Static forward networks and recurrent networks with feedback connections are the two most common types of networks applied to dynamical systems. Recurrent networks possessing memory and having dynamics can overcome the drawbacks and limitations of forward networks when applied to dynamical systems. This paper analyzes the implementation and on-line learning of recurrent networks for the identification and optimal control of nonlinear dynamical systems. An efficient procedure to improve and accelerate the on-line neuro-identification and optimal neuro-controller training process is presented. The analytical results are applied to the optimal control of a nonlinear high-speed ground vehicle.