Study of Q-learning and deep Q-network learning control for a rotary inverted pendulum system

Abstract The rotary inverted pendulum system (RIPS) is an underactuated mechanical system with highly nonlinear dynamics and it is difficult to control a RIPS using the classic control models. In the last few years, reinforcement learning (RL) has become a popular nonlinear control method. RL has a powerful potential to control systems with high non-linearity and complex dynamics, such as RIPS. Nevertheless, RL control for RIPS has not been well studied and there is limited research on the development and evaluation of this control method. In this paper, RL control algorithms are developed for the swing-up and stabilization control of a single-link rotary inverted pendulum (SLRIP) and compared with classic control methods such as PID and LQR. A physical model of the SLRIP system is created using the MATLAB/Simscape Toolbox, the model is used as a dynamic simulation in MATLAB/Simulink to train the RL agents. An agent trainer system with Q-learning (QL) and deep Q-network learning (DQNL) is proposed for the data training. Furthermore, agent actions are actuating the horizontal arm of the system and states are the angles and velocities of the pendulum and the horizontal arm. The reward is computed according to the angles of the pendulum and horizontal arm. The reward is zero when the pendulum attends the upright position. The RL algorithms are used without a deep understanding of the classical controllers and are used to implement the agent. Finally, the outcome indicates the effectiveness of the QL and DQNL algorithms compared to the conventional PID and LQR controllers.