In rough set theory, not too much work pays attention to the acquisition of decision rules and to the uses of the obtained rule set as classifier to predict data. In fact, rough set theory also can be applied to train data and create classifiers and then complete data prediction. This paper systematically studies the problem of acquisition of decision rules in decision systems. The main outcomes of this research are as follows: (1) the specific definition of minimum rule set is given, and such a minimum rule set can be used as a classifier to predict new data; (2) a new approach to finding out all minimum rule sets for a decision system, Algorithm 1, is proposed based on discrimination function, but with relatively low execution efficiency; (3) By improving Algorithm 1, a heuristic approach to computing a special minimum rule set, Algorithm 3, is proposed, which works far more efficiently than Algorithm 1. The outcomes can form the foundation for applying rough set theory to data classification and offer a new resolution to data classification.