A learning procedure to estimate missing values in fuzzy preference relations based on additive consistency