Note: Project is currently undergoing a major overhaul.

With wacky-rl, you can create your own custom reinforcement learning agents. The library is modular and lets you customize everything by subclassing and plugging in different modules. Note that there are not many restrictions - you are free to incorporate any wacky idea you have, hence the name wacky-rl.

This library is the right choice if you want to create and test your ideas for custom RL agents quickly. Personally, I use this project for experiments, research, and to expand my understanding of RL. Check out the prebuilt agents to start; the documentation will follow soon. Note that the library is based on PyTorch.

• DQN to RAINBOW (Note: Currently broken due to major overhaul.) [code], [1], [2], [3], [4]
  • Double DQN
  • DuelingNet
  • PrioritizedExperienceReplay
  • Categorical DQN
  • NoisyNet
  • N-step Learning
• REINFORCE [code], [5]
• A2C (Note: Currently broken due to major overhaul.) [code], [6]
• SAC [code], [7], [8]
• PPO [code], [9], [10]

git clone https://github.com/maik97/wacky-rl.git
cd wacky-rl
python setup.py install

• torch
• gym
• numpy