We present a novel approach to classification, based on a tight coupling of instancebased learning and a genetic algorithm. In contrast to the usual instance-based learning setting, we do not rely on (parts of) the given training set as the basis of a nearestneighbor classifier, but we try to employ artificially generated instances as concept prototypes. The extremely hard problem of finding an appropriate set of concept prototypes is tackled by a genetic search procedure with the classification accuracy on the given training set as evaluation criterion for the genetic fitness measure. Experiments with artificial datasets show that - due to the ability to find concise and accurate concept descriptions that contain few, but typical instances - this classification approach is considerably robust against noise, untypical training instances and irrelevant attributes. These favorable (theoretical) properties are corroborated using a number of hard real-world classification problems.