A learning-based robust optimization framework for synchromodal freight transportation under uncertainty

Synchromodal freight transport is characterized by its inherent dynamicity, necessitating the need for optimal decision-making in the presence of uncertainties in the real world. However, most prior research has overlooked the complexities of uncertainty modeling, often relying on assumed probability distributions that may not accurately reflect real-world conditions. This study presents a learning-based robust optimization framework for synchromodal freight transportation to derive data-driven explainable decisions. The study proposes a predict-then-optimize framework, using a combination of the Bayesian Neural Network with uncertainty quantification and dynamic robust optimization modules to solve the shipment matching problem under the synchromodality concept. The integration of prediction and optimization modules is achieved through scenario-based adjustable uncertainty sets. Rather than generating a single optimal solution, this framework produces an optimal policy based on various scenarios, enabling decision-makers to evaluate trade-offs and make informed decisions. The framework is implemented for the Great Lakes region containing nine intermodal terminals using real-world data and the performance is evaluated under various scenarios. In addition, a preprocessing heuristic-based feasible path generation algorithm is developed that helps the framework to maintain linear solution time. Numerical experiments performed on large demand instances (up to 700 shipment requests) demonstrate that the upstream prediction module significantly impacts the downstream optimization module. This effect is primarily due to variations in road travel times across scenarios, which impact transshipment operations, storage, and delay costs.