Optimization Days 2026

HEC Montréal, Québec, Canada

May 11 — 13, 2026

WA7 - Machine Learning

May 13 2026 09:00 – 10:40

Location: Budapest (green)

Chaired by Laurent Alsène-Racicot

4 Presentations

09:00 - 09:25

Convex Training Method for Lipschitz-Regularized Shallow Neural Networks

  • Chao Yin, speaker, Polytechnique Montréal

In this work, we introduce a training procedure for shallow neural networks that promotes robustness against adversarial attacks. We solve a non-convex Lipschitz-regularized training program by introducing a convex restriction that can be efficiently solved to global optimality. Our approach can be used as a post-processing step by taking a pre-trained network as an initial solution then solving the convex program whose optimal network is guaranteed to be no worse than the initial one. We investigate the improvements of our training procedure with experiments using real world datasets for regression tasks under an adversarial setting. We show numerically that solving our proposed convex program yields networks with lower objective values on the Lipschitz-regularized program compared to existing methods. Additionally, we show that on certain datasets, networks obtained using our convex program are both more accurate and robust with respect to adversarial attacks.

09:25 - 09:50

Reinforcement Learning and Graph Neural Networks to control a Multi-Task Robot

  • Hedi Boukamcha, speaker, Université Laval
  • Anas Neumann, ULaval / Poly MTL
  • Monia Rekik, Full Professor of The Department of Operations and Decision Systems - FSA ULaval
  • Adnene Hajji, Full Professor of The Department of Operations and Decision Systems - FSA ULaval

We propose a scheduling agent to control a real case robotic cell. The proposed approach incorporates a Graph Neural Network trained with a DQN and a dedicated solving strategy based on a Q-values guided beam search. The agent outperforms, in terms of solution quality and computing time, several heuristic approaches.

Keywords (3):
Scheduling agent, Graph neural network, Deep Q-learning

09:50 - 10:15

Simultaneous Variable Selection and Outlier Detection for Function-on-Scalar Regression Models via Mixed-Integer Programming

  • Hedayat Fathi, speaker, Uniersite Laval
  • Bernardo D'Agostino, Sant’Anna School of Advanced Studies
  • Ana Kenney, University of California, Irvine
  • Luca Insolia, University of Geneva
  • Francesca Chiaromonte, Pennsylvania State Universit
  • Marzia A. Cremona, Université Laval

We propose a mixed-integer programming framework for simultaneous variable selection and outlier detection in regression with functional response. We prove the functional robust oracle property and evaluate performance through simulations and an application studying the effect of microbiome composition on infant growth curves, showing good performance with modern solvers.

10:15 - 10:40

Quantification de l’incertitude des prévisions court-terme de la demande en électricité

  • Laurent Alsène-Racicot, speaker, Université de Montréal
  • Olivier Milon, Hydro-Quebec
  • Rachel Bazile, Hydro-Quebec
  • Alain Marcotte, Hydro-Quebec
  • Fabian Bastin, Université de Montréal

Nous présenterons des méthodes de quantification de l'incertitude pour les prévisions de la demande d'électricité à court terme (J+1) à partir des données d'Hydro-Québec. L'objectif est de pallier les limites des modèles déterministes actuels en fournissant des intervalles de prédiction calibrés.

L'approche retenue repose sur la prévision conforme, un cadre statistique robuste offrant des garanties de couverture marginale sous l'hypothèse d'interchangeabilité des données. Différentes méthodes ont été proposées et testées afin de résoudre les problématiques découlant de la réalité de données issues de séries temporelles : shift de distribution, hétéroscédasticité, etc.

Les résultats, évalués sur quatre hivers de données réelles (2022-2025), démontrent que les méthodes conformes atteignent la cible de couverture marginale de 80% tout en produisant des intervalles environ 16% plus étroits que la référence opérationnelle actuelle.