Turner, J. D., L. H. Manring, and B. P. Mann. “Reinforcement learning for active damping of harmonically excited pendulum with highly nonlinear actuator”. Nonlinear Structures and Systems, Volume 1: Proceedings of the 37th IMAC, Jan. 2019, (2020), pp. 119–123. (extended abstract (PDF) and publisher webpage)
Active vibration dampers can reduce or eliminate unwanted vibrations, but determining a good control policy can be challenging for highly nonlinear systems. For these types of systems, reinforcement learning is one method to optimize a control policy with only limited prior knowledge of the system dynamics. An experimental system was constructed by attaching a permanent magnet to the end of a pendulum and positioning an electromagnetic actuator below the resting position of the pendulum. The pendulum was excited with a sinusoidal force applied horizontally at the pivot point, and the control input was the applied voltage across the electromagnet. Due to the geometric arrangement and the strong dependence of magnetic force on distance, the relationship between the position of the pendulum and the actuation torque for any control input was highly nonlinear. A generalized version of the PILCO reinforcement learning algorithm was used to optimize a control policy for the electromagnet with the objective of minimizing the distance between the end of the pendulum and the downward position. After 16 s of interaction with the experimental system, the resulting learned policy was able to substantially reduce the amplitude of oscillation. This experiment illustrates the applicability of reinforcement learning to highly nonlinear active vibration damping problems.
Turner, J. D., et al., Society for Experimental Mechanics, Inc., 2020.