Khev / RL-practice-keras

Deep Reinforcement Learning

Here I am implementing various RL algorithms, using python 2.7. I will use keras for the neurals nets. I'm going to use the OpenAI gym to test the algorithms. I list the methods below, which roughly divide into two categories.

I took / adjusted code from various online sources, which I inexhaustively list below (and in the code itself).

Value based methods

• Q-learning (tabular)
• Deep Q-Network (DQN)
• Double DQN (DDQN)
• DQN with prioritised replay
• Distributional bellman

Policy based methods

• Policy gradient -- REINFORCE & with baseline.
• Actor critic (A2C)
• Deep Deterministic Policy Gradient (DDPG)
• Proximal policy optimization (PPO)
• Soft Actor-Critic (soft AC)

Multi-agent

• Muti-agent deep deterministic policy gradient (MADDPG)
• Actor-Attention-Critic (AAC)
• Value Decompostion Networks (VDN)
• QMIX

Others

• Explore-and-go
• Curiosity driven learning (CDL)
• Rainbow (RB)

Resources

Papers

• Q-learning
• DQN
• Dueling DQN
• DQN with prioritized replay
• Distriubtional Bellman
• PPO
• Soft AC
• DDPG
• MADDPG
• AAC
• CDL
• RB
• EG
• MARL review article
• QMIX
• VDN

Blogs

• Arthur Juliana
• yanpanlau
• Gumble softmax trick
• review_blog
• TRPO intro

Textbooks

• Sutton

Acknowledgements

• @germain-hug
• @Keras-RL
• @keon