Zhuoli Yin
Modern business and societal systems rely heavily on mathematical optimization to ensure their efficient planning and operations. Traditionally, these optimization problems have been handled by commercial solvers or expert-engineered heuristics. Recent advances in artificial intelligence (AI) create new opportunities to augment optimization systems by learning from operational data, solver logs, and structured patterns in problem instances. However, deploying learning-augmented optimization systems in practice raises three central challenges. First, learned optimization policies often struggle to generalize beyond the instance sizes and distributions seen during training. Second, many learning-based methods remain computational-budget agnostic: runtime is treated as an external stopping rule rather than as a decision resource that could govern where and how long the system searches. Third, the net sustainability impact of AI use in optimization remains unclear. Existing evaluations often consider compute-related costs and downstream operational savings in isolation.This dissertation uses vehicle routing problems as primary testbeds to study how learning-augmented optimization systems can generalize across problem instances, adapt to computational budgets, and be evaluated at the system level by weighing the downstream operational benefits that AI models enable against their energy consumption and environmental impacts, thereby informing their deployment in real-world decision-making:(1) First, to mitigate the generalization and scalability challenges of learning-based optimization heuristics, this dissertation proposes ViTSP, a vision-guided framework for large-scale traveling salesman problems. ViTSP leverages pre-trained vision-language models to identify promising geometric subproblems and delegates these subproblems to an off-the-shelf solver for high-quality re-optimization. The experimental results demonstrate that ViTSP outperforms both state-of-the-art heuristic methods and learning-based methods. This work enables generative AI to supply generalizable decomposition guidance that complements mature operations research (OR) solvers.(2) Second, this dissertation proposes PACER, a time-budget-aware neural local search framework for capacitated vehicle routing problems. PACER formulates local search as a budget-augmented sequential decision problem, where the remaining wall-clock time is part of the policy state and each action jointly selects a subproblem and a solver-time allocation. The results show that PACER outperforms existing off-the-shelf solvers and learning-based methods, while also demonstrating how computational budgets reshape neighborhood search.(3) Third, to understand the net sustainability of AI use, this dissertation develops a marginal carbon accounting framework for evaluating AI-enabled routing from a system perspective. The framework links computing-side energy use and carbon emissions from training and inference, together with amortized embodied emissions from computing hardware, to application-side vehicle energy use and carbon emissions resulting from travel along the generated routes. By comparing AI methods against the realistic status quo they replace, the framework identifies the deployment and utilization conditions under which AI yields net carbon savings or costs.Together, these frameworks treat AI and optimization as an integrated system and contribute both methodological and analytical foundations for AI-augmented optimization under practical deployment constraints. The results show that AI is most valuable when it is used to (1) augment rather than replace reliable solvers; (2) adapt local-search effort to the available time budget; and (3) support sustainability claims only under marginal accounting and well-specified deployment and utilization conditions. These insights advance the design of scalable, resource-aware, and sustainable optimization systems for transportation, logistics, and broader business operations, and lay the groundwork for deploying such systems in high-stakes, real-world settings.