Simple, Good, Fast (SGF)

Official code of the paper Simple, Good, Fast: Self-Supervised World Models Free of Baggage.
Published as a conference paper at ICLR 2025.

If you find this code or paper helpful, please reference it using the following citation:

@inproceedings{
  robine2025simple,
  title={Simple, Good, Fast: Self-Supervised World Models Free of Baggage},
  author={Jan Robine and Marc H{\"o}ftmann and Stefan Harmeling},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=yFGR36PLDJ}
}

Setup

Make sure the following dependencies are installed:
torch, torchvision, gymnasium, numpy, wandb, ruamel.yaml
Tested with PyTorch 2.5.1.

Training

To start a training run, execute the following command:

$ python src/main.py --device cuda:0 --game Breakout --project sgf --config configs/default.yaml --amp --compile --seed 1

The training script will log all relevant information to Weights & Biases.

To change the hyperparameters, create a copy of the default.yaml file and adjust the values as needed.

You can speed up training by setting --agent_eval final, which will only evaluate the agent at the end of training. To train an additional decoder for debugging, set --wm_eval decoder.