KL-DRO
A distributionally robust optimization (DRO) paradigm that shifts the center of the KL divergence ball from the empirical distribution toward a learned density model. This preserves the ability to extrapolate structure while avoiding excessive hedging against impossible events. KL-DRO balances robustness to estimation error with the capacity to generalize beyond observed data.
Timeline (last 90 days)
Centering the KL ball on a learned model makes the problem difficult to solve: the problem can only accommodate a small number of scenarios, while independent sampling of the same size is very likely to miss the rare realizations in which the worst case is concentrated.
The researchers designed an adaptive scenario set construction method that alternates between solving a finite problem and drawing new realizations from the learned nominal distribution.
The drawing of new samples is guided toward the worst case under the current decision, and corrected back to the model distribution via importance weights.
Experiments on a multi-item newsvendor problem and a capacity expansion problem based on real power grid data show that when records are scarce and the model class can capture structure, setting the ball center at the learned nominal distribution rather than at the records can reduce total cost under distribution shift, while in-distribution cost remains comparable.
The advantage of the learned nominal distribution does not hold unconditionally: as the record size grows, or when the distribution shift deviates from the structure learned by the model, the gain gradually diminishes.
All Facts (5)
Centering the KL ball on a learned model makes the problem difficult to solve: the problem can only accommodate a small number of scenarios, while independent sampling of the same size is very likely to miss the rare realizations in which the worst case is concentrated.
50%UnverifiedThe researchers designed an adaptive scenario set construction method that alternates between solving a finite problem and drawing new realizations from the learned nominal distribution.
50%UnverifiedThe drawing of new samples is guided toward the worst case under the current decision, and corrected back to the model distribution via importance weights.
50%UnverifiedExperiments on a multi-item newsvendor problem and a capacity expansion problem based on real power grid data show that when records are scarce and the model class can capture structure, setting the ball center at the learned nominal distribution rather than at the records can reduce total cost under distribution shift, while in-distribution cost remains comparable.
50%UnverifiedThe advantage of the learned nominal distribution does not hold unconditionally: as the record size grows, or when the distribution shift deviates from the structure learned by the model, the gain gradually diminishes.
50%