4 présentations

• 
  10h45 - 11h10

  Scaling of the output in mesh adaptive direct search

  • Charles Audet, GERAD - Polytechnique Montréal
  • Gilles Caporossi, GERAD, HEC Montréal
  • Stéphane Jacquet, prés., Polytechnique Montréal

  In blackbox optimisation algorithms, no derivatives are available and getting the output of the blackbox can take a lot of time. In 2006, Audet and Dennis proposed the algorithm MADS to solve blackbox problems with constraints. In 2016, dynamic scaling on mesh has been added to MADS to scale the input of the blackbox. Moreover, wrong scaling of the output of the blackbox can lead MADS to overestimate some constraints and converge to a local solution with a worse local optimum. This presentation compares three different scalings and the numerical results on analytical and blackbox problems.
• 
  11h10 - 11h35

  Optimization of noisy blackbox with adaptive precision

  • Pierre-Yves Bouchet, prés., Polytechnique Montréal
  • Charles Audet, GERAD - Polytechnique Montréal

  Blackbox optimization problems are sometimes affected by noisy objective function values where noise magnitude depends of computational intensity. The higher efforts are, the higher the statistical guarantees of precision are. This talk proposes a way to converge towards an optimal solution in controlled computation time, exploiting this tradeoff.
• 
  11h35 - 12h00

  Estimation of the constraint violation function with binary, unrelaxable and hidden constraints in MADS.

  • Charles Audet, GERAD - Polytechnique Montréal
  • Gilles Caporossi, GERAD, HEC Montréal
  • Stéphane Jacquet, prés.,

  In some industrial problems, no analytical expressions of the functions are available. The output of the function come from a blackbox, which can be either an experiment in a laboratory or from computer simulations. In 2006, Audet and Dennis proposed the algorithm MADS to solve blackbox problems with constraints. At first, the extrem barrier was used to handle the constraints by rejecting all the infeasible elements. Since, more flexible methods have been made to handle them. However, the extrem barrier is still the approach used for binary, unrelaxable and hidden constraints. This presentation suggests a way to handle this constraints in MADS by giving a dynamic surrogate of binary constraints calculated through regression models that come from supervised classification. Numerical results will show the improvements on blackbox problems.
• 
  12h00 - 12h25

  Dynamic improvements of static surrogates in direct search optimization

  • Charles Audet, prés., GERAD - Polytechnique Montréal
  • Julien Côté-Massicotte, Polytechnique Montréal

  The present work involves direct search algorithms guided by surrogate models. These models are classified into two categories: static surrogates and dynamic models. We introduce the hybrid quadratic model that dynamically corrects information from a static surrogate.