CASTLE Lab presentations
From time to time, I have the opportunity to give presentations that may be of general interest. Here are the Powerpoint presentations from these talks. Please note that the talks were designed assuming that I am giving the presentation. For this reason, these presentations may be little more than teasers. If a talk looks interesting, email me at powell@princeton.edu to see if it would be appropriate for me to give the presentation in person.
So you want to get funding from industry? (powerpoint format, pdf format)
Presented (twice) to the future academicians workshop at Informs, this provides a perspective on how to handle work with industry through a university. There are unique opportunities taking on industrial problems, but also unique challenges.
Approximate dynamic programming in rail operations (powerpoint format, pdf format)
Presented at TRISTAN VI (June 2007), this is an overview of current research in rail operations (in particular, locomotive optimization) using approximate dynamic programming. We are involved in the development of a sequence of models for strategic, tactical and real-time locomotive optimization. ADP allows us to model operations at a very high level of detail, while retaining the ability to use commercial packages such as Cplex to optimally solve the assignment of locomotives to trains at a point in time (avoiding the need for heuristics). The modeling and algorithmic strategy allows us to handle various forms of uncertainty, while capturing all the complex operational constraints that govern freight rail operations in the United States.
This is a two hour tutorial given at the IEEE Workshop on Approximate Dynamic Programming and Reinforcement Learning. It provides a complete introduction to the essential ideas of approximate dynamic programming, but in particular introduces the power of using the post-decision state variable. This strategy makes it possible to solve large-scale stochastic resource allocation problems using commercial optimization packages (such as Cplex) to solve the decision problem at time t. Simulation is used to model the transition of the system between time periods. The ideas are illustrated using an inventory problem, the modeling of a single, complex entity, a blood management problem, and a large scale fleet management problem.
An overview of CASTLE Lab (powerpoint format, pdf format)
A quick overview of research at CASTLE Lab with pictures of control centers, a hint at the approximate dynamic programming methods we have been developing, and a peek at a few applications.
Given in an elegant setting in Brescia Italy, this talk is a tutorial on approximate dynamic programming in the context of problems in transportation and logistics. I introduce the key idea of pre- and post-decision state variables, and the role they play in the development of effective algorithmic strategies for the large-scale problems that arise in transportation.
The Optimizing-Simulator: Merging Optimization and Simulation Using Approximate Dynamic programming (powerpoint format, pdf format)
This talk was given at the Winter Simulation Conference in December, 2005. It describes a sequence of methods of simulating the types of large scale problems that arise in transportation and logistics, starting with simple rule-based logic, and extending through four classes of information. The third class represents value functions obtained using approximate dynamic programming, while the fourth class represents patterns of behavior.
This talk divides resource allocation problems into four classes based on whether we are managing a single entity (trucks, locomotives, aircraft, machines, people) or multiple entities, and whether each entity is characterized by simple attributes or complex attributes. Below is the paper that was prepared for the conference. The presentation describes approximations that allow us to approximate solve resource allocation problems with thousands of entities, characterized by complex attributes (the attribute space is effectively infinite).
Information, Noise and Lies: The evolving discovery of misinformation in rail transportation (powerpoint format, pdf format)
We have worked for several years on a model for managing freight cars for a major railroad. The project started off looking like a textbook stochastic programming problem. By the end, the cars had evolved from a simple car-type attribute to a vector of attributes; the information process had morphed into a series of parallel information processes on customer orders, order attributes, car characteristics and transit times; we had learned about biases in customer orders, and we learned that there are some things that we will simply never learn.
The optimizing simulator for defense logistics (powerpoint format)
This is a more recent version of my "optimizing simulator" talk given to the defence research group in Valcartier, Canada. The talk focuses on defense applications with some civilian examples, and has segments on emergency response, modeling hierarchical decision making, and dealing with the long-haul bias problem in the context of airbase capacity.
Modeling information in freight transportation (pdf format)
Classical OR models focus on modeling physical flows. Modeling the dynamics of freight transportation requires learning how to model the flow of information, as well as the organization of decisions and information. This talk, given at the Montreal Spring School on Transportation in May, 2004, focuses on the subtle modeling issues that arise when modeling information.
Approximate Dynamic Programming for High-Dimensional Asset Management (powerpoint format)
This talk was given at Ohio State at the generous invitation of Professor Nick Hall. It describes the three ``curses of dimensionality," a general strategy for overcoming them, and an application to three broad classes of asset management problems: i) managing large numbers of simple assets (such as box cars), ii) managing a single, complex asset (such as a truck driver or pilot), and iii) managing large numbers of complex assets.
The Locomotive Routing Model (LRM) represents a five year development effort with two major railroads to model locomotive operations at a high level of detail. LRM uses the optimizing-simulator technology developed at CASTLE Lab, and is able to plan the flows of locomotives taking into account the power requirements of individual trains (both horsepower and tractive effort), shop routing requirements, consist management, repositioning for downstream requirements, and leader locomotives. The system can run in real-time (updating every few minutes), responding to information as it comes in. It can also be used as a simulator for certain types of strategic forecasting problems. The system interfaces with the Pilotview diagnostic system, allowing users to validate the results in a simple, intuitive way.
Tutorial on approximate dynamic programming for resource management
This is a three-part tutorial given at Twente University in the Netherlands, January, 2004:
Part I: Modeling and applications (powerpoint format)
The first part introduces an approximate dynamic programming strategy for resource management. We then divide the world into two problem classes: resources with simple attributes (such as classical multicommodity flow problems), and resources with complex attributes (people, locomotives, aircraft).
Part II: Fleet management for simple assets (powerpoint format)
A simple asset is a box car or truck trailer which might be characterized by the equipment type and location. The key is that we are going to assume that we can generate all the possible attribute types without getting into computational difficulty (this is usually the case if the number of attributes is less than 10,000).
Part III: Fleet management for complex assets (powerpoint format)
In many problems, the assets are complex, and we cannot enumerate all the asset types. For example, a driver might be characterized by his location, his fleet class, his home location, how many days he has been away from home, and how many hours he has been driving. It is not possible to enumerate all the possible driver types, but we need to be able to ask "if we assign a driver to perform a task, what is the value of the driver in the future?".
Real Time Optimization for Real-World Problems (powerpoint presentation with sound!)
I have given this talk several times. This is a slightly shorter version that I gave to the Princeton Operations Research Society, where I experimented with recording the presentation. It worked better than expected, but please note that it is a 40 meg file.
An Information Theoretic Approach to Dynamic Resource Management (html version)
This was a plenary address at the Optimization Days conference in Montreal. It suggests that instead of optimizing decisions, we can optimize the information provided to decision makers.
Adaptive Dynamic Programming for Dynamic Resource Management (html version)
A lot of the work in CASTLE Lab is based on developing functions that approximate the impact of a decision in one part of the system on another. This presentation is based primarily on the doctoral research of Huseyin Topaloglu and Greg Godfrey, and shows that in certain problem classes, we can obtain very high quality solutions to large scale problems. In particular, we are getting near optimal solutions to large, integer multicommodity network flow problems. These techniques easily handle both deterministic and stochastic data.
Optimization Models for Truckload Motor Carriers (html version)
This talk illustrates the use of optimization models for strategic, tactical and operational planning in the truckload motor carrier industry. Designed for a general audience, the talk highlights the modeling issues that arise in this classic resource allocation problem.
Castle lab presentation June 21, 2006 Powerpoint pdf