Actor-critic reinforcement learning for the feedback control of a swinging chain^⁎

Reinforcement learning offers a multitude of algorithms allowing to learn a nonlinear controller by interacting with the system without the need for a model of the plant. In this paper we investigate the suitability of online learning algorithms for a control task with incomplete state information. The system under consideration is a swinging chain that needs to be stabilized at a desired position, a problem that is occurring e.g. with bridge cranes with each change in the crane position. The measurable states are the position, velocity, angle and angular velocity at the top of the chain. A solution of the control problem based on an approximation of the chain as a continuous cable exists in the literature, see d’ Andrea-Novel and Coron (2000), which is included in the comparison as a reference for the control performance of the learned controllers.