CASTLE Lab prides itself on its ability to have an impact on industry. We do this in several ways. Most important is our direct participation in the development of systems to be used by our corporate sponsors. This work is done completely through Princeton University by research staff working in the department of Operations Research and Financial Engineering. These projects range from real-time optimization, short-term tactical planning and strategic simulators for analyzing the dynamics of operations. While the details of these projects are often confidential, we do our best to write papers summarizing this work. Over $24 million in research funding (since the founding of CASTLE Lab) is testament to the faith that our sponsors have put into our efforts.
Consulting firms started
Strategic and tactical optimization of locomotives for Norfolk Southern
Changing the LTL trucking industry
Real time optimization and strategic planning for truckload trucking
Awards and testimonials
Major projects:
Strategic fleet simulator for Schneider National
Locomotive optimization for Norfolk Southern Railroad
Tactical and strategic fleet planning for Netjets
Spare parts management for Embraer
Strategic and operational planning at Yellow Freight System
Real-time dispatching for truckload trucking
Freight car distribution for Norfolk Southern Railroad
The multilayered-resource scheduling problem for Air Products and Chemicals
Interactive network optimization for LTL trucking
Other projects (not listed below)
Ryder Truck Lines – This was the first LTL carrier we worked with, which produced the predecessor of SuperSPIN
North American Van Lines – High Value Product division – Optimization model for LTL network
North American Van Lines – Commercial Transportation Division – Developed LoadMAP (predecessor of MicroMAP) – Second place in Edelman competition in 1987.
Burlington Motor Carriers – Implemented real-time dispatch system in 1993-1994. System was commercialized by Transport Dynamics
Triple Crown Services – First real-time implementation of a routing and scheduling tool for short-haul drayage
Two consulting firms have been started by students who worked under the supervision of Warren Powell. These include
Princeton Transportation Consulting Group, Inc. – Founded in 1988, PTCG was started with two models developed at Princeton University. The first was SuperSPIN, an interactive optimization model for planning networks for less-than-truckload motor carriers. The second was MicroMAP, the first commercially successful model for real-time load matching for truckload motor carriers. The first management team consisted of David Cape ’87 as president, and Ken Nickerson ’84 who was the lead software engineer.
PTCG is now a subdivision of Manhattan Associates.
As of 2011, Manhattan Associates is still selling “MicroMAP” (as “Drivers and Loads”) and SUPERSPIN. MicroMAP is used by 20 of the largest truckload carriers where it is being used to dispatch over 66,000 drivers, while SUPERSPIN is used by 20 of the largest LTL carriers.
Transport Dynamics, Inc.- TDI was founded in 1995 by Derek Gittoes who received his MSE while working in CASTLE Lab. TDI undertook a number of projects with major freight transportation companies, primarily under the leadership of Paul Stephens who came to us from BNSF Rail (and formerly Sabre). TDI was taken over in 2004 by Princeton Consultants, Inc.
Strategic and tactical optimization of locomotives for Norfolk Southern:
PLASMA (Princeton Locomotive and Shop Management System) is the first production optimization model for locomotives for freight railroads in the U.S. Developed for Norfolk Southern Railroad, PLASMA is based on the modeling and algorithmic technology of approximate dynamic programming. It is used for strategic planning (primarily fleet size and mix) and short-term operational planning.
Click here for a complete description
Changing the LTL trucking industry:
It is not often that a model has the effect of actually restructuring an industry. SuperSPIN was such a model. SuperSPIN was developed in the 1980’s, as the LTL trucking industry was undergoing rapid change as a result of deregulation. Developed initially for Ryder Truck Lines (where it was called “APOLLO”), it matured at Yellow Freight under the guidance of John Braklow (where it was known as “SYSNET”). Under John Braklow’s leadership, “SYSNET” was used to completely restructure Yellow’s network, in addition to providing a steady series of operational improvements. The project at Yellow was recognized as a finalist by the prestigious Edelman competition:
An article in CIO Insight on Yellow Freight noted
But the crown jewel of [CEO] Zollars’ and [CIO] Caddell’s technology overhaul has been SYSNET, a state-of-the-art computer system designed jointly by Yellow and CASTLE Lab, an engineering, operations and finance research center at Princeton University. SYSNET is the central nervous system of Yellow Transportation. Click here for a copy of the article from CIO Insight.
SuperSPIN was marketed by the Princeton Transportation Consulting Group to the rest of the LTL industry. It was used in the early planning of Roadway Package System, which would later become FedEx Ground, as well as the LTL carriers that would later become Fedex Freight (the LTL division of FedEx). SuperSPIN was used at Overnight Transportation, which is now UPS Freight. PTCG was eventually bought out by Manhattan Associates. For a description of how the model changed the industry, click here.
