This is a compilation of reinforcement learning algorithms.
** There are no shortage of RL repos on github (see the Related works section). If you are looking for state-of-the-art implementations I recommend you look into these. **
Learn about RL by implementing algorithms by myself while using good practices (OOP, TDD)
I noticed how most repos lack a test suite and good abstractions.
- OpenAI baseline (https://github.com/openai/baselines/tree/master/baselines )
- rllab (https://github.com/rll/rllab)
- Implementation of selected reinforcement learning algorithms in Tensorflow. A3C, DDPG, REINFORCE, DQN, etc. (https://github.com/stormmax/reinforcement_learning)
- Implementation of Reinforcement Learning Algorithms. Python, OpenAI Gym, Tensorflow. (https://github.com/dennybritz/reinforcement-learning)
- A set of Deep Reinforcement Learning Agents implemented in Tensorflow. (https://github.com/awjuliani/DeepRL-Agents)
- TensorFlow implementation of Deep Reinforcement Learning papers (https://github.com/carpedm20/deep-rl-tensorflow)
- Pytorch Implementation of DQN / DDQN / Prioritized replay/ noisy networks/ distributional values/ Rainbow/ hierarchical RL (https://github.com/higgsfield/RL-Adventure)
- Highly modularized implementation of popular deep RL algorithms in PyTorch (https://github.com/ShangtongZhang/DeepRL)
- Action-value method (to solve the k-armed Bandit problem)
- Q-Learning (using a linear model and a deep neural network)
- Monte Carlo Control
- Actor-Critic with eligibility traces
- DDPG
- PPO (WIP)
This repo also includes an abstraction (HyperFitter) that allows a Grid or Random search on the hyperparameter space of the agent. I was inspired by HyperOpt and sklearn GridSearchCV. I had to make my own because they are not flexible enough for my needs.
- PyTorch (for neural networks)
- TensorFlow
- OpenAI Gym (for the environments)
- pytest (for testing)
- sklearn (for feature mapping)
- Matplotlib (for plotting)
- NumPy (for everything else)
Testing is an interesting issue in RL. It is not clear what assertions should be made about an agent. We know that it should learn, but it is hard to predict what the learning curve should look like. See this great talk by Dr. Pineau : https://www.youtube.com/watch?v=-0G98MYUtjI
This repo mainly use integration testing (testing the general behavior instead of every method of every class). To do so, tests check if the return of the agent (the metric we try to maximize) is greater then the return of an agent that takes random decisions.
PYTHONPATH=. pytest
You can add the flag --pdb to debug