ADVANCES IN TACTICAL PLANNING DECISION PROBLEMS IN TRUCKING SERVICE NETWORKS

The growth in e-commerce has spurred demand for parcel and less-than-truckload (LTL) freight services, and carriers providing these services compete by improving shipment delivery speed and reliability. Much of e-commerce relies on home delivery of small packages or parcels and other boxed freight. Key parcel carriers like UPS and FedEx con- tinually seek to redesign and operate profitable logistic networks that meet e-commerce customer service expectations. Beyond physical network design, including the location and sizing of various freight processing terminals, these companies need help with challenging service network design problems. This thesis addresses load or trailer (capacity) planning and trailer scheduling problems faced by planners at terminals in a parcel delivery service network. The research described in this thesis is conducted directly with a leading global parcel carrier that operates a massive network moving large volumes of packages each day. In Chapter 2, we address the Outbound Load Planning Problem (OLPP) to determine (1) how many loads or trailers to operate from a terminal to different destinations and (2) how to allocate the package volume processed at the terminal to outbound loads such that capacity constraints are satisfied, and packages are loaded into their compatible (primary and alternate) outbound destinations which are known apriori. We build a hierarchical optimization model to generate load plans with high utilization that are consistent in practice; this approach is beneficial to planners as they are more amenable to consistent and easy- to-evaluate recommendations. We develop an optimization-based learning methodology that blends machine learning (ML) and feasibility restoration frameworks. The ML model imitates the hierarchical optimization model to learn consistent solution patterns, and the feasibility restoration framework recovers a feasible solution from the ML predictions. An extensive computational study on industrial instances shows that the optimization-based learning approach is 10× faster than a commercial solver in obtaining the same quality solutions. In Chapter 3, we address the Robust Outbound Load Planning Problem (ROLPP). We formulate ROLPP as a two-stage robust optimization model with a relatively complete recourse. ROLPP aims to generate an outbound load plan for a terminal that remains feasible for all possible demand realizations that belong to a pre-defined uncertainty set. In stage 1, a planner determines the number of trailers to operate for each outbound destination from the terminal. Once an adversary chooses a demand scenario from a given interval budgeted uncertainty set, the planner has the flexibility to allocate package volume to their primary and/or alternate routing options to minimize the total cost of adding fractional trailer capacity for outbound destinations in the second (recourse) stage. We use a column-and-constraint generation algorithm that iterates between solving a master problem and a MILP reformulation of the bilinear subproblem to solve the ROLPP. We exploit the problem structure to carefully choose demand scenarios and propose a mountain climbing heuris- tic to generate local optimal solutions to the subproblem. These solutions can be used as warm starts as it is sometimes difficult to get feasible solutions to the subproblem using a commercial solver. Furthermore, we propose static and iterative approaches to generate tight lower bounds to the optimal objective value of ROLPP by carefully constructing a subset of demand scenarios that belong to the uncertainty set. We highlight that simple scenario-generation heuristics can often be used to produce tight lower bounds to the cost of robust load plans. In Chapter 4, we address the Integrated Inbound-Outbound Load Planning Problem (IOLPP) for a cluster of terminals in a parcel delivery service network. Each terminal in the cluster has a set of inbound trailers, a planned package processing capacity, and a set of planned outbound trailers. The planner aims to determine a cost-effective outbound load (capacity) plan for the cluster of terminals while ensuring that the package processing capacity is respected at each terminal in the cluster. The planner has the flexibility to shift an inbound trailer, planned to arrive at a terminal, to a nearby terminal in the same cluster to ensure that the package processing capacity is satisfied and to achieve a better outbound trailer consolidation. We formulate IOLPP as a MILP and solve it using a commercial solver. Through an extensive computational study on real-life instances, we show that it is often necessary to shift inbound trailers to ensure that the processing capacity at individual terminals is respected. Furthermore, shifting inbound trailers reduces outbound trailer capacity by 0.77% �- 2.10% due to better package consolidation. In Chapter 5, we address the Crossdock Trailer Scheduling Problem with Workforce constraints (XDTS-W). Given an arrival plan of inbound trailers and a fixed number of workers, planners need to determine an unloading schedule for the trailers that minimizes the total delay in loading the outbound trailers; the unloading schedule includes trailer-to-door assignments, trailer unloading sequence at each door, and worker-to-trailer assignment. We formulate XDTS-W as a MILP over a complete time-expanded network. We prove that if XDTS-W is formulated over a partial time-expanded network satisfying some properties, then the corresponding MILP is guaranteed to yield a lower bound to the optimal objective value of XDTS-W. We propose an exact dynamic discretization discovery (DDD) algorithm to solve XDTS-W. DDD iterates between a lower-bound and an upper-bound model and refines the partial network in each iteration until it converges to a provably optimal solution. Through computational experiments, we highlight the instances where the DDD algorithm outperforms existing state-of-the-art interval scheduling approaches in the literature and the instances where it is better to use a commercial solver to solve practical instances of XDTS-W.

Date

2024-09-09

Resource Type

Text

Resource Subtype

Dissertation

Full item page