ensemble methods.