Abstract: In this paper, two kinds of learning methods of fuzzy systems are analyzed first,using genetic algorithm and the gradient descent method. They have completeness of membership functions and fuzzy rules and other problems, such as the damage of the shapes of membership functions. The damage of completeness of membership functions leads to no useful rules which are available when some data are inputted. Then, a method that guarantees the ε\|completeness of membership functions and consistence of fuzzy sets semantics is proposed. Moreover, a new fast learning method of fuzzy systems both are based on genetic algorithms and gradient descent method is proposed. Some experiments are also made and the simulation result is presented to show the high effectiveness and some other advantages of guaranteeing the completeness of membership functions and consistence of fuzzy sets semantics.