PhD Thesis Proposal: Benjamin Amoh

"Decision-Centric Optimization for Operational Systems: Robust Infrastructure, Delayed Learning, and Multi-Agent Mechanisms"

Abstract

Operational decision systems increasingly rely on predictive models whose outputs feed optimization models, dispatch routines, market mechanisms, and capital plans. This thesis proposal studies how learning and optimization should be designed when the governing criterion is downstream decision quality rather than prediction accuracy alone. The dissertation is organized around three connected arcs in decision-centric operations research.

The first arc develops robust optimization models for airport energy infrastructure investment. It jointly sizes solar, wind, battery storage, and vehicle-to-grid infrastructure over a multi-decade horizon under short-term weather distributional uncertainty and long-term demand and cost uncertainty. A two-stage stochastic linear program is combined with conformal prediction-calibrated Wasserstein distributionally robust optimization and Information-Gap Decision Theory. Across Denver, Seattle, and San Francisco airports, the framework selects airport-specific robustness policies: Denver and San Francisco obtain lifecycle-cost reductions of 0.82B and 1.42B relative to uncalibrated baselines, while Seattle shows that additional robustness is not economically justified.

The second arc studies online decision-focused learning under delayed feedback. The NeurIPS manuscript establishes the foundational theory: stale bilevel feedback can amplify drift through inner-solution sensitivity, producing a quadratic delay penalty. Implicit Gradient Transport Online Mirror Descent corrects this failure mode by transporting stale implicit gradients to the current iterate, recovering the single-level delayed-online-learning rate up to an unavoidable inner-solver precision term. The Management Science manuscript develops this theory into practical guidance for deciding when gradient transport is worth its computational and organizational cost.

The third arc extends decision-focused learning to cooperative multi-agent mechanisms. In assignment-based procurement settings with disjoint private information, it proves that joint decision regret decomposes exactly across agents, links the training signal to Vickrey-Clarke-Groves externalities, and gives a centralized-training/decentralized-execution algorithm with $O(N/\sqrt{T})$ regret. Together, the chapters argue that the optimization layer is not a post-processing step; it is the structure that determines what learning, robustness, delay correction, and coordination should mean.

Thesis Committee

• Geoffrey Parker (Chair)
• Vikrant Vaze
• Wesley Marrero

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu .