The variance of the calculation is minimized on the basis of parameters generated by a learning technique. The optimum is obtained if sampling is biased proportionally to the expected root-mean-square score. The method is compared with existing methods, which bias proportionally to the expected score.