Increasing the number of classifiers: Experimental investigation on multi-classifier systems

In this paper the performance of multi-classifier systems, as the number of individual classifiers increases, is investigated Two approaches are considered The first one infers the expected performance of the multi-classifier system on the basis of the degree of similarity among the individual classifiers. The second one uses different sets of real classifiers to evaluate experimentally the performance of the multi-classifier system. The result demonstrates that the analysis of similarity among the individual classifiers provides useful information on the complex behaviour of multi-classifier systems.