658. Computing Counterfactual Explanations for Linear Optimization: A New Class of Bilevel Models and a Tailored Penalty Alternating Direction Method
Invited abstract in session WB-48: New Approaches in Explainable Optimization, stream Transparent and Fair Decision Making with Mathematical Optimization.
Wednesday, 10:30-12:00Room: Parkinson B09
Authors (first author is the speaker)
| 1. | Henri Lefebvre
|
| Department of Mathematics, Universität Trier | |
| 2. | Martin Schmidt
|
| Department of Mathematics, Trier University |
Abstract
Explainable artificial intelligence is one of the most important trends in modern machine-learning research. The idea is to explain the outcome of a model by presenting a certain change in the input of the model so that the outcome changes significantly. In this talk, we study this question for linear optimization problems as an automated decision-making tool. This leads to a new class of linear bilevel optimization problems that have more nonlinearities in their single-level reformulations compared to traditionally studied linear bilevel problems. For this class of problems, we present a tailored penalty alternating direction method and present its convergence theory that mainly ensures that we compute stationary points of the single-level reformulation. Finally, we illustrate the applicability of this method using the example of a real-world energy system model as well as by computing counterfactual explanations for a large set of linear optimization problems from the NETLIB as it has been proposed in the recent literature.
Keywords
- Artificial Intelligence
- Programming, Linear
Status: accepted
Back to the list of papers