Not Your Average Agent: Play-style Conditioned Decision Transformers
Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 108-117
DOI: 10.1609/aiide.v22i1.43569 ↗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 Sara Karimi und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Automated playtesting agents can help game designers evaluate balance, difficulty, and player experience by simulating diverse player behaviors. However, many existing reinforcement learning approaches collapse these behaviors into a single policy, limiting their ability to expose how different play styles interact with game mechanics and level design. In this work, we introduce StyleDT, a style-conditioned Decision Transformer designed to generate controllable, diverse gameplay behaviors. StyleDT models player strategies as a continuous latent style representation learned from full gameplay trajectories. A conditional prior maps designer-specified behavioral controls to distributions over latent styles, which are injected as prompt tokens into a Decision Transformer policy. This architecture enables inference-time control over agent behavior without requiring reference trajectories. We evaluate StyleDT in two MiniGrid-based playtesting environments of increasing complexity featuring multiple viable strategies, including combat, stealth, adventure, and avoidance behaviors. Experimental results show that StyleDT accurately reproduces different strategic behaviors and achieves strong style adherence, matching or exceeding specialized single-style policies while operating as a single unified model. Furthermore, the learned latent style representation enables reliable behavioral steering and generalizes to interpolated control configurations not present during training.
Zitieren
APA
Karimi, S., Asadi, S., & Payberah, A. H. (2026). Not Your Average Agent: Play-style Conditioned Decision Transformers. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 108-117. https://doi.org/10.1609/aiide.v22i1.43569
Vancouver
Karimi S, Asadi S, Payberah AH. Not Your Average Agent: Play-style Conditioned Decision Transformers. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):108-117. doi:10.1609/aiide.v22i1.43569
BibTeX
@article{Karimi2026,
author = {Karimi, Sara and Asadi, Sahar and Payberah, Amir H.},
title = {Not Your Average Agent: Play-style Conditioned Decision Transformers},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
year = {2026},
volume = {22},
number = {1},
pages = {108--117},
doi = {10.1609/aiide.v22i1.43569}
}
EUVIMED