A dual neural network architecture for linear and nonlinear control of inverted pendulum on a cart

The use of a self-contained dual neural network architecture for the solution of nonlinear optimal control problems is investigated in this study. The network structure solves the dynamic programming equations in stages and at the convergence, one network provides the optimal control and the second network provides a fault tolerance to the control system. We detail the steps in design and solve a linearized and a nonlinear, unstable, four-dimensional inverted pendulum on a cart problem. Numerical results are presented and compared with linearized optimal control. Unlike the previously published neural network solutions, this methodology does not need any external training, solves the nonlinear problem directly and provides a feedback control.