Research

Parallel machine scheduling problems are often solved with column generation, where the slow step is repeatedly solving a dynamic program to find new job sequences worth adding to the plan. This work replaces that step with a transformer-based neural network that learns to propose good sequences directly, cutting computation time by 45% on small-to-medium instances while a dynamic program still verifies optimality. On large-scale instances, the learned approach improves solution quality by 80% within minutes, and the model generalizes to problem sizes and distributions it was never trained on.

Supply chain planning problems have multiple interconnected layers of decision-making. In paper manufacturing, these layers include production scheduling, trimming decisions, vehicle loading, multi-period demand fulfillment, and inventory management. To solve the fully integrated problem at industrial scale, we exploit the structure of the problem and use a hybrid method combining column generation and dynamic programming for the supply-side with Benders decomposition for downstream fulfillement, giving the first exact model that ties all four decisions together at industrial scale. On real data it cuts costs by 24% and shrinks solve times from over five hours to under one hour compared to the prior approach.

Knowing not just a single expected delivery date but a full range of likely delivery times, with a real statistical guarantee, is valuable for setting customer expectations. This paper applies Conformal Predictive Systems and Cross Venn-Abers Predictors to forecast the full distribution of order fulfillment times, using spatiotemporal features about fulfillment location and carrier performance. On a large industrial dataset, the resulting point predictions beat the existing rule-based system by up to 14% in accuracy and up to 75% in catching late deliveries, while still providing rigorous coverage guarantees.

Online retailers typically decide which warehouse and carrier should fulfill each order using simple cost-minimization rules that ignore how orders could be batched together or how reliable a given carrier actually is. This paper develops a Contextual Stochastic Optimization framework that folds distributional forecasts of delivery delays directly into the fulfillment decision, tested on a real dataset with tens of thousands of products and up to thousands of fulfillment options. The framework is the first to combine multi-courier, omnichannel fulfillment with delivery-time uncertainty, and it substantially improves on-time delivery accuracy while letting retailers trade off cost against delivery risk.

Assigning locomotives to trains across a freight rail network is a large, cyclical planning problem shaped by power imbalances and complex operating rules. Working with a major Class I freight railroad, this paper models the problem as an integer program over a space-time network and introduces new reduction rules that make exact solutions possible for networks that were previously too large to solve. The resulting tool produces repeatable, high-quality weekly assignment plans that serve as a practical decision-support system for the railroad.

Consolidating smaller freight loads into larger shipments saves money and reduces emissions, but picking the right consolidation points has to respect real constraints like driver schedules and terminal operations. SPOT combines machine learning, spatio-temporal clustering, and frequent- pattern mining to discover recurring, operationally feasible consolidation points, then an optimization stage selects the most cost-effective routes each day. On industrial freight data, SPOT cuts travel distance and cost by roughly 50% compared to current practice, while also revealing recurring patterns that support proactive tactical planning.

Delivery load plans made days in advance often need last-minute adjustments as actual conditions unfold. This paper builds a two-stage deep learning system, refined operationally with the freshest data and paired with conformal prediction, to automate these load plan adjustments for a large logistics company. The approach outperforms traditional machine learning baselines and gives planners calibrated, trustworthy confidence sets rather than single point predictions, and it has already been deployed with strong accuracy in production.