Design of Optimal Controllers Applying Reinforcement Learning on an Inverted Pendulum Using Co-simulation NX/Simulink

Abstract

Reinforcement learning techniques allow obtaining optimal controllers without the necessity of a model of the system. These techniques primarily require data obtained from real or simulated systems. Data from real systems are very complicated and costly to obtain due to the constraints involved in the model, so in order to obtain data is necessary to use other tools such as virtual modeling, simulations, or co-simulations. This paper proposes the design of optimal controllers for an inverted pendulum system with the help of Siemens NX software and Matlab/Simulink in co-simulation mode. The co-simulation allows obtaining data for reinforcement learning algorithms to obtain optimal controllers. The controller obtained in the region where the exploration was performed is stabilizing. Likewise, reinforcement learning techniques provided an optimal controller employing obtained data from the co-simulation of the system.