nicolivain/rld

Deep Reinforcement Learning Agents in Pytorch in a modular framework

RLD

Reinforcement Learning and Advanced Deep Learning

Course from Master M2A (DAC) @ Sorbonne Université Paris.

This course covers reinforcement learning algorithms and generative deep learning methods.
Course website: https://dac.lip6.fr/master/rld-2021-2022/

Getting started:

main.py is a simple way to call other mains in command line, scheduler.py contains an hyperparameters search tool. A main file for each algorithm/TME is available under 'TP' folder. Hyperparameters of each algorithm can be tuned in 'Config/model_parameters', and then executed through the associated main function in 'TP'.

Implemented Algorithms:

• UCB and LinUCB Bandits
• Policy and Value Iteration
• QLearning, SARSA, DynaQ
• Deep Q Learning (minDQN), DuelingDQN, TargetDQN, Double VanillaDQN
• Goal VanillaDQN, Hindsight Experience Replay, Iterative Goal Sampling
• Actor Critic A2C
• Trusted Region Actor Critic PPO and Clipped PPO
• DDPG, Multi Agent DDPG
• SAC, Adaptative Temperature SAC
• Imitation Learning (GAIL)
• GAN, VAE
• Normalizing Flow: GLOW

Environnement:

Grid World, Cartpole, Lunar Lander, Pendulum, Continuous Lunar Lander, Mountain Car, MultiAgent

Ressources and references:

• [1] Lilian Weng's blog on Policy Algorithm
• [2] Alexandre Thomas' RL cheat sheet
• [3] the very useful minimalRL from seungeunrho
• [4] Sutton's reference textbook