Schneider modeled 6,000 drivers with approximate dynamic programming

Schneider National runs one of North America's largest truckload fleets, with over 6,000 long-haul drivers deciding where to base new drivers, whether to change work rules, and how to get drivers home.

The naive approach was to rely on dispatcher judgment or simple spreadsheets, but the truckload network has complex dynamics and multiple forms of uncertainty. Schneider and Princeton built a simulation that models drivers and loads at high detail.

The trick was approximate dynamic programming, which makes the model decide like a senior dispatcher — anticipating the future impact of a routing or hiring decision rather than acting greedily. The simulated policy calibrated closely to Schneider's historical performance, so managers trusted it.

The policy studies it enabled avoided $30M of costs by catching a faulty driver-management policy, saved $5M a year by picking the best driver domiciles, cut late deliveries by more than 50%, and saved $3.8M a year in border-crossing training costs.