EUVIMEDEUVIMED entdecken
Crossref

Reconstructing Plausible Game Histories via Legal-Projected Reverse Generation: A Shogi Case Study

Hiroki Taniai, Hidetsugu Irie

Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 394-402

DOI: 10.1609/aiide.v22i1.43597 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, online erschienen am 08.11.2026. Verfasst von Hiroki Taniai und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Game AI research has concentrated on the forward direction: evaluating positions, choosing moves, and searching toward future outcomes. Yet when a narrative creator needs a natural-looking game that culminates in a desired dramatic position, or a training system must generate a playable lead-in exercise for a target lesson position, the question reverses: what plausible sequence of legal moves could have produced this position? We study this plausible history reconstruction problem in shogi (Japanese chess), whose drops, promotions, captures, and large branching factors make backward reasoning particularly demanding. Direct neural predecessor generation provides useful local predictions, but its output structure does not guarantee that a generated position can legally reach the successor. We propose legal-projected reverse generation: a rule-based generator enumerates predecessor candidates that reach the successor by a legal move, and a transformer model ranks them by consistency with real game data. Across 96,023 test positions, the model ranks the true predecessor first in 85.8% of cases at moves 1-10 and 41.4% at moves 91-100, despite an average of 121 legal candidates. With iterative widening beam search and budgets up to 20,000 expansions, multi-step rollback reaches the initial position in 96.8%, 75.5%, and 76.9% of games from moves 40, 80, and 100, respectively, and remains above 70% from moves 120-160 beyond the 100-move training horizon. At move 40, generated histories and the corresponding recorded histories receive nearly the same average score from a trained AlphaZero-style policy model. Together, these results show that legal-projected reverse generation can reconstruct plausible shogi histories from target positions while preserving legality over long rollback sequences.

Zitieren

APA

Taniai, H., & Irie, H. (2026). Reconstructing Plausible Game Histories via Legal-Projected Reverse Generation: A Shogi Case Study. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 394-402. https://doi.org/10.1609/aiide.v22i1.43597

Vancouver

Taniai H, Irie H. Reconstructing Plausible Game Histories via Legal-Projected Reverse Generation: A Shogi Case Study. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):394-402. doi:10.1609/aiide.v22i1.43597

BibTeX
@article{Taniai2026,
  author = {Taniai, Hiroki and Irie, Hidetsugu},
  title = {Reconstructing Plausible Game Histories via Legal-Projected Reverse Generation: A Shogi Case Study},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
  year = {2026},
  volume = {22},
  number = {1},
  pages = {394--402},
  doi = {10.1609/aiide.v22i1.43597}
}
Kostenlos starten →