v2.11.0 (5919)

Cours scientifiques - CSC_53439_EP : Deep Reinforcement Learning

Domaine > Informatique.

Descriptif

Reinforcement Learning (RL) is of increasing relevance today, including in games, complex energy systems, recommendation engines, finance, logistics, transport, and for auto-tuning the parameters of large learning frameworks such as LLMs. In this course we study modern state-of-the-art reinforcement learning algorithms and related approaches (decision transformers, transfer learning, imitation learning, inverse reinforcement learning, ...). A focus of the course is on RL solutions with deep neural architectures that can scale to modern applications, and other aspects that are concerned in real-world deployments (safety, interpretability, ...).

Objectifs pédagogiques

Students should learn to design and deploy deep learning methodologies and architectures for a variety of contexts in involving reinforcement learning algorithms (and related algorithms), and develop a strong intuition, based on theoretical and practical insights, to their performance and limitations.

Pour les étudiants du diplôme MScT-Artificial Intelligence and Advanced Visual Computing

This course assumes familiarity with the foundations of RL and its main paradigms (temporal-difference learning, Monte Carlo, and policy-gradient methods), covered in a course such as CSC_52081_EP Reinforcement Learning and Autonomous Agents (in the 3A programme).

Format des notes

Numérique sur 20

Littérale/grade réduit

Pour les étudiants du diplôme MScT-Artificial Intelligence and Advanced Visual Computing

Vos modalités d'acquisition :

Oral exam (presentation and discussion format), which covers a discussion of lab assignments and lecture material, and a choice of research paper, throughout the course.

Le rattrapage est autorisé (Max entre les deux notes)
    L'UE est acquise si note finale transposée >= C
    • Crédits ECTS acquis : 2 ECTS

    La note obtenue rentre dans le calcul de votre GPA.

    Programme détaillé

    The course is roughly organised into four approaches to the theme of depth in Deep Reinforcement Learning:,

    1. Depth in value function (DQN and variants, distributional RL, ...),
    2. Depth in policy (PPO, SAC, imitation learning, ...),
    3. Depth in environment model (Monte Carlo Tree Search, model-based reinforcement learning),
    4. Depth in reward model (reward shaping, inverse reinforcement learning, transfer learning ...).

    Mots clés

    apprentissage par renforcement, agents autonomes, prise de décision probabiliste, agents intelligents, prise de décision séquentielle
    Veuillez patienter