Large scale optimization methods for fleet management in long-haul transportation networks
MetadataShow full item record
In this thesis, we explore topics in large-scale deterministic and stochastic optimization methods in transportation networks, with particular emphasis on methods of addressing uncertainty in fleet management problems. It consists of studies concentrating on emerging problems in fleet management and optimization under uncertainty. The first study, "A Robust Rolling Horizon Framework for Empty Repositioning", presents a robust optimization framework for an empty repositioning problem with demand uncertainty and provides insights on how to benefit from robustness while avoiding overconservatism. In this study, we demonstrate that by using the robust framework, significant improvements in service level can be attained, when compared to nominal solution approaches. We also analyze how the level of conservatism affects solution quality. Additionally, we experiment on ways to divide large transportation networks into "sharing groups" and limiting the repositioning moves to be mostly within sharing groups. We substantiate that the design of sharing groups can generate significant savings. The second study, "An Integrated Fleet Management Model Introducing Alternative Fuel Trucks", addresses the challenge of smoothly introducing alternative fuel trucks (AFTs) into an existing fleet while making necessary structural changes and maintaining feasible operations during the transition. In this study, we develop an integrated fleet management model and demonstrate that this model finds non-obvious solutions by making use of information that is often overlooked by most fleet replacement strategies. For finding optimal or good heuristic solutions to the integrated fleet management model efficiently, we propose a Benders’ decomposition framework and a Variable Neighborhood Search (VNS) algorithm; and demonstrate the performance of these solution methods. The third study, "Scenario Set Partition Dual Bounds for Multistage Stochastic Programming", focuses on finding dual bounds, called "partition bounds", in multistage stochastic programming problems with a finite number of scenarios. In this work, we propose a partition sampling approach for finding good partition bounds. We also present ways of choosing partitions more intelligently to further improve this sampling method in terms of bound quality and computational burden. With a comprehensive computational study, we demonstrate that by using the partition sampling approach, better dual bounds can be obtained with less computational effort, when compared to a Sample Average Approximation.