Aiming at problems existed in air-to-air missile attacking highly maneuvering target with variable structure guidance lawsuch as the divergence of line-of-sight angular rate and indetermination of variable structures parametersa new 3-dimensional guidance law was designed based on varying exponential reaching lawwhich could adjust the variable structure parameters on line by using multi-input multi-output RBF neural network (RBFNN).The simulation results show that the new guidance law can effectively reduce the buffeting and improve the guidance accuracy.It has strong adaptive capability and robustnessand can meet the tactical requirements for air-to-air missiles.