To meet the requirement of intelligent prosthetic leg control to fast follow the knee torque in different walking modes,a new cerebellar model articulation controller(CMAC)with variable struc- ture was designed.The structure of CMAC can be self-organized to a relatively stable status by utiliz- ing RRC learning algorithm to sufficiently search to connect the inputs of CMAC.The simulation re- sult indicates that the input connections in this status are unevenly distributed and follow power-law distribution so as to play a role of optimizing CMAC structure.The designed CMAC can significantly improve the quality of gait-following for intelligent prosthetic legs.