Adaptive-critic-based neural networks for aircraft optimal control

A dual neural network architecture for the solution of aircraft control problems is presented. The neural network structure, consisting of an action network and a critic network, is used to approximately solve the dynamic programming equations associated with optimal control with a high degree of accuracy. Numerical results from applying this methodology to optimally control the longitudinal dynamics of an aircraft are presented. The novelty in this synthesis of the optimal controller network is that it needs no external training inputs; it needs no a priori knowledge of the form of control. Numerical experiments with neural-network-based control as well as other pointwise optimal control techniques are presented. These results show that this network architecture yields optimal control over the entire range of training. In other words, the neural network can function as an autopilot. A scalar problem is also used in this study for easier illustration of the solution development.