RLplayground

Playing around reinforcement learning algorithms

This is a project acompanies my catching of the recent years developements in the Reinforecement learning domain. It gathers simple implementations of state-of-art RL algorithms under the same API.

Available algorithms

• DQN: With variants (Double, Dueling PriorityMemory, NoisyNets)
• VanilaPG
• A2C
• PPO: With value and norm clipping variants
• DDPG
• TD3
• ICM: In progress
• Evolution Methods: (ES CEM Etc..) Todo

Agent are sorted by the the action space types they can work with: Discrete, Continuous and Hybrid (Both)

Enviroments

CartPole

• All discrete agents

Pendulum

• TD3
• PPO

LunarLander

• DQN:
• PPO:

LunarLanderContinuous

• PPO: Agent can land but does'n larn to stop using throtle and get the final bonus
• DDPG: soved in ~900 epsiodes
• TD3:

BipedalWalker:

• TD3:
• PPO:

BipedalWalkerHardcore

• TD3 fine tuned from agent trained on BipedalWalker)

AtariPong:

• DQN:
• PPO:

AtariBreakout

• PPO:

Credits

Credit to all those Github repository I aspired from. I consulted a lot of repositories in order while implementing the algorithms, defining architectures and finetuning hyperparameters

• https://github.com/sfujim/TD3.git
• https://github.com/iKintosh/DQN-breakout-Pytorch.git
• https://github.com/Anjum48/rl-examples.git
• https://github.com/chagmgang/pytorch_ppo_rl.git
• https://github.com/shivaverma/OpenAIGym.git
• https://github.com/plopd/deep-reinforcement-learning.git
• https://github.com/adik993/ppo-pytorch.git
• https://github.com/higgsfield/RL-Adventure.git
• https://github.com/Adriel-M/OpenAI-Gym-Solutions.git
• https://github.com/nikhilbarhate99/PPO-PyTorch.git