Study of Methods for Combining Classifiers within the Theory of Evidence

This paper describes a research the effectiveness of the use of rules for combining evidence for the task of aggregating classifiers. The theory of evidence (Dempster-Shafer theory) has powerful tools for combining evidence, taking into account their conflict, reliability, etc. This research considers the information on the training of the particular classifier as evidence. This information is aggregated using the rules for combining evidence. The parameters of aggregation are the discounting factors, which are taken into account as the effectiveness of individual classifiers. This classifiers fusion model was tested on artificial and real data. The results of the performance of such classifier combining was compared with other known combining techniques and singleton classifiers. Some modifications of the model are considered. An effective way of combining classifiers is implemented, using the method of discounting the evidence and improving the generalizing ability.