Rough sets proposed by professor Pawlak in 1982 is an important tool to process the uncertainty of a set’s boundary, and it describes the uncertainty of set X (or concept) with two crisp boundaries that are upperapproximation set and lower-approximation set of X. However, a rough set does not give out the method for precisely, or approximately describe the uncertain set X (or concept) with existing knowledge base. In this paper, the similaritybetween two sets is proposed at first, the disadvantages of using upper-approximation set R(X) or lower- approximation set R(X) as an approximation set of the uncertain set X (or concept) are analyzed, and then amethod for building an approximation set of the uncertain set X is presented, the conclusion that the set R_0.5(X) is the optimal approximation set is proved. Finally, the changing regularities of similarity between R_0.5(X) and X with the change of knowledge granulatity in knowledge space are disscussed in detail. From the new viewpoint, this paper presents a new method for building an approximation set of the uncertain set X, and it will promote the development of rough set model.