Real-time Reinforcement Learning of Vibration Machine PI-controller

Abstract

Controller tuning is a standard engineering task. To quickly adjust the controller settings in real-time, it becomes necessary to use intelligent control algorithms. In this paper, we propose an approach to tuning the speed controller of a vibration machine, which will ensure its maximum performance, using the reinforcement learning method. In this context of problem solving, the policy is presented in a parametric family of controller gains. In this case, the agent interacts with the virtual environment and the PI controller is implemented the software. The effectiveness of the proposed approach has been verified by real-time simulation and experiments on the two-rotor vibration unit. The advantage of the described learning algorithm is that the complex system is considered a black box. Thus, it is required to know the reference drive speed and measure the output speed.