Maximizing the discriminating power of a multiple-score test