Um estudo da aplicação de aprendizagem por reforço no controle de sistemas não lineares

Abstract

The inverted pendulum on top of a cart (cartpole problem) is one of the best-known problems in control theory. There are several ways to

control it, from more classic techniques, such as linearizing the problem, to more elaborate techniques, such as using computational intelligence. This work, therefore, will present two ways to solve this problem: the implementation of an LQR control and the application of Reinforcement Learning. Due to the increase in interest in Machine Learning and new control techniques, the objective will be to compare a method more widespread in the literature, linearization of the system using state space variables, with a more recent method, showing the possibility of a system learning the control law just by interacting with the plant. A bibliographic review will be carried out on the most relevant subjects for understanding the work, then the methodology applied will be explained, the problem set up using reinforcement learning agents and LQR controller, and finally the results will be presented and compared. All work will be carried out through simulations using Matlab software.