The Control of Magnetic Levitation System Based on Improved Q-network

摘要

The magnetic system is a typical nonlinear system with the problem of uncertain parameters. Q-network is a model-free reinforcement learning method, which takes reward function as feedback and finds the optimal strategy through iterative learning. In this paper, to solve the problem, we first design the control based on the Q-network which realizes the stable control of the system without depending on the system model, and more importantly, we propose the retraining algorithm to further improve the accuracy of controller. Finally, the effectiveness of the Q-network controller in the control of a magnetic levitation system is verified by numerical simulation. The simulation results show that the network retraining algorithm can effectively reduce the steady-state error of the Q-network controller.