Ensembles generation; Neural networks; One-cycle adaptation; Weak learners diversity; Weak learners generation

In the paper we describe an experimental study which was aimed to explore the possibility of using neural network as the base algorithms for so called weak classifiers. We were interested in the possibility of using neural networks in place of commonly used decision trees. Our analysis is based on the idea that it is more efficient to create a number of imperfectly adapted networks small in the topology than one perfectly adapted a sophisticated network. Moreover we assume, that such simple networks will still be more accurate than decision trees. In our experiment neural networks went only through one adaptation cycle. We have generated experimental ensembles of 100 weak classifiers based on five types of neural networks. A total of 3,240 of such ensembles have been created and tested. We have proposed filtering of input as the new diversity-achieving method. We have tested this method in common with two others methods. The experiment has been conducted over the MNIST database [1].