ABSTRACT
We present a general-purpose neural network architecture capable of controlling nonlinear plants. The network is composed of dynamic, parallel, linear maps gated by nonlinear switches. Using a recurrent form of the back-propagation algorithm, we achieve control by optimizing the linear gains and the nonlinear switch parameters. A mean quadratic cost function computed across a nominal plant trajectory is minimized along with performance constraint penalties. The approach is demonstrated for a control task consisting of landing a commercial aircraft from a position on the glideslope to the ground position in difficult wind conditions. We show that the network learns how to control the aircraft in extreme wind conditions, yielding performance comparable to or better than that of a “traditional” autoland system while remaining within acceptable response characteristics constraints. Furthermore, we show that this performance is achieved not only through learning of control gains in the linear maps but also through learning of task-adapted gain schedules in the nonlinear switches.
