Flight Attitude Simulator Control System Design based on Model-free Reinforcement Learning Method

Two-degree-of-freedom flight attitude simulator is a commonly used experimental platform in the control theory research. Due to the unknown environment and complicated constructions, it is difficult to build an accurate dynamical model. Reinforcement learning (RL) has been considered as a new means of solving complex and unpredictable control problems in recent years compared with traditional control methods. This is because RL algorithm is independent on the system model, and the controller parameters can be optimized by the iteration of the algorithm itself. In this paper, three model-free RL methods including Deep Q Network (DQN), Policy-Gradient (PG) and Deep Deterministic Policy (DDPG) are used to design the controller of the 2-DOF flight attitude simulator, and the performances of the methods are analyzed respectively. The results of theoretical analysis and experimental verification show good robustness in the control system. Moreover, the contribution of the RL method to the control field is pointed out and the problems that still need to be solved are listed at the end of this paper.