Integrating optimal control with rules using neural networks

A recurrent neural network architecture augmented with rules capable of controlling nonlinear plants are presented. Using a recurrent form of the backpropagation algorithm, control is achieved by optimizing the network weights in the presence of task-adapted subnetworks representing rules. A quadratic cost function of endpoint trajectory values is minimized along with performance constraint penalties. The approach is demonstrated for a control task consisting of an aircraft flight path transition problem. It is shown that the network yields excellent performance while remaining within acceptable system constraints and while observing typical flight rules