Controlling an Inverted Pendulum by Neural Plant Identification

This paper shows, how wellknown supervised learning techniques can be applied to learn control of unstable systems. This is presently done in two steps: In a first step, a neural network is trained to predict the dynamic behavior of an arbitrary plant as accurate as possible. In a second step, the neural model is expanded by a control part in order to learn a control strategy that leads the system's state variables on given trajectories. I. Introduction Recently there has been great interest in applying neural networks to the domain of control problems. A main reason that led to this tendency probably lies in the difficulties and constraints of conventional controller synthesis. Conventional techniques are often restricted to linear, time-invariant plants. In practice, these assumptions are at most true for a small region around a working point of the system. Another drawback lies in the fact that the mathematical model underlying the plant has to be identified, a usually very hard or..