22.2K ·GitHub
What is this?
A collection of Reinforcement Learning algorithm implementations, exercises, and solutions meant to accompany Sutton's 'Reinforcement Learning: An Introduction' textbook and David Silver's RL course.
A collection of Reinforcement Learning algorithm implementations, exercises, and solutions meant to accompany Sutton's 'Reinforcement Learning: An Introduction' textbook and David Silver's RL course.Is this a library I import, or a set of standalone notebooks/scripts?It's a set of standalone Jupyter notebooks organized into folders, each corresponding to one or more chapters of the textbook/course, not an installable library.
It's a set of standalone Jupyter notebooks organized into folders, each corresponding to one or more chapters of the textbook/course, not an installable library.How do I get started with the exercises?Pick a folder matching the chapter/topic you're studying; each one includes learning goals, a concept summary, reading links, and an exercise notebook with a corresponding solution.
Pick a folder matching the chapter/topic you're studying; each one includes learning goals, a concept summary, reading links, and an exercise notebook with a corresponding solution.What do I need installed to run the code?Python 3, OpenAI Gym for the RL environments, and TensorFlow for the neural-network-based algorithms (Deep Q-Learning, A3C, DDPG, etc.).
Python 3, OpenAI Gym for the RL environments, and TensorFlow for the neural-network-based algorithms (Deep Q-Learning, A3C, DDPG, etc.).Are all the algorithms fully finished?No, several advanced sections are explicitly marked work-in-progress, including Deep Q Learning, Policy Gradient Methods, Prioritized Experience Replay, and DDPG.
No, several advanced sections are explicitly marked work-in-progress, including Deep Q Learning, Policy Gradient Methods, Prioritized Experience Replay, and DDPG.What algorithms are actually implemented here?Dynamic Programming (policy/value iteration), Monte Carlo prediction and control, SARSA, Q-Learning, Deep Q-Learning for Atari, Double DQN, REINFORCE, Actor-Critic, and A3C, among others.
Dynamic Programming (policy/value iteration), Monte Carlo prediction and control, SARSA, Q-Learning, Deep Q-Learning for Atari, Double DQN, REINFORCE, Actor-Critic, and A3C, among others.Does it cover Atari-style deep RL or just toy problems?Both: it starts with classic tabular methods like Dynamic Programming and Monte Carlo, then moves to Deep Q-Learning and Double DQN applied to Atari games using TensorFlow.
Both: it starts with classic tabular methods like Dynamic Programming and Monte Carlo, then moves to Deep Q-Learning and Double DQN applied to Atari games using TensorFlow.Where can I find extra background reading?The README links to Sutton's textbook, David Silver's UCL course, several university deep RL courses (Berkeley, Georgia Tech, Waterloo), key papers like the Deep Q-Network and DDPG papers, and related open-source RL projects.
The README links to Sutton's textbook, David Silver's UCL course, several university deep RL courses (Berkeley, Georgia Tech, Waterloo), key papers like the Deep Q-Network and DDPG papers, and related open-source RL projects.What are similar projects?cleanrl, easy-rl, tianshou, mit-deep-learning, DeepLearning, machine-learning-notes, ML-From-Scratch, handson-ml3
cleanrl, easy-rl, tianshou, mit-deep-learning, DeepLearning, machine-learning-notes, ML-From-Scratch, handson-ml3