4 présentations

• 
  15h30 - 15h55

  Expert based Learning for Sequential Decision Making

  • Prakash Gawas, prés., Polytechnique Montreal
  • Antoine Legrain, GERAD - Polytechnique Montréal
  • Louis-Martin Rousseau, Polytechnique Montréal

  Imitation learning, where an agent learns decision-making by mimicking expert behavior, has been widely applied in robotics and autonomous driving. This study introduces an innovative approach using DAgger, an imitation learning algorithm, to iteratively train policies for stochastic sequential decision problems (SDPs). These problems are particularly challenging when expert input is costly or unavailable. First, we define a taxonomy of experts that can be used for imitation in SDPs. We then focus on designing an effective expert within the DAgger framework, leveraging deterministic solutions derived from contextual scenarios at each decision point. A predictive model is subsequently trained to replicate the expert’s behavior, enabling real-time decision-making. However, imperfect experts may provide faulty demonstrations, slowing down learning. To mitigate this, we propose a strategy incorporating multiple experts within the DAgger framework.
  To demonstrate the methodology’s applicability, we address a dynamic employee notification timing problem for scheduling casual personnel in on-call shifts.

• 
  15h55 - 16h20

  An Efficient Scaled spectral preconditioner for sequences of symmetric positive definite linear systems

  • Oussama Mouhtal, prés., CERFACS and GERAD
  • Youssef Diouane, Polytechnique Montréal
  • Dominique Orban, GERAD - Polytechnique Montréal
  • Selime Gürol, CERFACS

  In this talk, we explore a scaled spectral preconditioner for the efficient solution of sequences of symmetric and positive-definite linear systems. We design the scaled preconditioner not only as an approximation of the inverse of the linear system but also with consideration of its use within the conjugate gradient (CG) method. We present three different strategies for selecting a scaling parameter, which aims to position the eigenvalues of the preconditioned matrix in a way that reduces the energy norm of the error, the quantity that CG monotonically decreases at each iteration. Numerical experiments provide in data assimilation confirm that the scaled spectral preconditioner can significantly improve early CG convergence with negligible computational cost.

• 
  16h20 - 16h45

  A Mixed Integer Programming Approach for the Erasmus+ placement problem

  • Marina Resta, University of Genoa
  • Luca Murazzano, prés., Université Laval
  • Paolo Landa, Université Laval
  • Elena Tànfani, University of Genova

  This study investigates a Mixed Integer Programming (MIP) approach to optimize the Erasmus+ student placement process. The Erasmus+ program supports international academic mobility, allowing students to study in European universities. The proposed mathematical model integrates students' preferences, academic performance indicators (such as GPA and ECTS credits), and language proficiency levels while addressing constraints from host universities, including capacity limits and program compatibility.
  The model prioritizes first-time Erasmus+ applicants to promote equitable opportunity distribution and considers backup placements for students when preferred destinations are unavailable. Empirical analysis, using anonymized data from the Department of Economics of the University of Genoa (DIEC), highlights the model's effectiveness compared to manual allocation methods. The findings reveal significant improvements in the alignment of student aspirations with host university offerings, resulting in enhanced satisfaction and fairness.
  By leveraging operations research techniques, this study offers a novel contribution to the field of academic mobility placement systems. The study demonstrates the potential of optimization frameworks to address complex allocation challenges, providing valuable insights for decision-makers in higher education institutions.

• 
  16h45 - 17h10

  Decentralized Mixed-Integer Predictive Control for Multi-zone Building HVAC Systems using a Physics-Informed Data-Driven Model

  • Étienne Tremblay, prés., Polytechnique Montréal
  • Antoine Lesage-Landry, Polytechnique Montréal

  Commercial buildings have a significant impact on global energy consumption with their heating, ventilation, and air conditioning (HVAC) systems accounting for the largest share of energy use. Optimizing HVAC control strategies can improve energy efficiency, reduce the impact on the electric grid, and enable responsive adjustments to grid events through demand response programs. This should, however, not be done at the users’ comfort expense, as failure to do so may prevent enrolment in the program. The thermal dynamics are unique to every building and require involved modelling and characterization to be readily embedded in a control strategy. Moreover, commercial buildings, commonly use rooftop unit (RTU) HVAC systems which rely on discrete operation modes, e.g., several cooling and heating stages in addition to an on/off fan. In this work, we propose a decentralized mixed-integer model predictive control (MPC) approach for large-scale commercial buildings. We employed a data-driven approach based on a physics-informed linear regression model to estimate the building thermal dynamics using both simulated or historical data and forecasted data.
  The computational efficiency of the approach combined with its data-driven aspect makes our model amenable to most buildings without the need for significant and costly upgrades. Finally, we evaluate the performance of our distribution optimization approach in a state-of-the-art building simulation.