4 Presentations

• 
  10:30 AM - 10:55 AM

  A Decomposition-Coordination Framework for Solving Combined Algorithm Selection and Mixed Hyperparameter Optimization Problems

  • Ahmed Bayoumy, presenter, Siemens Digital Industries Software
  • Michael Kokkolaras, McGill University

  The selection of learning algorithms and tuning their hyperparameters can significantly impact the size and characteristics of the hyperparameter configuration space. In machine learning, this is known as the combined algorithm selection and hyperparameter optimization (CASH) problem. Typically, derivative-free optimizers are used to solve this problem when the parameter configuration space has a computationally-tractable number of dimensions. The learning algorithm's configuration space may include various conditional parameter domains that require extending the optimizer's global/local search to exploit neighbor points on the conditional domain. However, improper coordination of obtained solutions among coupled conditional parameter domains may lead to suboptimal algorithm performance. We address this issue and propose a decomposition-coordination framework that uses the alternating directions method of multipliers as a coordinator for solving the CASH problem. Our framework handles conditional parameter domains by setting their target value based on the evaluated condition and updating/activating their linked parameter domain accordingly in the coordination loop. This helps reduce the complexity of the CASH automation process, reduce its consumed runtime, and avoid reaching a suboptimal performance of the selected algorithm. We demonstrate the effectiveness of our framework through a benchmarking example that utilizes the MNIST database.
• 
  10:55 AM - 11:20 AM

  Developing a Simulation-based optimization approach to evaluate the performance of a closed-loop food supply chain: A case study in the food company.

  • Mina kazemi miyangaskary, Polytechnique Montréal
  • Samira Keivanpour, Polytechnique Montréal
  • Asad Yarahmadi, presenter, Polytechnique Montréal

  Abstract
  The evaluation of supply chain performance has the potential to enhance overall performance, minimize expenses, and augment profitability. Simulation is an effective tool for evaluating the performance of the supply chain due to its flexibility, accuracy, and ability to model complex systems. This study aims to evaluate and optimize the performance of a closed-loop food supply chain using simulation-based optimization techniques. This research uses real data from a food company in Iran. The proposed method evaluates the supply chain performance and finds the best values for its decision variables. It considers various factors, such as demand, costs, inventory, suppliers, customers, profits, time, and capacity. This study contributes to the literature by considering uncertainty for model parameters in the developed approach.
  Keywords: Supply chain performance, Simulation-based Optimization, Closed-loop food supply chain.
• 
  11:20 AM - 11:45 AM

  Revisiting column generation based matheuristic for learning classification trees

  • Krunal Patel, presenter, CERC, Polytechnique Montréal, Montréal, Canada
  • Andrea Lodi, CERC, Polytechnique Montréal, Montréal, Canada and Jacobs Technion-Cornell Institute, Cornell Tech and Technion - IIT, New York, USA
  • Guy Desaulniers, GERAD - Polytechnique Montréal

  Decision trees are highly interpretable models for solving classification problems in machine learning (ML). The standard ML algorithms for training decision trees are fast but generate suboptimal trees in terms of accuracy. Other discrete optimization models in the literature address the optimality problem but only work well on smaller datasets. Firat et al. (2020) proposed a column generation based heuristic approach for learning decision trees. This approach improves scalability and can work with larger datasets. In this presentation, we describe improvements in the column generation approach. First, we modify the subproblem model to significantly reduce the number of subproblems in multiclass classification instances. Next, we show that the data-dependent constraints in the master problem are implied, and use them as cutting planes. Furthermore, we describe a separation model to generate data points for which the LP relaxation solution violates their corresponding constraints. Finally, we introduce a preprocessing and initialization routine that reduces the size of the master and subproblems. We conclude by presenting the computational results that show that these modifications result in better scalability.
• 
  11:45 AM - 12:10 PM

  Réseaux de Planifications Distribués: combiner l'algorithme des directions alternées non convexe et l'apprentissage profond pour contrôler les charges thermiques des bâtiments.

  • Vincent Taboga, presenter, Ecole Polytechnique de Montréal
  • Hanane Dagdougui, Ecole Polytechnique de Montréal

  Alors que 4 foyers sur 5 se chauffent à l’électricité au Québec, la consommation des systèmes de chauffage représente la moitié de notre consommation en énergie. Une bonne gestion des systèmes de chauffage, ventilation et climatisation (CVC) est donc primordiale pour garantir l’équilibre du réseau électrique.
  Ce travail propose une approche distribuée pour contrôler les systèmes CVC dans les bâtiments à l’aide de réseaux de planification. L’algorithme proposé est conçu pour être facilement adaptable à plusieurs bâtiments, quelle que soit leur taille. Il répond ainsi à deux des principaux défis du développement de tels systèmes dans l’optique d’un déploiement à grande échelle.
  La structure du système de contrôle proposé est hiérarchique avec deux couches. A la couche supérieure, un coordinateur gère la consommation de puissance à l'échelle du bâtiment. À la couche inférieure, des contrôleurs locaux gèrent la température dans chaque zone thermique. Au niveau du bâtiment, le coordonnateur doit s'assurer qu'une limite de puissance maximale est respectée en traduisant la puissance totale du bâtiment en objectifs de puissance locaux pour chaque zone. Les contrôleurs locaux peuvent modifier les points de consigne de température en fonction des objectifs de puissance locaux. L’algorithme des directions alternées (ou ADMM) est utilisé pour résoudre le problème de manière distribuée, permettant le calcul en parallèle.
  L’approche proposée a été testé en simulation sur un bâtiment de 18 zones modélisé avec EnergyPlus. L’algorithme permet de répartir l’effort d’économie d’énergie entre chaque zone lorsque le bâtiment participe à un programme de réponse à la demande.