Authors:
Dennis J. N. J. Soemers, Vegard Mella, Eric Piette, Matthew Stephenson, Cameron Browne, Olivier Teytaud

Venue:
BNAIC/BeNeLearn 2024 (encore abstract), 2024

Topics:
transfer learning, reinforcement learning, policy-value networks, general game playing, zero-shot transfer

Links: PDF · Related journal version (TMLR)

This paper studies how to transfer policy-value networks between games with different board sizes, board shapes, rules, and action spaces.

Using the Ludii general game system and its domain-specific language, the authors propose a simple baseline transfer approach able to handle differences in both state and action representations.

The experiments on many pairs of game variants show multiple cases of successful zero-shot transfer as well as successful transfer with fine-tuning, across a broader and more varied set of game pairings than prior work.

This short paper is an encore abstract of the longer journal article published in Transactions on Machine Learning Research.

It presents the core idea in a concise conference format: using shared semantics from a high-level game description language to transfer policy-value networks across different games and variants.

The paper is especially relevant for research on more general AI systems, because it addresses transfer between domains with changing action spaces, board geometries, and rules rather than only level variations within a single fixed task.

Soemers, D. J. N. J., Mella, V., Piette, E., Stephenson, M., Browne, C., Teytaud, O. (2024). Towards a General Transfer Approach for Policy-Value Networks. BNAIC/BeNeLearn 2024. Encore abstract.

@inproceedings{soemers2024transfer_bnaic,
  author    = {Soemers, Dennis J. N. J. and Mella, Vegard and Piette, Eric and Stephenson, Matthew and Browne, Cameron and Teytaud, Olivier},
  title     = {Towards a General Transfer Approach for Policy-Value Networks},
  booktitle = {BNAIC/BeNeLearn 2024},
  note      = {Encore abstract},
  year      = {2024},
  url       = {https://bnaic2024.sites.uu.nl/wp-content/uploads/sites/986/2024/10/Towards-a-General-Transfer-Approach-for-Policy-Value-Networks.pdf}
}