Introducing Reordering Algorithms to Classic Well-Known Ensembles to Improve Their Performance

Most of the well-known ensemble techniques use the same training algorithm and the same sequence of patterns from the learning set to adapt the trainable parameters (weights) of the neural networks in the ensemble. In this paper, we propose to replace the traditional training algorithm in which the sequence of patterns is kept unchanged during learning. With the new algorithms we want to add diversity to the ensemble and increase its accuracy by altering the sequence of patterns for each concrete network. Two new training set reordering strategies are proposed: Static reordering and Dynamic reordering. The new algorithms have been successfully tested with six different ensemble methods and the results show that reordering is a good alternative to traditional training.