IEEE Conference on Games 2023
Zhao Yang, Thomas Moerland, Mike Preuss, Aske Plaat
To learn more:
If you find our paper or code useful, please reference us:
@article{yang2023two,
title={Two-Memory Reinforcement Learning},
author={Yang, Zhao and Moerland, Thomas and Preuss, Mike and Plaat, Aske},
booktitle={IEEE Conference on Games},
year={2023}
}
Create the conda environment by running:
conda env create -f environment.yml
In order to run experiments on MinAtar tasks, you need to install MinAtar correctly by following instructions provided.
The code base uses wandb for logging all the results, for using it, you need to register as a user. Then you can pass --wandb
to enable wandb logging.
You can simply run the code python train_2m.py --wandb
, tabular experiments presented in the paper python tabular/train_tab.py --wandb
.
Please be noted hyper-parameters in this work are quite senstive, in order to fully reproduce the results presented in the paper, you need to set hyper-parameters the same as in file.
The structure of the code base.
2m/
|- train_2m.py # start training
|- DQN.py # implementation of DQN agent
|- MFEC_atari.py # implementation of model-free episodic control agent for MinAtar tasks
|- tabular/ # folder of tabular implementations
|- RB.py # implementation of replay buffers
|- utils.py # utils functions
2M builds on many prior works, and we thank the authors for their contributions.