Uncertainty-Aware Reinforcement Learning for Adaptive Educational Assessment under Item-Calibration Uncertainty
International Journal of Information and Communication Technology Trends · 2026
DOI: 10.66525/ijictt134 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Information and Communication Technology Trends, 2026. Verfasst von Caroline Mercer und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Adaptive educational assessment must make sequential item-selection decisions while both learner proficiency and item parameters are uncertain. Conventional maximum-information rules typically treat calibrated item difficulty as fixed, whereas reinforcement-learning policies often encode uncertainty only implicitly in a latent state. This paper proposes UA-DoubleQ, an uncertainty-aware reinforcement-learning framework that exposes posterior learner uncertainty and item-calibration uncertainty directly to the decision process. A Bayesian ability belief is maintained on a discretized latent-proficiency grid. At each step, a Double Q-learning controller selects a risk coefficient that modulates a robust Fisher-information score; the same controller can terminate the assessment when continued measurement is not worth its cost. This design separates uncertainty estimation from risk scheduling and allows the policy to become more conservative when item-parameter uncertainty is consequential. We evaluate the method in an ASSISTments-calibrated semi-synthetic benchmark, required because historical logs do not provide counterfactual responses to unadministered items. Across three independent seeds with 500 test examinees per seed, UA-DoubleQ used 14.02±1.25 questions, achieved ability RMSE 0.444±0.042, 92.1±1.1% empirical 95% interval coverage, and selected items with mean difficulty standard error 0.237. Relative to a vanilla Double-Q policy without uncertainty in the state, UA-DoubleQ shortened assessments by 17.0%, improved coverage by 1.8 percentage points, and reduced selected-item calibration uncertainty by 4.8%, with an explicitly quantified RMSE trade-off. Relative to point-estimate maximum Fisher information, it reduced RMSE by 14.9% and improved coverage by 4.3 percentage points. The results support uncertainty-conditioned risk scheduling as a practical middle ground between aggressive information maximization and conservative uncertainty-aware selection.
Zitieren
APA
Mercer, C., & Brooks, E. (2026). Uncertainty-Aware Reinforcement Learning for Adaptive Educational Assessment under Item-Calibration Uncertainty. International Journal of Information and Communication Technology Trends. https://doi.org/10.66525/ijictt134
Vancouver
Mercer C, Brooks E. Uncertainty-Aware Reinforcement Learning for Adaptive Educational Assessment under Item-Calibration Uncertainty. International Journal of Information and Communication Technology Trends. 2026. doi:10.66525/ijictt134
BibTeX
@article{Mercer2026,
author = {Mercer, Caroline and Brooks, Ethan},
title = {Uncertainty-Aware Reinforcement Learning for Adaptive Educational Assessment under Item-Calibration Uncertainty},
journal = {International Journal of Information and Communication Technology Trends},
year = {2026},
doi = {10.66525/ijictt134}
}
EUVIMED