This repository contains all the Reinforcement Learning work done for the Flatlands challenge by the CatLads team. Several approaches have been experimented, mostly based on the DQN algorithm, with different improvements and network structures.
This project needs the custom observations found in the CatLads/Observations repository, so remember to install that. In the dqn folder, you can find several algorithms/methods to train a reinforcement learning agent for the Flatlands challenge.
WandB was added to the project, meaning that you can use the platform to automate and customize the training.
To use this project, clone the repository and install the requirements found in requirements.txt:
$ pip install -r requirements.txt
Then, simply run the train.py script with the desired args. To discover the possible args, run:
python train.py --help
Through these you can customize the algorithm, the environment size and number of agents, the hyper parameters and so much more.
This code is licensed under GPL3, meaning you can edit, utilize and redistribute it. Feel free to do so.