A merge-based condensing strategy for multiple prototype classifiers