Documents disponibles écrits par cet auteur

Deep Reinforcement Learning for Autonomous Surveillance Quadrotors / ACHRAF OUSSIDI

Public ISBD Titre : Deep Reinforcement Learning for Autonomous Surveillance Quadrotors Type de document : texte imprimé Auteurs : ACHRAF OUSSIDI, Auteur Langues : Français (fre) Catégories : SDBD Index. décimale : mast 83/18 Résumé : DEEP reinforcement learning has been widely used in robotics and automated driving/flight. The reason behind this surge in popularity is the simplicity of its algorithms and its efficiency to deal with this type of problems. In this work, we review the literature of reinforcement learning, we present the advantages and limitations of the widely used algorithms; notably dynamic programming, Monte Carlo and the TD-Lambda method. We dive into their unique structures and the way they address the exploration vs exploitation problem. The second part of this work proposes a way for autonomous flight for quadrotors; one of the simplest designs for UAVs with a relatively complex and unstable dynamics and a non-linear system. We base our approach on RARL, an algorithm that mixes between adversary optimization scenarios and multi-agent reinforcement learning to produce an autonomous flight system for quadrotors that is robust to changes in the environment, generalizes better and performs well when the learned policy is transferred from a simulation to the real-world. The experimental results show a better performance compared to the baseline algorithm TD-Lambda and performs better in an environment under external forces that mimic the “imperfections” of the real-world. Deep Reinforcement Learning for Autonomous Surveillance Quadrotors [texte imprimé] / ACHRAF OUSSIDI, Auteur . - [s.d.]. Langues : Français (fre) Catégories : SDBD Index. décimale : mast 83/18 Résumé : DEEP reinforcement learning has been widely used in robotics and automated driving/flight. The reason behind this surge in popularity is the simplicity of its algorithms and its efficiency to deal with this type of problems. In this work, we review the literature of reinforcement learning, we present the advantages and limitations of the widely used algorithms; notably dynamic programming, Monte Carlo and the TD-Lambda method. We dive into their unique structures and the way they address the exploration vs exploitation problem. The second part of this work proposes a way for autonomous flight for quadrotors; one of the simplest designs for UAVs with a relatively complex and unstable dynamics and a non-linear system. We base our approach on RARL, an algorithm that mixes between adversary optimization scenarios and multi-agent reinforcement learning to produce an autonomous flight system for quadrotors that is robust to changes in the environment, generalizes better and performs well when the learned policy is transferred from a simulation to the real-world. The experimental results show a better performance compared to the baseline algorithm TD-Lambda and performs better in an environment under external forces that mimic the “imperfections” of the real-world.

Réservation

Réserver ce document

Exemplaires

Code barre	Cote	Support	Localisation	Section	Disponibilité
mast 83/18	mast 83/18 ACH	Texte imprimé	Unité des masters	Mast/18	Disponible