Authors:
Graham Todd, Alexander Padula, Matthew Stephenson, Eric Piette, Dennis J. N. J. Soemers, Julian Togelius

Venue:
Neural Information Processing Systems (NeurIPS), 2024

Topics:
generative AI, procedural content generation, game design, evolutionary algorithms, large language models

Links: PDF · OpenReview · ACM · arXiv

Automatically generating novel and interesting games is a long-standing challenge in artificial intelligence.

This paper introduces GAVEL, a system that combines large language models with evolutionary quality-diversity algorithms to generate new board games expressed in the Ludii game description language.

The approach trains a code language model to mutate game rules and integrates it into a MAP-Elites evolutionary framework to explore a diverse space of possible games.

Experimental results show that GAVEL generates playable, diverse, and novel games, including games outside the distribution of known human-designed games.

This work represents a major step toward automated game design, combining advances in generative AI and evolutionary computation.

It leverages the expressive power of the Ludii system to represent over 1000 existing games and uses this data to train models capable of recombining mechanics in meaningful ways.

The results highlight the potential of AI systems not only to play games, but to create new ones, opening new directions for research in creativity, procedural content generation, and general AI.

Todd, G., Padula, A., Stephenson, M., Piette, E., Soemers, D. J. N. J., Togelius, J. (2024). GAVEL: Generating Games Via Evolution and Language Models. In Neural Information Processing Systems (NeurIPS).

@inproceedings{todd2024gavel,
  author    = {Todd, Graham and Padula, Alexander and Stephenson, Matthew and Piette, Eric and Soemers, Dennis J. N. J. and Togelius, Julian},
  title     = {GAVEL: Generating Games Via Evolution and Language Models},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year      = {2024},
  url       = {https://arxiv.org/abs/2407.09388}
}