Key points are not available for this paper at this time.
A neural network is constructed to represent the input-output relation of a dynamical model. The parameters are calculated by means of a second-order training algorithm. Then, a nonlinear predictive controller is designed on the basis of a neural network plant model using the receding-horizon control approach. Based on the neural model, the control is calculated by minimizing a projected cost function that penalizes future tracking errors. As an illustration of the approach, the nonlinear dynamics of a planar two-joint arm with a flexible forearm are modeled using a sigmoidal network and an offline estimation procedure for a range of motions. The applicability of the approach is illustrated through computer simulations.
Song et al. (1999) studied this question.