The Design of Controller for Swing Up Cart Pole Problem with Strongly Constrained Input

This paper presents an approach to designing the controller for the swing up cart pole problem. The approach used to solve the problem is a method based on Deep Reinforcement Learning. We taught an artificial neural network to control the plant as a controller. Furthermore, the input saturation is strongly limited to check the maximum possibilities of the controller. We investigated the control accuracy and the actor policy if the control signal was reduced. Our control goal was to reach the unstable equilibrium and keep that state starting from the stable state of the cart pole. That task was performed by one controller tested in an environment prepared by us, where the parameters of the cart pole are based on the real device. The results come from tests of many neural networks trained for different seed values.