Co-operative Training in Classifier Ensembles

As the possibilities of combining experts become a more important direction in intelligent systems, difficulties arise in ways of generating these various experts and how to effectively use them concurrently. This is very evident in designing multiple classifier systems. The degree and method by which multiple classifier systems share training resources among their components can be a measure of co-operation. Training resources that are sharable in a multiple classifier system are training patterns, algorithms or information. In this paper we present co-operative training as a means of sharing training information amongst an ensemble during training. Improved classification accuracy demonstrates that sharing, or co-operation, amongst classifiers during their training is useful in a multiple classifier system.