UC Berkeley

This dissertation presents three projects that develop data-driven optimization frameworks for sequential decision-making under uncertainty. Each project addresses a distinct application domain---healthcare, environmental resource management, and robotics---but shares a common methodological foundation: combining statistical learning with mathematical optimization to derive actionable policies from observational or streaming data.The first project develops an interpretable reinforcement learning methodology for mechanical ventilation control in intensive care units. Using a causal, matching-based off-policy evaluation framework, we assess ventilator policies for their ability to improve patient-specific oxygenation outcomes while avoiding lung injury. Experiments on real-world ICU data demonstrate that an interpretable decision tree policy achieves performance comparable to deep reinforcement learning methods. The second project proposes a differential tax mechanism for groundwater management, modeled as a Stackelberg game between a water agency and water consumers. We derive a single-level reformulation of the resulting bilevel optimization problem, characterize the structure of optimal tax policies, and establish that the proposed mechanism can replicate the first-best social welfare outcome. Numerical simulations calibrated to the Western La Mancha aquifer validate the approach. The third project addresses profit maximization for a Robotics-as-a-Service operator who must jointly decide on job pricing and robot replacement under stochastic equipment degradation. We formulate this as a semi-Markov decision process and develop an online learning framework that integrates survival analysis and inverse optimization to estimate customer and degradation models. Numerical experiments show that the learned policy converges to near-optimal profitability.