Fabian Paischer1 2, Thomas Adler1, Vihang Patil1, Angela Bitto-Nemling1 3, Markus Holzleitner1, Sebastian Lehner1 2, Hamid Eghbal-zadeh1, Sepp Hochreiter1 2 3
1 LIT AI Lab, Institute for Machine Learning, Johannes Kepler University Linz, Austria
2 ELLIS Unit Linz
3 Institute of Advanced Research in Artificial Intelligence (IARAI)
This is the repository for the paper: History Compression via Language Models in Reinforcement Learning.
Detailed blog post on this paper at this link.
To reproduce our results, first clone the repository and install the conda environment by
git clone https://github.com/ml-jku/helm.git
cd helm
conda env create -f environment.yml
After installing the conda environment you can train HELM on the KeyCorridor environment by
python main.py
A new directory ./experiments/HELM/MiniGrid-KeyCorridorS3R1-v0 will be created in which all log files and checkpoints will be stored.
All changeable parameters are stored in the config.json file and can be adjusted via command line arguments as:
python main.py --var KEY=VALUE
For example, if you would like to train on RandomMaze-v0:
python main.py --var env=RandomMaze-v0
or on the Procgen environment maze:
python main.py --var env=maze
Note that by default the Procgen environments are created in the memory distribution mode, thus only the six environments as mentioned in the paper can be trained on, all others do not support the memory mode.
By default a Tensorboard log is created which can be visualized by
tensorboard --logdir ./experiments
MIT LICENSE