Learning Uncertainty Representations for Reliable Decision Making

Many ML models are not used as predictors in isolation. A forecast is passed to an optimizer, controller, or planning routine, and the system is judged by the resulting decision. In such settings, a confidence estimate is only useful if it helps the downstream decision rule manage the risks that matter, such as tail loss, constraint violation, false negatives, or robustness to uncertain futures.

I will describe a line of work on learning uncertainty representations with the decision rule in the loop. The examples include conformal risk training for finite-sample tail-risk control, end-to-end conformal calibration for robust optimization, and diffusion models trained for stochastic optimization. The primary motivating applications include energy storage and grid operations.