Trustworthy Reinforcement Learning for Dynamic Pricing and Large Language Model Alignment

<p dir="ltr">As personalized decision-making systems become more prevalent, ensuring their trustworthiness—in terms of robustness, fairness, and alignment with human intent—has become increasingly critical. This dissertation investigates trustworthy reinforcement learning (RL) in the context of dynamic pricing and human-in-the-loop learning.</p><p><br></p><p dir="ltr">In Chapter 2, we study the challenge of strategic buyer behavior in personalized pricing, where users may manipulate their reported features to gain lower prices. Such strategic manipulation can significantly degrade a seller’s revenue and undermine system integrity. We propose a dynamic pricing policy that accounts for this adversarial behavior and adapts accordingly, ensuring robust revenue performance under such strategic interactions.</p><p><br></p><p dir="ltr">In Chapter 3, we extend the study of dynamic pricing by introducing fairness constraints. Contextual pricing decisions that lead to group-based disparities—such as those by race or gender—can trigger negative perceptions or legal violations. We design a fairness-aware pricing framework that balances revenue optimization with social responsibility, even when buyers strategically manipulate their sensitive attributes.</p><p><br></p><p dir="ltr">In Chapter 4, we address the challenge of data efficiency and alignment in reinforcement learning from human feedback (RLHF), a key component in aligning large language models (LLMs) with human preferences. Human feedback is often costly and noisy. To this end, we propose a dual active learning framework to identify the most informative data and labelers. We further develop a pessimistic RL approach that ensures safe and reliable policy learning based on uncertain feedback.</p><p><br></p><p dir="ltr">Together, these three lines of work contribute to the foundation of trustworthy reinforcement learning in real-world decision-making systems, spanning economic robustness, ethical fairness, and human-centric learning.</p>