Sequential decision-making under uncertainty is a fundamental problem in artificial intelligence.
Real-world environments rarely provide well-defined rewards or complete information, and feedback is often qualitative, subjective, or inconsistent.
As A...
Sequential decision-making under uncertainty is a fundamental problem in artificial intelligence.
Real-world environments rarely provide well-defined rewards or complete information, and feedback is often qualitative, subjective, or inconsistent.
As AI systems are increasingly deployed in high-stakes domains such as finance, autonomous driving, and human–robot interaction, it becomes crucial to develop principled algorithms that can act reliably under uncertainty and align with human intentions.
However, existing reinforcement learning (RL) paradigms, which lack explicit modeling of uncertainty, rely primarily on expectation-based objectives and handcrafted rewards, leaving a substantial gap between theoretical optimality and human-aligned behavior.
This dissertation addresses these challenges through two complementary perspectives—distributional reinforcement learning (DistRL) and reinforcement learning from human feedback (RLHF)—and unifies them under a common theoretical lens of regret minimization.
The central goal is to establish a reliable foundation for learning human-aligned decision-making by interpreting the probabilistic nature inherent in human feedback.
The first part revisits the exploration problem in DistRL.
Existing approaches based on “optimism under uncertainty” rely on estimates of return variance but conflate epistemic and aleatoric uncertainties, which induces persistent risk-seeking bias and distorted data collection. To address this, we propose the Perturbed Quantile Regression (PQR) algorithm, which introduces randomized perturbations of distorted risk measures to guide action selection.
We theoretically establish that PQR avoids biased exploration and converges to the true optimum, and empirically show that it outperforms variance-based exploration methods across diverse benchmarks, including 55 Atari games.
The second part tackles the fundamental challenge of infinite dimensionality in DistRL. Prior work introduced the notion of Bellman closedness, but this fails to guarantee unbiased updates from finite samples in online learning. We propose the concept of Bellman Unbiasedness, which characterizes functionals that are not only preserved under Bellman updates but also estimable without bias from finite samples.
Our analysis shows that only moment functionals satisfy both conditions. Building on this result, we design the first provably efficient DistRL algorithm under general value function approximation—Statistical Functional Least-Squares Value Iteration (SF-LSVI)—which achieves a tight regret bound of $\tilde{O}(d_E H^{3/2}\sqrt{K})$, improving upon prior results.
The third part turns to RLHF, where agents learn from preference feedback instead of handcrafted rewards.
Recent frameworks such as Direct Preference Optimization (DPO) optimize policies directly without an explicit reward model but implicitly assume that all preference data are generated by the optimal policy, leading to a likelihood mismatch.
To overcome this, we reinterpret preferences through the lens of regret and propose Policy-labeled Preference Learning (PPL), which explicitly integrates policy labels into the learning process.
Our method introduces contrastive KL regularization that aligns policies with preferred data while contrasting against less-preferred data. We theoretically show that PPL characterizes an equivalence class of reward models consistent with a given optimal policy and establishes statistical robustness via uniquely defined regret. Empirically, PPL substantially improves RLHF performance in offline robotic manipulation tasks and demonstrates robustness in online learning.
Collectively, these contributions establish regret minimization as a unifying theoretical principle that bridges distributional modeling and human feedback, linking the mathematical efficiency of RL with the behavioral realism of human decision-making.
This work contributes to the foundation of trustworthy and human-aligned artificial intelligence, providing theoretical and algorithmic insights for robust decision-making under uncertainty.