Explaining the optimizer's answer
In the operations research journals, 1,178 recent papers carry an optimization keyword, and 36 of them carry an explanation or interpretability keyword, while interpretability appears on 379 papers in the archive, almost all about machine learning. Planners reject optimized schedules they cannot understand. What counts as an explanation of a solution, and do explanations raise adoption?
What prompted this
Since October 2025, 1,178 papers in the operations research journals carry an optimization, mathematical-programming or mixed-integer keyword, and 36 of them also carry an explanation, interpretability, counterfactual, transparency, human-in-the-loop or trust keyword. "Interpretability" appears on 379 papers in the archive, mostly in AI journals about models rather than decisions; "decision support system" appears on 70.
Questions
- Which forms of explanation (which constraints bind, what would change the answer, why a nearby plan is worse) help planners accept or correct a solution, and which do they ignore?
- Do explanations change the solutions planners choose, and do the changes improve or worsen outcomes?
- Can explanations be generated for large mixed-integer models without solving many variants?
Data that could help
Experiments with planners choosing between optimized plans with different explanations, and logs from decision support systems that record overrides.
Next step
Run an experiment with professional planners comparing acceptance of the same optimized schedule presented with three kinds of explanation.