Predict, Tune and Optimize for Data-Driven Shift Scheduling with Uncertain Demands

Michael Römer, Felix Hagemann, Till Frederik Porrmann

Abstract: When it comes to data-driven optimization under uncertainty, it is well known that a naïve predict-then-optimize pipeline in which point forecasts are plugged into a deterministic optimization model typically leads to a poor expected decision quality. In stochastic programming, one aims at obtaining better decisions by explicitly representing the joint probability distribution in the optimization model, e.g. in form of a sample approximation. A downside of that approach is that it gives rise to large-scale model instances that are hard to solve. An alternative approach that recently attracted considerable interest aims to train prediction models in a way that the expected decision quality obtained with the (prediction-informed) deterministic model is maximized, this approach is referred to as decision-focused learning or predict and optimize in the literature. In this paper, we propose to generalize this idea by optimizing not only parameters affecting the prediction but also additional parameters influencing other (non-stochastic) parts of the optimization model. Specifically, we propose to simultaneously optimize both types of parameters with the goal of maximizing expected decision quality and refer to this approach as predict, tune and optimize. We demonstrate the usefulness of the approach for a multi-activity shift scheduling problem under demand uncertainty. Specifically, we show that while decision-oriented tuning of point forecasts usually yields better results than a simple predict-then-optimize approach, adding the possibility to modify additional parameters considerably improves the expected performance which becomes competitive with a stochastic programming approach.