Computação paralela para controle em tempo real: controle preditivo baseado em modelos de aprendizado por reforço de sistemas de pêndulo invertido

Abstract

This work investigates the application of parallel computing in control systems, proposing a Reinforcement Learning-Based Model Predictive Control (RLMPC) framework. The focus lies on addressing the challenge of controlling an inverted pendulum system by integrating Reinforcement Learning (RL) and Model Predictive Control (MPC) to optimize performance and ensure real-time stability. As part of this dissertation, a physical inverted pendulum system was designed and assembled using industrial grade components that ensure robustness, reliability, and realistic operating conditions. This experimental platform played a key role in the validation of the proposed control structure in practice. The RLMPC framework employs parallelization to solve online optimization problems, enabling fast responses and effective decision-making. The parallel implementation proved to be efficient in meeting computational demands, allowing deployment in systems with real-time constraints. This study contributes to control theory by combining model-based and data-driven approaches, enhancing adaptability in uncertain industrial environments.