On the adaptability of ensemble methods for distributed classification systems: A comparative analysis

In this work, a two-stage architecture is used to analyze the information collected from several sensors. The first stage makes classifications from partial information of the entire target (i.e. from different points of view or from different kind of measures) using a simple artificial neural network as a classifier. In addition, the second stage aggregates all the estimations given by the ensemble in order to obtain the final classification. Four different ensembles methods are compared in the second stage: artificial neural network, plurality majority, basic weighted majority, and stochastic weighted majority. However, not only reliability is an important factor but also adaptation is critical when the ensemble is working in changing environments. Therefore, the artificial neural network and the plurality majority algorithm are compared against our two proposed adaptive algorithms. Unlike artificial neural network, majority methods do not require previous training. The effects of improving the first stage and how the system behaves when different perturbations are presented have been measured. Results have been obtained from two applications: a realistic one and another simpler one, with more training examples for a more accurate comparison. These results show that artificial neural network is the most accurate proposal, whereas the most innovative proposed stochastic weighted voting is the most adaptive one.