A new set membership identification algorithm was proposed for the robust identification problem of nonlinear dynamic systems with unknown but bounded noises. Radial basis function (RBF) networks were used to approximate unknown nonlinear dynamic systems utilizing their approximation ability according to input and output data of systems. The weights of the RBF network of the unknown nonlinear dynamic system were estimated using a linear-in-parameter set membership identification algorithm considering that the RBF network was a linear-in-parameter model and the modeling errors and system noises were bounded after the centers of the RBF network were determined. Since the result of the estimation was a set of the weights of the RBF network, it could be easily used to predict the interval of the practical system output. Simulation shows that the set membership algorithm is less affected by the distribution of the noises of the unknown nonlinear dynamic system than the least squares algorithm.