Title: Large-Scale Optimization for On-Demand Transportation Platforms and E-Commerce Supply Chains

Date: Monday, May 4th, 2026
Time: 2:00 pm – 3:30 pm ET
Location: Groseclose 303 and Teams 

Committee:
Dr. Alejandro Toriello (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Alan Erera, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Chelsea White, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
Dr. Myunseok Cheon, Supply Chain Optimization Technologies, Amazon
Dr. Chiwei Yan, Department of Industrial Engineering and Operations Research, University of California, Berkeley

Abstract:
This thesis studies three classes of optimization problems that arise in on-demand transportation platforms and e-commerce supply chains. We develop scalable and practical algorithms for systems operating at large scale under uncertainty or with complex combinatorial structure.

Chapter 2 studies a dynamic non-bipartite stochastic matching problem, where nodes appear following a type-specific independent distribution and wait in the system for a given sojourn time. This problem is motivated by applications in ride-sharing and freight transportation marketplaces. We study the asymptotic properties of two widely used policies, batching and greedy, by analyzing a single-pair case and then converting to the general counterpart using a fluid relaxation and randomization. We also extend our model to an impatient setting in which each unmatched node leaves at the end of each period with a type-dependent probability. We show that the results for the two policies still hold under different assumptions about the nodes' patience. Finally, we conduct a computational study simulating freight transportation and ride-sharing marketplaces to assess the empirical effectiveness of the policies. Our results suggest that managers can achieve near-optimal performance by using greedy or batching policies, with only a reasonably small maximum waiting time guarantee, and even in the presence of potentially impatient nodes.

Chapter 3 studies a network design problem with service time guarantees at industrial scale, motivated by the parcel delivery industry. This tactical service network design problem determines primary paths and delivery schedules for packages to minimize transportation and handling costs while ensuring committed service times. To construct a solution for a real-world instance with over 1,000 nodes, one million arcs, and 40,000 commodities, we propose a recursive graph partitioning and batching method. This method partitions the network into smaller regions and solves the problem for each region, considering only commodities whose origin and destination are in the same region. For commodities crossing regions, we first determine an appropriate sub-network, then solve a restricted model on this sub-network. To handle the large number of commodities, we divide these into batches based on schedule slack (intuitively, how flexibly a commodity can be scheduled) and solve the problem sequentially over each batch. Finally, we reoptimize commodities in low-utilization trailers. We demonstrate the scalability and efficiency of our approach through computational studies on real-world instances from an industry partner. Our method finds high-quality solutions after several hours of computing time, while a commercial solver is unable to even build a model for a much smaller instance.

Chapter 4 studies an inventory distribution problem for e-commerce supply chains. Despite extensive research on e-commerce supply chain optimization, some operational constraints remain underexplored. One critical factor is case-breaking at warehouses, where goods shipped in large quantities (case-packs) are disaggregated into individual units to meet retailers’ preference for small, frequent replenishments. Limited processing capacity may prevent all cases from being broken upstream, making it important to decide which cases to disaggregate and how to distribute both individual units and intact cases. Motivated by this, we study a case-pack allocation problem with case-breaking capacity constraints. Assuming deterministic warehouse inventory and target inventory levels at retailers, we optimize a concave, nondecreasing surrogate objective capturing product urgency, current inventory, and expected demand. We formulate an integer program and propose a column generation algorithm with efficient pricing heuristics. Computational experiments on realistic instances show the method finds near-optimal solutions with an average optimality gap of 0.004\% in 136 minutes, outperforming a commercial solver both in solution quality and runtime. We also analyze how the inventory distribution plan changes under tighter capacity constraints and varying warehouse constraints.