Learn-to-Optimize Frameworks for Networks: Models and Applications

• Large-scale network optimization problems arise in transportation, logistics, energy, and infrastructure systems, where the objective is to determine optimal operational or restoration decisions within highly interconnected systems. These problems, typically formulated as combinatorial or mixed-integer programs, are computationally challenging due to their exponential complexity and large decision spaces. To address this challenge, this research develops Learn-to-Optimize frameworks that integrate machine learning with advanced optimization algorithms to improve scalability and convergence. The proposed methodology leverages problem structure and network features—including spatial, temporal, and topological measures—to guide optimization search strategies. Specifically, the frameworks incorporate supervised and non-parametric learning mechanisms into decomposition-based algorithms, such as Branch-and-Price, Evolutionary Optimization, and Nested Benders Decomposition, to accelerate improvements in solution quality and computational efficiency. From a managerial perspective, this work enables faster and more reliable decision-making for large-scale infrastructure systems, supporting post-disruption restoration planning, resource allocation, and multi-objective trade-off analysis. Overall, the integration of learning and optimization offers a scalable and customizable pathway for tackling complex network problems across diverse application domains.