Straightforward reinforcement learning Python framework. It will provide all the boilerplate code needed to implement RL logic (see diagram below) with different publicly available environments and own agents (plus e.g. logging).
It's not a deep learning framework! It's designed to work with them to build agents (e.g. PyTorch or TensorFlow).
Work in progress! It's not officially released yet. Contributions are welcome 😄
- Unit-tested for Python 3.5 and 3.6 versions. Python 2 isn't supported.
- See
requirements.txtfor rest of dependencies (note thatpytestis needed only to run tests).
Run pytest in repo root directory.
- Run
pip install -e .in repo root directory.
We are currently working on research project "Transfer Learning in Reinforcement Learning" and we are developing this small tool as we go. Right now you can find usage examples in samples directory. You can also look at AlphaZero and World Models implementations in this framework.
The most important things now are to improve logging and visualization capabilities, but also add support for more environments. Visualizing and training supervision should be easy-peasy and one should be able to run experiments in many many environments with only minor changes in code! We are waiting for you contribution 😄
If you use HumbleRL in your research, you can cite it as follows:
@misc{humblerl,
author = {Grzegorz Beringer and Mateusz Jablonski and Piotr Januszewski},
title = {HumbleRL - Straightforward reinforcement learning Python framework},
year = {2019},
publisher = {GitHub},
journal = {GitHub repository},
url = {https://github.com/piojanu/humblerl}
}