Transportation and logistics

CASTLE Lab has a long history of working in the area of transportation. Our sponsors have spanned truckload and less-than-truckload trucking, rail, military air cargo and shipping companies, in addition to business jets and cryogenic gases. Our corporate partners have included many of the largest transportation companies in the world, who have presented us with a range of challenging problems. From this work, we developed a powerful set of modeling and algorithmic tools including the DRMS modeling library and approximate dynamic programming for solving these complex problems.

For a summary of some of our industrial projects, see our link for impact on industry. In the 1980's, we developed a model for planning the operations of less-than-truckload carriers at a time when this industry was undergoing significant change. It was originally developed as "APOLLO" and then evolved to "Sysnet" (Sysnet is a registered trademark of Yellow), and was then marketed by Princeton Transportation Consulting Group (now a part of Manhattan Associates) as "Superspin," At one point, Superspin was used by virtually every LTL carrier in the industry. In 1995, new research led to our second consulting firm, Transport Dynamics.

Most of our work has focused on stochastic, dynamic models that arise in real-time operations, or as part of more detailed planning models. These problems were motivated by the need to plan fleets for truckload motor carriers, drivers for Yellow Freight, freight cars and locomotives for railroads and cargo aircraft for the airlift mobility command of the Air Force. These problems formed the motivation for our research in approximate dynamic programming.

In a recent project, a model based on approximate dynamic programming was shown to closely simulate the dispatching decisions of Schneider National, one of the nation's largest truckload motor carrier with a fleet of over 15,000 drivers. Modeling this process required solving a dynamic program which featured a state variable with millions of dimensions. In addition, the model had to handle the complex operational details required to properly model the management of drivers and loads. For a paper on the project to appear in Transportation Science, click here (c) Informs. This project was the winner of the 2009 Daniel H. Wagner Prize in operations research, which recognizes contributions to new methodology in the context of real applications. To view the short paper submitted for the award, click here.

Using ADP, we have also developed models for managing freight cars at Norfolk Southern, and recently we completed the first production-quality optimization model (using ADP) for planning and management locomotives at Norfolk Southern. This is the first formal optimization model (called "PLASMA") that can handle all of the operational details of a North American freight railroad, including penalties for breaking locomotive consists, delaying trains when necessary, and while simultaneously routing power toward shop. It will reposition locomotives to locations where power may be needed in the future, either as extra locomotives on a train or as light engine moves. It can handle locomotive attributes such as being leader-qualified, managing foreign power, or the handling the proper management of road and yard units. PLASMA can be as either a strategic planning tool or a real-time operational system.

We have also been developing models for Netjets, the first and largest operator in the fractional-jet ownership business (customers own a fraction of a jet). Netjets manages over 600 jets which are operated by over 2000 pilots. The problem requires simultaneously planning the assignment of pilots to aircraft to customers, while capturing a wide range of operating policies. Our model is used by Netjets to study new operating policies.

In a separate project, we developed SPARES, a model based on approximate dynamic programming, to guide decisions on what parts to order, when, and where they should be stored. The logic handles hundreds of spare parts, including many high-value spare parts where there might be five spares spread among 20 locations. SPARES produces a simple-to-use policy that describes how spares should be allocated among a network of facilities. The logic handles both expendable and rotable parts, and also handles lead times that can extend to six or nine months.

In addition to our work with civilian companies, we have in addition worked on several problems for the air force. A recent project used approximate dynamic programming to produce a realistic, intelligent simulator that adaptly learns how to use tankers for mid-air refueling. Another project involved developing an airlift simulator which introduces the novel idea of modeling the information available to a controller. This model was used in a project for the Canadian Air Force to estimate the cost of uncertainty, and demonstrate the power of approximate dynamic programming to produce robust solutions.