Effective learning in noisy environment using neural network ensemble

We have previously (1999) proposed a model of a neural network ensemble composed of a number of multi layer perceptrons (MLP). The ensemble is trained so that each member has a unique expertise. It is also provided with a mechanism to automatically select the most relevant member with respect to the given environment, enabling the ensemble to adapt effectively in changing environment. In this research we trained the ensemble with a noisy training data set, which is a training set that contains a particular percentage of contradictionary (false) data. Based on the members´ expertise the ensemble has the ability to distinguish contradictionary data and treat such a kind of data set as one unique environment that differs from the clean environment formed by correct data. In the training process the ensemble will automatically select one of its member to be trained in the clean environment and switch to another member whenever a contradictionary data is given, resulting that one of the ensemble member will be successfully adapting the clean environment