Abstract: In this paper, for the case that the state variables can be separated from the control variables in a nonlinear dynamic system, an improved regression estimation algorithm based on SVMs is presented and then applied to nonlinear system identification. This kind of learning machines includes two nonlinear functions whose variables are the state and the control ones, respectively, and they are used to identify the two nonlinear functions in the separable-variable nonlinear dynamic system. The simulation results validate the efficiency of our method.