Reinforcement Learning Controller Trained with Dynamic Performance Index and Its Application to 3-DOF Helicopter

摘要

In this paper, the dynamic performance index in the classical control theory is introduced into the reward of reinforcement learning. The incompatibility of large initial error with small operation error when using binary reward are solved. In the simulation experiment based on 3-DOF helicopter platform, the feasibility is proved, and the comparison result between this method and binary rewarding method is given. Experiments show that this method enables the 3-DOF helicopter to track time-varying signals. It can also cope with the large initial error and has better dynamic performance.