DETERMINATION OF Q-FUNCTION OPTIMUM GRID APPLIED ON ACTIVE MAGNETIC BEARING CONTROL TASK

DETERMINATION OF Q-FUNCTION OPTIMUM GRID APPLIED ON ACTIVE MAGNETIC BEARING CONTROL TASK

Stať ve sborníku v databázi WoS či Scopus

Active magnetic bearing control task can be successfully solved using reinforcement learning based method called Q-learning. The main problem to solve is the convergence speed. Two-phase Q-learning can be used to speed up the learning process [2]. Efficient prelearning phase uses mathematical model, following tutorage phase runs on real system and uses conventional Q-learning. This method can increase learning speed significantly, however there are still certain issues remaining to solve in order to improve the overall performance of the controllers based on Q-learning. When the table is used as Q-function approximation, the learning speed and precision of found controllers depend highly on the Q-function table grid properties. The paper is denoted to the determination of optimum grid with respect to the properties of controllers found by given method. Comparison of the results with performance of referential PID controller is included. Obtained results indicate that using nonlinear grid of Q-function table approximation improves the performance in terms of quadratic control quality criterion values, however, regarding the robustness against variables observation error and action delay the effect of nonlinear grid is questionable, improving the robustness for reduced state definitions only and degrading the robustness for common state definition which considers rotor deflection, velocity and acceleration as system state variables

Active magnetic bearing control task can be successfully solved using reinforcement learning based method called Q-learning. The main problem to solve is the convergence speed. Two-phase Q-learning can be used to speed up the learning process [2]. Efficient prelearning phase uses mathematical model, following tutorage phase runs on real system and uses conventional Q-learning. This method can increase learning speed significantly, however there are still certain issues remaining to solve in order to improve the overall performance of the controllers based on Q-learning. When the table is used as Q-function approximation, the learning speed and precision of found controllers depend highly on the Q-function table grid properties. The paper is denoted to the determination of optimum grid with respect to the properties of controllers found by given method. Comparison of the results with performance of referential PID controller is included. Obtained results indicate that using nonlinear grid of Q-function table approximation improves the performance in terms of quadratic control quality criterion values, however, regarding the robustness against variables observation error and action delay the effect of nonlinear grid is questionable, improving the robustness for reduced state definitions only and degrading the robustness for common state definition which considers rotor deflection, velocity and acceleration as system state variables

Reinforcement learning, Q-learning, Active Magnetic Bearing, Control

Reinforcement learning, Q-learning, Active Magnetic Bearing, Control

BŘEZINA, T., KREJSA, J.

24.03.2003

Institute of Mechanics of Solids Faculty of Mechanical Engineering Brno University of Technologi

Brno

80-214-2312-9

Mechtronics, Robotics and Biomechanics 2003

55

2