EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Causality-Based Data-Augmentation for Offline Reinforcement Learning

Wei Zhang, Azlaan Mustafa SAMAD, Yuan Xue, Daniel Kudenko

IEEE Access · 2026 · 14 · 76639-76648

DOI: 10.1109/access.2026.3694183 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Wei Zhang und 3 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Offline reinforcement learning (RL) seeks to derive an optimal decision-making policy from a static dataset, addressing scenarios in which interaction with the environment is impossible or data collection is costly or risky. The inherent limitation of the static dataset, incomplete state-action space coverage, leads to extrapolation errors when out-of-distribution data points are encountered during policy training. One way to mitigate this problem is to augment the static dataset by generating new data points, thereby broadening the state-action space available for policy training. These new data points should follow the same distribution as the original static dataset. The static dataset is collected by a behavioral policy whose decision-making logic determines the distribution. In this work, we propose Causality Based data-Augmentation (CBA), a novel framework that models the Markov Decision Process (MDP) in the form of a Structural Causal Model (SCM) and augments the static dataset while explicitly accounting for the decision-making logic of the behavior policy. Experimental results demonstrate that CBA enhances policy performance by augmenting static datasets with higher reward transitions, highlighting its potential for advancing research in offline RL.

Zitieren

APA

Zhang, W., SAMAD, A. M., Xue, Y., & Kudenko, D. (2026). Causality-Based Data-Augmentation for Offline Reinforcement Learning. IEEE Access, 14, 76639-76648. https://doi.org/10.1109/access.2026.3694183

Vancouver

Zhang W, SAMAD AM, Xue Y, Kudenko D. Causality-Based Data-Augmentation for Offline Reinforcement Learning. IEEE Access. 2026;14:76639-76648. doi:10.1109/access.2026.3694183

BibTeX
@article{Zhang2026,
  author = {Zhang, Wei and SAMAD, Azlaan Mustafa and Xue, Yuan and Kudenko, Daniel},
  title = {Causality-Based Data-Augmentation for Offline Reinforcement Learning},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {76639--76648},
  doi = {10.1109/access.2026.3694183}
}
Kostenlos starten →