Real-time optimization and strategic planning for truckload trucking
We have also worked extensively in the area of real-time driver scheduling for the truckload industry. MicroMAP was the first real-time model which would handle driver assignments at a very high level of detail, but which was also able to capture the impact of decisions now on the future. This system was installed at 20 of the nation’s largest truckload carriers, including J.B. Hunt and Swift. As of 2011, MicroMAP (now marketed as “Drivers and Loads” by Manhattan Associates), is being used to dispatch over 66,000 drivers. A technical research paper describing this work can be downloaded from:
A more readable summary, including a case study with some behind-the-scenes insights, is given in
We recently completed a detailed strategic fleet planning system for Schneider National that has allowed Schneider to save tens of millions of dollars by allowing it to model, at a high level of detail, the operations of its fleet. The model allows Schneider to understand the impact of policies before trying them in the field. As of this writing, Swift, Schneider and Hunt are the three largest truckload carriers in the U.S. Click here for more information.
-
- Winner, 2009 Donald H. Wagner Prize, given by Informs for advancement of theory in the context of an application. The award was given for the development of a new class of approximate dynamic programming algorithms for infinite-dimensional states variables for optimizing the fleet of 6,000 drivers for Schneider National. The optimization model was found to closely mimic the behavior of Schneider’s team of dispatchers, while also producing estimates of the marginal value of each type of driver, similar to the dual variables of a linear program. One testimonial read:
“Schneider National is able to use the TPS to assess the myriad impacts of proposed changes to the HOS rules,” added Don Osterberg, senior vice president of Safety and Driver Training. “We were given the HOS proposal on a Tuesday morning and had to provide projections to the American Trucking Associations by Thursday. We ran the simulator and were able to provide timely and credible projections on a critical topic. The TPS provides decision support tools that project the first-, second- and third-order effects of contemplated decisions and help remove emotion and conjecture from the debate. It’s a powerful tool to have.”
- PLASMA is the first locomotive optimization model that can handle extended planning horizons, while modeling locomotive operations at a high level of detail. In addition, it has the ability to delay trains, and NS found that the model produced estimates of train delay very close to historical performance, making it possible to estimate fleet requirements for projected train schedules. The model is used at Norfolk Southern for fleet sizing. NS issued the following testimonial:
“PLASMA has recently completed the user acceptance test at Norfolk Southern as a strategic planning system. It has been used to assist the decision making in the 2009 locomotive road fleet requirement. Reaching this critical plateau required producing a carefully calibrated system that captured NS operating policies while matching their performance on important measures such as train delay.”
- In the 1990’s, we developed an operational planning system to plan the flows of drivers and loads for Yellow Freight System. The system runs continuously during the day, providing planners with information to control the movements at a network level. This system was discussed in an article in CIO Insight (Nov. 2003) which included the note:
But the crown jewel of [CEO] Zollars’ and [CIO] Caddell’s technology overhaul has been SYSNET, a state-of-the-art computer system designed jointly by Yellow and CASTLE Lab, an engineering, operations and finance research center at Princeton University. SYSNET is the central nervous system of Yellow Transportation.
Click here for the complete article (or here for a pdf copy from CIO Insight).
- In 2001, Yellow received its third consecutive CIO-100 Award for innovative business practices and services. In its citation (see here) was the note
Yellow system network experts used a network optimization model – a collaboration between Yellow and Princeton University – to simulate and analyze traffic flow on thousands of virtual shipment lanes. They created the most efficient relay and team network to move shipments faster, with reduced handling.
- Finalist, Franz Edelman competition (1992) – In recognition of the work with Yellow Freight System in the implementation of an interactive network design system. Click here for more information.
- Second place, Franz Edelman competition (1987) – In recognition of the work with North American Van Lines in the implementation of an operational fleet planning system. Click here for more information.
Major Projects
All of the projects below were undertaken at CASTLE Laboratory, and produced systems that were delivered to the sponsor and put into production.
Strategic fleet simulator for Schneider National
Using the techniques of approximate dynamic programming, we have developed the first production-quality fleet simulator for truckload trucking. Funded by Schneider National, the system handles the high level of detail required to accurately simulate the physics of drivers with the ability of ADP to optimize over time. This capability was critical in order to match the performance of their dispatchers and planners. In one project, the simulator was credited with savings of over $20 million annually. A summary of the methodology and the project may be found in “An Approximate Dynamic Programming Algorithm for Large-Scale Fleet Management: A Case Application“.
Locomotive optimization for Norfolk Southern Railroad
Sponsored by the Norfolk Southern and Burlington Northern Sante Fe Railroads, we solved the problem of assigning locomotives to trains over a planning horizon (a week for real-time applications, and a month for strategic planning applications) using approximate dynamic programming. ADP allowed us to assign locomotives to trains “here and now” capturing a high level of detail about both locomotives and trains, as well as a variety of complex business rules. ADP allowed us to optimize over time (approximately) while providing very high-quality solutions at a point in time. The system simultaneously handled the problem of routing locomotives to shop locations. As of this writing (2007), this system is still in develoment with Norfolk Southern. More information is available on the PLASMA page.
