This repository proposes an introduction du Q-Learning on some toy exemples
The notebooks are readable on GitHUb but some features cannot be displayed on the site and it is better to experiment with the algorithms and the parameters.
You can directly clone the repository and install the packages listed in the requirements.txt file. Here are the commands using Anaconda:
git clone https://github.com/JbRemy/Reinforcement_Toys
conda create -n <NAME> python=3.7
conda activate <NAME>
pip install -r requirements.txt
For Mac OsX you might also need to install ffmpeg:
brew install ffmpeg
Note that this is usefull to replicate the exact setting I use to develop and test the repository. However there is not much specific interaction and it might work perfectly with a lot of configuration as long as you have the libraries listed in the requirements.txt file.
Each notebook presents some teoritical concepts and a new application. Here we describe the notebooks in the order they are meant to be read.
This notebook proposes a theoritical introducation to Q-Learning, a short presentation of the classes I implemented to make the experimentationo seamless to the reader and application of the basic Q-Learning algorithm on maze problem.
This notebook introduces the discretization of the Q-table and demonstrates the importance of the hyper-parameters. The problem is applied to the CartPole-v1 environment of the Gym AI library.
- Verify and update the description of the classes in the first notebook
- Add explration vs exploitation trade-off visualization on the first task
- Add more complex tasks, and algorithms (DeepQ, Lunar lander)
- Complete README for codes
- Update Notebook description (images)