Implementación y evaluación de algoritmos de aprendizaje por refuerzo para tareas de control de drones en un ambiente de simulación realista

The recent success of the Deep Reinforcement Learning (DRL) algorithms opened its use in different environments and dynamical systems. We present the behavior of a complex dynamic system (quadrotor) in basic tasks as hovering and X-Y displacement in a realistic simulator. The DRL algorithms used were for continuous spaces, such as Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO). We probed dense and sparse reward functions and changed the negative component to demonstrate the impact of these parameters on a fast and repeatable learning process. We found that there is a major impact on the agent learning process because of the reward function, a correct selection could make the training times shorter and the repeatability higher. Contrasted with dense rewards, sparse rewards have less repetitive results and have poor results on tasks such as hovering and reaching X-Y points. Negative rewards directly affect the learning process when using PPO.

• Autores: Garzón Albarracin, Juan Felipe
• Fecha de publicación: 2021
• Institución: Universidad de los Andes
• Repositorio: Séneca: repositorio Uniandes
• Idioma: eng
• OAI Identifier: oai:repositorio.uniandes.edu.co:1992/50934
• Acceso en línea: http://hdl.handle.net/1992/50934
• Palabra clave: Drones; Vehículos piloteados de forma remota; Aprendizaje por refuerzo (Aprendizaje automático); Redes neurales (Computadores); Control automático; Ingeniería
• Rights: openAccess
• License: https://repositorio.uniandes.edu.co/static/pdf/aceptacion_uso_es.pdf