This paper presents a neural multi-step predictive control strategy for on-line ship control application. The strategy adopts both the neural network information processing mode and the predictive control mechanism, and can on-line control a nonlinear dynamical system whose exact mathematical model is not available. The multi-step predictor is performed by a variable radial basis function network (RBFN), which is realized by the proposed minimal regularized orthogonal least squares (MROLS) algorithm for growing or pruning basis functions according to designed strategy. With a parsimonious structure of variable RBFN, the strategy enables on-line control of a nonlinear dynamical ship model under random disturbances and measurement noises. Simulation results demonstrate the strong adaptive ability and effectiveness of the proposed strategy