The Study About Control Of The Inverted Pendulum Based On Reinforcement Learning

The control of the inverted pendulum is a traditional problem in complex system control. The inverted pendulum system is a multivariable, highly nonlinear and absolutely unstable dynamic system. In the controlling process, it can effectively reflect many key problems such as robustness and tracking in control. Thus it is a ideal model to prove many control theories. In the area of modern mechanical system control, the stability control of inverted pendulum is similar to the problem involved in the emitting of rocket, the operating of satellite, weight lifting and moving of robot. So the researches of the inverted pendulum have substantial meaning not only to the theory, but to the project. In recent years, reinforcement learning, which is a cross-courses research direction, has become one of the hottest areas in artificial intelligence and machine learning and attracted many attentions of the researchers from other fields, including operation research, control theory and robotics. Reinforcement learning is different from supervised learning in that there are no teacher signals needed. It emphasizes learning through interacting with the environment and the learning target which maximize (or minimize) the evaluative feedback from the environment. This paper begins with learning the cognition of motor balance control, choosing the reinforcement learning methods as research objective, and using the inverted pendulum as experiment model. Through agent studying the control of the inverted pendulum, the intelligent system can learn from the running process and have motor control skills similar to people and animals. Based on reinforcement learning and dynamic programming algorithms which had existed, the paper presents an improved reinforcement learning system using double BP networks. Although the reinforcement learning system has no prior experiences, agent can cognize motor balance control skill of the inverted pendulum by adjusting itself online and control the inverted pendulum at last. We made several simulations under various conditions, and proved that the proposed reinforcement learning system has the ability to cognize the motor balance control skill of the inverted pendulum.