Abstract
Although interpretability is gaining traction in Artificial Intelligence and Machine Learning, it remains largely overlooked in Optimization, where black-box models are the standard. We introduce a method to enhance the interpretability of optimization models by focusing on the most influential variables and constraints, effectively reducing complexity with controlled impact on accuracy.
Our approach systematically eliminates extraneous variables and redundant constraints using filters inspired by existing presolve techniques. To ensure feasibility, constraints are carefully adjusted while redundant elements are removed.
We validated this method on 60 linear problems from the GAMS library and openTEPES, a real-world optimization model for long-term energy system expansion. Sensitivity analysis across different thresholds demonstrated that our approach significantly reduces problem size while preserving optimality and feasibility. While preserving the essential structure of the problem, the analysis of the automatic simplication process makes optimization models more transparent and easier to interpret.
