based on the distance between the decision surface and training patterns lying on the boundary of the class cluster