On using prototype reduction schemes to optimize dissimilarity-based classification