Robustness Measures for Stochastic Parallel Machine Scheduling and Train Unit Shunting

TL;DR

This paper evaluates and compares various robustness measures for stochastic scheduling problems, introducing new measures and demonstrating their effectiveness in improving schedule stability and reducing delays in practical applications.

Contribution

It identifies and tests 18 robustness measures, including 4 new ones, and demonstrates their effectiveness in enhancing schedule robustness across different problem settings.

Findings

Up to 90% delay reduction using robustness measures

Certain measures strongly correlate with schedule stability

Robustness measures can be effectively integrated into local search algorithms

Abstract

In many real world scheduling problems, the processing times of tasks are subject to uncertainty. This makes it essential to design schedules that are robust and able to handle potential disruptions. Therefore, we investigate measures that give us information about the robustness of a schedule. Although many measures can be found in literature, there is no consensus on which measures are the best. We identify 14 robustness measures from the literature, as well as introduce 4 new ones. To find out which of these measures are best used for generating robust schedules, we perform an elaborate simulation study to investigate how well these robustness measures correlate with the stability of the objective function under disturbances (quality robustness) and with the stability of the schedule itself (solution robustness). We first consider the Stochastic Parallel Machine Scheduling Problem (SPMSP) with precedence constraints, which is a very general setting that is relevant for many practical situations. We then perform a second simulation study by taking the best performing measures from the first experiment, and using them for the Train Unit Shunting Problem with Service Scheduling (TUSPwSS). After establishing the correlation with quality and solution robustness, we included the measures as objective in a local search algorithm. We make a comparison between the theoretical setting of the SPMSP and the TUSPwSS, and identify a set of robustness measures that can be applied in many different settings. We show that we can achieve up to 90% decreases in delays compared to using no robustness measures. Lastly, we also identify properties that can be used to predict the effectiveness of such a robustness measure.