2018 Optimization Days

HEC Montréal, Québec, Canada, 7 — 9 May 2018

Schedule Authors My Schedule

WB9 On the integration of machine learning and mathematical optimization III

May 9, 2018 03:30 PM – 05:10 PM

Location: Quebecor (80)

Chaired by Louis-Martin Rousseau

3 Presentations

  • 03:30 PM - 03:55 PM

    Data-driven transplantation science

    • Margarida Carvalho, presenter, Polytechnique Montréal
    • Andrea Lodi, Polytechnique Montreal
    • Yoshua Bengio, MILA, Université de Montréal
    • Héloise Cardinal, Centre Hospitalier de l'Université de Montréal
    • Jean-Noël Weller Weller, Polytechnique Montréal
    • Margaux Luck, Montreal Institute for Learning Algorithms Montreal
    • Tristan Sylvain, Montreal Institute for Learning Algorithms

    There is a rising incidence of patients suffering from renal failure. The life quality of these patients would be significantly improved through a kidney transplant from a compatible donor. Thus, it is of crucial importance to properly plan the system that allocates donated organs to patients. In this project, we use machine learning models to predict graft survival and analyze how to introduce this qualitative information on the allocation systems.

  • 03:55 PM - 04:20 PM

    Learning representations of uncertainty for decision making processes

    • Louis-Martin Rousseau, presenter, Polytechnique Montréal
    • Adulyasak Yossiri, HEC Montreal
    • Laurent Charlin, HEC Montréal
    • Christian Dorion, HEC Montreal
    • Alexandre Jeanneret, HEC Montréal

    Decision support and optimization-based software are being deployed and used in a constantly growing number of areas affecting our lives. Most of the existing software/programs are today based on data which are approximated using a set of a priori models. In most cases, such software take important decisions based on a weak representation of the uncertainty of the underlying data. Fundamental questions such as “what is the evolution of a tumor during treatment?”, “which market conditions have an impact on a given stock?”, or “how does the demand for a new product fluctuates over time?” are often answered by point forecasts, which means ignoring variance and adverse scenarios. In some other cases, one makes use of probabilistic forecasts based on choices made by human modelers. However, as our economic and social environment is constantly and rapidly evolving, we believe that it is crucial to base decisions on data-driven representations of uncertainty rather than relying on a priori beliefs. We propose to address this challenge using probabilistic deep neural network (DNN), which can learn complex hierarchical representation based on historical and current data to generate possible scenarios of future outcomes. Such scenarios can then be used by optimization methods under uncertainty such as multistage stochastic programming, robust optimization, and chance constrained programming. The challenge will consist of understanding the tradeoff between the quantity and diversity of such scenarios and the complexity of the decision model. We aim to use this new approach to address key paradigms in the decision making processes and methods in several domains, i.e., health care, logistics, and finance, which will overcome the limitations of the tradition methods in their respective applications. This proposal is thus at the core of the research program of IVADO’s apogee grant, both from methodological and application standpoints. This research aims to bridge the gap between the field of ML and OR in a practical yet fundamental way. We first describe the methodology and then present some areas in which we will apply the proposed techniques.

  • 04:20 PM - 04:45 PM

    Optimization of hydropower: Dynamic programming techniques, present and future

    • Michel Denault, presenter, GERAD - HEC Montréal
    • Pascal Côté, Rio Tinto
    • Dominique Orban, GERAD - Polytechnique Montréal
    • Jean-Guy Simonato, HEC Montréal

    We will present our current results based on simulation-and-regression dynamic programming techniques, as well as our projects for the future, which are based on reinforcement learning and Q-learning approaches. Our reference application is hydropower optimization, but we may also discuss portfolio optimization issues.