SkaRL: A Single Self-Play Policy Outperforms Search-Based Skat Programs
Skat is a three-player card game which combines hidden cards, an auction, game types with opposite goals and a two-against-one partnership that changes in every deal. The strongest Skat programs search over the hidden cards at play time, learn from human games, or both. We present SkaRL, one transformer policy that makes every decision, from bidding to card play, with one forward pass and no search over hidden cards. Its parameters are learned by self-play with proximal policy optimization from random initialization, without supervised pretraining. In 10,000 live games on the International Skat Server against two instances of Kermit, a search-based program reported at human expert level, SkaRL scored 4.11 Seeger-Fabian (SF) points per game above their mean (standard error 0.82). In an evaluation against the strongest program known to us (SkatZero), SkaRL led by 1.51 SF per game (standard error 0.21). It also beat all other evaluated programs: TheCount, Zoot, Fox and the open-source programs JSkat and SkatKlar. SkaRL needs a median of 7.2 ms per move.
Authors
- Nils Purschke (ORCID: https://orcid.org/0009-0008-9470-5795)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24393691
- Primary Topic
- Artificial Intelligence in Games
- Type
- article
- Field-Weighted Citation Impact
- 0.00