2009 Daniel H. Wagner Prize

for

Application of Approximate Dynamic Programming to Model Fleet Operations

at Schneider National

CASTLE Lab is pleased to announce that a project involving the adaptation of approximate dynamic programming to model fleet operations at Schneider National received the 2009 Daniel H. Wagner Prize, given by the College for the Practice of Management Science at Informs. The Wagner prize recognizes the development of new methodology in the context of real-world applications.

The team contributing to the project includes Hugo P. Simao, Abraham George and Warren B. Powell from Princeton University, and Ted Gifford, Jeff Day and John Nienow from Schneider National.

The model, known as the Tactical Planning Simulator (TPS) at Schneider, produces a highly detailed "simulation" of fleet operations, capturing the dynamics of drivers and loads at a very high level of detail. The problem is formulated as a very large-scale stochastic optimization problem. We used an optimization framework to model the collective intelligence of the dispatch center (at the right). When we modeled operations at the level of detail required both for proper calibration and to provide sensitivity to key policy studies, the result was a stochastic linear programming with an effectively infinite number of rows.

The problem was solved using approximate dynamic programming, which overcomes the high dimensional state variables using methods from machine learning. This logic captured the critical ability of the dispatch center to anticipate the impact of decisions now on the future. This logic made it possible to capture not only a 15-dimensional attribute describing each driver, but also uncertainties in loads (demands) and travel times.

Considerable attention was put into capturing a host of real-world details so that the model closely matched a number of performance metrics. This includes issue such as: