Research: machine learning for decision making. Noah Schutte, PhD researcher at TU Delft. Research Machine Learning forDecision Making

1. 2026
   Preprint

   Scalable Decision-Focused Learning through Cost-Sensitive Regression

   Noah Schutte, Senne Berden, Tias Guns, Krzysztof Postek, Neil Yorke-Smith

   arXiv:2605.18005 · May 2026

   Decision-focused learning trains a prediction model on how good the final decision is. The catch: it usually re-solves the whole optimization problem over and over during training, which gets slow fast. This paper recasts the task as cost-sensitive regression, where over- and under-predictions are penalised by what they do to the decision. It needs at most one solve per training example, matches the best existing methods on decision quality, and scales to problem sizes decision-focused learning had not handled before.

   • Scalability
   • Cost-sensitive regression
   • Combinatorial optimization

   PDF ↗arXiv ↗

   Cite (BibTeX)

   @misc{schutte2026scalable,
     title         = {Scalable Decision-Focused Learning through Cost-Sensitive Regression},
     author        = {Schutte, Noah and Berden, Senne and Guns, Tias and Postek, Krzysztof and Yorke-Smith, Neil},
     year          = {2026},
     eprint        = {2605.18005},
     archivePrefix = {arXiv},
     primaryClass  = {cs.LG}
   }

   Copy
2. 2026
   IJCAI 2026

   Sufficient Decision Proxies for Decision-Focused Learning

   Noah Schutte, Grigorii Veviurko, Krzysztof Postek, Neil Yorke-Smith

   Proceedings of the 35th International Joint Conference on Artificial Intelligence, IJCAI-26, pp. 2337-2345

   To plan under uncertainty, a model can predict one scenario (a single best guess) or a full probability distribution. This paper looks at which features of a problem decide what kind of prediction, a "decision proxy", is enough to make a good decision, and proposes new options that add little computation. It is tested on continuous and discrete problems where both the goal and the constraints are uncertain.

   • Decision proxies
   • Uncertainty
   • IJCAI 2026

   PDF ↗arXiv ↗

   Cite (BibTeX)

   @inproceedings{schutte2026sufficient,
     title     = {Sufficient Decision Proxies for Decision-Focused Learning},
     author    = {Schutte, Noah and Veviurko, Grigorii and Postek, Krzysztof and Yorke-Smith, Neil},
     booktitle = {Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, {IJCAI-26}},
     pages     = {2337--2345},
     year      = {2026}
   }

   Copy
3. 2024
   IJCAI 2024

   Robust Losses for Decision-Focused Learning

   Noah Schutte, Krzysztof Postek, Neil Yorke-Smith

   Proceedings of the 33rd International Joint Conference on Artificial Intelligence, IJCAI-24, pp. 4868-4875

   Decision-focused learning usually trains on "regret": how much worse a decision is than the best possible one for each past example. But the best decision for a single noisy example can be far from the best decision on average, so the model learns the wrong lesson. The paper measures how randomness and lack of knowledge both distort this signal, then proposes three more robust loss functions. They improve results on unseen data without extra training time, on shortest-path, travelling-salesperson and energy-scheduling problems.

   • Robust optimization
   • Loss functions
   • IJCAI 2024

   PDF ↗DOI ↗arXiv ↗

   Cite (BibTeX)

   @inproceedings{schutte2024robust,
     title     = {Robust Losses for Decision-Focused Learning},
     author    = {Schutte, Noah and Postek, Krzysztof and Yorke-Smith, Neil},
     booktitle = {Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, {IJCAI-24}},
     pages     = {4868--4875},
     year      = {2024},
     doi       = {10.24963/ijcai.2024/538}
   }

   Copy
4. 2024
   IJCAI 2024

   Optimization Under Epistemic Uncertainty Using Prediction

   Noah Schutte

   Proceedings of the 33rd International Joint Conference on Artificial Intelligence, IJCAI-24, Doctoral Consortium, pp. 8504-8505

   His PhD agenda in two pages. Optimization models often pretend the world is certain so they can be solved. Uncertainty that comes from missing knowledge, called epistemic uncertainty, is especially hard to model. The aim: make decisions more robust in practice without making the models harder to solve, with decision-focused learning as the main tool.

   • PhD agenda
   • Epistemic uncertainty
   • Sole author

   PDF ↗DOI ↗

   Cite (BibTeX)

   @inproceedings{schutte2024optimization,
     title     = {Optimization Under Epistemic Uncertainty Using Prediction},
     author    = {Schutte, Noah},
     booktitle = {Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, {IJCAI-24}},
     pages     = {8504--8505},
     year      = {2024},
     doi       = {10.24963/ijcai.2024/967}
   }

   Copy

Sources: IJCAI proceedings, arXiv, TU Delft. New papers are added automatically from dblp.