Applying Backpropagation through Time to a Real Inverted Pendulum Problem

The “inverted pendulum problem” is perhaps the most widely used benchmarking study to assess the effectiveness of emerging control design techniques. In this paper the “Backpropagation Through Time” learning method is used to train a multilayer perceptron neural network to control an actual physical system, consisting in a pendulum free to pivot on a cart, which is moved by a DC motor within a space of 60 cm on a pair of slide guides. The goal of the neural controller is to maintain the inverted pendulum balanced at the middle of the slide guides. The neural network has been implemented both with a DSP and with a Fast Prototipyng Neural System (FPNS) hardware. The experimental results show the effectiveness of the used technique: the network is able to balance the pendulum also from difficult initial conditions.