Simple Ensemble of Extreme Learning Machine

In this paper, a novel approach for neural network ensemble called Simple Ensemble of Extreme Learning Machine (SE-ELM) is proposed. It is proved theoretically in this study that the generalization ability of an ensemble is determined by the diversity of its components´ output space. Therefore SE-ELM regards the diversity of components´ output space as a target during the training process. In the first phase, SE-ELM initializes each component with different input weights and analytically determines the output weights through generalized inverse operation of the hidden layer output matrices. The difference among components´ input weights forces those components to have different output space thus increasing the diversity of the ensemble. Experiments carried on four real world problems show that SE-ELM not only runs much faster but also presents better generalization performance than some classic ensemble algorithms.