Tactical and strategic fleet planning for Netjets
Netjets operates over 500 jets, consisting of approximately 15 different types of jets. We have developed two models used by Netjets. The first is a very detailed optimization model that “simulates” (through optimization) the assignment of pilots to aircraft to serve customers. This model is able to handle work rules and operating policies at a very high level of detail, allowing Netjets to understand how changes in policies affect operations.
The second model focuses on fleet size and mix, capturing daily assignments of aircraft to customers over 10 years or more. This model is able to predict the substitution of different types of aircraft among customers, while modeling the daily, weekly and seasonal variations in demands.
Managing high value spare parts for Embraer
Embraer faced the problem of designing an inventory policy for its high value spare parts for a new line of aircraft. Using approximate dynamic programming, the model was able to capture tradeoffs between placing parts in field locations which provide fast, but limited, response times, versus keeping parts in distribution centers while incurring high transportation costs to respond to last minute demands. For this reason, the model used daily time increments, yet could handle lead times of six to nine months. It can produce a plan for over 1,000 spare parts, recommending the spatial allocation of inventories (which might involve allocating, say, six engines over 20 locations) while observing budget constraints and level of service targets at an aggregate level.
Strategic and operational planning at Yellow Freight System
Over a 15 year period, we developed a complete suite of models for strategic planning of the network (load planning) and operational planning for drivers. The operational planning system, which runs continuously during the day, was recognized in an article in CIO Insight, which included the quote:
But the crown jewel of [CEO] Zollars’ and [CIO] Caddell’s technology overhaul has been SYSNET, a state-of-the-art computer system designed jointly by Yellow and CASTLE Lab, an engineering, operations and finance research center at Princeton University. SYSNET is the central nervous system of Yellow Transportation. – CIO Insight
Real-time dispatching for truckload trucking
Triple Crown Services – Over the 1980’s and 90’s, we developed a number of technologies for optimizing the assignment of drivers to loads in real-time for a number of companies. In 1993-1995, we developed and implement a real-time driver scheduling system for Triple Crown Services, which needed a system for scheduling short-haul driver movements. This was the first system to solve, in real-time, the short-haul driver scheduling system, which required routing a driver (capturing full work rules and time windows) through multiple loads at the same time. The system responded to updates within a second. The system became the foundation of Transport Dynamics. Research based on this technology is summarized in:
Burlington Motor Carriers – Over the 1994 -1996 time frame, we undertook an implementation of our system at Burlington Motor Carriers, a long-haul truckload motor carrier with approximately 300 trucks. The project is described in full in the paper below.
Roberts Express (now FedEx Custom Critical) – In 1995, Transport Dynamics, using software developed at CASTLE Lab, implemented a real-time driver assignment system for Roberts, which guaranteed immediate pickup and fast delivery of critical orders.
The multilayered-resource scheduling problem
Working with Air Products and Chemicals, we developed a system for simultaneously routing drivers, tractors, trailers and chemical product to serve customers. Known within the academic community as the inventory routing problem, we found that the real issue focused on the complex operations surrounding the detailed scheduling of individual drivers, tractors and trailers. A cornucopia of business rules govern what drivers can do on a particular day, which affected what product should be delivered and the equipment required to deliver it. Developed over an eight year period, the system produced schedules for every driver and piece of equipment, handling every conceivable business rule.
Freight car distribution for Norfolk Southern Railroad
Managing the flows of freight cars for railroads requires making decisions about where to move cars empty to handle demands that have not yet become known. More than just demands, this problem has to be solved in the presence of random travel times, uncertainty about customer acceptance of freight cars (“it is too dirty!”) and equipment failures. Based on the Ph.D. dissertation of Huseyin Topaloglu using advances in approximate dynamic programming, this is the first production system (it has been adopted by the sponsor, Norfolk Southern Railroad) to explicitly solve a multistage stochastic optimization problem to manage its freight cars. The system is in production. A summary of the methodology is contained in
Interactive network optimization for LTL trucking
In the 1980’s, we created the first interactive-optimization model for designing service networks for less-than-truckload motor carriers. First funded by IU International (jointly at Princeton and MIT), the system evolved from APOLLO (Advanced Planner of LTL Operations), to SYSNET(r) (funded by Yellow Freight System), to SUPERSPIN which was marketed by the Princeton Transportation Consulting Group, founded in 1988 in Princeton NJ by two former undergraduates of Princeton University. The project at Yellow Freight was a finalist at the prestigious Franz Edelman competition sponsored by Informs:
This system allowed trucking companies, for the first time, to understand the tradeoff between the size of an end of line terminal and network costs. Prior to this system, major carriers such as Yellow had over 700 “end of line” terminals where freight originated and terminated. More end of line terminals were thought to be better because it reduced the expensive costs of deliving freight around the city. The system allowed companies to understand that with fewer, larger end of lines, it was possible to fill trucks to move destinations, reducing handling costs. For example, by the mid 1990’s Yellow Freight had reduced its network to around 400 end of lines. The system is still marketed and maintained by Manhattan Associates, and is used throughout the LTL trucking industry.