Abstract :
[en] In recent years, the deployment of artificial intelligence (AI) in high-stakes domains such as law, medicine, finance, and customer profiling has intensified the need for trustworthy AI; systems whose outputs remain safe, reliable, and actionable under real-world constraints. In such settings, predictive performance on historical data is not sufficient: models can still produce incorrect predictions with high apparent confidence, particularly under label noise, limited evidence, severe class imbalance, and distribution shift between training and deployment data. This miscalibration and overconfidence are operationally costly: they can trigger unwarranted automated actions, hide failure modes that emerge out-of-distribution, and create a false sense of high performance. Decision-makers therefore need signals of when a prediction is trustworthy and when it should be treated cautiously, deferred, or supported by additional information. This thesis therefore focuses on uncertainty quantification (UQ) as a practical mechanism for trustworthy decision-making by making explicit when model outputs are reliable enough and when caution or human oversight is warranted.
The thesis investigates uncertainty across both supervised machine learning and
reinforcement learning. In supervised learning, we study uncertainty estimation in
operational prediction problems where data are noisy and imbalanced. We examine distribution-free and evidence-theoretic approaches, including conformal prediction and belief functions, and identify conditions under which they lose actionability (e.g., inflated prediction sets under weak separability or high residual ignorance under conflicting evidence). Motivated by these limitations, we develop an ensemble-based approach tailored to real-world decision support: a heterogeneous ensemble framework that combines diverse learning paradigms to improve calibration and to operationalize epistemic uncertainty as disagreement across heterogeneous hypotheses, while separately characterizing aleatoric uncertainty arising from irreducible data noise. We further analyze how feature transformations and model composition affect both predictive performance and the stability of uncertainty estimates. Beyond supervised prediction, the thesis extends uncertainty-aware learning to sequential decision-making. We propose a lightweight uncertainty-aware reinforcement learning framework that leverages masksembles to obtain epistemic uncer-
tainty estimates without requiring large ensembles. By incorporating uncertainty into temporal-difference learning, the resulting agents improve training stability and robustness under extreme observation noise and perturbations, addressing a key gap in uncertainty-aware reinforcement learning where uncertainty is often used primarily for exploration rather than for stabilizing value-function learning. From a managerial perspective, the thesis demonstrates how calibrated uncertainty enables decision policies that are more dependable than score-based automation alone. Uncertainty quantification supports selective prediction and human-in-loop workflows, risk-aware thresholding, and resource allocation strategies such as sub-portfolio selection, where limited operational capacity requires prioritizing actions on cases that are both high-impact and sufficiently reliable. Overall, the thesis provides a cross-paradigm foundation for uncertainty-aware learning systems that translate predictive uncertainty into safer, more robust, and operationally actionable decision support.
