New method for making decisions when data is limited

A new approach uses a learned density model to handle errors in estimated distributions.

Distributionally robust optimization (DRO) protects against errors in distributions estimated from a finite record. This method hedges against errors in the estimated distribution.

A divergence ball contains only reweightings of the record. A transport ball contains distributions supported off the record. A Kullback-Leibler (KL) ball is centered at a learned density model fitted to the record.

The KL ball lets an adversary up-weight costly unobserved realizations. This happens in proportion to the probability the model assigns them. The method builds the scenario set adaptively.

The process alternates between solving the finite problem and drawing new realizations. These draws are steered toward the worst case at the current decision. Importance weights correct the draws back to the model.

The advantage fades as the record grows. It also fades if the shift departs from the structure the model has learned. The method lowers cost under distribution shift when the record is scarce.

Key facts

  • Centering the ball at the learned nominal lowers cost under distribution shift when the record is scarce.
  • The method works when the model class captures structure that lets it extrapolate beyond the record.
  • The approach was tested on a multi-item newsvendor problem and a capacity expansion problem.

All sections

Edition