Modulbeschreibung
Deep Reinforcement Learning
Kurzzeichen:
Unterrichtssprache:
ECTS-Credits:
Arbeitsaufwand (h):
Leitidee:
n this module, we introduce the foundations of deep reinforcement learning and will apply different algorithms to solve concrete problems in robotics and/or games.
After successful completion of this module the students:
- know how to apply deep reinforcement learning (RL) to solve problems in robotics and automation.
- know how to implement RL in Python using state-of-the-art libraries (for example TensorFlow, Keras)
- know how to formally describe a task as an RL problem
- understand what a computational-graph is and how it is used by neural networks
- understand the link between RL and Deep Learning
- know the limitations of RL
- understand the role of simulators and know how to use a robotics simulator (for example openAI-Gym)
- know how to qualitatively and quantitatively describe the progress of the training process.
Modulverantwortung:
Lehrpersonen:
Standort (angeboten):
Vorausgesetzte Module:
Zusätzliche Eingangskompetenzen:
Ebenfalls vorausgesetzt sind die Module Elektrotechnik und Lineare Algebra I und II sowie Informatik.
Modultyp:
Bemerkungen:
This module is taught in English. Most of the AI-literature is in English. An intermediate level is recommended. The students are free to write reports in German or English. For the oral exam, students can choose between English and German
Modulbewertung:
Leistungsnachweise und deren Gewichtung
Während der Unterrichtsphase:
Während der Unterrichtsphase wird eine Projektarbeit mittels Bericht und einem Fachgespräch bewertet.
Bewertungsart:
Gewichtung:
Während der Unterrichtsphase wird eine Projektarbeit mittels Bericht (Gewicht 50%) und einem Fachgespräch (Gewicht 50%) bewertet.
Bemerkungen:
Es findet keine abgesetzte Modulschlussprüfung statt.
Inhalte
Angestrebte Lernergebnisse (Abschlusskompetenzen):
In this module, we introduce the foundations of deep reinforcement learning and will apply different algorithms to solve concrete problems in robotics and/or games.
After successful completion of this module the students:
- know how to apply deep reinforcement learning (RL) to solve problems in robotics and automation.
- know how to implement RL in Python using state-of-the-art libraries (for example TensorFlow, Keras)
- know how to formally describe a task as an RL problem
- understand what a computational-graph is and how it is used by neural networks
- understand the link between RL and Deep Learning
- know the limitations of RL
- understand the role of simulators and know how to use a robotics simulator (for example openAI-Gym)
- know how to qualitatively and quantitatively describe the progress of the training process.
Modul- und Lerninhalt:
Learn how to describe a problem:
- describe a problem in terms of a Markov decision making task
- Fundamentals: state, action, reward, policy, Q-value
- Bellman equations
Learn algorithms to solve RL problems:
- SARSA, Deep-Q (and some of their variants)
- Backprop (Deep RL)
- optional: model-based RL
Learn to use Python and its libraries:
- Libraries: (for example Keras, TensorFlow)
- Simulation: (for example: openAI-Gym)
Apply your skills in a (graded) project
- Implementation of a project in the domain of robotics (e.g. https://gym.openai.com/envs/#robotics) or games (e.g. Connect Four)
Lehr- und Lernmethoden:
Unterrichtsgespräch im Klassenverband
Selbststudium (Übungen, Vor- und Nachbereitung der Lerninhalte)
Projektarbeit (Einzelarbeit oder im Zweierteam)
Lehrmittel/-materialien:
- Online resources
- Students’ notes
- Excerpts from R. Sutton, A. Barto, “Reinforcement Learning: An Introduction” ISBN-13: 978-0262039246
- Articles, e.g.: “Human-level control through deep reinforcement learning”, Mnih et. al, 2015, Nature