강화학습 기반 제어 알고리즘을 통한 시간 지연을 갖는 구조물의 진동 제어 연구

Abstract

The vibration control of a one-degree-of-freedom system was performed in this study using Deep Deterministic Policy Gradient (DDPG), a reinforcement learning method. A delayed control force compared to the target control force is applied to the system due to the dynamic characteristics of an actuator, such as a pneumatic spring. Reinforcement learning is a learning method that finds better behavior by learning by itself according to a reward function that is directly related to the learning goal without using a complex mathematical model for the system. Since the accelerometer is the most commonly used sensor in vibration measurement, we proposed a suitable learning excitation force and compensation function based on the acceleration data. The final learned policy was used to simulate the superior performance of the control force for various external forces. It was found from the numerical simulation that the vibration control based on the DDPG and reinforced learning is effective in suppressing vibrations.