Usually fuzzy control rules are generated by translating the operator's experience into fuzzy linguistic form directly. But this paper presents one learning method which generates fuzzy control rules by numeric Input/Output data. The membership function which is described by several values and the consequent part of inference rules are tuned by means of descent method. The learing speed and the learing capacity of this method are higher than conventional learing method by descent method. And this method has a capability to express the knowledge acquired from Input-Output data in form of fuzzy inference rules. The simulation shows the proposed fuzzy controller can learn nonlinear functions.