Journées de l'optimisation 2022

HEC Montréal, Québec, Canada, 16 — 18 mai 2022

Horaire Auteurs Mon horaire

WC8 - Derivative-free and Blackbox Optimization VI

18 mai 2022 15h30 – 17h10

Salle: METRO INC. (jaune)

Présidée par Stéphane Alarie

4 présentations

  • 15h30 - 15h55

    Escaping unknown discontinuous regions in blackbox optimization

    • Solène Kojtych, prés., Polytechnique Montreal
    • Charles Audet, GERAD - Polytechnique Montréal
    • Alain Batailly, Polytechnique Montréal

    We study constrained blackbox optimization problems involving discontinuities whose positions are not known apriori. We propose a MADS algorithm, called DiscoMads, revealing and escaping these regions progressively. A convergence analysis supports the algorithm and numerical tests are conducted on analytical and engineering problems, among which a mechanical engineering problem related to the vibration of aircraft engine blades.

  • 15h55 - 16h20

    Constrained stochastic blackbox optimization using a progressive barrier and probabilistic estimates

    • Kwassi Joseph Dzahini, prés., Argonne National Laboratory
    • Sébastien Le Digabel, GERAD, Polytechnique Montréal
    • Michael Kokkolaras, Université McGill

    This work introduces StoMADS-PB, a direct-search method using a progressive barrier approach for constrained stochastic blackbox optimization. The values of the objective and constraint functions can only be computed with random noise whose distribution is unknown. Since the deterministic computable version of the blackbox is unavailable, StoMADS-PB uses estimates and introduces probabilistic bounds (for its constraint violation function), required to be accurate and reliable with high, but fixed, probabilities. Using Clarke calculus and martingale theory, Clarke stationarity convergence results for the objective and the violation function are derived with probability one.

  • 16h20 - 16h45

    NOMAD 4: development of a new version of a blackbox optimization solver

    • Christophe Tribes, prés., Polytechnique Montréal
    • Charles Audet, GERAD - Polytechnique Montréal
    • Sébastien Le Digabel, GERAD, Polytechnique Montréal
    • Viviane Rochon Montplaisir,

    NOMAD is an open source optimization software well suited for costly blackbox problems with inequality constraints, non-smooth and noisy design spaces. Derivative free algorithms such as mesh adaptive direct search (MADS) or Nelder Mead are available in NOMAD.
    The version 3 of NOMAD is available since 2008 with several major revisions up to the current 3.9 version. However, the design of NOMAD 3 is strongly oriented by MADS and show its limits when maintenance is required or when adding new algorithms or new features. Also, the parallel evaluation capabilities of NOMAD 3 are not adapted to today's computers computational capabilities.
    NOMAD 4 has been developed from scratch with a complete new architecture and design.
    The presentation will describe the new design and the improvements that allow for better parallel evaluations, easier maintenance and faster addition of new features.
    Benchmarking NOMAD 3 and NOMAD 4 on test suites show that the performance is maintained in the new version.

  • 16h45 - 17h10

    Blackbox optimization with hierarchical constraints

    • Stéphane Alarie, prés., Institut de recherche d'Hydro-Québec
    • Charles Audet, GERAD - Polytechnique Montréal
    • Paulin Jacquot, GERAD - Polytechnique Montréal
    • Sébastien Le Digabel, GERAD, Polytechnique Montréal

    Some complex and expensive blackbox evaluations can be structured as a sequence of constraint verifications. One can then interrupt any evaluation as soon as it appears unfeasible, especially if one already has a feasible solution in hand. By doing so, less time should be needed to solve the problem. Two different approaches are proposed here to define the hierarchical order of the constraints and to manage the interruption of the evaluations throughout the optimization process. Numerical experiments will be presented on a closed-form test problem.