02/03/2023: ALOP Colloquium

Talk: Models for Interpretable Optimization (Marc Goerigk, University of Siegen)

In the context of the colloquium of the Research Training Group Algorithmic Optimization on

Monday, February 13, 2023
4:00 p.m. c.t.
Lecture room 9

the following lecture will take place:

Models for Interpretable Optimization

Marc Goerigk, University Siegen

While explainability and interpretability has become a major research area in machine learning over the past years, in operations research and mathematical programming the comprehensibility of solutions is hardly every scrutinized. Likely due to the existence of explicitly stated models and well defined solution processes, experts have high confidence in the correctness and usefulness of found solutions. However, even understanding a mathematical model can quickly become a challenging task, especially for people not familiar with modeling techniques. Users with less mathematical and computer background, e.g. the planner using the optimization software and the workers in charge of implementing the result, may consider the solver a black box. We present and experimentally validate a modeling framework that inherently provides an interpretable decision rule, trying to clarify the causal effect of occurring scenario and selected solution. By restricting the number of eligible solutions and providing an interpretable rule, e.g. in form of a decision tree, we shift the focus towards comprehensibility and transparency. Our experiments indicate that the costs of interpretability can be small, i.e., only a small percentage of nominal performance needs to be sacrificed.