Open Source Robotics

Una raccolta curata di progetti, strumenti e software di robotica open source di alta qualità per la comunità della robotica.

CS 285: Deep Reinforcement Learning

  • Erogato da: UC Berkeley
  • Docente: Sergey Levine
  • Livello: advanced
  • Formato: video-lectures
  • Durata: One Semester
  • Categoria: reinforcement-learning
  • Lingua: English
  • Prezzo: Gratuito
  • Tag: reinforcement-learning, deep-learning, policy-gradient, actor-critic, imitation-learning, meta-learning

Questo corso tratta l'intersezione tra controllo, apprendimento per rinforzo e deep learning. Gli argomenti comprendono una trattazione avanzata del formalismo dell'apprendimento per rinforzo, gli algoritmi di RL model-free (policy gradient, funzioni di valore, Q-learning, actor-critic), l'RL model-based, l'apprendimento per imitazione e temi avanzati tra cui esplorazione, predizione video, transfer learning, apprendimento multi-task e meta-learning. Tra i prerequisiti: familiarità con l'RL, l'ottimizzazione numerica, il machine learning e le reti neurali profonde.

Perché seguirlo

Sergey Levine's CS 285 at UC Berkeley is widely considered the best deep reinforcement learning course available. It goes far beyond basic RL, covering the latest research in model-based RL, offline RL, meta-learning, and their applications to robotics. Many of the methods taught here are directly used in state-of-the-art robot learning systems today.

Cosa imparerai

  • Policy gradient methods and their variants (REINFORCE, PPO, TRPO)
  • Value function methods and deep Q-learning
  • Actor-critic architectures for continuous control
  • Model-based reinforcement learning for sample efficiency
  • Imitation learning and inverse reinforcement learning
  • Exploration strategies, meta-learning, and transfer learning

Programma

A full semester course with video lectures, written homework, and programming assignments. Progresses from imitation learning through policy gradients, value-based methods, and actor-critic architectures, then covers advanced topics including model-based RL, exploration, offline RL, and meta-learning. Assignments involve implementing algorithms from scratch and training policies in simulated environments.

Argomenti principali

  • Supervised learning of behaviors: imitation learning
  • Policy gradients and variance reduction
  • Actor-critic methods (A2C, A3C, SAC)
  • Value function methods and Q-learning (DQN, TD3)
  • Model-based RL and world models
  • Exploration and exploitation
  • Offline RL and batch RL
  • Transfer learning and meta-learning
  • Goal-conditioned RL and multi-task learning

Prerequisiti

  • Machine learning fundamentals (CS 189 level or equivalent)
  • Deep learning basics (experience with PyTorch or TensorFlow)
  • Probability and statistics
  • Comfort with mathematical notation and optimization

A chi si rivolge

  • Graduate students researching RL or robot learning
  • ML engineers applying RL to real-world control problems
  • Robotics researchers wanting to understand modern learning-based control
  • Anyone who wants a deep, research-level understanding of deep RL

Link

  • Sito web: https://rail.eecs.berkeley.edu/deeprlcourse/

Questo sito richiede JavaScript. Attiva JavaScript per esplorare i nostri progetti di robotica open source.